Add files using upload-large-folder tool

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-040dc4c3-65f8-4130-b808-b9afeb8440701759216917788-2025_09_30-09.22.06.63/source.csv

@@ -0,0 +1,11 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,4,"slurm/jobs/franz/berlin/atari/data_generation/spawner.sh",0,0,"sbatch slurm/jobs/franz/berlin/atari/data_generation/alien.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/amidar.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/assault.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/asterix.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/bank_heist.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/battle_zone.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/boxing.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/breakout.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/chopper_command.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/crazy_climber.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/demon_attack.sh\nsbatch slurm/jobs/franz/berlin/atari/data_generation/pong.sh",shellscript,tab
+2,333,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"9:22:06 AM [info] Activating crowd-code\n9:22:06 AM [info] Recording started\n9:22:06 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,446,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"9:22:06 AM [info] Git repository found\n9:22:06 AM [info] Git provider initialized successfully\n9:22:06 AM [info] Initial git state: [object Object]\n",Log,content
+4,2100,"slurm/jobs/franz/berlin/atari/data_generation/spawner.sh",0,0,"",shellscript,tab
+5,3700,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+6,3701,"TERMINAL",0,0,"",,terminal_command
+7,5061,"slurm/jobs/franz/berlin/atari/data_generation/spawner.sh",0,0,"",shellscript,tab
+8,8021,"TERMINAL",0,0,"squeue",,terminal_command
+9,8087,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  29653   xiao.liu interacti     1    64  R 2025-09-29T15:54:10 2025-09-29T15:54:10 17:28:04   23:59:00 hai002\r\n  29652   xiao.liu interacti     1    64  R 2025-09-29T15:53:48 2025-09-29T15:53:48 17:28:26   23:59:00 hai006\r\n  29651   xiao.liu interacti     1    64  R 2025-09-29T15:53:13 2025-09-29T15:53:13 17:29:01   23:59:00 hai007\r\n  29693 franz.sram  standard     1   128  R 2025-09-29T23:35:46 2025-09-30T09:19:38     2:36 1-00:00:00 hai001\r\n  29692 franz.sram  standard     1   128  R 2025-09-29T23:35:46 2025-09-30T08:57:44    24:30 1-00:00:00 hai003\r\n  29691 franz.sram  standard     1   128  R 2025-09-29T23:35:46 2025-09-30T08:35:32    46:42 1-00:00:00 hai005\r\n  29690 franz.sram  standard     1   128  R 2025-09-29T23:35:46 2025-09-30T08:08:11  1:14:03 1-00:00:00 hai006\r\n  29689 franz.sram  standard     1   128  R 2025-09-29T23:35:46 2025-09-30T07:32:17  1:49:57 1-00:00:00 hai004\r\n  29688 franz.sram  standard     1   128  R 2025-09-29T23:35:46 2025-09-30T07:30:01  1:52:13 1-00:00:00 hai008\r\n  29624 alfred.ngu  standard     1    16  R 2025-09-29T17:00:52 2025-09-29T17:02:56 16:19:18 1-00:00:00 hai003\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+10,10916,"TERMINAL",0,0,"",,terminal_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-2b5d7b91-020b-49f2-bc14-8ee144ef52891764346477780-2025_11_28-17.14.44.858/source.csv

@@ -0,0 +1,49 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,624,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:14:44 PM [info] Activating crowd-code\n5:14:45 PM [info] Recording started\n5:14:45 PM [info] Initializing git provider using file system watchers...\n5:14:45 PM [info] No workspace folder found\n",Log,tab
+3,2304,"extension-output-pdoom-org.crowd-code-#1-crowd-code",194,0,"5:14:47 PM [info] Retrying git provider initialization...\n5:14:47 PM [info] No workspace folder found\n",Log,content
+4,3273,"Untitled-1",0,0,"",plaintext,tab
+5,5949,"Untitled-1",0,0,"t",plaintext,content
+6,5952,"Untitled-1",1,0,"",plaintext,selection_keyboard
+7,6015,"Untitled-1",1,0,"h",plaintext,content
+8,6017,"Untitled-1",2,0,"",plaintext,selection_keyboard
+9,6053,"Untitled-1",2,0,"i",plaintext,content
+10,6055,"Untitled-1",3,0,"",plaintext,selection_keyboard
+11,6195,"Untitled-1",3,0,"s",plaintext,content
+12,6197,"Untitled-1",4,0,"",plaintext,selection_keyboard
+13,6268,"Untitled-1",4,0,"i",plaintext,content
+14,6270,"Untitled-1",5,0,"",plaintext,selection_keyboard
+15,6423,"Untitled-1",5,0,"s",plaintext,content
+16,6425,"Untitled-1",6,0,"",plaintext,selection_keyboard
+17,6744,"Untitled-1",5,1,"",plaintext,content
+18,6885,"Untitled-1",4,1,"",plaintext,content
+19,6951,"Untitled-1",4,0," ",plaintext,content
+20,6953,"Untitled-1",5,0,"",plaintext,selection_keyboard
+21,6973,"Untitled-1",5,0,"i",plaintext,content
+22,6975,"Untitled-1",6,0,"",plaintext,selection_keyboard
+23,7114,"Untitled-1",6,0,"s",plaintext,content
+24,7116,"Untitled-1",7,0,"",plaintext,selection_keyboard
+25,7175,"Untitled-1",7,0," ",plaintext,content
+26,7176,"Untitled-1",8,0,"",plaintext,selection_keyboard
+27,7288,"Untitled-1",8,0,"a",plaintext,content
+28,7290,"Untitled-1",9,0,"",plaintext,selection_keyboard
+29,7363,"Untitled-1",9,0," ",plaintext,content
+30,7365,"Untitled-1",10,0,"",plaintext,selection_keyboard
+31,8310,"Untitled-1",9,0,"",plaintext,selection_command
+32,8912,"Untitled-1",10,0,"",plaintext,selection_command
+33,12503,"Untitled-1",9,0,"",plaintext,selection_command
+34,52510,"Untitled-1",10,0,"",plaintext,selection_command
+35,52777,"Untitled-1",10,0,"t",plaintext,content
+36,52780,"Untitled-1",11,0,"",plaintext,selection_keyboard
+37,52826,"Untitled-1",11,0,"e",plaintext,content
+38,52829,"Untitled-1",12,0,"",plaintext,selection_keyboard
+39,52997,"Untitled-1",12,0,"s",plaintext,content
+40,52999,"Untitled-1",13,0,"",plaintext,selection_keyboard
+41,53002,"Untitled-1",13,0,"t",plaintext,content
+42,53003,"Untitled-1",14,0,"",plaintext,selection_keyboard
+43,68816,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+44,70900,"TERMINAL",0,0,"",,terminal_focus
+45,70903,"Untitled-1",0,0,"",plaintext,tab
+46,72355,"TERMINAL",0,0,"squeue",,terminal_command
+47,72363,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  35094 franz.sram interacti     1    20  R 2025-11-28T17:12:32 2025-11-28T17:12:32     3:25 1-00:00:00 hai001\r\n  35071   xiao.liu interacti     1   128  R 2025-11-28T12:35:47 2025-11-28T12:35:47  4:40:10   23:59:00 hai006\r\n  35044   xiao.liu interacti     1   128  R 2025-11-28T04:21:56 2025-11-28T04:21:56 12:54:01   23:59:00 hai005\r\n  35070   xiao.liu  standard     1   128  R 2025-11-28T12:32:45 2025-11-28T13:49:11  3:26:46   23:59:00 hai004\r\n  35058 kalyan.nad  standard     1    64  R 2025-11-28T11:52:36 2025-11-28T11:52:39  5:23:18 1-00:00:00 hai003\r\n]0;franz.srambical@hai-login2:~",,terminal_output
+48,653160,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+49,654290,"Untitled-1",0,0,"",plaintext,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-3aefeec5-b2f0-4bf1-a63e-7fcacb754dce1765226802084-2025_12_08-21.46.48.843/source.csv

@@ -0,0 +1,232 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,420,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"9:46:48 PM [info] Activating crowd-code\n9:46:49 PM [info] Recording started\n9:46:49 PM [info] Initializing git provider using file system watchers...\n9:46:49 PM [info] No workspace folder found\n",Log,tab
+3,2297,"extension-output-pdoom-org.crowd-code-#1-crowd-code",194,0,"9:46:51 PM [info] Retrying git provider initialization...\n9:46:51 PM [info] No workspace folder found\n",Log,content
+4,3700,"Untitled-1",0,0,"",plaintext,tab
+5,21578,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+6,22412,"Untitled-1",0,0,"",plaintext,tab
+7,26953,"Untitled-1",0,0,"""",plaintext,content
+8,26955,"Untitled-1",1,0,"",plaintext,selection_keyboard
+9,27141,"Untitled-1",1,0,"""",plaintext,content
+10,27142,"Untitled-1",2,0,"",plaintext,selection_keyboard
+11,27272,"Untitled-1",2,0,"""",plaintext,content
+12,27274,"Untitled-1",3,0,"",plaintext,selection_keyboard
+13,27842,"Untitled-1",3,0,"""",plaintext,content
+14,27844,"Untitled-1",4,0,"",plaintext,selection_keyboard
+15,28012,"Untitled-1",4,0,"""",plaintext,content
+16,28013,"Untitled-1",5,0,"",plaintext,selection_keyboard
+17,28173,"Untitled-1",5,0,"""",plaintext,content
+18,28175,"Untitled-1",6,0,"",plaintext,selection_keyboard
+19,28491,"Untitled-1",5,0,"",plaintext,selection_command
+20,28607,"Untitled-1",4,0,"",plaintext,selection_command
+21,28771,"Untitled-1",3,0,"",plaintext,selection_command
+22,29408,"Untitled-1",3,0,"h",plaintext,content
+23,29409,"Untitled-1",4,0,"",plaintext,selection_keyboard
+24,29488,"Untitled-1",4,0,"e",plaintext,content
+25,29490,"Untitled-1",5,0,"",plaintext,selection_keyboard
+26,29647,"Untitled-1",5,0,"l",plaintext,content
+27,29648,"Untitled-1",6,0,"",plaintext,selection_keyboard
+28,29789,"Untitled-1",6,0,"l",plaintext,content
+29,29791,"Untitled-1",7,0,"",plaintext,selection_keyboard
+30,29872,"Untitled-1",7,0,"o",plaintext,content
+31,29874,"Untitled-1",8,0,"",plaintext,selection_keyboard
+32,30094,"Untitled-1",8,0," ",plaintext,content
+33,30096,"Untitled-1",9,0,"",plaintext,selection_keyboard
+34,30114,"Untitled-1",9,0,"w",plaintext,content
+35,30115,"Untitled-1",10,0,"",plaintext,selection_keyboard
+36,30355,"Untitled-1",10,0,"o",plaintext,content
+37,30356,"Untitled-1",11,0,"",plaintext,selection_keyboard
+38,30465,"Untitled-1",11,0,"r",plaintext,content
+39,30466,"Untitled-1",12,0,"",plaintext,selection_keyboard
+40,30691,"Untitled-1",12,0,"l",plaintext,content
+41,30692,"Untitled-1",13,0,"",plaintext,selection_keyboard
+42,30814,"Untitled-1",13,0,"d",plaintext,content
+43,30816,"Untitled-1",14,0,"",plaintext,selection_keyboard
+44,30906,"Untitled-1",14,0," ",plaintext,content
+45,30907,"Untitled-1",15,0,"",plaintext,selection_keyboard
+46,31041,"Untitled-1",15,0,"f",plaintext,content
+47,31043,"Untitled-1",16,0,"",plaintext,selection_keyboard
+48,31056,"Untitled-1",16,0,"u",plaintext,content
+49,31058,"Untitled-1",17,0,"",plaintext,selection_keyboard
+50,31213,"Untitled-1",17,0,"n",plaintext,content
+51,31215,"Untitled-1",18,0,"",plaintext,selection_keyboard
+52,31369,"Untitled-1",18,0,"c",plaintext,content
+53,31371,"Untitled-1",19,0,"",plaintext,selection_keyboard
+54,31609,"Untitled-1",19,0,"t",plaintext,content
+55,31611,"Untitled-1",20,0,"",plaintext,selection_keyboard
+56,31674,"Untitled-1",20,0,"i",plaintext,content
+57,31675,"Untitled-1",21,0,"",plaintext,selection_keyboard
+58,31706,"Untitled-1",21,0,"o",plaintext,content
+59,31708,"Untitled-1",22,0,"",plaintext,selection_keyboard
+60,31780,"Untitled-1",22,0,"n",plaintext,content
+61,31782,"Untitled-1",23,0,"",plaintext,selection_keyboard
+62,34007,"Untitled-1",22,0,"",plaintext,selection_command
+63,40214,"Untitled-1",26,0,"",plaintext,selection_mouse
+64,40217,"Untitled-1",25,0,"",plaintext,selection_command
+65,43157,"Untitled-1",0,26,"""""""hello world function""""""\n",plaintext,content
+66,44854,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+67,47319,"TERMINAL",0,0,"",,terminal_focus
+68,47321,"Untitled-1",0,0,"",plaintext,tab
+69,49221,"TERMINAL",0,0,"bash",,terminal_focus
+70,50321,"TERMINAL",0,0,"bash",,terminal_focus
+71,51844,"TERMINAL",0,0,"bash",,terminal_focus
+72,55399,"Untitled-1",26,0,"\n",plaintext,content
+73,58113,"Untitled-1",28,0,"",plaintext,selection_command
+74,58502,"Untitled-1",27,0,"",plaintext,selection_command
+75,60534,"Untitled-1",27,0,"d",plaintext,content
+76,60535,"Untitled-1",28,0,"",plaintext,selection_keyboard
+77,60659,"Untitled-1",28,0,"e",plaintext,content
+78,60660,"Untitled-1",29,0,"",plaintext,selection_keyboard
+79,60820,"Untitled-1",29,0,"f",plaintext,content
+80,60821,"Untitled-1",30,0,"",plaintext,selection_keyboard
+81,60899,"Untitled-1",30,0," ",plaintext,content
+82,60901,"Untitled-1",31,0,"",plaintext,selection_keyboard
+83,61406,"Untitled-1",31,0,"h",plaintext,content
+84,61408,"Untitled-1",32,0,"",plaintext,selection_keyboard
+85,61470,"Untitled-1",32,0,"e",plaintext,content
+86,61472,"Untitled-1",33,0,"",plaintext,selection_keyboard
+87,61637,"Untitled-1",33,0,"l",plaintext,content
+88,61639,"Untitled-1",34,0,"",plaintext,selection_keyboard
+89,61852,"Untitled-1",34,0,"l",plaintext,content
+90,61853,"Untitled-1",35,0,"",plaintext,selection_keyboard
+91,61855,"Untitled-1",35,0,"o",plaintext,content
+92,61856,"Untitled-1",36,0,"",plaintext,selection_keyboard
+93,62176,"Untitled-1",36,0,"_",plaintext,content
+94,62177,"Untitled-1",37,0,"",plaintext,selection_keyboard
+95,62713,"Untitled-1",37,0,"w",plaintext,content
+96,62714,"Untitled-1",38,0,"",plaintext,selection_keyboard
+97,62853,"Untitled-1",38,0,"o",plaintext,content
+98,62854,"Untitled-1",39,0,"",plaintext,selection_keyboard
+99,62960,"Untitled-1",39,0,"r",plaintext,content
+100,62961,"Untitled-1",40,0,"",plaintext,selection_keyboard
+101,63169,"Untitled-1",40,0,"l",plaintext,content
+102,63170,"Untitled-1",41,0,"",plaintext,selection_keyboard
+103,63253,"Untitled-1",41,0,"d",plaintext,content
+104,63255,"Untitled-1",42,0,"",plaintext,selection_keyboard
+105,63559,"Untitled-1",42,0,"()",plaintext,content
+106,63560,"Untitled-1",43,0,"",plaintext,selection_keyboard
+107,63692,"Untitled-1",43,1,")",plaintext,content
+108,63693,"Untitled-1",44,0,"",plaintext,selection_keyboard
+109,64312,"Untitled-1",44,0,"\n",plaintext,content
+110,65332,"Untitled-1",44,1,"",plaintext,content
+111,65704,"Untitled-1",44,0,":",plaintext,content
+112,65705,"Untitled-1",45,0,"",plaintext,selection_keyboard
+113,69226,"Untitled-1",45,0,"\n",plaintext,content
+114,70999,"Untitled-1",27,0,"",plaintext,selection_command
+115,71279,"Untitled-1",46,0,"",plaintext,selection_command
+116,73051,"Untitled-1",27,0,"",plaintext,selection_command
+117,73687,"Untitled-1",46,0,"",plaintext,selection_command
+118,75517,"Untitled-1",27,19,"def hello_world():\n",plaintext,content
+119,77141,"TERMINAL",0,0,"sed -i '2,2c\def hello_world()' extension-output-pdoom-org.crowd-code-#1-crowd-code",,terminal_command
+120,78969,"Untitled-1",27,19,"def hello_world():\n",plaintext,content
+121,79747,"TERMINAL",0,0,"sed -i '2,2c\def hello_world()' extension-output-pdoom-org.crowd-code-#1-crowd-code",,terminal_command
+122,80246,"Untitled-1",46,0,"    ",plaintext,content
+123,83939,"Untitled-1",46,5,"    \n",plaintext,content
+124,85698,"Untitled-1",31,0,"",plaintext,selection_command
+125,85994,"Untitled-1",50,0,"",plaintext,selection_command
+126,86771,"Untitled-1",31,0,"",plaintext,selection_command
+127,87212,"Untitled-1",50,0,"",plaintext,selection_command
+128,87418,"Untitled-1",31,0,"",plaintext,selection_command
+129,88729,"Untitled-1",50,0,"",plaintext,selection_command
+130,89017,"Untitled-1",51,0,"",plaintext,selection_command
+131,89774,"Untitled-1",46,0,"",plaintext,selection_command
+132,90395,"Untitled-1",47,0,"",plaintext,selection_command
+133,90570,"Untitled-1",48,0,"",plaintext,selection_command
+134,90743,"Untitled-1",49,0,"",plaintext,selection_command
+135,90887,"Untitled-1",50,0,"",plaintext,selection_command
+136,92262,"Untitled-1",31,0,"",plaintext,selection_command
+137,92686,"Untitled-1",50,0,"",plaintext,selection_command
+138,92995,"Untitled-1",31,0,"",plaintext,selection_command
+139,93708,"Untitled-1",50,0,"",plaintext,selection_command
+140,94934,"Untitled-1",31,0,"",plaintext,selection_command
+141,95272,"Untitled-1",50,0,"",plaintext,selection_command
+142,95624,"Untitled-1",31,0,"",plaintext,selection_command
+143,96025,"Untitled-1",50,0,"",plaintext,selection_command
+144,96316,"Untitled-1",31,0,"",plaintext,selection_command
+145,96782,"Untitled-1",50,0,"",plaintext,selection_command
+146,96973,"Untitled-1",31,0,"",plaintext,selection_command
+147,97399,"Untitled-1",50,0,"",plaintext,selection_command
+148,97606,"Untitled-1",31,0,"",plaintext,selection_command
+149,98252,"Untitled-1",50,0,"",plaintext,selection_command
+150,98450,"Untitled-1",31,0,"",plaintext,selection_command
+151,99053,"Untitled-1",50,0,"",plaintext,selection_command
+152,99230,"Untitled-1",31,0,"",plaintext,selection_command
+153,100215,"Untitled-1",50,0,"",plaintext,selection_command
+154,103706,"Untitled-1",49,0,"",plaintext,selection_command
+155,107964,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",0,0,"#!/usr/bin/env python3\nimport sys\nfrom pathlib import Path\nimport pandas as pd\n\ninput_dir = sys.argv[1]\nfor parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)\n    for text in df['text']:\n        print(text)\n",python,tab
+156,115203,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",23,0,"",python,selection_command
+157,115596,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",34,0,"",python,selection_command
+158,115603,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",59,0,"",python,selection_command
+159,115605,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",79,0,"",python,selection_command
+160,115608,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",80,0,"",python,selection_command
+161,115611,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,0,"",python,selection_command
+162,115615,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",173,0,"",python,selection_command
+163,115640,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",212,0,"",python,selection_command
+164,115675,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",240,0,"",python,selection_command
+165,115709,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"",python,selection_command
+166,116067,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"\n",python,content
+167,117091,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"d",python,content
+168,117093,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"",python,selection_keyboard
+169,117214,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"e",python,content
+170,117217,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0,"",python,selection_keyboard
+171,117407,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0,"f",python,content
+172,117409,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"",python,selection_keyboard
+173,117471,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0," ",python,content
+174,117474,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_keyboard
+175,127817,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"h",python,content
+176,127820,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",266,0,"",python,selection_keyboard
+177,127910,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",266,0,"e",python,content
+178,127913,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",267,0,"",python,selection_keyboard
+179,128089,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",267,0,"l",python,content
+180,128092,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",268,0,"",python,selection_keyboard
+181,128217,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",268,0,"l",python,content
+182,128220,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",269,0,"",python,selection_keyboard
+183,128320,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",269,0,"o",python,content
+184,128322,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",270,0,"",python,selection_keyboard
+185,128772,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",270,0,"_",python,content
+186,128774,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",271,0,"",python,selection_keyboard
+187,129253,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",271,0,"w",python,content
+188,129255,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",272,0,"",python,selection_keyboard
+189,129337,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",272,0,"o",python,content
+190,129340,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"",python,selection_keyboard
+191,129434,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"r",python,content
+192,129437,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",274,0,"",python,selection_keyboard
+193,129667,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",274,0,"l",python,content
+194,129670,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",275,0,"",python,selection_keyboard
+195,129865,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",275,0,"d",python,content
+196,129867,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",276,0,"",python,selection_keyboard
+197,129958,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",276,0,"()",python,content
+198,129960,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",277,0,"",python,selection_keyboard
+199,130057,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",277,1,")",python,content
+200,130059,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",278,0,"",python,selection_keyboard
+201,130836,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",278,0,":",python,content
+202,130838,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",279,0,"",python,selection_keyboard
+203,131211,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",279,0,"\n    ",python,content
+204,131458,"TERMINAL",0,0,"sed -i '12,12c\ndef' /home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py && cat -n /home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py | sed -n '1,12p'",,terminal_command
+205,131475,"TERMINAL",0,0,"]633;C     1\t#!/usr/bin/env python3\r\n     2\timport sys\r\n     3\tfrom pathlib import Path\r\n     4\timport pandas as pd\r\n     5\t\r\n     6\tinput_dir = sys.argv[1]\r\n     7\tfor parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\r\n     8\t    df = pd.read_parquet(parquet_file)\r\n     9\t    for text in df['text']:\r\n    10\t        print(text)\r\n]0;franz.srambical@hai-login1:~",,terminal_output
+206,133387,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_command
+207,133707,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",284,0,"",python,selection_command
+208,141513,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_command
+209,141820,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",284,0,"",python,selection_command
+210,141997,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_command
+211,142369,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",284,0,"",python,selection_command
+212,143758,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",284,0,"p",python,content
+213,143760,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",285,0,"",python,selection_keyboard
+214,143956,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",285,0,"r",python,content
+215,143959,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",286,0,"",python,selection_keyboard
+216,144089,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",286,0,"i",python,content
+217,144092,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",287,0,"",python,selection_keyboard
+218,144128,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",287,0,"n",python,content
+219,144130,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",288,0,"",python,selection_keyboard
+220,144204,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",288,0,"t",python,content
+221,144207,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",289,0,"",python,selection_keyboard
+222,144606,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",289,0,"()",python,content
+223,144608,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",290,0,"",python,selection_keyboard
+224,144689,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",290,1,")",python,content
+225,144691,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",291,0,"",python,selection_keyboard
+226,145114,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",290,0,"",python,selection_command
+227,145433,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",290,0,"""""",python,content
+228,145435,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",291,0,"",python,selection_keyboard
+229,145574,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",291,1,"""",python,content
+230,145577,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",292,0,"",python,selection_keyboard
+231,153755,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",280,13,"    print("""")\n",python,content
+232,175879,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",293,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-4f0212e3-f158-4084-9812-b477f28dea1e1763058437317-2025_11_13-19.27.56.165/source.csv

@@ -0,0 +1,5 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,6,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"#!/usr/bin/env python3\n""""""\nCommon utilities for dataset serialization scripts.\n""""""\n\nfrom __future__ import annotations\n\nfrom dataclasses import dataclass\nfrom pathlib import Path\nfrom typing import List, Optional, Tuple, Dict\n\nimport difflib\nimport re\nimport pandas as pd\nfrom datasets import Dataset, load_dataset\n\n\n_ANSI_CSI_RE = re.compile(r""\x1b\[[0-9;?]*[ -/]*[@-~]"")\n_ANSI_OSC_TERMINATED_RE = re.compile(r""\x1b\][\s\S]*?(?:\x07|\x1b\\)"")\n_ANSI_OSC_LINE_FALLBACK_RE = re.compile(r""\x1b\][^\n]*$"")\n_BRACKETED_PASTE_ENABLE = ""\x1b[?2004h""\n_BRACKETED_PASTE_DISABLE = ""\x1b[?2004l""\n_OSC_633 = ""\x1b]633;""\n_OSC_0 = ""\x1b]0;""\n\n\n@dataclass\nclass SerializeConfig:\n    output_dir: str\n    shard_size: int\n    target_chars: int\n    overlap_chars: int\n    min_session_chars: int\n    max_docs: Optional[int]\n    long_pause_threshold_ms: int\n    csv_root: Optional[str]\n    val_ratio: float\n    arrayrecord_group_size: Optional[int] = None\n\n\ndef _clean_text(text: str) -> str:\n    # Normalize line endings and strip trailing spaces; preserve tabs/newlines.\n    return text.replace(""\r\n"", ""\n"").replace(""\r"", ""\n"").rstrip()\n\n\ndef _fenced_block(path: str, language: Optional[str], content: str) -> str:\n    lang = (language or """").lower()\n    return f""```{lang}\n{content}\n```\n""\n\n\ndef _apply_change(content: str, offset: int, length: int, new_text: str) -> str:\n    # Mirrors crowd_code_player.replay_file.apply_change\n    base = str(content)\n    text = str(new_text) if pd.notna(new_text) else """"\n    text = text.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n    if offset > len(base):\n        base = base + ("" "" * (offset - len(base)))\n    return base[:offset] + text + base[offset + length:]\n\n\ndef _apply_backspaces(text: str) -> str:\n    out: List[str] = []\n    for ch in text:\n        if ch == ""\b"":  # \x08\n            if out:\n                out.pop()\n        else:\n            out.append(ch)\n    return """".join(out)\n\n\ndef _normalize_terminal_output(raw: str) -> str:\n    """"""\n    Normalize PTY/terminal output for training:\n      - Apply backspaces (\x08)\n      - Strip OSC (window title/shell integration) first, keeping BEL/ST terminators intact\n      - Resolve carriage returns (\r) by keeping the last rewrite per line\n      - Strip CSI (coloring etc.)\n      - Finally drop any remaining BEL (\x07)\n    """"""\n    if not raw:\n        return raw\n    s = _apply_backspaces(raw)\n    # Remove OSC sequences that are properly terminated (BEL or ST)\n    s = _ANSI_OSC_TERMINATED_RE.sub("""", s)\n    # Fallback: drop any unterminated OSC up to end-of-line only\n    s = ""\n"".join(_ANSI_OSC_LINE_FALLBACK_RE.sub("""", line) for line in s.split(""\n""))\n    # Resolve carriage returns per line:\n    # - If there are multiple rewrites, keep the last non-empty chunk\n    # - If it's CRLF (ending with '\r' before '\n'), keep the content before '\r'\n    resolved_lines: List[str] = []\n    for seg in s.split(""\n""):\n        parts = seg.split(""\r"")\n        chosen = """"\n        # pick last non-empty part if available; else last part\n        for p in reversed(parts):\n            if p != """":\n                chosen = p\n                break\n        if chosen == """" and parts:\n            chosen = parts[-1]\n        resolved_lines.append(chosen)\n    s = ""\n"".join(resolved_lines)\n    # Strip ANSI escape sequences\n    s = _ANSI_CSI_RE.sub("""", s)\n    # Remove any remaining BEL beeps\n    s = s.replace(""\x07"", """")\n    return s\n\n\ndef _line_numbered_output(content: str, start_line: Optional[int] = None, end_line: Optional[int] = None) -> str:\n    # TODO (f.srambical): check whether this corresponds **exactly** to the output of cat -n {file_path} | sed -n '{vstart},{vend}p'\n    lines = content.splitlines()\n    total = len(lines)\n    if total == 0:\n        return """"\n    s = 1 if start_line is None else max(1, min(start_line, total))\n    e = total if end_line is None else max(1, min(end_line, total))\n    if e < s:\n        # FIXME (f.srambical): If this does not happen, remove the condition\n        raise ValueError(""This should never happen!"")\n        e = s\n    buf: List[str] = []\n    for idx in range(s, e + 1):\n        buf.append(f""{idx:6}\t{lines[idx - 1]}"")\n    return ""\n"".join(buf)\n\n\ndef _compute_viewport(total_lines: int, center_line: int, radius: int) -> Tuple[int, int]:\n    if total_lines <= 0:\n        return (1, 0)\n    start = max(1, center_line - radius)\n    end = min(total_lines, center_line + radius)\n    if end < start:\n        # FIXME (f.srambical): If this does not happen, remove the condition\n        raise ValueError(""This should never happen!"")\n    return (start, end)\n\n\ndef _escape_single_quotes_for_sed(text: str) -> str:\n    # Close quote, add an escaped single quote, reopen quote: '""'""'\n    return text.replace(""'"", ""'\""'\""'"")\n\n\ndef _compute_changed_block_lines(before: str, after: str) -> Tuple[int, int, List[str]]:\n    """"""\n    Return 1-based start and end line numbers in 'before' that should be replaced,\n    and the replacement lines from 'after'.\n    For pure deletions, the replacement list may be empty.\n    """"""\n    before_lines = before.splitlines()\n    after_lines = after.splitlines()\n    sm = difflib.SequenceMatcher(a=before_lines, b=after_lines, autojunk=False)\n    opcodes = [op for op in sm.get_opcodes() if op[0] != ""equal""]\n    if not opcodes:\n        # FIXME (f.srambical): clean this up\n        raise ValueError(""No diff opcodes found for content change"")\n                # No visible change; choose a safe single-line replace at end of file\n        start_line = max(1, len(before_lines))\n        end_line = start_line\n        repl = after_lines[start_line - 1:start_line] if after_lines else [""""]\n        return (start_line, end_line, repl)\n\n    first = opcodes[0]\n    last = opcodes[-1]\n    # i1/i2 refer to 'before' indices, j1/j2 to 'after'\n    start_line = (first[1] + 1) if (first[1] + 1) > 0 else 1\n    end_line = last[2] # no increment since we go from 'exclusive' to 'inclusive' indexing\n    replacement_lines = after_lines[first[3]:last[4]]\n    return (start_line, end_line, replacement_lines)\n\n\ndef _session_to_transcript(\n    df: pd.DataFrame,\n    long_pause_threshold_ms: int,\n) -> str:\n\n    file_states: Dict[str, str] = {}\n    terminal_state: str = """"\n    per_file_event_counts: Dict[str, int] = {}\n    per_file_cursor_positions: Dict[str, Tuple[int, int]] = {}  # (offset, length) for each file\n    last_time_ms: Optional[int] = None\n\n    parts: List[str] = []\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_time: int = row[""Time""]\n        language: Optional[str] = row[""Language""]\n\n        # Long pause detection\n        if last_time_ms is not None:\n            delta = event_time - last_time_ms\n            if delta > long_pause_threshold_ms:\n                # TODO (f.srambical): think about whether we want to emit this as an observation or not\n                parts.append(f""<obs long_pause ms=\""{delta}\"" />"")\n        last_time_ms = event_time\n\n        event_type = row[""Type""]\n\n        match event_type:\n            case ""tab"":\n                # File switch event\n                parts.append(f""<act focus file=\""{file_path}\"" />"")\n                \n                # If Text is present, this is the first time opening the file\n                # and the entire file content is captured\n                text = row[""Text""]\n                if pd.notna(text):\n                    file_content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = file_content\n                    parts.append(f""// observation: file={file_path}"")\n                    parts.append(_fenced_block(file_path, language, _clean_text(file_content)))\n\n            case ""terminal_command"":\n                # Terminal command execution\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<act terminal_command />"")\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(command_str)))\n\n            case ""terminal_output"":\n                # Terminal output capture\n                output = row[""Text""]\n                output_str = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<obs terminal_output />"")\n                parts.append(_fenced_block(file_path, None, _clean_text(output_str)))\n\n            case ""terminal_focus"":\n                # Terminal focus event\n                parts.append(f""<act focus target=\""terminal\"" />"")\n\n            case ""git_branch_checkout"":\n                # Git branch checkout event\n                branch_info = row[""Text""]\n                branch_str = str(branch_info).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<act git_branch_checkout />"")\n                parts.append(f""// git: {_clean_text(branch_str)}"")\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                # Handle cursor movement\n                offset = row[""RangeOffset""]\n                length = row[""RangeLength""]\n                old_cursor = per_file_cursor_positions.get(file_path, (0, 0))\n                new_cursor = (offset, length)\n                per_file_cursor_positions[file_path] = new_cursor\n                \n                # Emit cursor movement observation if position changed\n                if old_cursor != new_cursor:\n                    parts.append(f""<act cursor file=\""{file_path}\"" offset=\""{offset}\"" len=\""{length}\"" />"")\n\n            case ""content"":\n                # Handle file edit events\n                offset = row[""RangeOffset""]\n                length = row[""RangeLength""]\n                new_text = row[""Text""]\n                new_text_str = str(new_text) if pd.notna(new_text) else """"\n\n                operation = ""noop""\n                if length == 0 and new_text_str:\n                    operation = ""insert""\n                elif length > 0 and not new_text_str:\n                    operation = ""delete""\n                elif length > 0 and new_text_str:\n                    operation = ""replace""\n\n                parts.append(f""<act {operation} file=\""{file_path}\"" offset=\""{offset}\"" len=\""{length}\"" />"")\n\n                if new_text_str and (operation == ""insert"" or operation == ""replace""):\n                    parts.append(_fenced_block(file_path, language, _clean_text(new_text_str)))\n\n                before = file_states.get(file_path, """")\n                after = _apply_change(before, offset, length, new_text)\n                file_states[file_path] = after\n                per_file_event_counts[file_path] = per_file_event_counts.get(file_path, 0) + 1\n\n                # Update cursor position after edit (cursor moves to end of inserted/replaced text)\n                per_file_cursor_positions[file_path] = (offset + len(new_text_str), 0)\n\n            case _:\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    return ""\n"".join(parts).strip()\n\n\ndef session_to_bash_formatted_transcript(\n    df: pd.DataFrame,\n    viewport_radius: int = 10,\n    normalize_terminal_output: bool = True,\n    coalesce_radius: int = 5,\n) -> str:\n    r""""""\n    Serialize a session to a bash-like transcript comprised of:\n      - Commands (bash fenced blocks): cat -n, sed -i 'S,Ec\...' && cat -n | sed -n 'VSTART,VENDp'\n      - Outputs (<stdout>...</stdout>) that reflect the file state after each action\n    Tracks per-file state and a per-file viewport. Viewport only shifts when selection moves out of bounds\n    or when first initialized.\n    """"""\n    file_states: Dict[str, str] = {}\n    per_file_viewport: Dict[str, Optional[Tuple[int, int]]] = {}\n\n    parts: List[str] = []\n    terminal_output_buffer: List[str] = []\n    pending_edits_before: Dict[str, Optional[str]] = {}\n    pending_edit_regions: Dict[str, Optional[Tuple[int, int]]] = {}\n\n    def _flush_terminal_output_buffer() -> None:\n        if not terminal_output_buffer:\n            return\n        aggregated = """".join(terminal_output_buffer)\n        out = aggregated\n        if normalize_terminal_output:\n            out = _normalize_terminal_output(out)\n        cleaned = _clean_text(out)\n        if cleaned.strip():\n            parts.append(f""<stdout>\n{cleaned}\n</stdout>"")\n        terminal_output_buffer.clear()\n\n    def _flush_pending_edit_for_file(target_file: str) -> None:\n        before_snapshot = pending_edits_before.get(target_file)\n        if before_snapshot is None:\n            return\n        after_state = file_states.get(target_file, """")\n        try:\n            start_line, end_line, repl_lines = _compute_changed_block_lines(before_snapshot, after_state)\n        except ValueError:\n            pending_edits_before[target_file] = None\n            return\n        before_total_lines = len(before_snapshot.splitlines())\n        if end_line < start_line:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            if start_line <= max(1, before_total_lines):\n                sed_cmd = f""sed -i '{start_line}i\\\n{sed_payload}' {target_file}""\n            else:\n                sed_cmd = f""sed -i '$a\\\n{sed_payload}' {target_file}""\n        elif not repl_lines:\n            sed_cmd = f""sed -i '{start_line},{end_line}d' {target_file}""\n        else:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            sed_cmd = f""sed -i '{start_line},{end_line}c\\\n{sed_payload}' {target_file}""\n        total_lines = len(after_state.splitlines())\n        center = (start_line + end_line) // 2\n        vp = _compute_viewport(total_lines, center, viewport_radius)\n        per_file_viewport[target_file] = vp\n        vstart, vend = vp\n        chained_cmd = f""{sed_cmd} && cat -n {target_file} | sed -n '{vstart},{vend}p'""\n        breakpoint()\n        parts.append(_fenced_block(target_file, ""bash"", _clean_text(chained_cmd)))\n        viewport_output = _line_numbered_output(after_state, vstart, vend)\n        parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n        pending_edits_before[target_file] = None\n        pending_edit_regions[target_file] = None\n\n    def _flush_all_pending_edits() -> None:\n        for fname in list(pending_edits_before.keys()):\n            _flush_pending_edit_for_file(fname)\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_type = row[""Type""]\n        \n        match event_type:\n            case ""tab"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                text = row[""Text""]\n                if pd.notna(text):\n                    content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = content\n                    # First open with full file capture\n                    cmd = f""cat -n {file_path}""\n                    parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                    output = _line_numbered_output(content)\n                    parts.append(f""<stdout>\n{output}\n</stdout>"")\n                else:\n                    # File switch without content snapshot: show current viewport only\n                    content = file_states.get(file_path, """")\n                    total_lines = len(content.splitlines())\n                    vp = per_file_viewport.get(file_path)\n                    if not vp or vp[1] == 0:\n                        vp = _compute_viewport(total_lines, 1, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                    if vp:\n                        vstart, vend = vp\n                        cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                        parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                        viewport_output = _line_numbered_output(content, vstart, vend)\n                        parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n\n            case ""content"":\n                _flush_terminal_output_buffer()\n                offset = int(row[""RangeOffset""])\n                length = int(row[""RangeLength""])\n                new_text = row[""Text""]\n                before = file_states.get(file_path, """")\n                # Approximate current edit region in line space\n                new_text_str = str(new_text) if pd.notna(new_text) else """"\n                start_line_current = before[:offset].count(""\n"") + 1\n                deleted_chunk = before[offset:offset + length]\n                lines_added = new_text_str.count(""\n"")\n                lines_deleted = deleted_chunk.count(""\n"")\n                region_start = start_line_current\n                region_end = start_line_current + max(lines_added, lines_deleted, 0)\n                # Flush pending edits if this edit is far from the pending region\n                current_region = pending_edit_regions.get(file_path)\n                if current_region is not None:\n                    rstart, rend = current_region\n                    if region_start < (rstart - coalesce_radius) or region_start > (rend + coalesce_radius):\n                        _flush_pending_edit_for_file(file_path)\n                        current_region = None\n                after = _apply_change(before, offset, length, new_text)\n                if pending_edits_before.get(file_path) is None:\n                    pending_edits_before[file_path] = before\n                # Update/initialize region union\n                if current_region is None:\n                    pending_edit_regions[file_path] = (region_start, max(region_start, region_end))\n                else:\n                    rstart, rend = current_region\n                    pending_edit_regions[file_path] = (min(rstart, region_start), max(rend, region_end))\n                file_states[file_path] = after\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                # During an edit burst (pending edits), suppress flush and viewport emissions\n                if pending_edits_before.get(file_path) is None:\n                    _flush_terminal_output_buffer()\n                else:\n                    # Skip emitting viewport while edits are pending to avoid per-keystroke sed/cat spam\n                    continue\n                offset = int(row[""RangeOffset""])\n                content = file_states.get(file_path, """")\n                total_lines = len(content.splitlines())\n                target_line = content[:offset].count(""\n"") + 1\n                vp = per_file_viewport.get(file_path)\n                should_emit = False\n                if not vp or vp[1] == 0:\n                    vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                    per_file_viewport[file_path] = vp\n                    should_emit = True\n                else:\n                    vstart, vend = vp\n                    if target_line < vstart or target_line > vend:\n                        vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                        should_emit = True\n                if should_emit and vp:\n                    vstart, vend = vp\n                    cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                    parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                    viewport_output = _line_numbered_output(content, vstart, vend)\n                    parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n\n            case ""terminal_command"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(command_str)))\n\n            case ""terminal_output"":\n                output = row[""Text""]\n                raw_output = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                terminal_output_buffer.append(raw_output)\n\n            case ""terminal_focus"" | ""git_branch_checkout"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                # FIXME (f.srambical): handle these events \n                pass\n\n            case _:\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    _flush_all_pending_edits()\n    _flush_terminal_output_buffer()\n    return ""\n"".join(parts).strip()\n\ndef load_hf_csv(hf_path: str, split: str) -> Dataset:\n    loaded = load_dataset(hf_path, split=split)\n\n    assert isinstance(loaded, Dataset), ""Expected a Dataset from load_dataset""\n    return loaded\n\n\ndef _discover_local_sessions(root: Path) -> List[Path]:\n    # Recursively find all CSV files\n    paths: List[Path] = []\n    for p in root.rglob(""*.csv""):\n        if p.is_file():\n            paths.append(p)\n    paths.sort()\n    return paths\n\n\ndef _chunk_text(text: str, target_chars: int, overlap_chars: int) -> List[str]:\n    """"""Split a long text into overlapping chunks near target length.""""""\n    if target_chars <= 0:\n        return [text]\n    n = len(text)\n    if n <= target_chars:\n        return [text]\n\n    chunks: List[str] = []\n    start = 0\n    # Ensure sane overlap\n    overlap = max(0, min(overlap_chars, target_chars // 2))\n    while start < n:\n        end_target = min(start + target_chars, n)\n        if end_target < n:\n            end = end_target\n        else:\n            end = n\n        chunk = text[start:end].strip()\n        chunks.append(chunk)\n        if end == n:\n            break\n        # advance with overlap\n        start = max(0, end - overlap)\n        if start >= n:\n            break\n    return chunks\n\n\n",python,tab
+2,395,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:27:56 PM [info] Activating crowd-code\n7:27:56 PM [info] Recording started\n7:27:56 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,704,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"7:27:56 PM [info] Git repository found\n7:27:56 PM [info] Git provider initialized successfully\n7:27:56 PM [info] Initial git state: [object Object]\n",Log,content
+4,139066,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"",python,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-51002999-75fd-453b-98e2-6003b6e5c8e61755511610416-2025_08_18-12.06.53.598/source.csv

@@ -0,0 +1,19 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"megatron/core/optimizer/optimizer_config.py",0,0,"# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.\n\nfrom dataclasses import dataclass\nfrom typing import Callable, Optional\n\nimport torch\n\nfrom ..utils import is_te_min_version\n\n\n@dataclass\nclass OptimizerConfig:\n    """"""Configuration for optimizer.""""""\n\n    ##############\n    # General\n    ##############\n    optimizer: str = 'adam'\n    """"""Optimizer to use (one of Adam or SGD).""""""\n\n    lr: Optional[float] = None\n    """"""Initial learning rate. Depending on decay style and initial warmup, the learning rate at each\n       iteration would be different.\n    """"""\n\n    min_lr: Optional[float] = None\n    """"""Minumum value for learning rate. The scheduler clip values below this threshold.""""""\n\n    decoupled_lr: Optional[float] = None\n    """"""Separate learning rate for the input and output layer.""""""\n\n    decoupled_min_lr: Optional[float] = None\n    """"""Minimum value for learning rate for the input and output layer. The scheduler clip values\n       below this threshold.\n    """"""\n\n    weight_decay: float = 0.01\n    """"""Weight decay coefficient for L2 regularization.""""""\n\n    ##############\n    # Precision\n    ##############\n    fp8_recipe: Optional[str] = None\n    """"""The type of fp8 recipe will affect the processing logic inside distributed optimizer.""""""\n\n    fp16: bool = False\n    """"""If true, train with fp16 mixed precision training. Defaults to False.""""""\n\n    bf16: bool = False\n    """"""If true, train with bf16 mixed precision training. Defaults to False.""""""\n\n    reuse_grad_buf_for_mxfp8_param_ag: bool = False\n    """"""If true, reuse the grad buffer for param AG when using mxfp8 recipe. Should be \n       set to True only when fp8_recipe is mxfp8 and fp8_param_gather is True.""""""\n\n    params_dtype: torch.dtype = torch.float32\n    """"""dtype used when intializing the weights. Defaults to torch.float32.""""""\n\n    use_precision_aware_optimizer: bool = False\n    """"""If true, allows optimizer-related tensors (master_param, gradients and optimizer states)\n    to be set to lower precision. Defaults to False.\n    """"""\n\n    store_param_remainders: bool = True\n    """"""If true, store the 16-bit FP32 parameter remainders in the optimizer state, excluding the\n        16 bits shared with the BF16 parameters. This lowers GPU memory usage. Defaults to True.\n    """"""\n\n    main_grads_dtype: torch.dtype = torch.float32\n    """"""dtype of main grads when enabling precision-aware-optimizer""""""\n\n    main_params_dtype: torch.dtype = torch.float32\n    """"""dtype of main params when enabling precision-aware-optimizer""""""\n\n    exp_avg_dtype: torch.dtype = torch.float32\n    """"""dtype of exp_avg when enabling precision-aware-optimizer""""""\n\n    exp_avg_sq_dtype: torch.dtype = torch.float32\n    """"""dtype of exp_avg_sq when enabling precision-aware-optimizer""""""\n\n    ###############\n    # Loss scaling\n    ###############\n    loss_scale: Optional[float] = None\n    """"""Static loss scaling, positive power of 2 values can improve fp16 convergence. If None,\n       dynamic loss scaling is used.\n    """"""\n\n    initial_loss_scale: float = 2**32\n    """"""Initial loss-scale for dynamic loss scaling.""""""\n\n    min_loss_scale: float = 1.0\n    """"""Minimum loss scale for dynamic loss scaling.""""""\n\n    loss_scale_window: float = 1000\n    """"""Window over which to raise/lower dynamic scale.""""""\n\n    hysteresis: int = 2\n    """"""Hysteresis for dynamic loss scaling.""""""\n\n    ##############\n    # Optimizer\n    ##############\n    # Adam\n    adam_beta1: float = 0.9\n    """"""First coefficient for computing running averages of gradient and its square in Adam\n    optimizer.\n    """"""\n\n    adam_beta2: float = 0.999\n    """"""Second coefficient for computing running averages of gradient and its square in Adam\n    optimizer.\n    """"""\n\n    adam_eps: float = 1e-08\n    """"""Term added to the denominator to improve numerical stability in Adam optimizer.""""""\n\n    # SGD.\n    sgd_momentum: float = 0.9\n    """"""Momentum factor for SGD optimizer.""""""\n\n    #######################\n    # Distributed optimizer\n    #######################\n    use_distributed_optimizer: bool = False\n    """"""Distribute optimizer state over data-parallel replicas.""""""\n\n    overlap_param_gather: bool = False\n    """"""If true, overlap param all-gather with forward compute. \n        This argument is intended to have the same value as the ""overlap_param_gather"" argument \n        in the ""distributed_data_parallel_config.py"" file. In the optimizer, this argument is \n        only used when ""reuse_grad_buf_for_mxfp8_param_ag=True & fp8_param_gather=True"".\n    """"""\n\n    overlap_param_gather_with_optimizer_step: bool = False\n    """"""If true, overlap param all-gather of first bucket with optimizer step.""""""\n\n    #######################\n    # Optimizer Offload\n    #######################\n\n    optimizer_cpu_offload: bool = False\n    """"""If True, offload optimizer states tensor and compute to CPU.""""""\n\n    optimizer_offload_fraction: float = 0.0\n    """"""Specifies the fraction of optimizer states to offload from GPU memory to CPU.""""""\n\n    use_torch_optimizer_for_cpu_offload: bool = False\n    """"""If True, use torch.optim.Optimizer for CPU offload.""""""\n\n    overlap_cpu_optimizer_d2h_h2d: bool = False\n    """"""\n    When set to `True`, this flag enables overlapping of the CPU optimizer\n    update process with the data transfer operations. This can help improve\n    overall training efficiency by reducing idle time during data movement,\n    allowing the optimizer to perform updates while gradients and parameters\n    are being transferred between devices.\n    """"""\n\n    pin_cpu_grads: bool = True\n    """"""If True, pin the optimizer gradients to CPU memory.""""""\n\n    pin_cpu_params: bool = True\n    """"""If True, pin the optimizer parameters to CPU memory.""""""\n\n    ################\n    # Miscellaneous\n    ################\n    clip_grad: float = 1.0\n    """"""Gradient clipping based on global L2 norm.""""""\n\n    log_num_zeros_in_grad: bool = False\n    """"""If true, calculate and log the number of zeros in gradient.""""""\n\n    barrier_with_L1_time: bool = False\n    """"""If true, use barrier with level 1 time measurements.""""""\n\n    timers: Optional[Callable] = None\n    """"""Function to get timers.""""""\n\n    config_logger_dir: str = """"\n    """"""When non-empty, dumps entry-point configs to config_logger_dir""""""\n\n    def __post_init__(self):\n        """"""Check the validity of the config.""""""\n\n        # The following condition is used to avoid repetition in distrib_optimizer.py.\n        # This is because in distrib_optimizer.py, the process to handle parameters are\n        # different for different training precision settings. FP8 cases require different\n        # handling while FP8 delayed scaling is an exception because the Adam optimizer in\n        # TransformerEngine supports it in the kernel computation.\n        # This is also the flag to determine the usage of param.grad or param.decoupled_grad\n        self.use_precision_aware_optimizer_no_fp8_or_ds_fp8 = (\n            self.use_precision_aware_optimizer\n            and (\n                self.main_params_dtype != torch.float32\n                or (self.fp8_recipe is None or self.fp8_recipe == ""delayed"")\n                or self.optimizer_cpu_offload\n            )\n        )\n\n        if self.fp8_recipe == ""mxfp8"":\n            if not self.reuse_grad_buf_for_mxfp8_param_ag:\n                import warnings\n\n                warnings.warn(\n                    ""mxfp8 without using reuse_grad_buf_for_mxfp8_param_ag and fp8_param_gather""\n                    ""will use significant amount additional GPU memory.""\n                    ""Setting --reuse-grad-buf-for-mxfp8-param-ag and --fp8-param-gather is ""\n                    ""recommended for mxfp8 training.""\n                )\n\n        if self.use_precision_aware_optimizer:\n            assert (\n                self.optimizer == 'adam'\n            ), '--use-precision-aware-optimizer only supported with adam'\n            assert (\n                self.use_distributed_optimizer\n            ), '--use-precision-aware-optimizer only supported with distributed optimizer'\n\n            if not is_te_min_version(""2.1.0""):\n                self.store_param_remainders = False\n\n            # Only the FusedAdam in TE and HybridDeviceOptimizer supports\n            # --use-precision-aware-optimizer.\n            # TODO: Remove this check when apex's FusedAdam is no longer used.\n            if self.optimizer_cpu_offload:\n                return\n            try:\n                import inspect\n\n                from transformer_engine.pytorch.optimizers import FusedAdam as Adam\n\n                adam_args = inspect.signature(Adam).parameters\n                arg_names = [\n                    'master_weight_dtype',\n                    'exp_avg_dtype',\n                    'exp_avg_sq_dtype',\n                    'use_decoupled_grad',\n                ]\n                for name in arg_names:\n                    assert name in adam_args, (\n                        ""Current FusedAdam of TE doesn't support --use-precision-aware-optimizer, ""\n                        ""please update TE version.""\n                    )\n            except ImportError:\n                raise RuntimeError(\n                    '--use-precision-aware-optimizer requires FusedAdam from TransformerEngine, '\n                    'but not found.'\n                )\n        else:\n            assert (\n                self.main_grads_dtype == torch.float32\n            ), ""main_grads_dtype can only be fp32 when not using precision-aware optimizer""\n            assert (\n                self.main_params_dtype == torch.float32\n            ), ""main_params_dtype can only be fp32 when not using precision-aware optimizer""\n            assert (\n                self.exp_avg_dtype == torch.float32\n            ), ""exp_avg_dtype can only be fp32 when not using precision-aware optimizer""\n            assert (\n                self.exp_avg_sq_dtype == torch.float32\n            ), ""exp_avg_sq_dtype can only be fp32 when not using precision-aware optimizer""\n",python,tab
+2,65,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"12:06:53 PM [info] Activating crowd-code\n12:06:53 PM [info] Recording started\n12:06:53 PM [info] Initializing git provider using file system watchers...\n12:06:53 PM [info] Git repository found\n12:06:53 PM [info] Git provider initialized successfully\n",Log,tab
+3,187,"extension-output-pdoom-org.crowd-code-#1-crowd-code",250,0,"12:06:53 PM [info] Initial git state: [object Object]\n",Log,content
+4,5044208,"megatron/core/optimizer/optimizer_config.py",0,0,"",python,tab
+5,5332056,"TERMINAL",0,0,"",,terminal_focus
+6,5466765,"megatron/core/transformer/module.py",0,0,"# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.\n\n""""""Megatron Module.""""""\nfrom typing import Optional, Tuple\n\nimport torch\nfrom torch.autograd import Variable\nfrom torch.nn.parameter import Parameter\n\nfrom megatron.core import parallel_state\nfrom megatron.core.dist_checkpointing.mapping import ShardedStateDict\nfrom megatron.core.transformer.transformer_config import TransformerConfig\nfrom megatron.core.transformer.utils import (\n    make_sharded_tensors_for_checkpoint,\n    sharded_state_dict_default,\n)\n\n_FLOAT_TYPES = (torch.FloatTensor, torch.cuda.FloatTensor)\n_HALF_TYPES = (torch.HalfTensor, torch.cuda.HalfTensor)\n_BF16_TYPES = (torch.BFloat16Tensor, torch.cuda.BFloat16Tensor)\n\n\ndef param_is_not_shared(param):  # pylint: disable=missing-function-docstring\n    return not hasattr(param, 'shared') or not param.shared\n\n\nclass MegatronModule(torch.nn.Module):\n    """"""Base Megatron module inhertied by all Models.\n\n    Megatron specific extensions of torch Module with support\n    for pipelining\n\n    Args:\n        config (TransformerConfig): Transformer config\n    """"""\n\n    # def __init__(self, config: TransformerConfig, share_word_embeddings=True):\n    def __init__(self, config: TransformerConfig):\n        super().__init__()\n        self.config = config\n\n    def state_dict_for_save_checkpoint(self, prefix: str = '', keep_vars: bool = False):\n        """"""Override state dict for saving checkpoints Use this function to override the\n        state dict for saving checkpoints.\n\n        Args:\n            prefix (str, optional): _description_. Defaults to ''.\n            keep_vars (bool, optional): _description_. Defaults to False.\n\n        Returns:\n            _type_: _description_\n        """"""\n\n        return self.state_dict(prefix=prefix, keep_vars=keep_vars)\n\n    def sharded_state_dict(\n        self,\n        prefix: str = '',\n        sharded_offsets: Tuple[Tuple[int, int, int]] = (),\n        metadata: Optional[dict] = None,\n    ) -> ShardedStateDict:\n        """"""Default implementation for sharded state dict for distributed checkpointing.\n\n        General definition of sharded_state_dict simply calls `sharded_state_dict_default`\n        (which call sharded_state_dict method if possible or a default implementation otherwise)\n        recursively on all submodules.\n\n        Args:\n            prefix (str): prefix for the state dict keys\n            sharded_offsets (Tuple[Tuple[int, int, int]], optional): sharding already\n                applied (e.g. PP related) by sup-modules. Passed along to ShardedTensor\n            metadata (dict, optional): metadata passed recursively to sharded_state_dict methods\n\n        Returns:\n            dict: dictionary of state dict keys mapped to ShardedTensors\n        """"""\n        sharded_state_dict = {}\n        # Save parameters\n        self._save_to_state_dict(sharded_state_dict, '', keep_vars=True)\n        sharded_state_dict = make_sharded_tensors_for_checkpoint(\n            sharded_state_dict, prefix, sharded_offsets=sharded_offsets\n        )\n        # Recurse into submodules\n        for name, module in self.named_children():\n            sharded_state_dict.update(\n                sharded_state_dict_default(module, f'{prefix}{name}.', sharded_offsets, metadata)\n            )\n        return sharded_state_dict\n\n    def set_is_first_microbatch(self):\n        """"""Sets the is_first_microbatch flag if it exists and config.fp8==True.\n        When this flag is set, TE modules will update their fp8 parameter cache.\n        If kitchen is being used, kitchen controls quantization level.\n        """"""\n        if self.config.fp8 is not None or getattr(self.config, 'use_kitchen', False):\n            if not hasattr(self, ""modules_with_is_first_microbatch""):\n                self.modules_with_is_first_microbatch = []\n                for m in self.modules():\n                    if hasattr(m, ""is_first_microbatch""):\n                        self.modules_with_is_first_microbatch.append(m)\n            for m in self.modules_with_is_first_microbatch:\n                m.is_first_microbatch = True\n\n    def set_symmetric_ar(self, set_to: Optional[str] = None) -> None:\n        """"""\n        Set symmetric all-reduce functionality across all eligible modules.\n\n        This method traverses the model's module hierarchy to find all modules\n        with the 'symmetric_ar_type' attribute, caches them, and then sets their\n        '_symmetric_ar_cache' attribute to the specified value to enable or\n        disable symmetric all-reduce operations.\n\n        Args:\n            set_to (Any, optional): Value to set for the 'symmetric_ar_type' to.\n            Allowed choices ['two_shot', ""one_shot"", ""multimem_all_reduce"", None]\n        """"""\n        assert set_to in ['two_shot', ""one_shot"", ""multimem_all_reduce"", None]\n\n        # Recursive function to find all modules with our target attributes\n        def create_ar_cache(module):\n            # Check if this module has any of our target attributes\n            if hasattr(module, ""symmetric_ar_type""):\n                self._symmetric_ar_cache.append(module)\n\n            # Check all children modules recursively\n            for child in module._modules.values():\n                if child is not None:\n                    create_ar_cache(child)\n\n        if not hasattr(self, ""_symmetric_ar_cache""):\n            self._symmetric_ar_cache = []\n            create_ar_cache(self)\n\n        for module in self._symmetric_ar_cache:\n            module._symmetric_ar_cache = set_to\n\n\ndef conversion_helper(val, conversion):\n    """"""Recursively applies a conversion function to values in nested data structures.\n\n    Args:\n        val: A single value or a nested structure (tuple/list) of values to convert\n        conversion (callable): A function that performs the desired conversion on a single value\n\n    Returns:\n        The converted value, maintaining the same nested structure as the input.\n        If input is a single value, returns the converted value.\n        If input is a tuple/list, returns a tuple/list with all elements converted.\n    """"""\n    if not isinstance(val, (tuple, list)):\n        return conversion(val)\n    rtn = [conversion_helper(v, conversion) for v in val]\n    if isinstance(val, tuple):\n        rtn = tuple(rtn)\n    return rtn\n\n\ndef fp32_to_float16(val, float16_convertor):\n    """"""Converts floating-point values from fp32 to fp16.\n\n    Args:\n        val: The value to convert. Can be a single number, a tuple, or a list.\n        float16_convertor: A function that converts a single fp32 value to fp16\n    """"""\n\n    def half_conversion(val):\n        val_typecheck = val\n        if isinstance(val_typecheck, (Parameter, Variable)):\n            val_typecheck = val.data\n        if isinstance(val_typecheck, _FLOAT_TYPES):\n            val = float16_convertor(val)\n        return val\n\n    return conversion_helper(val, half_conversion)\n\n\ndef float16_to_fp32(val):\n    """"""Converts floating-point values from fp16 to fp32.\n\n    Args:\n        val: The value to convert. Can be a single number, a tuple, or a list.\n    """"""\n\n    def float_conversion(val):\n        val_typecheck = val\n        if isinstance(val_typecheck, (Parameter, Variable)):\n            val_typecheck = val.data\n        if isinstance(val_typecheck, (_BF16_TYPES, _HALF_TYPES)):\n            val = val.float()\n        return val\n\n    return conversion_helper(val, float_conversion)\n\n\nclass Float16Module(MegatronModule):\n    """"""Float 16 Module.\n\n    Attributes:\n        config (TransformerConfig): Transformer config\n        fp16 (bool) : Specifies if the model runs in fp16 mode\n        bf16 (bool) : Specifies if the model runs in bf16 mode\n\n    Args:\n        config (TransformerConfig): The transformer config used to initalize the model\n    """"""\n\n    def __init__(self, config: TransformerConfig, module: torch.nn.Module):\n        super(Float16Module, self).__init__(config)\n        self.config = config\n        self.fp16 = config.fp16\n        self.bf16 = config.bf16\n        self.vp_stage = getattr(module, 'vp_stage', None)\n\n        if self.fp16:\n            self.add_module('module', module.half())\n\n            def float16_convertor(val):\n                return val.half()\n\n        elif self.bf16:\n            self.add_module('module', module.bfloat16())\n\n            def float16_convertor(val):\n                return val.bfloat16()\n\n        else:\n            raise Exception('Either config.fp16 or config.bf16 should be True.')\n\n        self.float16_convertor = float16_convertor\n\n    def set_input_tensor(self, input_tensor):  # pylint: disable=missing-function-docstring\n        return self.module.set_input_tensor(input_tensor)\n\n    def forward(self, *inputs, **kwargs):  # pylint: disable=missing-function-docstring\n        if parallel_state.is_pipeline_first_stage(ignore_virtual=False, vp_stage=self.vp_stage):\n            inputs = fp32_to_float16(inputs, self.float16_convertor)\n        outputs = self.module(*inputs, **kwargs)\n        if parallel_state.is_pipeline_last_stage(ignore_virtual=False, vp_stage=self.vp_stage):\n            outputs = float16_to_fp32(outputs)\n        return outputs\n\n    def state_dict(\n        self, destination=None, prefix='', keep_vars=False\n    ):  # pylint: disable=missing-function-docstring\n        return self.module.state_dict(destination=destination, prefix=prefix, keep_vars=keep_vars)\n\n    def state_dict_for_save_checkpoint(self, prefix='', keep_vars=False):\n        """"""Retrieve state_dict from the module being wrapped.""""""\n        return self.module.state_dict_for_save_checkpoint(prefix=prefix, keep_vars=keep_vars)\n\n    def sharded_state_dict(self, prefix='', *args, **kwargs):\n        """"""Retrieve sharded_state_dict from the module being wrapped.""""""\n        return self.module.sharded_state_dict(prefix, *args, **kwargs)\n\n    def load_state_dict(\n        self, state_dict, strict=True\n    ):  # pylint: disable=missing-function-docstring\n        self.module.load_state_dict(state_dict, strict=strict)\n",python,tab
+7,5466770,"megatron/core/transformer/module.py",8635,445,"    def forward(self, *inputs, **kwargs):  # pylint: disable=missing-function-docstring\n        if parallel_state.is_pipeline_first_stage(ignore_virtual=False, vp_stage=self.vp_stage):\n            inputs = fp32_to_float16(inputs, self.float16_convertor)\n        outputs = self.module(*inputs, **kwargs)\n        if parallel_state.is_pipeline_last_stage(ignore_virtual=False, vp_stage=self.vp_stage):\n            outputs = float16_to_fp32(outputs)",python,selection_command
+8,5472247,"megatron/core/transformer/module.py",9080,0,"",python,selection_mouse
+9,5472251,"megatron/core/transformer/module.py",9079,0,"",python,selection_command
+10,5474047,"megatron/core/transformer/module.py",8851,0,"",python,selection_mouse
+11,5474950,"megatron/core/transformer/module.py",9080,0,"",python,selection_mouse
+12,5474964,"megatron/core/transformer/module.py",9079,0,"",python,selection_command
+13,5476116,"megatron/core/transformer/module.py",9063,0,"",python,selection_mouse
+14,5476600,"megatron/core/transformer/module.py",9080,0,"",python,selection_mouse
+15,5476610,"megatron/core/transformer/module.py",9079,0,"",python,selection_command
+16,5962098,"megatron/core/transformer/module.py",7378,0,"",python,selection_command
+17,5962732,"megatron/core/transformer/module.py",6869,0,"",python,selection_command
+18,5963526,"megatron/core/transformer/module.py",6868,0,"",python,selection_mouse

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-55c207a6-ee6d-464d-8f77-e1220855a4f41754396844541-2025_08_05-14.27.33.594/source.csv

@@ -0,0 +1,81 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,631,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"2:27:33 PM [info] Activating crowd-code\n2:27:33 PM [info] Recording started\n2:27:33 PM [info] Initializing git provider using file system watchers...\n2:27:33 PM [info] Git repository found\n2:27:33 PM [info] Git provider initialized successfully\n2:27:33 PM [info] Initial git state: [object Object]\n",Log,tab
+3,4851,"train_dynamics.py",0,0,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    use_gt_actions: bool = False\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n\n\nargs = tyro.cli(Args)\n\n\ndef dynamics_loss_fn(\n    model: Genie, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n    """"""Compute masked dynamics loss""""""\n    gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    gt = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]), size=args.num_patch_latents, fill_value=0\n    )\n    if args.use_gt_actions:\n        _, index_counts_lam = jnp.unique_counts(\n            jnp.ravel(outputs[""lam_indices""]), size=args.num_actions, fill_value=0\n        )\n        codebook_usage_lam = (index_counts_lam != 0).mean()\n    else:\n        codebook_usage_lam = None\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_lam=codebook_usage_lam,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    return ce_loss, (outputs[""recon""], metrics)\n\n\n@nnx.jit\ndef train_step(\n    model: Genie, optimizer: nnx.Optimizer, inputs: dict\n) -> tuple[jax.Array, jax.Array, dict]:\n    """"""Update state and compute metrics""""""\n\n    def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        return dynamics_loss_fn(model, inputs)\n\n    (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(model)\n    optimizer.update(grads)\n    if args.log_gradients:\n        metrics[""gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""dynamics""]\n        )\n    return loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_actions=args.num_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        # Dynamics\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        use_gt_actions=args.use_gt_actions,\n        rngs=rngs,\n    )\n\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.dtype,\n    )\n    optimizer = nnx.Optimizer(genie, tx)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        use_gt_actions=args.use_gt_actions,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for videos, actions in dataloader:\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(videos=videos, actions=actions, mask_rng=_rng_mask)\n            loss, recon, metrics = train_step(optimizer.model, optimizer, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                            recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+4,5815,"train_dynamics.py",0,0,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    use_gt_actions: bool = False\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n\n\nargs = tyro.cli(Args)\n\n\ndef dynamics_loss_fn(\n    model: Genie, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n    """"""Compute masked dynamics loss""""""\n    gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    gt = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]), size=args.num_patch_latents, fill_value=0\n    )\n    if args.use_gt_actions:\n        _, index_counts_lam = jnp.unique_counts(\n            jnp.ravel(outputs[""lam_indices""]), size=args.num_actions, fill_value=0\n        )\n        codebook_usage_lam = (index_counts_lam != 0).mean()\n    else:\n        codebook_usage_lam = None\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_lam=codebook_usage_lam,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    return ce_loss, (outputs[""recon""], metrics)\n\n\n@nnx.jit\ndef train_step(\n    model: Genie, optimizer: nnx.Optimizer, inputs: dict\n) -> tuple[jax.Array, jax.Array, dict]:\n    """"""Update state and compute metrics""""""\n\n    def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        return dynamics_loss_fn(model, inputs)\n\n    (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(model)\n    optimizer.update(grads)\n    if args.log_gradients:\n        metrics[""gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""dynamics""]\n        )\n    return loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_actions=args.num_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        # Dynamics\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        use_gt_actions=args.use_gt_actions,\n        rngs=rngs,\n    )\n\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.dtype,\n    )\n    optimizer = nnx.Optimizer(genie, tx)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        use_gt_actions=args.use_gt_actions,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for videos, actions in dataloader:\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(videos=videos, actions=actions, mask_rng=_rng_mask)\n            loss, recon, metrics = train_step(optimizer.model, optimizer, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                            recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+5,39197,"train_dynamics.py",1557,10990,"    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_type: str = ""maskgit""  # supported options: maskgit, causal\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n\n\nargs = tyro.cli(Args)\n\n\ndef dynamics_loss_fn(\n    model: Genie, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n    """"""Compute masked dynamics loss""""""\n    gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    gt = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    _, index_counts_lam = jnp.unique_counts(\n        jnp.ravel(outputs[""lam_indices""]), size=args.num_latent_actions, fill_value=0\n    )\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]), size=args.num_patch_latents, fill_value=0\n    )\n    codebook_usage_lam = (index_counts_lam != 0).mean()\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_lam=codebook_usage_lam,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    return ce_loss, (outputs[""recon""], metrics)\n\n\n@nnx.jit\ndef train_step(\n    model: Genie, optimizer: nnx.Optimizer, inputs: dict\n) -> tuple[jax.Array, jax.Array, dict]:\n    """"""Update state and compute metrics""""""\n\n    def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        return dynamics_loss_fn(model, inputs)\n\n    (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(model)\n    optimizer.update(grads)\n    if args.log_gradients:\n        metrics[""gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""dynamics""]\n        )\n    return loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        # Dynamics\n        dyna_type=args.dyna_type,\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        decode=False,\n        rngs=rngs,\n    )\n\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.dtype,\n    )\n    optimizer = nnx.Optimizer(genie, tx)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(videos=videos, mask_rng=_rng_mask)\n",python,content
+6,40187,"train_dynamics.py",0,0,"Switched from branch 'gt-action-prepending-support-maskgit' to 'causal-spatiotemporal-kv-cache'",python,git_branch_checkout
+7,47544,"train_dynamics.py",10520,0,"",python,selection_keyboard
+8,50404,"utils/nn.py",0,0,"import math\nfrom typing import Tuple, Callable, List\n\nfrom flax import nnx\nimport jax\nimport jax.numpy as jnp\nimport einops\n\n\nclass PositionalEncoding(nnx.Module):\n    """"""https://uvadlc-notebooks.readthedocs.io/en/latest/tutorial_notebooks/JAX/tutorial6/Transformers_and_MHAttention.html""""""\n\n    def __init__(self, d_model: int, max_len: int = 5000):\n        self.d_model = d_model\n        self.max_len = max_len\n\n        pe = jnp.zeros((self.max_len, self.d_model))\n        position = jnp.arange(0, self.max_len, dtype=jnp.float32)[:, None]\n        div_term = jnp.exp(\n            jnp.arange(0, self.d_model, 2) * (-math.log(10000.0) / self.d_model)\n        )\n        pe = pe.at[:, 0::2].set(jnp.sin(position * div_term))\n        pe = pe.at[:, 1::2].set(jnp.cos(position * div_term))\n        self.pe = nnx.Variable(pe)\n\n    def __call__(self, x: jax.Array) -> jax.Array:\n        x = x + self.pe[: x.shape[2]]\n        return x\n\n\nclass STBlock(nnx.Module):\n    def __init__(\n        self,\n        dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.dim = dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.spatial_pos_enc = PositionalEncoding(self.dim)\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=False\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.temporal_pos_enc = PositionalEncoding(self.dim)\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array) -> jax.Array:\n        # --- Spatial attention ---\n        z_BTNM = self.spatial_pos_enc(x_BTNM)\n        z_BTNM = self.spatial_norm(z_BTNM)\n        z_BTNM = self.spatial_attention(z_BTNM)\n        x_BTNM = x_BTNM + z_BTNM\n\n        # --- Temporal attention ---\n        x_BNTM = x_BTNM.swapaxes(1, 2)\n        z_BNTM = self.temporal_pos_enc(x_BNTM)\n        z_BNTM = self.temporal_norm(z_BNTM)\n        z_BNTM = self.temporal_attention(z_BNTM)\n        x_BNTM = x_BNTM + z_BNTM\n        x_BTNM = x_BNTM.swapaxes(1, 2)\n\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n\n        return x_BTNM\n\n\nclass STTransformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        O: number of output features\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n        self.blocks = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                STBlock(\n                    dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    rngs=rngs,\n                )\n            )\n\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM)\n\n        x_BTNO = self.output_dense(x_BTNM)\n        return x_BTNO\n\nclass TransformerBlock(nnx.Module):\n    def __init__(\n        self,\n        model_dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.decode = decode\n\n        self.temporal_pos_enc = PositionalEncoding(self.model_dim)\n        self.spatial_pos_enc = PositionalEncoding(self.model_dim)\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        # --- Spatial attention ---\n        B, T, N, M = x_BTNM.shape\n        z_FNM = einops.rearrange(x_BTNM, ""b t n m -> (b t) n m"")\n        z_FNM = self.spatial_norm(z_FNM)\n        if self.decode:\n            assert pos_index is not None\n            z_FM = z_FNM[:, pos_index[1]]\n            z_F1M = jnp.reshape(z_FM, (B * T, 1, M))\n            z_F1M = self.spatial_attention(z_F1M)\n            z_FM = jnp.reshape(z_F1M, (B * T, M))\n            z_FNM = z_FNM.at[:, pos_index[1], :].set(z_FM)\n        else:\n            z_FNM = self.spatial_attention(z_FNM)\n        z_BTNM = einops.rearrange(z_FNM, ""(b t) n m -> b t n m"", t=T)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Temporal attention ---\n        z_PTM = einops.rearrange(x_BTNM, ""b t n m -> (b n) t m"")\n        z_PTM = self.temporal_norm(z_PTM)\n        if self.decode:\n            assert pos_index is not None\n            z_PM = z_PTM[:, pos_index[0]]\n            z_P1M = jnp.reshape(z_PM, (B * N, 1, M))\n            z_P1M = self.temporal_attention(z_P1M)\n            z_PM = jnp.reshape(z_P1M, (B * N, M))\n            z_PTM = z_PTM.at[:, pos_index[0], :].set(z_PM)\n        else:\n            z_PTM = self.temporal_attention(z_PTM)\n        z_BTNM = einops.rearrange(z_PTM, ""(b n) t m -> b t n m"", n=N)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n\n        return x_BTNM\n\nclass Transformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        O: number of output features\n        F: number of frames in batch\n        P: number of patch positions in batch\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.pos_enc = PositionalEncoding(self.model_dim)\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n        self.blocks: List[TransformerBlock] = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                TransformerBlock(\n                    model_dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    decode=decode,\n                    rngs=rngs,\n                )\n            )\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM, pos_index)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        return x_BTNV\n\ndef normalize(x: jax.Array) -> jax.Array:\n    return x / (jnp.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-8)\n\n\nclass VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n    def __init__(\n        self, latent_dim: int, num_latents: int, dropout: float, rngs: nnx.Rngs\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(self.codebook.value)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = self.codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_DL + jax.lax.stop_gradient(z_DL - x_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n\n    def get_codes(self, indices_E: jax.Array) -> jax.Array:\n        return self.codebook[indices_E]\n\n\ndef _create_flash_attention_fn(use_flash_attention: bool, is_causal: bool) -> Callable:\n    """"""\n    Create an attention function that uses flash attention if enabled.\n\n    flax.nnx.MultiHeadAttention provides tensors with shape (batch..., length, num_heads, head_dim),\n    but jax.nn.dot_product_attention expects (batch, length, num_heads, head_dim). We reshape to\n    ensure compatibility. cuDNN's flash attention additionally requires a sequence length that\n    is a multiple of 4. We pad the sequence length to the nearest multiple of 4 and mask\n    accordingly. Note that cuDNN requires the mask to be broadcast before calling the attention\n    function due to strict shape checking.\n    """"""\n\n    def attention_fn(query_BTHD, key_BSHD, value_BSHD, bias=None, mask_B111=None, **kwargs):\n        implementation = ""cudnn"" if use_flash_attention else None\n\n        def _merge_batch_dims(x):\n            return einops.rearrange(x, ""... l h k -> (...) l h k"")\n\n        def _pad(x, pad_size):\n            return jnp.pad(x, ((0, 0), (0, pad_size), (0, 0), (0, 0)))\n\n        original_shape = query_BTHD.shape\n        T = query_BTHD.shape[-3]\n        S = key_BSHD.shape[-3]\n\n        # Pad to nearest multiple of 4\n        Q = ((T + 3) // 4) * 4\n        pad_size_Q = Q - T\n        K = ((S + 3) // 4) * 4\n        pad_size_K = K - S\n\n        query_BQHD = _pad(_merge_batch_dims(query_BTHD), pad_size_Q)\n        key_BKHD = _pad(_merge_batch_dims(key_BSHD), pad_size_K)\n        value_BKHD = _pad(_merge_batch_dims(value_BSHD), pad_size_K)\n        B = query_BQHD.shape[0]\n\n        attention_mask = jnp.ones((Q, K), dtype=jnp.bool_)\n        attention_mask = attention_mask.at[Q:, :].set(False)\n        attention_mask = attention_mask.at[:, K:].set(False)\n\n        # Handle causal mask for cached decoder self-attention (from nnx.MultiHeadAttention)\n        if mask_B111 is not None:\n            # FIXME (f.srambical): Why do we need this?\n            mask_B111 = _merge_batch_dims(mask_B111)\n            # We need to broadcast T and S dimensions to target_seq_len since cudnn attention strictly checks the mask shape\n            # https://github.com/jax-ml/jax/issues/28974\n            # https://github.com/jax-ml/jax/blob/08c7677393672ccb85c10f1ed0bd506905c3c994/jax/_src/cudnn/fused_attention_stablehlo.py#L1830\n            # https://github.com/jax-ml/jax/blob/08c7677393672ccb85c10f1ed0bd506905c3c994/jax/_src/cudnn/fused_attention_stablehlo.py#L337\n            mask_B1TS = einops.repeat(mask_B111, ""... 1 1 -> ... t s"", t=Q, s=K)\n            mask_B1TS = mask_B111.astype(jnp.bool)\n        else:\n            mask_11TS = attention_mask[jnp.newaxis, jnp.newaxis, :, :]\n            mask_B1TS = jnp.broadcast_to(mask_11TS, (B, 1, Q, K))\n\n        bias_4d = _merge_batch_dims(bias) if bias is not None else None\n\n        # NOTE: jax.nn.dot_product_attention does not support dropout\n        output_4d = jax.nn.dot_product_attention(\n            query=query_BQHD,\n            key=key_BKHD,\n            value=value_BKHD,\n            bias=bias_4d,\n            mask=mask_B1TS,\n            implementation=implementation,\n            is_causal=is_causal,\n        )\n        return output_4d[..., :T, :, :].reshape(original_shape)\n\n    return attention_fn\n",python,tab
+9,51623,"utils/nn.py",4285,0,"",python,selection_command
+10,52308,"utils/nn.py",6901,0,"",python,selection_command
+11,177589,"utils/nn.py",7044,0,"",python,selection_command
+12,179216,"utils/nn.py",7320,0,"",python,selection_command
+13,179299,"utils/nn.py",7330,0,"",python,selection_command
+14,185317,"utils/nn.py",7508,0,"",python,selection_command
+15,193669,"utils/nn.py",11253,0,"",python,selection_command
+16,196652,"utils/nn.py",12324,0,"",python,selection_command
+17,198509,"utils/nn.py",12311,58,"",python,content
+18,198539,"utils/nn.py",12319,0,"",python,selection_command
+19,199494,"utils/nn.py",12910,0,"",python,selection_command
+20,200159,"utils/nn.py",12912,0,"",python,selection_command
+21,200399,"utils/nn.py",12921,0,"",python,selection_command
+22,200449,"utils/nn.py",12970,0,"",python,selection_command
+23,200463,"utils/nn.py",13011,0,"",python,selection_command
+24,200504,"utils/nn.py",13043,0,"",python,selection_command
+25,200533,"utils/nn.py",13077,0,"",python,selection_command
+26,200553,"utils/nn.py",13123,0,"",python,selection_command
+27,200598,"utils/nn.py",13165,0,"",python,selection_command
+28,200775,"utils/nn.py",13211,0,"",python,selection_command
+29,200911,"utils/nn.py",13253,0,"",python,selection_command
+30,201154,"utils/nn.py",13303,0,"",python,selection_command
+31,201188,"utils/nn.py",13341,0,"",python,selection_command
+32,201214,"utils/nn.py",13407,0,"",python,selection_command
+33,201254,"utils/nn.py",13442,0,"",python,selection_command
+34,201289,"utils/nn.py",13473,0,"",python,selection_command
+35,201319,"utils/nn.py",13491,0,"",python,selection_command
+36,201354,"utils/nn.py",13505,0,"",python,selection_command
+37,201392,"utils/nn.py",13545,0,"",python,selection_command
+38,201424,"utils/nn.py",13585,0,"",python,selection_command
+39,201469,"utils/nn.py",13624,0,"",python,selection_command
+40,201504,"utils/nn.py",13666,0,"",python,selection_command
+41,201522,"utils/nn.py",13696,0,"",python,selection_command
+42,201570,"utils/nn.py",13719,0,"",python,selection_command
+43,201584,"utils/nn.py",13721,0,"",python,selection_command
+44,201624,"utils/nn.py",13730,0,"",python,selection_command
+45,201673,"utils/nn.py",13840,0,"",python,selection_command
+46,201705,"utils/nn.py",13882,0,"",python,selection_command
+47,201724,"utils/nn.py",13924,0,"",python,selection_command
+48,201761,"utils/nn.py",13966,0,"",python,selection_command
+49,201810,"utils/nn.py",13996,0,"",python,selection_command
+50,201834,"utils/nn.py",14005,0,"",python,selection_command
+51,201863,"utils/nn.py",14039,0,"",python,selection_command
+52,201966,"utils/nn.py",14077,0,"",python,selection_command
+53,201966,"utils/nn.py",14086,0,"",python,selection_command
+54,201967,"utils/nn.py",14129,0,"",python,selection_command
+55,201993,"utils/nn.py",14143,0,"",python,selection_command
+56,202049,"utils/nn.py",14152,0,"",python,selection_command
+57,202090,"utils/nn.py",14194,0,"",python,selection_command
+58,202105,"utils/nn.py",14260,0,"",python,selection_command
+59,202149,"utils/nn.py",14261,0,"",python,selection_command
+60,202184,"utils/nn.py",14270,0,"",python,selection_command
+61,202194,"utils/nn.py",14303,0,"",python,selection_command
+62,202235,"utils/nn.py",14313,0,"",python,selection_command
+63,202266,"utils/nn.py",14333,0,"",python,selection_command
+64,202389,"utils/nn.py",14313,0,"",python,selection_command
+65,202644,"utils/nn.py",14303,0,"",python,selection_command
+66,202684,"utils/nn.py",14270,0,"",python,selection_command
+67,202706,"utils/nn.py",14261,0,"",python,selection_command
+68,202746,"utils/nn.py",14260,0,"",python,selection_command
+69,202774,"utils/nn.py",14194,0,"",python,selection_command
+70,202840,"utils/nn.py",14152,0,"",python,selection_command
+71,202859,"utils/nn.py",14143,0,"",python,selection_command
+72,202879,"utils/nn.py",14129,0,"",python,selection_command
+73,202902,"utils/nn.py",14086,0,"",python,selection_command
+74,202948,"utils/nn.py",14077,0,"",python,selection_command
+75,202973,"utils/nn.py",14039,0,"",python,selection_command
+76,203009,"utils/nn.py",14005,0,"",python,selection_command
+77,203049,"utils/nn.py",13996,0,"",python,selection_command
+78,203404,"utils/nn.py",13966,0,"",python,selection_command
+79,204040,"utils/nn.py",13958,38,"",python,content
+80,211356,"utils/nn.py",0,0,"",python,tab
+81,211357,"utils/nn.py",12311,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-59277073-3efa-40d2-8b9d-49c0c0aba00f1758640879243-2025_09_23-17.21.58.23/source.csv

@@ -0,0 +1,20 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,8,"train_dynamics.py",0,0,"import os\nimport time\n\n\nos.environ.setdefault(""XLA_PYTHON_CLIENT_MEM_FRACTION"", ""0.98"")\n\nfrom dataclasses import dataclass, field\nimport itertools\nfrom typing import cast, Optional\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_type: str = ""maskgit""  # supported options: maskgit, causal\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    use_gt_actions: bool = False\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    val_data_dir: str = """"\n    val_interval: int = 20_000\n    val_steps: int = 50\n    eval_full_frame: bool = False\n    val_maskgit_steps: int = 25\n    val_temperature: float = 1\n    val_sample_argmax: bool = False\n    wandb_id: str = """"\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[Genie, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_actions=args.num_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        use_gt_actions=args.use_gt_actions,\n        # Dynamics\n        dyna_type=args.dyna_type,\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        decode=False,\n        rngs=rngs,\n    )\n    if args.use_gt_actions:\n        assert (\n            not args.lam_checkpoint\n        ), ""Cannot use LAM when using ground-truth actions.""\n    else:\n        assert genie.lam is not None\n        del genie.lam.decoder\n    return genie, rng\n\n\ndef build_optimizer(genie: Genie, args: Args) -> nnx.ModelAndOptimizer:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.ModelAndOptimizer(genie, tx)\n    return optimizer\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    actions_sharding = NamedSharding(mesh, PartitionSpec(""data"", None))\n    return mesh, replicated_sharding, videos_sharding, actions_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.ModelAndOptimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args, data_dir: str) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(data_dir, x)\n        for x in os.listdir(data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> Optional[ocp.CheckpointManager]:\n    if args.restore_ckpt or args.save_ckpt:\n        handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n        handler_registry.add(\n            ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n        )\n        handler_registry.add(\n            ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n        )\n        handler_registry.add(\n            ""train_dataloader_state"",\n            grain.checkpoint.CheckpointSave,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        handler_registry.add(\n            ""train_dataloader_state"",\n            grain.checkpoint.CheckpointRestore,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        if args.val_data_dir:\n            handler_registry.add(\n                ""val_dataloader_state"",\n                grain.checkpoint.CheckpointSave,\n                cast(\n                    ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler\n                ),\n            )\n            handler_registry.add(\n                ""val_dataloader_state"",\n                grain.checkpoint.CheckpointRestore,\n                cast(\n                    ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler\n                ),\n            )\n        checkpoint_options = ocp.CheckpointManagerOptions(\n            save_interval_steps=args.log_checkpoint_interval,\n            max_to_keep=3,\n            keep_period=args.log_checkpoint_keep_period,\n            step_format_fixed_length=6,\n            cleanup_tmp_directories=True,\n        )\n        checkpoint_manager = ocp.CheckpointManager(\n            args.ckpt_dir,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        return checkpoint_manager\n    else:\n        return None\n\n\ndef restore_or_initialize_components(\n    args: Args,\n    checkpoint_manager: Optional[ocp.CheckpointManager],\n    optimizer: nnx.ModelAndOptimizer,\n    train_iterator: grain.DataLoaderIterator,\n    rng: jax.Array,\n    replicated_sharding: NamedSharding,\n    val_iterator: Optional[grain.DataLoaderIterator],\n    restore_step: Optional[int] = None,\n) -> tuple[\n    int,\n    nnx.ModelAndOptimizer,\n    grain.DataLoaderIterator,\n    grain.DataLoaderIterator,\n    jax.Array,\n]:\n    step = 0\n    if checkpoint_manager and restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        assert checkpoint_manager is not None\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        if val_iterator:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n                val_dataloader_state=grain.checkpoint.CheckpointRestore(val_iterator),  # type: ignore\n            )\n        else:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n            )\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(), args=restore_args\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        train_iterator = restored[""train_dataloader_state""]\n        if val_iterator:\n            val_iterator = restored[""val_dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        rng, _rng = jax.random.split(rng)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, _rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n    return step, optimizer, train_iterator, val_iterator, rng\n\n\ndef _calculate_step_metrics(\n    outputs: dict[str, jax.Array],\n    gt: jax.Array,\n    num_actions: int,\n    num_patch_latents: int,\n) -> tuple[jax.Array, dict]:\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    gt_val = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt_val, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt_val, recon)).mean()\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]),\n        size=num_patch_latents,\n        fill_value=0,\n    )\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    if ""lam_indices"" in outputs.keys():\n        _, index_counts_lam = jnp.unique_counts(\n            jnp.ravel(outputs[""lam_indices""]),\n            size=num_actions,\n            fill_value=0,\n        )\n        codebook_usage_lam = (index_counts_lam != 0).mean()\n        metrics[""codebook_usage_lam""] = codebook_usage_lam\n    return ce_loss, metrics\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    genie, rng = build_model(args, rng)\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer = build_optimizer(genie, args)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    _, replicated_sharding, videos_sharding, actions_sharding = build_mesh_and_sharding(\n        num_devices\n    )\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    train_iterator = build_dataloader(args, args.data_dir)\n    val_iterator = None\n    if args.val_data_dir:\n        val_iterator = build_dataloader(args, args.val_data_dir)\n\n    # --- Restore checkpoint ---\n    step, optimizer, train_iterator, val_iterator, rng = (\n        restore_or_initialize_components(\n            args,\n            checkpoint_manager,\n            optimizer,\n            train_iterator,\n            rng,\n            replicated_sharding,\n            val_iterator,\n        )\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def dynamics_loss_fn(\n        model: Genie,\n        inputs: dict,\n        training: bool = False,\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        outputs = model(inputs, training=training)\n        ce_loss, metrics = _calculate_step_metrics(\n            outputs, gt, args.num_actions, args.num_patch_latents\n        )\n        return ce_loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.ModelAndOptimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            model.train()\n            return dynamics_loss_fn(model, inputs, training=True)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""dynamics""]\n            )\n        return loss, recon, metrics\n\n    @nnx.jit\n    def val_step(genie: Genie, inputs: dict) -> dict:\n        """"""Evaluate model and compute metrics""""""\n        genie.eval()\n        (loss, (recon, metrics)) = dynamics_loss_fn(genie, inputs, training=False)\n        val_output = {""loss"": loss, ""recon"": recon, ""metrics"": metrics}\n\n        # --- Evaluate full frame prediction (sampling) ---\n        if args.eval_full_frame:\n            tokenizer_outputs = genie.tokenizer.vq_encode(\n                inputs[""videos""], training=False\n            )\n            tokens_full_frame = tokenizer_outputs[""indices""]\n            lam_indices = None\n            if not args.use_gt_actions:\n                lam_indices = genie.vq_encode(inputs, training=False)\n                inputs[""latent_actions""] = lam_indices\n            gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n            inputs[""videos""] = gt[:, :-1].astype(\n                args.dtype\n            )  # remove last frame for generation\n            recon_full_frame, logits_full_frame = genie.sample(\n                inputs,\n                args.seq_len,\n                args.val_temperature,\n                args.val_sample_argmax,\n                args.val_maskgit_steps,\n            )\n            step_outputs = {\n                ""recon"": recon_full_frame,\n                ""token_logits"": logits_full_frame,\n                ""video_tokens"": tokens_full_frame,\n                ""mask"": jnp.zeros_like(tokens_full_frame).at[:, -1].set(True),\n            }\n            if lam_indices is not None:\n                step_outputs[""lam_indices""] = lam_indices\n            loss_full_frame, metrics_full_frame = _calculate_step_metrics(\n                step_outputs, gt, args.num_actions, args.num_patch_latents\n            )\n            val_output.update(\n                {\n                    ""loss_full_frame"": loss_full_frame,\n                    ""recon_full_frame"": recon_full_frame,\n                    ""metrics_full_frame"": metrics_full_frame,\n                }\n            )\n        return val_output\n\n    def calculate_validation_metrics(val_dataloader, genie, rng):\n        step = 0\n        loss_per_step = []\n        metrics_per_step = []\n        loss_full_frame_per_step = []\n        metrics_full_frame_per_step = []\n        batch = None\n        recon = None\n        recon_full_frame = None\n        for batch in val_dataloader:\n            rng, _rng_mask = jax.random.split(rng, 2)\n            batch[""rng""] = _rng_mask\n            val_outputs = val_step(genie, batch)\n            loss_per_step.append(val_outputs[""loss""])\n            metrics_per_step.append(val_outputs[""metrics""])\n            recon = val_outputs[""recon""]\n            if args.eval_full_frame:\n                loss_full_frame_per_step.append(val_outputs[""loss_full_frame""])\n                metrics_full_frame_per_step.append(val_outputs[""metrics_full_frame""])\n                recon_full_frame = val_outputs[""recon_full_frame""]\n            step += 1\n            if step > args.val_steps:\n                break\n\n        if step < args.val_steps:\n            print(\n                f""Warning: Your validation dataset is too small to make val_steps many steps. Made {step} steps, expected {args.val_steps}""\n            )\n\n        val_metrics = {\n            f""val_{key}"": np.mean([float(m[key]) for m in metrics_per_step])\n            for key in metrics_per_step[0].keys()\n        }\n        val_metrics[""val_loss""] = np.mean(loss_per_step)\n        if args.eval_full_frame:\n            val_metrics_full_frame = {\n                f""val_full_frame_{key}"": np.mean(\n                    [float(m[key]) for m in metrics_full_frame_per_step]\n                )\n                for key in metrics_full_frame_per_step[0].keys()\n            }\n            val_metrics.update(val_metrics_full_frame)\n            val_metrics[""val_loss_full_frame""] = np.mean(loss_full_frame_per_step)\n        return val_metrics, batch, recon, recon_full_frame\n\n    # --- TRAIN LOOP ---\n    dataloader_train = (\n        {\n            ""videos"": jax.make_array_from_process_local_data(\n                videos_sharding, local_data=elem[""videos""]\n            ),\n            ""actions"": (\n                jax.make_array_from_process_local_data(\n                    actions_sharding, elem[""actions""]\n                )\n                if args.use_gt_actions\n                else None\n            ),\n        }\n        for elem in train_iterator\n    )\n    dataloader_val = None\n    if val_iterator:\n        dataloader_val = (\n            {\n                ""videos"": jax.make_array_from_process_local_data(\n                    videos_sharding, elem[""videos""]\n                ),\n                ""actions"": (\n                    jax.make_array_from_process_local_data(\n                        actions_sharding, elem[""actions""]\n                    )\n                    if args.use_gt_actions\n                    else None\n                ),\n            }\n            for elem in val_iterator\n        )\n    if jax.process_index() == 0:\n        first_batch = next(dataloader_train)\n        first_batch[""rng""] = rng  # type: ignore\n        compiled = train_step.lower(optimizer, first_batch).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader_train = itertools.chain([first_batch], dataloader_train)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    # --- Timing instrumentation for first 50 train steps ---\n    timing_steps = 0\n    timing_start = 0.0\n    while step < args.num_steps:\n        for batch in dataloader_train:\n            # --- Train step ---\n            if timing_steps < 50:\n                if timing_steps == 0:\n                    timing_start = time.perf_counter()\n                rng, _rng_mask = jax.random.split(rng, 2)\n                batch[""rng""] = _rng_mask\n                loss, recon, metrics = train_step(optimizer, batch)\n                timing_steps += 1\n                if timing_steps == 50:\n                    _ = loss.block_until_ready()\n                    total_s = time.perf_counter() - timing_start\n                    if jax.process_index() == 0:\n                        avg_ms = (total_s / 50.0) * 1000.0\n                        print(\n                            f""Average train step time over first 50 steps: {avg_ms:.2f} ms""\n                        )\n            else:\n                rng, _rng_mask = jax.random.split(rng, 2)\n                batch[""rng""] = _rng_mask\n                loss, recon, metrics = train_step(optimizer, batch)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            metrics[""lr""] = get_lr_schedule(\n                args.lr_schedule,\n                args.init_lr,\n                args.max_lr,\n                args.decay_end,\n                args.num_steps,\n                args.warmup_steps,\n                args.wsd_decay_steps)(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Validation loss ---\n            val_results = {}\n            if dataloader_val and step % args.val_interval == 0:\n                rng, _rng_mask_val = jax.random.split(rng, 2)\n                print(""Calculating validation metrics..."")\n                val_metrics, val_gt_batch, val_recon, val_recon_full_frame = (\n                    calculate_validation_metrics(\n                        dataloader_val, optimizer.model, _rng_mask_val\n                    )\n                )\n                print(f""Step {step}, validation loss: {val_metrics['val_loss']}"")\n                val_results = {\n                    ""metrics"": val_metrics,\n                    ""gt_batch"": val_gt_batch,\n                    ""recon"": val_recon,\n                    ""full_frame"": val_recon_full_frame,\n                }\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    log_dict = {""loss"": loss, ""step"": step, **metrics}\n                    if val_results:\n                        log_dict.update(val_results[""metrics""])\n                    wandb.log(log_dict)\n                if step % args.log_image_interval == 0:\n                    gt_seq = batch[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if val_results:\n                        val_results[""gt_seq_val""] = (\n                            val_results[""gt_batch""][""videos""][0].astype(jnp.float32)\n                            / 255.0\n                        )\n                        val_results[""recon_seq_val""] = val_results[""recon""][0].clip(\n                            0, 1\n                        )\n                        val_comparison_seq = jnp.concatenate(\n                            (val_results[""gt_seq_val""], val_results[""recon_seq_val""]),\n                            axis=1,\n                        )\n                        val_results[""val_comparison_seq""] = einops.rearrange(\n                            val_comparison_seq * 255, ""t h w c -> h (t w) c""\n                        )\n                        if args.eval_full_frame:\n                            val_results[""full_frame_seq_val""] = val_results[\n                                ""full_frame""\n                            ][0].clip(0, 1)\n                            val_results[""val_full_frame_comparison_seq""] = (\n                                jnp.concatenate(\n                                    (\n                                        val_results[""gt_seq_val""],\n                                        val_results[""full_frame_seq_val""],\n                                    ),\n                                    axis=1,\n                                )\n                            )\n                            val_results[""val_full_frame_comparison_seq""] = (\n                                einops.rearrange(\n                                    val_results[""val_full_frame_comparison_seq""] * 255,\n                                    ""t h w c -> h (t w) c"",\n                                )\n                            )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                            recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        if val_results:\n                            log_images.update(\n                                dict(\n                                    val_image=wandb.Image(\n                                        np.asarray(\n                                            val_results[""gt_seq_val""][args.seq_len - 1]\n                                        )\n                                    ),\n                                    val_recon=wandb.Image(\n                                        np.asarray(\n                                            val_results[""recon_seq_val""][\n                                                args.seq_len - 1\n                                            ]\n                                        )\n                                    ),\n                                    val_true_vs_recon=wandb.Image(\n                                        np.asarray(\n                                            val_results[""val_comparison_seq""].astype(\n                                                np.uint8\n                                            )\n                                        )\n                                    ),\n                                )\n                            )\n                            if args.eval_full_frame:\n                                log_images.update(\n                                    dict(\n                                        val_full_frame=wandb.Image(\n                                            np.asarray(\n                                                val_results[""full_frame_seq_val""][\n                                                    args.seq_len - 1\n                                                ]\n                                            )\n                                        ),\n                                        val_true_vs_full_frame=wandb.Image(\n                                            np.asarray(\n                                                val_results[\n                                                    ""val_full_frame_comparison_seq""\n                                                ].astype(np.uint8)\n                                            )\n                                        ),\n                                    )\n                                )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                assert checkpoint_manager is not None\n                optimizer_state = nnx.state(optimizer)\n                if val_iterator:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                        val_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            val_iterator  # type: ignore\n                        ),\n                    )\n                else:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                    )\n                checkpoint_manager.save(step, args=ckpt_manager_args)\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    if checkpoint_manager:\n        checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+2,1095,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:21:58 PM [info] Activating crowd-code\n5:21:58 PM [info] Recording started\n5:21:58 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,1231,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"5:21:58 PM [info] Git repository found\n5:21:58 PM [info] Git provider initialized successfully\n5:21:58 PM [info] Initial git state: [object Object]\n",Log,content
+4,4937,"TERMINAL",0,0,"",,terminal_command
+5,11973,"TERMINAL",0,0,"",,terminal_command
+6,88445,"TERMINAL",0,0,"",,terminal_command
+7,275820,"train_dynamics.py",0,0,"",python,tab
+8,353481,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+9,354825,"train_dynamics.py",0,0,"",python,tab
+10,359668,"TERMINAL",0,0,"ls /fast/project/HFMI_SynergyUnit/jafar_ws/data/breakout/train/",,terminal_command
+11,359669,"TERMINAL",0,0,"]633;C",,terminal_output
+12,359775,"TERMINAL",0,0,"data_0000.array_record  data_0111.array_record  data_0222.array_record  data_0333.array_record  data_0444.array_record  data_0555.array_record\r\ndata_0001.array_record  data_0112.array_record  data_0223.array_record  data_0334.array_record  data_0445.array_record  data_0556.array_record\r\ndata_0002.array_record  data_0113.array_record  data_0224.array_record  data_0335.array_record  data_0446.array_record  data_0557.array_record\r\ndata_0003.array_record  data_0114.array_record  data_0225.array_record  data_0336.array_record  data_0447.array_record  data_0558.array_record\r\ndata_0004.array_record  data_0115.array_record  data_0226.array_record  data_0337.array_record  data_0448.array_record  data_0559.array_record\r\ndata_0005.array_record  data_0116.array_record  data_0227.array_record  data_0338.array_record  data_0449.array_record  data_0560.array_record\r\ndata_0006.array_record  data_0117.array_record  data_0228.array_record  data_0339.array_record  data_0450.array_record  data_0561.array_record\r\ndata_0007.array_record  data_0118.array_record  data_0229.array_record  data_0340.array_record  data_0451.array_record  data_0562.array_record\r\ndata_0008.array_record  data_0119.array_record  data_0230.array_record  data_0341.array_record  data_0452.array_record  data_0563.array_record\r\ndata_0009.array_record  data_0120.array_record  data_0231.array_record  data_0342.array_record  data_0453.array_record  data_0564.array_record\r\ndata_0010.array_record  data_0121.array_record  data_0232.array_record  data_0343.array_record  data_0454.array_record  data_0565.array_record\r\ndata_0011.array_record  data_0122.array_record  data_0233.array_record  data_0344.array_record  data_0455.array_record  data_0566.array_record\r\ndata_0012.array_record  data_0123.array_record  data_0234.array_record  data_0345.array_record  data_0456.array_record  data_0567.array_record\r\ndata_0013.array_record  data_0124.array_record  data_0235.array_record  data_0346.array_record  data_0457.array_record  data_0568.array_record\r\ndata_0014.array_record  data_0125.array_record  data_0236.array_record  data_0347.array_record  data_0458.array_record  data_0569.array_record\r\ndata_0015.array_record  data_0126.array_record  data_0237.array_record  data_0348.array_record  data_0459.array_record  data_0570.array_record\r\ndata_0016.array_record  data_0127.array_record  data_0238.array_record  data_0349.array_record  data_0460.array_record  data_0571.array_record\r\ndata_0017.array_record  data_0128.array_record  data_0239.array_record  data_0350.array_record  data_0461.array_record  data_0572.array_record\r\ndata_0018.array_record  data_0129.array_record  data_0240.array_record  data_0351.array_record  data_0462.array_record  data_0573.array_record\r\ndata_0019.array_record  data_0130.array_record  data_0241.array_record  data_0352.array_record  data_0463.array_record  data_0574.array_record\r\ndata_0020.array_record  data_0131.array_record  data_0242.array_record  data_0353.array_record  data_0464.array_record  data_0575.array_record\r\ndata_0021.array_record  data_0132.array_record  data_0243.array_record  data_0354.array_record  data_0465.array_record  data_0576.array_record\r\ndata_0022.array_record  data_0133.array_record  data_0244.array_record  data_0355.array_record  data_0466.array_record  data_0577.array_record\r\ndata_0023.array_record  data_0134.array_record  data_0245.array_record  data_0356.array_record  data_0467.array_record  data_0578.array_record\r\ndata_0024.array_record  data_0135.array_record  data_0246.array_record  data_0357.array_record  data_0468.array_record  data_0579.array_record\r\ndata_0025.array_record  data_0136.array_record  data_0247.array_record  data_0358.array_record  data_0469.array_record  data_0580.array_record\r\ndata_0026.array_record  data_0137.array_record  data_0248.array_record  data_0359.array_record  data_0470.array_record  data_0581.array_record\r\ndata_0027.array_record  data_0138.array_record  data_0249.array_record  data_0360.array_record  data_0471.array_record  data_0582.array_record\r\ndata_0028.array_record  data_0139.array_record  data_0250.array_record  data_0361.array_record  data_0472.array_record  data_0583.array_record\r\ndata_0029.array_record  data_0140.array_record  data_0251.array_record  data_0362.array_record  data_0473.array_record  data_0584.array_record\r\ndata_0030.array_record  data_0141.array_record  data_0252.array_record  data_0363.array_record  data_0474.array_record  data_0585.array_record\r\ndata_0031.array_record  data_0142.array_record  data_0253.array_record  data_0364.array_record  data_0475.array_record  data_0586.array_record\r\ndata_0032.array_record  data_0143.array_record  data_0254.array_record  data_0365.array_record  data_0476.array_record  data_0587.array_record\r\ndata_0033.array_record  data_0144.array_record  data_0255.array_record  data_0366.array_record  data_0477.array_record  data_0588.array_record\r\ndata_0034.array_record  data_0145.array_record  data_0256.array_record  data_0367.array_record  data_0478.array_record  data_0589.array_record\r\ndata_0035.array_record  data_0146.array_record  data_0257.array_record  data_0368.array_record  data_0479.array_record  data_0590.array_record\r\ndata_0036.array_record  data_0147.array_record  data_0258.array_record  data_0369.array_record  data_0480.array_record  data_0591.array_record\r\ndata_0037.array_record  data_0148.array_record  data_0259.array_record  data_0370.array_record  data_0481.array_record  data_0592.array_record\r\ndata_0038.array_record  data_0149.array_record  data_0260.array_record  data_0371.array_record  data_0482.array_record  data_0593.array_record\r\ndata_0039.array_record  data_0150.array_record  data_0261.array_record  data_0372.array_record  data_0483.array_record  data_0594.array_record\r\ndata_0040.array_record  data_0151.array_record  data_0262.array_record  data_0373.array_record  data_0484.array_record  data_0595.array_record\r\ndata_0041.array_record  data_0152.array_record  data_0263.array_record  data_0374.array_record  data_0485.array_record  data_0596.array_record\r\ndata_0042.array_record  data_0153.array_record  data_0264.array_record  data_0375.array_record  data_0486.array_record  data_0597.array_record\r\ndata_0043.array_record  data_0154.array_record  data_0265.array_record  data_0376.array_record  data_0487.array_record  data_0598.array_record\r\ndata_0044.array_record  data_0155.array_record  data_0266.array_record  data_0377.array_record  data_0488.array_record  data_0599.array_record\r\ndata_0045.array_record  data_0156.array_record  data_0267.array_record  data_0378.array_record  data_0489.array_record  data_0600.array_record\r\ndata_0046.array_record  data_0157.array_record  data_0268.array_record  data_0379.array_record  data_0490.array_record  data_0601.array_record\r\ndata_0047.array_record  data_0158.array_record  data_0269.array_record  data_0380.array_record  data_0491.array_record  data_0602.array_record\r\ndata_0048.array_record  data_0159.array_record  data_0270.array_record  data_0381.array_record  data_0492.array_record  data_0603.array_record\r\ndata_0049.array_record  data_0160.array_record  data_0271.array_record  data_0382.array_record  data_0493.array_record  data_0604.array_record\r\ndata_0050.array_record  data_0161.array_record  data_0272.array_record  data_0383.array_record  data_0494.array_record  data_0605.array_record\r\ndata_0051.array_record  data_0162.array_record  data_0273.array_record  data_0384.array_record  data_0495.array_record  data_0606.array_record\r\ndata_0052.array_record  data_0163.array_record  data_0274.array_record  data_0385.array_record  data_0496.array_record  data_0607.array_record\r\ndata_0053.array_record  data_0164.array_record  data_0275.array_record  data_0386.array_record  data_0497.array_record  data_0608.array_record\r\ndata_0054.array_record  data_0165.array_record  data_0276.array_record  data_0387.array_record  data_0498.array_record  data_0609.array_record\r\ndata_0055.array_record  data_0166.array_record  data_0277.array_record  data_0388.array_record  data_0499.array_record  data_0610.array_record\r\ndata_0056.array_record  data_0167.array_record  data_0278.array_record  data_0389.array_record  data_0500.array_record  data_0611.array_record\r\ndata_0057.array_record  data_0168.array_record  data_0279.array_record  data_0390.array_record  data_0501.array_record  data_0612.array_record\r\ndata_0058.array_record  data_0169.array_record  data_0280.array_record  data_0391.array_record  data_0502.array_record  data_0613.array_record\r\ndata_0059.array_record  data_0170.array_record  data_0281.array_record  data_0392.array_record  data_0503.array_record  data_0614.array_record\r\ndata_0060.array_record  data_0171.array_record  data_0282.array_record  data_0393.array_record  data_0504.array_record  data_0615.array_record\r\ndata_0061.array_record  data_0172.array_record  data_0283.array_record  data_0394.array_record  data_0505.array_record  data_0616.array_record\r\ndata_0062.array_record  data_0173.array_record  data_0284.array_record  data_0395.array_record  data_0506.array_record  data_0617.array_record\r\ndata_0063.array_record  data_0174.array_record  data_0285.array_record  data_0396.array_record  data_0507.array_record  data_0618.array_record\r\ndata_0064.array_record  data_0175.array_record  data_0286.array_record  data_0397.array_record  data_0508.array_record  data_0619.array_record\r\ndata_0065.array_record  data_0176.array_record  data_0287.array_record  data_0398.array_record  data_0509.array_record  data_0620.array_record\r\ndata_0066.array_record  data_0177.array_record  data_0288.array_record  data_0399.array_record  data_0510.array_record  data_0621.array_record\r\ndata_0067.array_record  data_0178.array_record  data_0289.array_record  data_0400.array_record  data_0511.array_record  data_0622.array_record\r\ndata_0068.array_record  data_0179.array_record  data_0290.array_record  data_0401.array_record  data_0512.array_record  data_0623.array_record\r\ndata_0069.array_record  data_0180.array_record  data_0291.array_record  data_0402.array_record  data_0513.array_record  data_0624.array_record\r\ndata_0070.array_record  data_0181.array_record  data_0292.array_record  data_0403.array_record  data_0514.array_record  data_0625.array_record\r\ndata_0071.array_record  data_0182.array_record  data_0293.array_record  data_0404.array_record  data_0515.array_record  data_0626.array_record\r\ndata_0072.array_record  data_0183.array_record  data_0294.array_record  data_0405.array_record  data_0516.array_record  data_0627.array_record\r\ndata_0073.array_record  data_0184.array_record  data_0295.array_record  data_0406.array_record  data_0517.array_record  data_0628.array_record\r\ndata_0074.array_record  data_0185.array_record  data_0296.array_record  data_0407.array_record  data_0518.array_record  data_0629.array_record\r\ndata_0075.array_record  data_0186.array_record  data_0297.array_record  data_0408.array_record  data_0519.array_record  data_0630.array_record\r\ndata_0076.array_record  data_0187.array_record  data_0298.array_record  data_0409.array_record  data_0520.array_record  data_0631.array_record\r\ndata_0077.array_record  data_0188.array_record  data_0299.array_record  data_0410.array_record  data_0521.array_record  data_0632.array_record\r\ndata_0078.array_record  data_0189.array_record  data_0300.array_record  data_0411.array_record  data_0522.array_record  data_0633.array_record\r\ndata_0079.array_record  data_0190.array_record  data_0301.array_record  data_0412.array_record  data_0523.array_record  data_0634.array_record\r\ndata_0080.array_record  data_0191.array_record  data_0302.array_record  data_0413.array_record  data_0524.array_record  data_0635.array_record\r\ndata_0081.array_record  data_0192.array_record  data_0303.array_record  data_0414.array_record  data_0525.array_record  data_0636.array_record\r\ndata_0082.array_record  data_0193.array_record  data_0304.array_record  data_0415.array_record  data_0526.array_record  data_0637.array_record\r\ndata_0083.array_record  data_0194.array_record  data_0305.array_record  data_0416.array_record  data_0527.array_record  data_0638.array_record\r\ndata_0084.array_record  data_0195.array_record  data_0306.array_record  data_0417.array_record  data_0528.array_record  data_0639.array_record\r\ndata_0085.array_record  data_0196.array_record  data_0307.array_record  data_0418.array_record  data_0529.array_record  data_0640.array_record\r\ndata_0086.array_record  data_0197.array_record  data_0308.array_record  data_0419.array_record  data_0530.array_record  data_0641.array_record\r\ndata_0087.array_record  data_0198.array_record  data_0309.array_record  data_0420.array_record  data_0531.array_record  data_0642.array_record\r\ndata_0088.array_record  data_0199.array_record  data_0310.array_record  data_0421.array_record  data_0532.array_record  data_0643.array_record\r\ndata_0089.array_record  data_0200.array_record  data_0311.array_record  data_0422.array_record  data_0533.array_record  data_0644.array_record\r\ndata_0090.array_record  data_0201.array_record  data_0312.array_record  data_0423.array_record  data_0534.array_record  data_0645.array_record\r\ndata_0091.array_record  data_0202.array_record  data_0313.array_record  data_0424.array_record  data_0535.array_record  data_0646.array_record\r\ndata_0092.array_record  data_0203.array_record  data_0314.array_record  data_0425.array_record  data_0536.array_record  data_0647.array_record\r\ndata_0093.array_record  data_0204.array_record  data_0315.array_record  data_0426.array_record  data_0537.array_record  data_0648.array_record\r\ndata_0094.array_record  data_0205.array_record  data_0316.array_record  data_0427.array_record  data_0538.array_record  data_0649.array_record\r\ndata_0095.array_record  data_0206.array_record  data_0317.array_record  data_0428.array_record  data_0539.array_record  data_0650.array_record\r\ndata_0096.array_record  data_0207.array_record  data_0318.array_record  data_0429.array_record  data_0540.array_record  data_0651.array_record\r\ndata_0097.array_record  data_0208.array_record  data_0319.array_record  data_0430.array_record  data_0541.array_record  data_0652.array_record\r\ndata_0098.array_record  data_0209.array_record  data_0320.array_record  data_0431.array_record  data_0542.array_record  data_0653.array_record\r\ndata_0099.array_record  data_0210.array_record  data_0321.array_record  data_0432.array_record  data_0543.array_record  data_0654.array_record\r\ndata_0100.array_record  data_0211.array_record  data_0322.array_record  data_0433.array_record  data_0544.array_record  data_0655.array_record\r\ndata_0101.array_record  data_0212.array_record  data_0323.array_record  data_0434.array_record  data_0545.array_record  data_0656.array_record\r\ndata_0102.array_record  data_0213.array_record  data_0324.array_record  data_0435.array_record  data_0546.array_record  data_0657.array_record\r\ndata_0103.array_record  data_0214.array_record  data_0325.array_record  data_0436.array_record  data_0547.array_record  data_0658.array_record\r\ndata_0104.array_record  data_0215.array_record  data_0326.array_record  data_0437.array_record  data_0548.array_record  data_0659.array_record\r\ndata_0105.array_record  data_0216.array_record  data_0327.array_record  data_0438.array_record  data_0549.array_record  data_0660.array_record\r\ndata_0106.array_record  data_0217.array_record  data_0328.array_record  data_0439.array_record  data_0550.array_record  data_0661.array_record\r\ndata_0107.array_record  data_0218.array_record  data_0329.array_record  data_0440.array_record  data_0551.array_record  data_0662.array_record\r\ndata_0108.array_record  data_0219.array_record  data_0330.array_record  data_0441.array_record  data_0552.array_record  data_0663.array_record\r\ndata_0109.array_record  data_0220.array_record  data_0331.array_record  data_0442.array_record  data_0553.array_record  data_0664.array_record\r\ndata_0110.array_record  data_0221.array_record  data_0332.array_record  data_0443.array_record  data_0554.array_record\r\n]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+13,363127,"train_dynamics.py",22709,16,"            ",python,content
+14,365245,"train_dynamics.py",22721,0,"    ",python,content
+15,365248,"train_dynamics.py",22744,0,"",python,selection_command
+16,368268,"experiments/dynamics_grain_tok_restore.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=2\n#SBATCH --ntasks-per-node=8\n#SBATCH --time=12:00:00\n#SBATCH --cpus-per-task=1\n#SBATCH --gres=gpu:8\n#SBATCH --mem=100GB\n\nsource .venv/bin/activate\n\ndata_dir=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/coinrun/array_records_10m""\nckpt_dir=""${PWD}/checkpoints/dynamics_openai_grain_tok_restore""\ntokenizer_ckpt_dir=""${PWD}/checkpoints/tokenizer_openai_grain_checkpointing""\n\nexport PYTHONUNBUFFERED=1\nsrun python train_dynamics.py \\n    --dyna_type 'causal' \\n    --batch_size 120 \\n    --image_height=64 \\n    --image_width=64 \\n    --patch_size=16 \\n    --tokenizer_checkpoint $tokenizer_ckpt_dir \\n    --log_checkpoint_interval 5 \\n    --ckpt_dir $ckpt_dir \\n    --num_steps 300000 \\n    --warmup_steps 10000 \\n    --seed 0 \\n    --init_lr=0.0000866 \\n    --max_lr=0.0000866 \\n    --data_dir $data_dir",shellscript,tab
+17,371533,"train_dynamics.py",0,0,"",python,tab
+18,371552,"train_dynamics.py",10,0,"",python,selection_command
+19,382441,"train_dynamics.py",10,23861,"\n\nos.environ.setdefault(""XLA_PYTHON_CLIENT_MEM_FRACTION"", ""0.98"")\n\nfrom dataclasses import dataclass, field\nimport itertools\nfrom typing import cast, Optional\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_type: str = ""maskgit""  # supported options: maskgit, causal\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    use_gt_actions: bool = False\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    val_data_dir: str = """"\n    val_interval: int = 20_000\n    val_steps: int = 50\n    eval_full_frame: bool = False\n    val_maskgit_steps: int = 25\n    val_temperature: float = 1\n    val_sample_argmax: bool = False\n    wandb_id: str = """"\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[Genie, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_actions=args.num_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        use_gt_actions=args.use_gt_actions,\n        # Dynamics\n        dyna_type=args.dyna_type,\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        decode=False,\n        rngs=rngs,\n    )\n    if args.use_gt_actions:\n        assert (\n            not args.lam_checkpoint\n        ), ""Cannot use LAM when using ground-truth actions.""\n    else:\n        assert genie.lam is not None\n        del genie.lam.decoder\n    return genie, rng\n\n\ndef build_optimizer(genie: Genie, args: Args) -> nnx.ModelAndOptimizer:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.ModelAndOptimizer(genie, tx)\n    return optimizer\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    actions_sharding = NamedSharding(mesh, PartitionSpec(""data"", None))\n    return mesh, replicated_sharding, videos_sharding, actions_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.ModelAndOptimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args, data_dir: str) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(data_dir, x)\n        for x in os.listdir(data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> Optional[ocp.CheckpointManager]:\n    if args.restore_ckpt or args.save_ckpt:\n        handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n        handler_registry.add(\n            ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n        )\n        handler_registry.add(\n            ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n        )\n        handler_registry.add(\n            ""train_dataloader_state"",\n            grain.checkpoint.CheckpointSave,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        handler_registry.add(\n            ""train_dataloader_state"",\n            grain.checkpoint.CheckpointRestore,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        if args.val_data_dir:\n            handler_registry.add(\n                ""val_dataloader_state"",\n                grain.checkpoint.CheckpointSave,\n                cast(\n                    ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler\n                ),\n            )\n            handler_registry.add(\n                ""val_dataloader_state"",\n                grain.checkpoint.CheckpointRestore,\n                cast(\n                    ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler\n                ),\n            )\n        checkpoint_options = ocp.CheckpointManagerOptions(\n            save_interval_steps=args.log_checkpoint_interval,\n            max_to_keep=3,\n            keep_period=args.log_checkpoint_keep_period,\n            step_format_fixed_length=6,\n            cleanup_tmp_directories=True,\n        )\n        checkpoint_manager = ocp.CheckpointManager(\n            args.ckpt_dir,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        return checkpoint_manager\n    else:\n        return None\n\n\ndef restore_or_initialize_components(\n    args: Args,\n    checkpoint_manager: Optional[ocp.CheckpointManager],\n    optimizer: nnx.ModelAndOptimizer,\n    train_iterator: grain.DataLoaderIterator,\n    rng: jax.Array,\n    replicated_sharding: NamedSharding,\n    val_iterator: Optional[grain.DataLoaderIterator],\n    restore_step: Optional[int] = None,\n) -> tuple[\n    int,\n    nnx.ModelAndOptimizer,\n    grain.DataLoaderIterator,\n    grain.DataLoaderIterator,\n    jax.Array,\n]:\n    step = 0\n    if checkpoint_manager and restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        assert checkpoint_manager is not None\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        if val_iterator:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n                val_dataloader_state=grain.checkpoint.CheckpointRestore(val_iterator),  # type: ignore\n            )\n        else:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n            )\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(), args=restore_args\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        train_iterator = restored[""train_dataloader_state""]\n        if val_iterator:\n            val_iterator = restored[""val_dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        rng, _rng = jax.random.split(rng)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, _rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n    return step, optimizer, train_iterator, val_iterator, rng\n\n\ndef _calculate_step_metrics(\n    outputs: dict[str, jax.Array],\n    gt: jax.Array,\n    num_actions: int,\n    num_patch_latents: int,\n) -> tuple[jax.Array, dict]:\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    gt_val = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt_val, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt_val, recon)).mean()\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]),\n        size=num_patch_latents,\n        fill_value=0,\n    )\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    if ""lam_indices"" in outputs.keys():\n        _, index_counts_lam = jnp.unique_counts(\n            jnp.ravel(outputs[""lam_indices""]),\n            size=num_actions,\n            fill_value=0,\n        )\n        codebook_usage_lam = (index_counts_lam != 0).mean()\n        metrics[""codebook_usage_lam""] = codebook_usage_lam\n    return ce_loss, metrics\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    genie, rng = build_model(args, rng)\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer = build_optimizer(genie, args)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    _, replicated_sharding, videos_sharding, actions_sharding = build_mesh_and_sharding(\n        num_devices\n    )\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    train_iterator = build_dataloader(args, args.data_dir)\n    val_iterator = None\n    if args.val_data_dir:\n        val_iterator = build_dataloader(args, args.val_data_dir)\n\n    # --- Restore checkpoint ---\n    step, optimizer, train_iterator, val_iterator, rng = (\n        restore_or_initialize_components(\n            args,\n            checkpoint_manager,\n            optimizer,\n            train_iterator,\n            rng,\n            replicated_sharding,\n            val_iterator,\n        )\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def dynamics_loss_fn(\n        model: Genie,\n        inputs: dict,\n        training: bool = False,\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        outputs = model(inputs, training=training)\n        ce_loss, metrics = _calculate_step_metrics(\n            outputs, gt, args.num_actions, args.num_patch_latents\n        )\n        return ce_loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.ModelAndOptimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            model.train()\n            return dynamics_loss_fn(model, inputs, training=True)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""dynamics""]\n            )\n        return loss, recon, metrics\n\n    @nnx.jit\n    def val_step(genie: Genie, inputs: dict) -> dict:\n        """"""Evaluate model and compute metrics""""""\n        genie.eval()\n        (loss, (recon, metrics)) = dynamics_loss_fn(genie, inputs, training=False)\n        val_output = {""loss"": loss, ""recon"": recon, ""metrics"": metrics}\n\n        # --- Evaluate full frame prediction (sampling) ---\n        if args.eval_full_frame:\n            tokenizer_outputs = genie.tokenizer.vq_encode(\n                inputs[""videos""], training=False\n            )\n            tokens_full_frame = tokenizer_outputs[""indices""]\n            lam_indices = None\n            if not args.use_gt_actions:\n                lam_indices = genie.vq_encode(inputs, training=False)\n                inputs[""latent_actions""] = lam_indices\n            gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n            inputs[""videos""] = gt[:, :-1].astype(\n                args.dtype\n            )  # remove last frame for generation\n            recon_full_frame, logits_full_frame = genie.sample(\n                inputs,\n                args.seq_len,\n                args.val_temperature,\n                args.val_sample_argmax,\n                args.val_maskgit_steps,\n            )\n            step_outputs = {\n                ""recon"": recon_full_frame,\n                ""token_logits"": logits_full_frame,\n                ""video_tokens"": tokens_full_frame,\n                ""mask"": jnp.zeros_like(tokens_full_frame).at[:, -1].set(True),\n            }\n            if lam_indices is not None:\n                step_outputs[""lam_indices""] = lam_indices\n            loss_full_frame, metrics_full_frame = _calculate_step_metrics(\n                step_outputs, gt, args.num_actions, args.num_patch_latents\n            )\n            val_output.update(\n                {\n                    ""loss_full_frame"": loss_full_frame,\n                    ""recon_full_frame"": recon_full_frame,\n                    ""metrics_full_frame"": metrics_full_frame,\n                }\n            )\n        return val_output\n\n    def calculate_validation_metrics(val_dataloader, genie, rng):\n        step = 0\n        loss_per_step = []\n        metrics_per_step = []\n        loss_full_frame_per_step = []\n        metrics_full_frame_per_step = []\n        batch = None\n        recon = None\n        recon_full_frame = None\n        for batch in val_dataloader:\n            rng, _rng_mask = jax.random.split(rng, 2)\n            batch[""rng""] = _rng_mask\n            val_outputs = val_step(genie, batch)\n            loss_per_step.append(val_outputs[""loss""])\n            metrics_per_step.append(val_outputs[""metrics""])\n            recon = val_outputs[""recon""]\n            if args.eval_full_frame:\n                loss_full_frame_per_step.append(val_outputs[""loss_full_frame""])\n                metrics_full_frame_per_step.append(val_outputs[""metrics_full_frame""])\n                recon_full_frame = val_outputs[""recon_full_frame""]\n            step += 1\n            if step > args.val_steps:\n                break\n\n        if step < args.val_steps:\n            print(\n                f""Warning: Your validation dataset is too small to make val_steps many steps. Made {step} steps, expected {args.val_steps}""\n            )\n\n        val_metrics = {\n            f""val_{key}"": np.mean([float(m[key]) for m in metrics_per_step])\n            for key in metrics_per_step[0].keys()\n        }\n        val_metrics[""val_loss""] = np.mean(loss_per_step)\n        if args.eval_full_frame:\n            val_metrics_full_frame = {\n                f""val_full_frame_{key}"": np.mean(\n                    [float(m[key]) for m in metrics_full_frame_per_step]\n                )\n                for key in metrics_full_frame_per_step[0].keys()\n            }\n            val_metrics.update(val_metrics_full_frame)\n            val_metrics[""val_loss_full_frame""] = np.mean(loss_full_frame_per_step)\n        return val_metrics, batch, recon, recon_full_frame\n\n    # --- TRAIN LOOP ---\n    dataloader_train = (\n        {\n            ""videos"": jax.make_array_from_process_local_data(\n                videos_sharding, local_data=elem[""videos""]\n            ),\n            ""actions"": (\n                jax.make_array_from_process_local_data(\n                    actions_sharding, elem[""actions""]\n                )\n                if args.use_gt_actions\n                else None\n            ),\n        }\n        for elem in train_iterator\n    )\n    dataloader_val = None\n    if val_iterator:\n        dataloader_val = (\n            {\n                ""videos"": jax.make_array_from_process_local_data(\n                    videos_sharding, elem[""videos""]\n                ),\n                ""actions"": (\n                    jax.make_array_from_process_local_data(\n                        actions_sharding, elem[""actions""]\n                    )\n                    if args.use_gt_actions\n                    else None\n                ),\n            }\n            for elem in val_iterator\n        )\n    if jax.process_index() == 0:\n        first_batch = next(dataloader_train)\n        first_batch[""rng""] = rng  # type: ignore\n        compiled = train_step.lower(optimizer, first_batch).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader_train = itertools.chain([first_batch], dataloader_train)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for batch in dataloader_train:\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            batch[""rng""] = _rng_mask\n            loss, recon, metrics = train_step(optimizer, batch)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n",python,content

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-60f50fcf-267d-433b-a431-0f34ef2ae0411755711838954-2025_08_20-19.44.03.411/source.csv

@@ -0,0 +1,277 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,178,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:44:03 PM [info] Activating crowd-code\n7:44:03 PM [info] Recording started\n7:44:03 PM [info] Initializing git provider using file system watchers...\n7:44:03 PM [info] Git repository found\n7:44:03 PM [info] Git provider initialized successfully\n7:44:03 PM [info] Initial git state: [object Object]\n",Log,tab
+3,111047,"MaxText/input_pipeline/_grain_data_processing.py",0,0,"# Copyright 2023–2025 Google LLC\n#\n# Licensed under the Apache License, Version 2.0 (the ""License"");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n#    https://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an ""AS IS"" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\n""""""Input pipeline using Grain.""""""\n\nimport glob\nfrom pathlib import Path\nimport functools\n\nimport ml_collections\n\nimport jax\n\nimport grain.python as grain\n\nfrom MaxText.input_pipeline import _input_pipeline_utils\nfrom MaxText.input_pipeline import _grain_tokenizer\nfrom MaxText import multihost_dataloading\nfrom MaxText import max_logging\nfrom MaxText import tokenizer\n\n\ndef find_data_files(data_file_pattern):\n  data_files = glob.glob(str(Path(data_file_pattern).expanduser().resolve()))\n  assert len(data_files) > 0, f""No file found with pattern {data_file_pattern}.""\n  max_logging.log(f""Found {len(data_files)} files for train/eval with grain"")\n  return data_files\n\n\ndef get_datasets(\n    data_file_pattern,\n    data_file_type,\n    shuffle,\n    shuffle_seed,\n    num_epoch,\n    dataloading_host_index,\n    dataloading_host_count,\n    grain_worker_count,\n):\n  """"""Load dataset from array_record files for using with grain""""""\n  if data_file_type == ""arrayrecord"":\n    if "";"" in data_file_pattern:\n      data_file_patterns, weights = zip(*[pattern.split("":"") for pattern in data_file_pattern.split("";"")])\n      assert len(data_file_patterns) == len(weights), ""Number of data file patterns and weights must match""\n      weights = [float(weight) for weight in weights]\n      weights = [round(weight / sum(weights), 4) for weight in weights]\n      dataset_list = [\n          grain.MapDataset.source(grain.ArrayRecordDataSource(find_data_files(pattern))) for pattern in data_file_patterns\n      ]\n      dataset = grain.MapDataset.mix(dataset_list, weights)\n    else:\n      data_files = find_data_files(data_file_pattern)\n      dataset = grain.MapDataset.source(grain.ArrayRecordDataSource(data_files))\n    if shuffle:\n      dataset = dataset.shuffle(seed=shuffle_seed)\n    dataset = dataset.repeat(num_epoch)\n    dataset = dataset[dataloading_host_index::dataloading_host_count]  # sharding\n    dataset = dataset.to_iter_dataset()\n  elif data_file_type == ""parquet"":\n    data_files = find_data_files(data_file_pattern)\n    dataset = grain.MapDataset.source(data_files)\n    if shuffle:\n      dataset = dataset.shuffle(seed=shuffle_seed)\n    dataset = dataset.repeat(num_epoch)\n    dataset = dataset[dataloading_host_index::dataloading_host_count]  # sharding\n    assert grain_worker_count <= len(dataset), (\n        f""grain worker count is currently {grain_worker_count}, exceeding the max allowable value {len(dataset)} ""\n        f""(file shard count of a data loading host) for your dataset. ""\n        f""Please lower grain_worker_count or increase file shard count.""\n    )\n    dataset = dataset.map(grain.experimental.ParquetIterDataset)\n    dataset = grain.experimental.InterleaveIterDataset(dataset, cycle_length=len(dataset))\n    dataset = grain.experimental.WindowShuffleIterDataset(dataset, window_size=100, seed=shuffle_seed)\n  else:\n    raise ValueError(f""grain pipeline supports (arrayrecord, parquet) as grain_file_type, but got {data_file_type}"")\n\n  return dataset\n\n\ndef pretrain_preprocessing_pipeline(dataset, config, data_columns, tokenize, grain_worker_count):\n  """"""Use grain pipeline to pre-process the dataset and return iterators for pretrain""""""\n  if config.grain_file_type == ""arrayrecord"":\n    dataset = dataset.map(_input_pipeline_utils.ParseFeatures(data_columns, tokenize))\n    dataset = dataset.map(_input_pipeline_utils.NormalizeFeatures(data_columns, tokenize))\n\n  assert len(data_columns) == 1\n  rekey_dict = {""inputs"": ""text"", ""targets"": ""text""}\n  dataset = dataset.map(_input_pipeline_utils.Rekey(rekey_dict))\n  data_columns = (""inputs"", ""targets"")\n\n  tokenizer_model = tokenizer.build_tokenizer(\n      config.tokenizer_path,\n      config.tokenizer_type,\n      config.add_bos,\n      config.add_eos,\n      config.hf_access_token,\n      config.dataset_type,\n  )\n  if tokenizer_model.pad_id is not None:\n    pad_id = tokenizer_model.pad_id\n  elif tokenizer_model.unk_id is not None:\n    pad_id = tokenizer_model.unk_id\n  else:\n    pad_id = -1\n\n  if tokenize:\n    dataset = dataset.map(\n        _grain_tokenizer.TokenizeAndTrim(\n            data_columns, config.max_target_length, config.add_bos, config.add_eos, tokenizer_model\n        )\n    )\n\n  # Pack and Batch examples.\n  if config.packing:\n    length_struct = {col: config.max_target_length for col in data_columns}\n    dataset = grain.experimental.FirstFitPackIterDataset(dataset, length_struct=length_struct, num_packing_bins=30)\n    rekey_dict = {\n        ""targets_segmentation"": ""targets_segment_ids"",\n        ""inputs_segmentation"": ""inputs_segment_ids"",\n        ""targets_position"": ""targets_positions"",\n        ""inputs_position"": ""inputs_positions"",\n    }\n    dataset = dataset.map(_input_pipeline_utils.Rekey(rekey_dict))\n  else:\n    dataset = dataset.map(_input_pipeline_utils.PadToMaxLength(config.max_target_length, pad_id))\n  dataset = dataset.batch(batch_size=config.global_batch_size_to_load // jax.process_count(), drop_remainder=False)\n\n  # Shift inputs for teacher-forced training\n  dataset = dataset.map(\n      _input_pipeline_utils.ShiftData(\n          ignored_ids=[pad_id],\n          axis=1,\n      )\n  )\n  dataset = dataset.mp_prefetch(grain.MultiprocessingOptions(num_workers=grain_worker_count))\n  return dataset\n\n\ndef dpo_preprocessing_pipeline(dataset, config, data_columns, tokenize, grain_worker_count):\n  """"""Use grain to pre-process the dataset and return iterators for dpo fine-tuning""""""\n  if config.grain_file_type == ""arrayrecord"":\n    dataset = dataset.map(_input_pipeline_utils.ParseFeatures(data_columns, tokenize))\n    dataset = dataset.map(_input_pipeline_utils.NormalizeFeatures(data_columns, tokenize))\n  tokenizer_model = tokenizer.build_tokenizer(\n      config.tokenizer_path,\n      config.tokenizer_type,\n      config.add_bos,\n      config.add_eos,\n      config.hf_access_token,\n      config.dataset_type,\n  )\n  if tokenizer_model.pad_id is not None:\n    pad_id = tokenizer_model.pad_id\n  elif tokenizer_model.unk_id is not None:\n    pad_id = tokenizer_model.unk_id\n  else:\n    pad_id = -1\n\n  if tokenize:\n    dataset = dataset.map(\n        _grain_tokenizer.TokenizeAndTrim(\n            data_columns, config.max_target_length, config.add_bos, config.add_eos, tokenizer_model\n        )\n    )\n\n  dataset = dataset.map(_input_pipeline_utils.PadToMaxLength(config.max_target_length, pad_id))\n  dataset = dataset.batch(batch_size=config.global_batch_size_to_load // jax.process_count(), drop_remainder=False)\n  dataset = dataset.mp_prefetch(grain.MultiprocessingOptions(num_workers=grain_worker_count))\n  return dataset\n\n\ndef make_grain_train_iterator(\n    config: ml_collections.ConfigDict,\n    global_mesh,\n    process_indices,\n):\n  """"""Load, preprocess dataset and return iterators""""""\n  assert (\n      config.global_batch_size_to_load % global_mesh.size == 0\n  ), ""Batch size should be divisible by number of global devices.""\n  if not config.colocated_python_data_input:\n    train_ds = get_datasets(\n        config.grain_train_files,\n        config.grain_file_type,\n        shuffle=config.enable_data_shuffling,\n        shuffle_seed=config.data_shuffle_seed,\n        num_epoch=config.num_epoch,\n        dataloading_host_index=process_indices.index(jax.process_index()),\n        dataloading_host_count=len(process_indices),\n        grain_worker_count=config.grain_worker_count,\n    )\n    if config.use_dpo:\n      train_dataloader = dpo_preprocessing_pipeline(\n          train_ds,\n          config,\n          data_columns=config.train_data_columns,\n          tokenize=config.tokenize_train_data,\n          grain_worker_count=config.grain_worker_count,\n      )\n    else:\n      train_dataloader = pretrain_preprocessing_pipeline(\n          train_ds,\n          config,\n          data_columns=config.train_data_columns,\n          tokenize=config.tokenize_train_data,\n          grain_worker_count=config.grain_worker_count,\n      )\n    return multihost_dataloading.MultiHostDataLoadIterator(train_dataloader, global_mesh)\n  else:\n    get_ds_fn = functools.partial(\n        get_datasets,\n        config.grain_train_files,\n        config.grain_file_type,\n        shuffle=config.enable_data_shuffling,\n        shuffle_seed=config.data_shuffle_seed,\n        num_epoch=config.num_epoch,\n        grain_worker_count=config.grain_worker_count,\n    )\n    if config.use_dpo:\n      preprocessing_fn = functools.partial(\n          pretrain_preprocessing_pipeline,\n          config=config,\n          data_columns=config.train_data_columns,\n          tokenize=config.tokenize_train_data,\n          grain_worker_count=config.grain_worker_count,\n      )\n    else:\n      preprocessing_fn = functools.partial(\n          pretrain_preprocessing_pipeline,\n          config=config,\n          data_columns=config.train_data_columns,\n          tokenize=config.tokenize_train_data,\n          grain_worker_count=config.grain_worker_count,\n      )\n    global_shape = (config.global_batch_size_to_load, config.max_target_length)\n    return multihost_dataloading.RemoteIterator(get_ds_fn, preprocessing_fn, global_mesh, global_shape)\n\n\ndef make_grain_eval_iterator(\n    config: ml_collections.ConfigDict,\n    global_mesh,\n    process_indices,\n):\n  """"""Load, preprocess dataset and return iterators""""""\n  assert (\n      config.global_batch_size_to_load_eval % global_mesh.size == 0\n  ), ""Batch size should be divisible by number of global devices.""\n  if not config.colocated_python_data_input:\n    eval_ds = get_datasets(\n        config.grain_eval_files,\n        config.grain_file_type,\n        shuffle=False,\n        shuffle_seed=config.data_shuffle_seed,\n        num_epoch=1,\n        dataloading_host_index=process_indices.index(jax.process_index()),\n        dataloading_host_count=len(process_indices),\n        grain_worker_count=config.grain_worker_count_eval,\n    )\n    if config.use_dpo:\n      eval_dataloader = dpo_preprocessing_pipeline(\n          eval_ds,\n          config,\n          data_columns=config.eval_data_columns,\n          tokenize=config.tokenize_eval_data,\n          grain_worker_count=config.grain_worker_count_eval,\n      )\n    else:\n      eval_dataloader = pretrain_preprocessing_pipeline(\n          eval_ds,\n          config,\n          data_columns=config.eval_data_columns,\n          tokenize=config.tokenize_eval_data,\n          grain_worker_count=config.grain_worker_count_eval,\n      )\n    return multihost_dataloading.MultiHostDataLoadIterator(eval_dataloader, global_mesh)\n  else:\n    get_ds_fn = functools.partial(\n        get_datasets,\n        config.grain_eval_files,\n        config.grain_file_type,\n        shuffle=False,  # No shuffle for eval\n        shuffle_seed=config.data_shuffle_seed,\n        num_epoch=1,\n        grain_worker_count=config.grain_worker_count_eval,\n    )\n    if config.use_dpo:\n      preprocessing_fn = functools.partial(\n          dpo_preprocessing_pipeline,\n          config=config,\n          data_columns=config.eval_data_columns,\n          tokenize=config.tokenize_eval_data,\n          grain_worker_count=config.grain_worker_count_eval,\n      )\n    else:\n      preprocessing_fn = functools.partial(\n          pretrain_preprocessing_pipeline,\n          config=config,\n          data_columns=config.eval_data_columns,\n          tokenize=config.tokenize_eval_data,\n          grain_worker_count=config.grain_worker_count_eval,\n      )\n    global_shape = (config.global_batch_size_to_load, config.max_target_length)\n    return multihost_dataloading.RemoteIterator(get_ds_fn, preprocessing_fn, global_mesh, global_shape)\n",python,tab
+4,123936,"MaxText/input_pipeline/_grain_data_processing.py",1267,0,"",python,selection_keyboard
+5,124218,"MaxText/input_pipeline/_grain_data_processing.py",0,0,"",python,selection_keyboard
+6,144625,"MaxText/input_pipeline/_grain_data_processing.py",1267,0,"",python,selection_keyboard
+7,145360,"MaxText/input_pipeline/_grain_data_processing.py",0,0,"",python,selection_keyboard
+8,168938,"MaxText/input_pipeline/_grain_data_processing.py",1267,0,"",python,selection_keyboard
+9,170117,"MaxText/input_pipeline/_grain_data_processing.py",0,0,"",python,selection_keyboard
+10,171861,"MaxText/input_pipeline/_grain_data_processing.py",1267,0,"",python,selection_keyboard
+11,177264,"MaxText/input_pipeline/_grain_data_processing.py",3212,0,"",python,selection_keyboard
+12,186320,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+13,190604,"MaxText/input_pipeline/_grain_data_processing.py",0,0,"",python,tab
+14,190606,"TERMINAL",0,0,"",,terminal_focus
+15,192314,"TERMINAL",0,0,"squeue",,terminal_command
+16,192321,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  21567 franz.sram interacti     1     1  R 2025-08-20T14:53:47 2025-08-20T14:53:47  4:53:28 1-00:00:00 hai005\r\n  20995   xiao.liu interacti     1    32  R 2025-08-20T05:38:54 2025-08-20T05:38:54 14:08:21   23:59:00 hai004\r\n  20993   xiao.liu interacti     2    64  R 2025-08-19T20:31:34 2025-08-19T20:31:34 23:15:41   23:59:00 hai[001-002]\r\n  21821 alfred.ngu  standard     1     8  R 2025-08-20T17:30:23 2025-08-20T17:30:23  2:16:52 1-00:00:00 hai006\r\n  21820 nishant.ku  standard     3   192  R 2025-08-20T16:59:48 2025-08-20T16:59:48  2:47:27 1-00:00:00 hai[003,007-008]\r\n  21806 alfred.ngu  standard     1     8  R 2025-08-20T16:29:13 2025-08-20T16:29:17  3:17:58 1-00:00:00 hai005\r\n  21578 alfred.ngu  standard     1     5  R 2025-08-20T15:44:35 2025-08-20T15:44:35  4:02:40 1-00:00:00 hai006\r\n  21312 alfred.ngu  standard     1     8  R 2025-08-20T11:54:10 2025-08-20T11:54:21  7:52:54 1-00:00:00 hai006\r\n  21311 alfred.ngu  standard     1     5  R 2025-08-20T11:51:51 2025-08-20T11:51:51  7:55:24 1-00:00:00 hai005\r\n]0;franz.srambical@hai-login1:~/maxtext",,terminal_output
+17,216776,"TERMINAL",0,0,"scancel 21567",,terminal_command
+18,216779,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login1:~/maxtext",,terminal_output
+19,219268,"TERMINAL",0,0,"sqaueue",,terminal_command
+20,219271,"TERMINAL",0,0,"]633;Cbash: sqaueue: command not found\r\n]0;franz.srambical@hai-login1:~/maxtext",,terminal_output
+21,220385,"TERMINAL",0,0,"squeue",,terminal_command
+22,220390,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  21567 franz.sram interacti     1     1 CG 2025-08-20T14:53:47 2025-08-20T14:53:47  4:53:53 1-00:00:00 hai005\r\n  20995   xiao.liu interacti     1    32  R 2025-08-20T05:38:54 2025-08-20T05:38:54 14:08:49   23:59:00 hai004\r\n  20993   xiao.liu interacti     2    64  R 2025-08-19T20:31:34 2025-08-19T20:31:34 23:16:09   23:59:00 hai[001-002]\r\n  21821 alfred.ngu  standard     1     8  R 2025-08-20T17:30:23 2025-08-20T17:30:23  2:17:20 1-00:00:00 hai006\r\n  21820 nishant.ku  standard     3   192  R 2025-08-20T16:59:48 2025-08-20T16:59:48  2:47:55 1-00:00:00 hai[003,007-008]\r\n  21806 alfred.ngu  standard     1     8  R 2025-08-20T16:29:13 2025-08-20T16:29:17  3:18:26 1-00:00:00 hai005\r\n  21578 alfred.ngu  standard     1     5  R 2025-08-20T15:44:35 2025-08-20T15:44:35  4:03:08 1-00:00:00 hai006\r\n  21312 alfred.ngu  standard     1     8  R 2025-08-20T11:54:10 2025-08-20T11:54:21  7:53:22 1-00:00:00 hai006\r\n  21311 alfred.ngu  standard     1     5  R 2025-08-20T11:51:51 2025-08-20T11:51:51  7:55:52 1-00:00:00 hai005\r\n]0;franz.srambical@hai-login1:~/maxtext",,terminal_output
+23,223217,"TERMINAL",0,0,"salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G",,terminal_command
+24,223269,"TERMINAL",0,0,"]633;Csalloc: Granted job allocation 21824\r\n",,terminal_output
+25,223365,"TERMINAL",0,0,"salloc: Nodes hai006 are ready for job\r\n",,terminal_output
+26,223826,"TERMINAL",0,0,"Running inside SLURM, Job ID 21824.\r\n",,terminal_output
+27,223912,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/maxtext[?2004h[franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+28,224460,"TERMINAL",0,0,"[franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+29,250424,"TERMINAL",0,0,"\r[franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+30,252357,"TERMINAL",0,0,"export XLA_PYTHON_CLIENT_MEM_FRACTION=0.94\r\n\r",,terminal_output
+31,253259,"TERMINAL",0,0,"export XLA_PYTHON_CLIENT_MEM_FRACTION=0.94\r\n\r\r\n[?2004l\r]0;franz.srambical@hai-login1:~/maxtext[?2004h[franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+32,261732,"TERMINAL",0,0,"export XLA_FLAGS=""--xla_gpu_enable_latency_hiding_scheduler=true --xla_disable_hlo_passes=rematerialization""",,terminal_output
+33,276191,"TERMINAL",0,0,"export XLA_FLAGS=""--xla_gpu_enable_latency_hiding_scheduler=true --xla_disable_hlo_passes=rematerialization""\r\n[?2004l\r]0;franz.srambical@hai-login1:~/maxtext[?2004h[franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+34,289710,"TERMINAL",0,0,"python3 -m MaxText.train MaxText/configs/base.yml \\r\n\r  run_name=h100_mfu_test \\r\n\r  hardware=gpu \\r\n\r  dataset_type=synthetic \\r\n\r  steps=60 \\r\n\r  log_period=1 \\r\n\r  enable_checkpointing=False \\r\n\r  gcs_metrics=False \\r\n\r  metrics_file=/tmp/h100_mfu_metrics.jsonl \\r\n\r  base_output_directory=/tmp/maxtext \\r\n\r  attention=cudnn_flash \\r\n\r  per_device_batch_size=4",,terminal_output
+35,290396,"TERMINAL",0,0,"python3 -m MaxText.train MaxText/configs/base.yml \\r\n\r  run_name=h100_mfu_test \\r\n\r  hardware=gpu \\r\n\r  dataset_type=synthetic \\r\n\r  steps=60 \\r\n\r  log_period=1 \\r\n\r  enable_checkpointing=False \\r\n\r  gcs_metrics=False \\r\n\r  metrics_file=/tmp/h100_mfu_metrics.jsonl \\r\n\r  base_output_directory=/tmp/maxtext \\r\n\r  attention=cudnn_flash \\r\n\r  per_device_batch_size=4\r\n[?2004l\r",,terminal_output
+36,290468,"TERMINAL",0,0,"Traceback (most recent call last):\r\n  File ""<frozen runpy>"", line 189, in _run_module_as_main\r\n  File ""<frozen runpy>"", line 112, in _get_module_details\r\n  File ""/fast/home/franz.srambical/maxtext/MaxText/__init__.py"", line 27, in <module>\r\n    from MaxText import maxtext_utils\r\n  File ""/fast/home/franz.srambical/maxtext/MaxText/maxtext_utils.py"", line 21, in <module>\r\n    from flax import linen as nn\r\nModuleNotFoundError: No module named 'flax'\r\n]0;franz.srambical@hai-login1:~/maxtext[?2004h[franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+37,294639,"TERMINAL",0,0,"uv",,terminal_output
+38,294852,"TERMINAL",0,0," ",,terminal_output
+39,386715,"TERMINAL",0,0,"v",,terminal_output
+40,386827,"TERMINAL",0,0,"e",,terminal_output
+41,386985,"TERMINAL",0,0,"n",,terminal_output
+42,387092,"TERMINAL",0,0,"v",,terminal_output
+43,388070,"TERMINAL",0,0," ",,terminal_output
+44,388298,"TERMINAL",0,0,"-",,terminal_output
+45,388420,"TERMINAL",0,0,"-",,terminal_output
+46,388894,"TERMINAL",0,0,"p",,terminal_output
+47,388950,"TERMINAL",0,0,"y",,terminal_output
+48,389173,"TERMINAL",0,0,"h",,terminal_output
+49,389600,"TERMINAL",0,0,"t",,terminal_output
+50,389696,"TERMINAL",0,0,"h",,terminal_output
+51,389750,"TERMINAL",0,0,"o",,terminal_output
+52,389869,"TERMINAL",0,0,"n",,terminal_output
+53,390053,"TERMINAL",0,0," ",,terminal_output
+54,390352,"TERMINAL",0,0,"3",,terminal_output
+55,390555,"TERMINAL",0,0,".",,terminal_output
+56,391084,"TERMINAL",0,0,"1",,terminal_output
+57,391143,"TERMINAL",0,0,"2",,terminal_output
+58,394023,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+59,394440,"TERMINAL",0,0,"Using CPython 3.12.9 interpreter at: /opt/miniforge3/bin/python3.12\r\nCreating virtual environment at: .venv\r\nActivate with: source .venv/bin/activate\r\n]0;franz.srambical@hai-login1:~/maxtext[?2004h[franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+60,395346,"TERMINAL",0,0,"\r(reverse-i-search)`': ",,terminal_output
+61,395635,"TERMINAL",0,0,"s': python3 -m MaxText.train MaxText/configs/base.yml   run_name=h100_mfu_test   hardware=gpu   dataset_type=synthetic   steps=60   log_period=1   enable_checkpointing=False   gcs_metrics=False   metrics_file=/tmp/h100_mfu_metrics.jsonl   base_output_directory=/tmp/maxtext   attention=cudnn_flash   per_device_batch_size=4\ro': python3 -m MaxText.train MaxText/configs/base.yml   run_name=h100_mfu_test   hardware=gpu  dataset_type=synthetic   steps=60   log_period=1   enable_checkpointing=False   gcs_metrics=False   metrics_file=/tmp/h100_mfu_metrics.jsonl   base_output_directory=/tmp/maxtext   attention=cudnn_flash   per_device_batch_size=4u': source /home/franz.srambical/jafar/.venv/bin/activate\r\n\r\r\n\r\r\n\r",,terminal_output
+62,395748,"TERMINAL",0,0,"[1@r': sour",,terminal_output
+63,396917,"TERMINAL",0,0,"\r[26@[franz.srambical@hai006.haicore.berlin:~/maxtext] $ sour",,terminal_output
+64,397549,"TERMINAL",0,0,"",,terminal_output
+65,398191,"TERMINAL",0,0,"so",,terminal_output
+66,398270,"TERMINAL",0,0,"u",,terminal_output
+67,398326,"TERMINAL",0,0,"r",,terminal_output
+68,398534,"TERMINAL",0,0,"c",,terminal_output
+69,398772,"TERMINAL",0,0,"e",,terminal_output
+70,398944,"TERMINAL",0,0," .",,terminal_output
+71,399070,"TERMINAL",0,0,"v",,terminal_output
+72,399150,"TERMINAL",0,0,"e",,terminal_output
+73,399283,"TERMINAL",0,0,"n",,terminal_output
+74,399553,"TERMINAL",0,0,"v/",,terminal_output
+75,399934,"TERMINAL",0,0,"v",,terminal_output
+76,400029,"TERMINAL",0,0,"i",,terminal_output
+77,400359,"TERMINAL",0,0,"",,terminal_output
+78,400565,"TERMINAL",0,0,"b",,terminal_output
+79,400642,"TERMINAL",0,0,"i",,terminal_output
+80,400811,"TERMINAL",0,0,"n",,terminal_output
+81,400986,"TERMINAL",0,0,"/",,terminal_output
+82,401236,"TERMINAL",0,0,"a",,terminal_output
+83,401601,"TERMINAL",0,0,"ctivate",,terminal_output
+84,402463,"TERMINAL",0,0,"\r\n[?2004l\r]0;franz.srambical@hai-login1:~/maxtext[?2004h(maxtext) [franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+85,412365,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_output
+86,412868,"TERMINAL",0,0,"",,terminal_output
+87,414950,"TERMINAL",0,0,"b",,terminal_output
+88,415254,"TERMINAL",0,0,"as",,terminal_output
+89,415368,"TERMINAL",0,0,"h",,terminal_output
+90,415703,"TERMINAL",0,0," ",,terminal_output
+91,416248,"TERMINAL",0,0,"set",,terminal_output
+92,416382,"TERMINAL",0,0,"up",,terminal_output
+93,416585,"TERMINAL",0,0,"",,terminal_output
+94,518944,"TERMINAL",0,0,"",,terminal_output
+95,521558,"TERMINAL",0,0,"u",,terminal_output
+96,521637,"TERMINAL",0,0,"v",,terminal_output
+97,521830,"TERMINAL",0,0," ",,terminal_output
+98,521934,"TERMINAL",0,0,"pi",,terminal_output
+99,522074,"TERMINAL",0,0,"p",,terminal_output
+100,522307,"TERMINAL",0,0," ",,terminal_output
+101,522463,"TERMINAL",0,0,"i",,terminal_output
+102,522688,"TERMINAL",0,0,"n",,terminal_output
+103,522774,"TERMINAL",0,0,"s",,terminal_output
+104,522836,"TERMINAL",0,0,"t",,terminal_output
+105,522937,"TERMINAL",0,0,"a",,terminal_output
+106,523021,"TERMINAL",0,0,"l",,terminal_output
+107,523143,"TERMINAL",0,0,"l",,terminal_output
+108,523322,"TERMINAL",0,0," ",,terminal_output
+109,523511,"TERMINAL",0,0,"-",,terminal_output
+110,523769,"TERMINAL",0,0,"r",,terminal_output
+111,523959,"TERMINAL",0,0," ",,terminal_output
+112,524084,"TERMINAL",0,0,"r",,terminal_output
+113,524139,"TERMINAL",0,0,"e",,terminal_output
+114,524239,"TERMINAL",0,0,"q",,terminal_output
+115,524351,"TERMINAL",0,0,"iu",,terminal_output
+116,524490,"TERMINAL",0,0,"r",,terminal_output
+117,524564,"TERMINAL",0,0,"e",,terminal_output
+118,524690,"TERMINAL",0,0,"m",,terminal_output
+119,525082,"TERMINAL",0,0,"",,terminal_output
+120,525225,"TERMINAL",0,0,"",,terminal_output
+121,525380,"TERMINAL",0,0,"",,terminal_output
+122,525728,"TERMINAL",0,0,"",,terminal_output
+123,525924,"TERMINAL",0,0,"r",,terminal_output
+124,526007,"TERMINAL",0,0,"e",,terminal_output
+125,526069,"TERMINAL",0,0,"u",,terminal_output
+126,526125,"TERMINAL",0,0,"i",,terminal_output
+127,526750,"TERMINAL",0,0,"",,terminal_output
+128,526913,"TERMINAL",0,0,"",,terminal_output
+129,526994,"TERMINAL",0,0,"q",,terminal_output
+130,527058,"TERMINAL",0,0,"u",,terminal_output
+131,527112,"TERMINAL",0,0,"i",,terminal_output
+132,527383,"TERMINAL",0,0,"rements",,terminal_output
+133,530897,"TERMINAL",0,0,".",,terminal_output
+134,531220,"TERMINAL",0,0,"txt ",,terminal_output
+135,581173,"TERMINAL",0,0,"-",,terminal_output
+136,581420,"TERMINAL",0,0,"c",,terminal_output
+137,581955,"TERMINAL",0,0," ",,terminal_output
+138,583874,"TERMINAL",0,0,"c",,terminal_output
+139,583958,"TERMINAL",0,0,"o",,terminal_output
+140,584183,"TERMINAL",0,0,"n",,terminal_output
+141,584270,"TERMINAL",0,0,"s",,terminal_output
+142,584354,"TERMINAL",0,0,"t",,terminal_output
+143,584495,"TERMINAL",0,0,"a",,terminal_output
+144,584865,"TERMINAL",0,0,"",,terminal_output
+145,585176,"TERMINAL",0,0,"r",,terminal_output
+146,585275,"TERMINAL",0,0,"a",,terminal_output
+147,585453,"TERMINAL",0,0,"ints_gpu.txt ",,terminal_output
+148,587670,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+149,587813,"TERMINAL",0,0,"⠋ Resolving dependencies...                                                                                      \r⠙ Resolving dependencies...                                                                                      ",,terminal_output
+150,588008,"TERMINAL",0,0,"\r⠹ Resolving dependencies...                                                                                      \r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n⠹ Resolving dependencies...                                                                                      \r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n⠹ Resolving dependencies...                                                                                      ",,terminal_output
+151,588209,"TERMINAL",0,0,"\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠹ Resolving dependencies...                                                                                      \r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠹ Resolving dependencies...                                                                                      \r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠸ Resolving dependencies...                                                                                      ",,terminal_output
+152,588410,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠼ Resolving dependencies...                                                                                      ",,terminal_output
+153,588610,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠴ Resolving dependencies...                                                                                      ",,terminal_output
+154,588813,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠦ Resolving dependencies...                                                                                      ",,terminal_output
+155,589012,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠧ Resolving dependencies...                                                                                      ",,terminal_output
+156,589211,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠇ Resolving dependencies...                                                                                      ",,terminal_output
+157,589411,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠋ Resolving dependencies...                                                                                      ",,terminal_output
+158,589612,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠙ Resolving dependencies...                                                                                      ",,terminal_output
+159,589816,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠹ Resolving dependencies...                                                                                      ",,terminal_output
+160,590008,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠸ Resolving dependencies...                                                                                      ",,terminal_output
+161,590211,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠼ Resolving dependencies...                                                                                      ",,terminal_output
+162,590412,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠴ Resolving dependencies...                                                                                      ",,terminal_output
+163,590609,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠦ Resolving dependencies...                                                                                      ",,terminal_output
+164,590810,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠧ Resolving dependencies...                                                                                      ",,terminal_output
+165,591009,"TERMINAL",0,0,"\r\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠇ Resolving dependencies...                                                                                      ",,terminal_output
+166,591211,"TERMINAL",0,0,"\r\r\r    Updated https://github.com/mlperf/logging.git (ed4e6d5f84e3439d6a86f9e63646bed3b0606292)\r\n   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠇ Resolving dependencies...                                                                                      \r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠋ Resolving dependencies...                                                                                      ",,terminal_output
+167,591410,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠙ Resolving dependencies...                                                                                      ",,terminal_output
+168,591610,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠹ Resolving dependencies...                                                                                      ",,terminal_output
+169,591805,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠸ Resolving dependencies...                                                                                      ",,terminal_output
+170,592020,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠼ Resolving dependencies...                                                                                      ",,terminal_output
+171,592209,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠴ Resolving dependencies...                                                                                      ",,terminal_output
+172,592410,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠦ Resolving dependencies...                                                                                      ",,terminal_output
+173,592612,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠧ Resolving dependencies...                                                                                      ",,terminal_output
+174,592810,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠇ Resolving dependencies...                                                                                      ",,terminal_output
+175,593015,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠋ Resolving dependencies...                                                                                      ",,terminal_output
+176,593245,"TERMINAL",0,0,"\r\r   Updating https://github.com/AI-Hypercomputer/JetStream.git (HEAD)\r\n⠙ Resolving dependencies...                                                                                      \r\r    Updated https://github.com/AI-Hypercomputer/JetStream.git (29329e8e73820993f77cfc8efe34eb2a73f5de98)\r\n⠙ Resolving dependencies...                                                                                      ",,terminal_output
+177,593414,"TERMINAL",0,0,"\r⠹ Resolving dependencies...                                                                                      ",,terminal_output
+178,593612,"TERMINAL",0,0,"\r⠸ Resolving dependencies...                                                                                      ",,terminal_output
+179,593810,"TERMINAL",0,0,"\r⠼ Resolving dependencies...                                                                                      ",,terminal_output
+180,594010,"TERMINAL",0,0,"\r⠴ Resolving dependencies...                                                                                      ",,terminal_output
+181,594212,"TERMINAL",0,0,"\r⠦ Resolving dependencies...                                                                                      ",,terminal_output
+182,594411,"TERMINAL",0,0,"\r⠧ Resolving dependencies...                                                                                      ",,terminal_output
+183,594613,"TERMINAL",0,0,"\r⠇ Resolving dependencies...                                                                                      ",,terminal_output
+184,594811,"TERMINAL",0,0,"\r⠋ Resolving dependencies...                                                                                      ",,terminal_output
+185,595011,"TERMINAL",0,0,"\r⠙ Resolving dependencies...                                                                                      ",,terminal_output
+186,595211,"TERMINAL",0,0,"\r⠹ Resolving dependencies...                                                                                      ",,terminal_output
+187,595412,"TERMINAL",0,0,"\r⠸ Resolving dependencies...                                                                                      ",,terminal_output
+188,595506,"TERMINAL",0,0,"\r⠋ Resolving dependencies...                                                                                      \r⠙ Resolving dependencies...                                                                                      ",,terminal_output
+189,595750,"TERMINAL",0,0,"\r⠹ Resolving dependencies...                                                                                      \r⠹                                                                                                                \r  × No solution found when resolving dependencies:\r\n  ╰─▶ Because you require sentencepiece==0.2.0 and sentencepiece==0.1.97, we can conclude that your\r\n      requirements are unsatisfiable.\r\n",,terminal_output
+190,597090,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/maxtext[?2004h(maxtext) [franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+191,617290,"constraints_gpu.txt",0,0,"Deprecated>=1.2.18\nJinja2>=3.1.6\nMarkdown>=3.7\nMarkupSafe>=3.0.2\nPyGObject>=3.42.1\nPyYAML>=5.4.1\nPygments==2.18.0\nWerkzeug==2.0.3\nabsl-py==2.1.0\naiohappyeyeballs==2.4.3\naiohttp==3.10.10\naiosignal==1.3.1\nannotated-types==0.7.0\nanyio==4.6.2.post1\naqtp==0.8.2\narray_record==0.7.2\nastroid==3.3.5\nastunparse==1.6.3\nasync-timeout==4.0.3\nattrs==24.2.0\nblack==24.8.0\nblobfile==3.0.0\nbuild==1.2.2.post1\ncachetools==5.5.0\ncertifi==2024.8.30\ncfgv==3.4.0\ncharset-normalizer==3.4.0\nchex==0.1.87\nclick==8.1.7\ncloud-accelerator-diagnostics==0.1.1\ncloud-tpu-diagnostics==0.1.5\ncloudpickle==3.1.0\nclu==0.0.12\ncontextlib2==21.6.0\ncoverage==7.6.4\ndatasets==3.0.2\ndbus-python==1.2.18\ndecorator==5.1.1\ndill==0.3.8\ndistlib==0.3.9\ndm-tree==0.1.8\ndocstring_parser==0.16\neditdistance==0.8.1\netils==1.10.0\nexceptiongroup==1.2.2\nfastapi==0.115.3\nfilelock==3.16.1\nflatbuffers==24.3.25\nflax==0.11.0\nfrozenlist==1.5.0\nfsspec==2024.9.0\ngast==0.6.0\ngcsfs==2024.9.0.post1\ngoogle-api-core==2.24.1\ngoogle-auth-oauthlib==1.2.1\ngoogle-auth==2.35.0\ngoogle-cloud-aiplatform==1.61.0\ngoogle-cloud-appengine-logging==1.4.5\ngoogle-cloud-audit-log==0.3.0\ngoogle-cloud-bigquery==3.26.0\ngoogle-cloud-core==2.4.1\ngoogle-cloud-logging==3.11.3\ngoogle-cloud-resource-manager==1.12.5\ngoogle-cloud-storage==2.18.2\ngoogle-crc32c==1.6.0\ngoogle-pasta==0.2.0\ngoogle-resumable-media==2.7.2\ngoogleapis-common-protos==1.65.0\ngrain-nightly==0.0.10\ngrpc-google-iam-v1==0.13.1\ngrpcio-status==1.48.2\ngrpcio==1.67.0\ngviz-api==1.10.0\nh11==0.14.0\nh5py==3.12.1\nhuggingface-hub==0.26.1\nhumanize==4.11.0\nidentify==2.6.1\nidna==3.10\nimmutabledict==4.2.0\nimportlab==0.8.1\nimportlib_metadata==8.4.0\nimportlib_resources==6.4.5\niniconfig==2.0.0\nisort\njax-cuda12-pjrt\njax-cuda12-plugin\njax==0.6.2\njaxlib==0.6.2\njaxtyping==0.3.2\njsonlines==4.0.0\nkeras==3.10.0\nlibclang==18.1.1\nlibcst==1.8.2\nlxml==5.4.0\nmarkdown-it-py==3.0.0\nmccabe==0.7.0\nmdurl==0.1.2\nml-collections==0.1.1\nml-dtypes==0.4.1\nml-goodput-measurement==0.0.10\nmore-itertools==10.5.0\nmsgpack==1.1.0\nmsgspec==0.18.6\nmultidict==6.1.0\nmultiprocess==0.70.16\nmypy-extensions==1.0.0\nnamex==0.0.8\nnest-asyncio==1.6.0\nnetworkx==3.4.2\nninja==1.11.1.1\nnodeenv==1.9.1\nnumpy==1.26.4\nnvidia-cublas-cu12==12.6.4.1\nnvidia-cuda-cupti-cu12==12.6.80\nnvidia-cuda-nvcc-cu12==12.6.85\nnvidia-cuda-runtime-cu12==12.6.77\nnvidia-cudnn-cu12==9.6.0.74\nnvidia-cufft-cu12==11.3.0.4\nnvidia-cusolver-cu12==11.7.1.2\nnvidia-cusparse-cu12==12.5.4.2\nnvidia-nccl-cu12==2.23.4\nnvidia-nvjitlink-cu12==12.6.85\noauthlib==3.2.2\nomegaconf\nopentelemetry-api==1.27.0\nopt_einsum==3.4.0\noptax==0.2.3\noptree==0.13.0\norbax-checkpoint==0.11.22\npackaging==24.1\npandas==2.2.3\nparameterized==0.9.0\npathspec==0.12.1\npip-tools==7.4.1\nplatformdirs==4.3.6\npluggy==1.5.0\nportpicker==1.6.0\npre_commit==4.0.1\nprometheus_client==0.21.0\npromise==2.3\npropcache==0.2.0\nproto-plus==1.25.0\nprotobuf==3.20.3\npsutil==6.1.0\npyarrow==17.0.0\npyasn1==0.6.1\npyasn1_modules==0.4.1\npybind11==2.13.6\npybind11_global==2.13.6\npycnite==2024.7.31\npycryptodomex==3.21.0\npydantic==2.9.2\npydantic_core==2.23.4\npydot==3.0.2\npyglove==0.4.4\npyink==24.10.0\npylint==3.3.1\npyparsing==3.2.0\npyproject_hooks==1.2.0\npytest==8.3.3\npython-dateutil==2.9.0.post0\npytype==2024.10.11\npytz==2024.2\nqwix@git+https://github.com/google/qwix.git\nregex==2024.9.11\nrequests-oauthlib==2.0.0\nrequests==2.32.3\nrich==13.9.3\nrsa==4.9\nsafetensors==0.4.5\nscipy==1.14.1\nsentencepiece==0.1.97\nseqio==0.0.19\nshapely==2.0.6\nshortuuid==1.0.13\nsimple-parsing==0.1.6\nsix==1.16.0\nsniffio==1.3.1\nstarlette==0.41.0\ntabulate==0.9.0\ntensorboard-data-server==0.7.2\ntensorboard-plugin-profile==2.17.0\ntensorboard==2.17.1\ntensorboardX==2.6.2.2\ntensorflow-datasets==4.9.6\ntensorflow-io-gcs-filesystem==0.37.1\ntensorflow-metadata==1.16.1\ntensorflow-text==2.17.0\ntensorflow==2.17.0\ntensorstore==0.1.68\ntermcolor==2.5.0\ntfds-nightly==4.9.2.dev202308090034\ntiktoken==0.8.0\ntokenizers==0.20.1\ntoml==0.10.2\ntomli==2.0.2\ntomlkit==0.13.2\ntoolz==1.0.0\ntqdm==4.66.5\ntransformer-engine==1.13.0+e5edd6c\ntransformers==4.46.0\ntunix@git+https://github.com/google/tunix.git\ntypeguard==2.13.3\ntyping_extensions==4.12.2\ntzdata==2024.2\nurllib3==2.2.3\nuvicorn==0.32.0\nvirtualenv==20.27.0\nwrapt==1.16.0\nxxhash==3.5.0\nyarl==1.16.0\nzipp==3.20.2\n",plaintext,tab
+192,623065,"constraints_gpu.txt",3342,0,"",plaintext,selection_command
+193,623720,"constraints_gpu.txt",3342,22,"",plaintext,content
+194,625870,"TERMINAL",0,0,"\r\n[?2004l\r]0;franz.srambical@hai-login1:~/maxtext[?2004h(maxtext) [franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+195,626999,"constraints_gpu.txt",3355,0,"\n",plaintext,content
+196,627736,"constraints_gpu.txt",3356,1,"",plaintext,content
+197,629721,"TERMINAL",0,0,"uv pip install -r requirements.txt -c constraints_gpu.txt ",,terminal_output
+198,629940,"TERMINAL",0,0,"\r\n[?2004l\r⠋ Resolving dependencies...                                                                                      \r⠙ Resolving dependencies...                                                                                      ",,terminal_output
+199,630142,"TERMINAL",0,0,"\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n⠙ Resolving dependencies...                                                                                      \r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n⠙ Resolving dependencies...                                                                                      \r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n⠹ Resolving dependencies...                                                                                      ",,terminal_output
+200,630342,"TERMINAL",0,0,"\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n⠸ Resolving dependencies...                                                                                      ",,terminal_output
+201,630542,"TERMINAL",0,0,"\r\r   Updating https://github.com/mlperf/logging.git (HEAD)\r\n⠼ Resolving dependencies...                                                                                      ",,terminal_output
+202,630645,"TERMINAL",0,0,"\r\r    Updated https://github.com/mlperf/logging.git (ed4e6d5f84e3439d6a86f9e63646bed3b0606292)\r\n⠼ Resolving dependencies...                                                                                      \r⠋ Resolving dependencies...                                                                                      \r⠙ Resolving dependencies...                                                                                      \r⠙                                                                                                                \r  × No solution found when resolving dependencies:\r\n  ╰─▶ Because you require datasets>=4.0.0 and datasets==3.0.2, we can conclude that your requirements are\r\n      unsatisfiable.\r\n]0;franz.srambical@hai-login1:~/maxtext[?2004h(maxtext) [franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+203,636725,"constraints_gpu.txt",3591,0,"",plaintext,selection_command
+204,644681,"constraints_gpu.txt",3591,7,"dataset",plaintext,selection_command
+205,644685,"constraints_gpu.txt",3591,8,"datasets",plaintext,selection_command
+206,645640,"constraints_gpu.txt",3598,0,"",plaintext,selection_command
+207,645854,"constraints_gpu.txt",628,0,"",plaintext,selection_command
+208,652518,"constraints_gpu.txt",628,16,"",plaintext,content
+209,1113605,"TERMINAL",0,0,"gi",,terminal_output
+210,1113740,"TERMINAL",0,0,"t",,terminal_output
+211,1113907,"TERMINAL",0,0," l",,terminal_output
+212,1114028,"TERMINAL",0,0,"o",,terminal_output
+213,1114084,"TERMINAL",0,0,"g",,terminal_output
+214,1114313,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+215,1114385,"TERMINAL",0,0,"[?1h=\rcommit 8def32a8a5b96fc6267636a8e58abfe4c178e161 (HEAD -> main, origin/main, origin/HEAD)\r\nMerge: dd7b6f77 7b03cc07\r\nAuthor: maxtext authors <google-ml-automation@google.com>\r\nDate:   Tue Aug 19 14:19:27 2025 -0700\r\n\r\n    Merge pull request #2199 from phu0ngng:fix_norm_axis\r\n    \r\n    PiperOrigin-RevId: 797018596\r\n\r\ncommit dd7b6f771e4341facfe55e256457c101b8e86f77\r\nMerge: d3763c23 146e8196\r\nAuthor: maxtext authors <google-ml-automation@google.com>\r\nDate:   Tue Aug 19 10:58:23 2025 -0700\r\n\r\n    Merge pull request #2201 from AI-Hypercomputer:sft_readme_with_ckpt_conversion\r\n    \r\n    PiperOrigin-RevId: 796941951\r\n\r\ncommit d3763c234e2f65358893c540c60831f82992d499\r\nMerge: a2cca69b 6e7789cc\r\nAuthor: maxtext authors <google-ml-automation@google.com>\r\nDate:   Tue Aug 19 09:46:02 2025 -0700\r\n\r\n    Merge pull request #2188 from AI-Hypercomputer:SelfDebuggingAgent\r\n    \r\n    PiperOrigin-RevId: 796911993\r\n\r\ncommit a2cca69b54c2c1c5ee95271eb9e80207220d58c7\r\nAuthor: Jesse Lu <jesselu@google.com>\r\nDate:   Tue Aug 19 09:29:34 2025 -0700\r\n\r\n    Introduce `_sort_activations()` with custom backward pass.\r\n    \r\n    PiperOrigin-RevId: 796906239\r\n\r\ncommit 21cb1132df8495466e0aa0a7c0e16e2507c7cc67\r\nMerge: aada9454 af1e9742\r\nAuthor: maxtext authors <google-ml-automation@google.com>\r\nDate:   Tue Aug 19 09:10:53 2025 -0700\r\n\r\n    Merge pull request #2000 from SamuelMarks:qa_MaxText.inference\r\n:",,terminal_output
+216,1119085,"TERMINAL",0,0,"\r[?1l>]0;franz.srambical@hai-login1:~/maxtext[?2004h(maxtext) [franz.srambical@hai006.haicore.berlin:~/maxtext] $ ",,terminal_output
+217,1160076,"TERMINAL",0,0,"git log",,terminal_output
+218,1160821,"TERMINAL",0,0,"uv pip install -r requirements.txt -c constraints_gpu.txt ",,terminal_output
+219,1166243,"TERMINAL",0,0,"",,terminal_output
+220,1167068,"TERMINAL",0,0,"\r\r\n\r",,terminal_output
+221,1167204,"TERMINAL",0,0,"",,terminal_output
+222,1167363,"TERMINAL",0,0,"",,terminal_output
+223,1167501,"TERMINAL",0,0,"",,terminal_output
+224,1167647,"TERMINAL",0,0,"",,terminal_output
+225,1167800,"TERMINAL",0,0,"",,terminal_output
+226,1169375,"constraints_gpu.txt",1027,0,"",plaintext,selection_command
+227,1169451,"constraints_gpu.txt",1535,0,"",plaintext,selection_command
+228,1169621,"constraints_gpu.txt",1881,0,"",plaintext,selection_command
+229,1169797,"constraints_gpu.txt",2352,0,"",plaintext,selection_command
+230,1169972,"constraints_gpu.txt",2757,0,"",plaintext,selection_command
+231,1170134,"constraints_gpu.txt",3136,0,"",plaintext,selection_command
+232,1170300,"constraints_gpu.txt",3542,0,"",plaintext,selection_command
+233,1170469,"constraints_gpu.txt",4021,0,"",plaintext,selection_command
+234,1170636,"constraints_gpu.txt",4141,0,"",plaintext,selection_command
+235,1171310,"constraints_gpu.txt",3788,0,"",plaintext,selection_command
+236,1171466,"constraints_gpu.txt",3326,0,"",plaintext,selection_command
+237,1171790,"constraints_gpu.txt",2936,0,"",plaintext,selection_command
+238,1172145,"constraints_gpu.txt",2563,0,"",plaintext,selection_command
+239,1172325,"constraints_gpu.txt",2078,0,"",plaintext,selection_command
+240,1172495,"constraints_gpu.txt",1705,0,"",plaintext,selection_command
+241,1172688,"constraints_gpu.txt",1317,0,"",plaintext,selection_command
+242,1172827,"constraints_gpu.txt",803,0,"",plaintext,selection_command
+243,1173591,"constraints_gpu.txt",412,0,"",plaintext,selection_command
+244,1173783,"constraints_gpu.txt",47,0,"",plaintext,selection_command
+245,1179912,"requirements.txt",0,0,"absl-py\naqtp\narray-record\ncloud-accelerator-diagnostics\ncloud-tpu-diagnostics\ndatasets>=4.0.0\nflax>=0.11.0\ngcsfs\ngoogle-api-python-client\ngoogle-cloud-aiplatform==1.61.0\ngoogle-cloud-monitoring\ngoogle-jetstream@git+https://github.com/AI-Hypercomputer/JetStream.git\ngrain[parquet]>=0.2.6\nhuggingface_hub\njax>=0.4.30\njaxlib>=0.4.30\njaxtyping\njsonlines\nml-collections\nml-goodput-measurement\nmlperf-logging@git+https://github.com/mlperf/logging.git\nnumpy\nomegaconf\noptax\norbax-checkpoint>=0.11.22\npathwaysutils>=0.1.1\npillow>=11.1.0\npre-commit\nprotobuf==3.20.3\npyink\npylint\npytest\npytype\nqwix@git+https://github.com/google/qwix.git\nsentencepiece==0.2.0\ntensorboard-plugin-profile\ntensorboardx>=2.6.2.2\ntensorflow-datasets\ntensorflow-text>=2.17.0\ntensorflow>=2.16.0\ntiktoken\ntransformers\ntunix@git+https://github.com/google/tunix.git\n",pip-requirements,tab
+246,1185362,"requirements.txt",303,0,"",pip-requirements,selection_command
+247,1185920,"requirements.txt",305,0,"",pip-requirements,selection_command
+248,1186148,"requirements.txt",306,0,"",pip-requirements,selection_command
+249,1187172,"requirements.txt",306,0,"(",pip-requirements,content
+250,1187173,"requirements.txt",307,0,"",pip-requirements,selection_keyboard
+251,1187841,"requirements.txt",306,1,"",pip-requirements,content
+252,1188234,"requirements.txt",306,0,":",pip-requirements,content
+253,1188234,"requirements.txt",307,0,"",pip-requirements,selection_keyboard
+254,1188827,"requirements.txt",306,1,"",pip-requirements,content
+255,1189184,"requirements.txt",306,0,"{",pip-requirements,content
+256,1189184,"requirements.txt",307,0,"",pip-requirements,selection_keyboard
+257,1189895,"requirements.txt",306,1,"",pip-requirements,content
+258,1190019,"requirements.txt",305,1,"",pip-requirements,content
+259,1190691,"requirements.txt",305,0,"x",pip-requirements,content
+260,1190691,"requirements.txt",306,0,"",pip-requirements,selection_keyboard
+261,1191074,"requirements.txt",306,0,"[",pip-requirements,content
+262,1191074,"requirements.txt",307,0,"",pip-requirements,selection_keyboard
+263,1191591,"requirements.txt",307,0,"c",pip-requirements,content
+264,1191591,"requirements.txt",308,0,"",pip-requirements,selection_keyboard
+265,1191664,"requirements.txt",308,0,"u",pip-requirements,content
+266,1191664,"requirements.txt",309,0,"",pip-requirements,selection_keyboard
+267,1191778,"requirements.txt",309,0,"d",pip-requirements,content
+268,1191778,"requirements.txt",310,0,"",pip-requirements,selection_keyboard
+269,1191846,"requirements.txt",310,0,"a",pip-requirements,content
+270,1191846,"requirements.txt",311,0,"",pip-requirements,selection_keyboard
+271,1192246,"requirements.txt",311,0,"1",pip-requirements,content
+272,1192247,"requirements.txt",312,0,"",pip-requirements,selection_keyboard
+273,1192247,"requirements.txt",312,0,"2",pip-requirements,content
+274,1192247,"requirements.txt",313,0,"",pip-requirements,selection_keyboard
+275,1193260,"requirements.txt",313,0,"]",pip-requirements,content
+276,1193261,"requirements.txt",314,0,"",pip-requirements,selection_keyboard
+277,1193882,"requirements.txt",313,0,"",pip-requirements,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-647bca65-0df4-418a-9a0e-4af535614fb51767713701593-2026_01_06-16.35.08.270/source.csv

@@ -0,0 +1,51 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,304,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"4:35:08 PM [info] Activating crowd-code\n4:35:08 PM [info] Recording started\n4:35:08 PM [info] Initializing git provider using file system watchers...\n4:35:08 PM [info] Git repository found\n4:35:08 PM [info] Git provider initialized successfully\n4:35:08 PM [info] Initial git state: [object Object]\n",Log,tab
+3,6280,"TERMINAL",0,0,"",,terminal_focus
+4,386348,"src/extension.ts",0,0,"// The module 'vscode' contains the VS Code extensibility API\n// Import the module and reference it with the alias vscode in your code below\nimport * as vscode from 'vscode';\n\n// This method is called when your extension is activated\n// Your extension is activated the very first time the command is executed\nexport function activate(context: vscode.ExtensionContext) {\n\n\t// Use the console to output diagnostic information (console.log) and errors (console.error)\n\t// This line of code will only be executed once when your extension is activated\n\tconsole.log('Congratulations, your extension ""test"" is now active!');\n\n\t// The command has been defined in the package.json file\n\t// Now provide the implementation of the command with registerCommand\n\t// The commandId parameter must match the command field in package.json\n\tconst disposable = vscode.commands.registerCommand('test.helloWorld', () => {\n\t\t// The code you place here will be executed every time your command is executed\n\t\t// Display a message box to the user\n\t\tvscode.window.showInformationMessage('Hello World from test!');\n\t});\n\n\tcontext.subscriptions.push(disposable);\n}\n\n// This method is called when your extension is deactivated\nexport function deactivate() {}\n",typescript,tab
+5,396361,"src/extension.ts",0,1229,"// The module 'vscode' contains the VS Code extensibility API\n// Import the module and reference it with the alias vscode in your code below\nimport * as vscode from 'vscode';\n\n// This method is called when your extension is activated\n// Your extension is activated the very first time the command is executed\nexport function activate(context: vscode.ExtensionContext) {\n\n\t// Use the console to output diagnostic information (console.log) and errors (console.error)\n\t// This line of code will only be executed once when your extension is activated\n\tconsole.log('Congratulations, your extension ""test"" is now active!');\n\n\t// The command has been defined in the package.json file\n\t// Now provide the implementation of the command with registerCommand\n\t// The commandId parameter must match the command field in package.json\n\tconst disposable = vscode.commands.registerCommand('test.helloWorld', () => {\n\t\t// The code you place here will be executed every time your command is executed\n\t\t// Display a message box to the user\n\t\tvscode.window.showInformationMessage('Hello World from test!');\n\t});\n\n\tcontext.subscriptions.push(disposable);\n}\n\n// This method is called when your extension is deactivated\nexport function deactivate() {}\n",typescript,selection_command
+6,396460,"src/extension.ts",0,1229,"import * as vscode from 'vscode'\n\nlet output: vscode.OutputChannel\n\nexport function activate(context: vscode.ExtensionContext) {\n    output = vscode.window.createOutputChannel('Edit Trace Probe')\n\n    const log = (msg: string) => {\n        output.appendLine(`[${new Date().toISOString()}] ${msg}`)\n    }\n\n    // ---- Command to show the output channel ----\n    const showOutputCmd = vscode.commands.registerCommand(\n        'editTraceProbe.showOutput',\n        () => {\n            output.show(true) // true = preserve focus\n            log('Output channel focused via command')\n        }\n    )\n\n    context.subscriptions.push(showOutputCmd)\n\n    // ---- Optional status bar button ----\n    const statusBar = vscode.window.createStatusBarItem(\n        vscode.StatusBarAlignment.Right,\n        1000\n    )\n    statusBar.text = '$(search) Edit Trace Probe'\n    statusBar.tooltip = 'Show Edit Trace Probe output'\n    statusBar.command = 'editTraceProbe.showOutput'\n    statusBar.show()\n    context.subscriptions.push(statusBar)\n\n    // ---- Text document changes ----\n    context.subscriptions.push(\n        vscode.workspace.onDidChangeTextDocument(event => {\n            const active =\n                vscode.window.activeTextEditor?.document === event.document\n            log(\n                `TEXT_CHANGE\n  uri=${event.document.uri.toString()}\n  version=${event.document.version}\n  changes=${event.contentChanges.length}\n  active=${active}`\n            )\n        })\n    )\n\n    // ---- Document open ----\n    context.subscriptions.push(\n        vscode.workspace.onDidOpenTextDocument(doc => {\n            log(\n                `DOC_OPEN\n  uri=${doc.uri.toString()}\n  version=${doc.version}\n  language=${doc.languageId}`\n            )\n        })\n    )\n\n    // ---- File system watcher ----\n    const watcher = vscode.workspace.createFileSystemWatcher('**/*')\n    context.subscriptions.push(watcher)\n\n    watcher.onDidChange(uri => log(`FS_CHANGE uri=${uri.toString()}`))\n    watcher.onDidCreate(uri => log(`FS_CREATE uri=${uri.toString()}`))\n    watcher.onDidDelete(uri => log(`FS_DELETE uri=${uri.toString()}`))\n\n    log('Edit Trace Probe activated')\n}\n\nexport function deactivate() {}\n",typescript,content
+7,397048,"src/extension.ts",797,0,"",typescript,selection_command
+8,397797,"src/extension.ts",0,0,"",typescript,selection_command
+9,401229,"TERMINAL",0,0," *  Executing task: npm run watch ",,terminal_command
+10,401425,"TERMINAL",0,0,"/bin/bash: line 1: npm: command not found\r\n",,terminal_output
+11,469589,"TERMINAL",0,0,"",,terminal_focus
+12,473859,"TERMINAL",0,0,"npm run compile",,terminal_command
+13,476617,"TERMINAL",0,0,"module load nodejs",,terminal_command
+14,476665,"TERMINAL",0,0,"]633;C",,terminal_output
+15,476847,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/test_extension/test",,terminal_output
+16,478419,"TERMINAL",0,0,"watch",,terminal_focus
+17,482613,"TERMINAL",0,0,"bash",,terminal_focus
+18,494696,".vscode/launch.json",0,0,"// A launch configuration that compiles the extension and then opens it inside a new window\n// Use IntelliSense to learn about possible attributes.\n// Hover to view descriptions of existing attributes.\n// For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387\n{\n\t""version"": ""0.2.0"",\n\t""configurations"": [\n\t\t{\n\t\t\t""name"": ""Run Extension"",\n\t\t\t""type"": ""extensionHost"",\n\t\t\t""request"": ""launch"",\n\t\t\t""args"": [\n\t\t\t\t""--extensionDevelopmentPath=${workspaceFolder}""\n\t\t\t],\n\t\t\t""outFiles"": [\n\t\t\t\t""${workspaceFolder}/out/**/*.js""\n\t\t\t],\n\t\t\t""preLaunchTask"": ""${defaultBuildTask}""\n\t\t}\n\t]\n}\n",jsonc,tab
+19,496015,".vscode/launch.json",92,0,"",jsonc,selection_command
+20,496155,".vscode/launch.json",148,0,"",jsonc,selection_command
+21,507324,"/home/franz.srambical/crowd-pilot-extension/.vscode/launch.json",0,0,"// A launch configuration that compiles the extension and then opens it inside a new window\n// Use IntelliSense to learn about possible attributes.\n// Hover to view descriptions of existing attributes.\n// For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387\n{\n\t""version"": ""0.2.0"",\n\t""configurations"": [\n\t\t{\n\t\t\t""name"": ""Run Extension"",\n\t\t\t""type"": ""extensionHost"",\n\t\t\t""request"": ""launch"",\n\t\t\t""args"": [\n\t\t\t\t""--extensionDevelopmentPath=${workspaceFolder}""\n\t\t\t],\n\t\t\t""timeout"": 20000, // Increase the timeout to 20 seconds\n\t\t\t""outFiles"": [\n\t\t\t\t""${workspaceFolder}/out/**/*.js""\n\t\t\t],\n\t\t\t""preLaunchTask"": ""${defaultBuildTask}""\n\t\t}\n\t]\n}\n",jsonc,tab
+22,508588,".vscode/launch.json",0,0,"",jsonc,tab
+23,509273,"/home/franz.srambical/crowd-pilot-extension/.vscode/launch.json",0,0,"",jsonc,tab
+24,510978,".vscode/launch.json",0,0,"",jsonc,tab
+25,511794,"/home/franz.srambical/crowd-pilot-extension/.vscode/launch.json",0,0,"",jsonc,tab
+26,513487,".vscode/launch.json",0,0,"",jsonc,tab
+27,514609,".vscode/tasks.json",0,0,"// See https://go.microsoft.com/fwlink/?LinkId=733558\n// for the documentation about the tasks.json format\n{\n\t""version"": ""2.0.0"",\n\t""tasks"": [\n\t\t{\n\t\t\t""type"": ""npm"",\n\t\t\t""script"": ""watch"",\n\t\t\t""problemMatcher"": ""$tsc-watch"",\n\t\t\t""isBackground"": true,\n\t\t\t""presentation"": {\n\t\t\t\t""reveal"": ""never""\n\t\t\t},\n\t\t\t""group"": {\n\t\t\t\t""kind"": ""build"",\n\t\t\t\t""isDefault"": true\n\t\t\t}\n\t\t}\n\t]\n}\n",jsonc,tab
+28,522163,".vscode/launch.json",0,0,"",jsonc,tab
+29,523449,".vscode/tasks.json",0,0,"",jsonc,tab
+30,532241,"/home/franz.srambical/crowd-pilot-extension/.vscode/launch.json",0,0,"",jsonc,tab
+31,533389,".vscode/tasks.json",0,0,"",jsonc,tab
+32,534076,"/home/franz.srambical/crowd-pilot-extension/.vscode/launch.json",0,0,"",jsonc,tab
+33,538593,".vscode/tasks.json",0,0,"",jsonc,tab
+34,550096,"/home/franz.srambical/crowd-pilot-extension/.vscode/tasks.json",0,0,"// See https://go.microsoft.com/fwlink/?LinkId=733558\n// for the documentation about the tasks.json format\n{\n\t""version"": ""2.0.0"",\n\t""tasks"": [\n\t\t{\n\t\t\t""label"": ""watch"",\n\t\t\t""type"": ""shell"",\n\t\t\t// FIXME (f.srambical): We should make this cluster-agnostic\n\t\t\t""command"": ""module load nodejs && npm run watch"",\n\t\t\t""problemMatcher"": ""$tsc-watch"",\n\t\t\t""isBackground"": true,\n\t\t\t""presentation"": {\n\t\t\t\t""reveal"": ""never""\n\t\t\t},\n\t\t\t""group"": {\n\t\t\t\t""kind"": ""build"",\n\t\t\t\t""isDefault"": true\n\t\t\t}\n\t\t}\n\t]\n}\n",jsonc,tab
+35,551662,".vscode/tasks.json",0,0,"",jsonc,tab
+36,552543,"/home/franz.srambical/crowd-pilot-extension/.vscode/tasks.json",0,0,"",jsonc,tab
+37,554045,".vscode/tasks.json",0,0,"",jsonc,tab
+38,554778,"/home/franz.srambical/crowd-pilot-extension/.vscode/tasks.json",0,0,"",jsonc,tab
+39,555232,"/home/franz.srambical/crowd-pilot-extension/.vscode/tasks.json",0,484,"// See https://go.microsoft.com/fwlink/?LinkId=733558\n// for the documentation about the tasks.json format\n{\n\t""version"": ""2.0.0"",\n\t""tasks"": [\n\t\t{\n\t\t\t""label"": ""watch"",\n\t\t\t""type"": ""shell"",\n\t\t\t// FIXME (f.srambical): We should make this cluster-agnostic\n\t\t\t""command"": ""module load nodejs && npm run watch"",\n\t\t\t""problemMatcher"": ""$tsc-watch"",\n\t\t\t""isBackground"": true,\n\t\t\t""presentation"": {\n\t\t\t\t""reveal"": ""never""\n\t\t\t},\n\t\t\t""group"": {\n\t\t\t\t""kind"": ""build"",\n\t\t\t\t""isDefault"": true\n\t\t\t}\n\t\t}\n\t]\n}\n",jsonc,selection_command
+40,555354,"/home/franz.srambical/crowd-pilot-extension/.vscode/tasks.json",484,0,"",jsonc,selection_command
+41,556081,".vscode/tasks.json",0,0,"",jsonc,tab
+42,556460,".vscode/tasks.json",0,0,"// See https://go.microsoft.com/fwlink/?LinkId=733558\n// for the documentation about the tasks.json format\n{\n\t""version"": ""2.0.0"",\n\t""tasks"": [\n\t\t{\n\t\t\t""label"": ""watch"",\n\t\t\t""type"": ""shell"",\n\t\t\t// FIXME (f.srambical): We should make this cluster-agnostic\n\t\t\t""command"": ""module load nodejs && npm run watch"",\n\t\t\t""problemMatcher"": ""$tsc-watch"",\n\t\t\t""isBackground"": true,\n\t\t\t""presentation"": {\n\t\t\t\t""reveal"": ""never""\n\t\t\t},\n\t\t\t""group"": {\n\t\t\t\t""kind"": ""build"",\n\t\t\t\t""isDefault"": true\n\t\t\t}\n\t\t}\n\t]\n}\n",jsonc,content
+43,556462,".vscode/tasks.json",0,479,"// See https://go.microsoft.com/fwlink/?LinkId=733558\n// for the documentation about the tasks.json format\n{\n\t""version"": ""2.0.0"",\n\t""tasks"": [\n\t\t{\n\t\t\t""label"": ""watch"",\n\t\t\t""type"": ""shell"",\n\t\t\t// FIXME (f.srambical): We should make this cluster-agnostic\n\t\t\t""command"": ""module load nodejs && npm run watch"",\n\t\t\t""problemMatcher"": ""$tsc-watch"",\n\t\t\t""isBackground"": true,\n\t\t\t""presentation"": {\n\t\t\t\t""reveal"": ""never""\n\t\t\t},\n\t\t\t""group"": {\n\t\t\t\t""kind"": ""build"",\n\t\t\t\t""isDefault"": true\n\t\t\t}\n\t\t}\n",jsonc,selection_command
+44,557194,".vscode/tasks.json",479,0,"",jsonc,selection_command
+45,557604,".vscode/tasks.json",0,850,"// See https://go.microsoft.com/fwlink/?LinkId=733558\n// for the documentation about the tasks.json format\n{\n\t""version"": ""2.0.0"",\n\t""tasks"": [\n\t\t{\n\t\t\t""label"": ""watch"",\n\t\t\t""type"": ""shell"",\n\t\t\t// FIXME (f.srambical): We should make this cluster-agnostic\n\t\t\t""command"": ""module load nodejs && npm run watch"",\n\t\t\t""problemMatcher"": ""$tsc-watch"",\n\t\t\t""isBackground"": true,\n\t\t\t""presentation"": {\n\t\t\t\t""reveal"": ""never""\n\t\t\t},\n\t\t\t""group"": {\n\t\t\t\t""kind"": ""build"",\n\t\t\t\t""isDefault"": true\n\t\t\t}\n\t\t}\n\t]\n}\n// See https://go.microsoft.com/fwlink/?LinkId=733558\n// for the documentation about the tasks.json format\n{\n\t""version"": ""2.0.0"",\n\t""tasks"": [\n\t\t{\n\t\t\t""type"": ""npm"",\n\t\t\t""script"": ""watch"",\n\t\t\t""problemMatcher"": ""$tsc-watch"",\n\t\t\t""isBackground"": true,\n\t\t\t""presentation"": {\n\t\t\t\t""reveal"": ""never""\n\t\t\t},\n\t\t\t""group"": {\n\t\t\t\t""kind"": ""build"",\n\t\t\t\t""isDefault"": true\n\t\t\t}\n\t\t}\n\t]\n}\n",jsonc,selection_command
+46,557909,".vscode/tasks.json",0,850,"// See https://go.microsoft.com/fwlink/?LinkId=733558\n// for the documentation about the tasks.json format\n{\n\t""version"": ""2.0.0"",\n\t""tasks"": [\n\t\t{\n\t\t\t""label"": ""watch"",\n\t\t\t""type"": ""shell"",\n\t\t\t// FIXME (f.srambical): We should make this cluster-agnostic\n\t\t\t""command"": ""module load nodejs && npm run watch"",\n\t\t\t""problemMatcher"": ""$tsc-watch"",\n\t\t\t""isBackground"": true,\n\t\t\t""presentation"": {\n\t\t\t\t""reveal"": ""never""\n\t\t\t},\n\t\t\t""group"": {\n\t\t\t\t""kind"": ""build"",\n\t\t\t\t""isDefault"": true\n\t\t\t}\n\t\t}\n\t]\n}\n",jsonc,content
+47,558806,".vscode/tasks.json",0,0,"",jsonc,selection_command
+48,566714,"TERMINAL",0,0,"npm run compile",,terminal_command
+49,566761,"TERMINAL",0,0,"]633;C",,terminal_output
+50,567775,"TERMINAL",0,0,"\r\n> test@0.0.1 compile\r\n> tsc -p ./\r\n\r\n",,terminal_output
+51,569368,"TERMINAL",0,0,"npm notice \r\nnpm notice New major version of npm available! 10.5.2 -> 11.7.0\r\nnpm notice Changelog: https://github.com/npm/cli/releases/tag/v11.7.0\r\nnpm notice Run npm install -g npm@11.7.0 to update!\r\nnpm notice \r\n]0;franz.srambical@hai-login1:~/test_extension/test",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-6773632c-44b1-4ba5-9d03-748e64873cca1767950934524-2026_01_09-10.29.15.409/source.csv

@@ -0,0 +1,8 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,13,"src/gitProvider.ts",0,0,"import * as vscode from 'vscode'\nimport { recording, addToFileQueue, buildCsvRow, appendToFile } from './recording'\nimport { ChangeType } from './types'\nimport { isCurrentFileExported } from './recording'\nimport * as child_process from 'child_process'\nimport * as util from 'util'\nimport { logToOutput } from './utilities'\n\ninterface LocalGitState {\n    branch: string\n    repository: string\n}\n\nlet currentGitState: LocalGitState | null = null\nlet gitWatcherInitialized = false\nlet lastKnownBranch: string | null = null\nlet gitStateCheckInterval: NodeJS.Timeout | undefined\nlet gitHeadWatcher: vscode.FileSystemWatcher | undefined\n\n/**\n * Initializes the git detection using file system watchers and git commands\n */\nexport function initializeGitProvider(): void {\n    logToOutput('Initializing git provider using file system watchers...', 'info')\n    \n    // Try to initialize immediately\n    tryInitializeGitProvider().catch(error => {\n        logToOutput(`Error in initial git provider initialization: ${error}`, 'error')\n    })\n    \n    // Also try after a delay in case git is not ready yet\n    setTimeout(() => {\n        if (!gitWatcherInitialized) {\n            logToOutput('Retrying git provider initialization...', 'info')\n            tryInitializeGitProvider().catch(error => {\n                logToOutput(`Error in retry git provider initialization: ${error}`, 'error')\n            })\n        }\n    }, 2000)\n    \n    // Listen for workspace changes\n    vscode.workspace.onDidChangeWorkspaceFolders(() => {\n        if (!gitWatcherInitialized) {\n            logToOutput('Workspace folders changed, retrying git provider initialization...', 'info')\n            tryInitializeGitProvider().catch(error => {\n                logToOutput(`Error in workspace change git provider initialization: ${error}`, 'error')\n            })\n        }\n    })\n}\n\n/**\n * Attempts to initialize the git provider\n */\nasync function tryInitializeGitProvider(): Promise<void> {\n    try {\n        const workspaceFolder = vscode.workspace.workspaceFolders?.[0]\n        if (!workspaceFolder) {\n            logToOutput('No workspace folder found', 'info')\n            return\n        }\n        \n        // Check if this is a git repository\n        const gitDir = vscode.Uri.joinPath(workspaceFolder.uri, '.git')\n        try {\n            await vscode.workspace.fs.stat(gitDir)\n            logToOutput('Git repository found', 'info')\n            \n            // Get initial state\n            updateGitState()\n            \n            // Set up periodic checking for branch changes\n            gitStateCheckInterval = setInterval(() => {\n                if (recording.isRecording) {\n                    checkForBranchChanges()\n                }\n            }, 5000) // Check every 5 seconds when recording\n            \n            // Watch for changes in .git/HEAD file\n            gitHeadWatcher = vscode.workspace.createFileSystemWatcher(\n                new vscode.RelativePattern(workspaceFolder, '.git/HEAD')\n            )\n            \n            gitHeadWatcher.onDidChange(() => {\n                logToOutput('Git HEAD file changed, checking for branch checkout...', 'info')\n                setTimeout(() => checkForBranchChanges(), 100) // Small delay to ensure file is written\n            })\n            \n            gitWatcherInitialized = true\n            logToOutput('Git provider initialized successfully', 'info')\n        } catch (error) {\n            logToOutput(`Not a git repository: ${error}`, 'error') \n        }\n        \n    } catch (error) {\n        logToOutput(`Error initializing git provider: ${error}`, 'error')\n    }\n}\n\n/**\n * Checks for branch changes\n */\nasync function checkForBranchChanges(): Promise<void> {\n    try {\n        const newBranch = await getCurrentGitBranchFromCommand()\n        if (newBranch && newBranch !== lastKnownBranch) {\n            const workspaceFolder = vscode.workspace.workspaceFolders?.[0]\n            const repository = workspaceFolder?.uri.fsPath || 'unknown'\n            \n            const newState = { branch: newBranch, repository }\n            \n            if (lastKnownBranch) {\n                logToOutput(`Branch checkout detected: ${lastKnownBranch} -> ${newBranch}`, 'info')\n                handleBranchCheckout(newState, { branch: lastKnownBranch, repository })\n            }\n            \n            lastKnownBranch = newBranch\n            currentGitState = newState\n        }\n    } catch (error) {\n        logToOutput(`Error checking for branch changes: ${error}`, 'error')\n\n    }\n}\n\n/**\n * Updates the current git state\n */\nfunction updateGitState(): void {\n    getCurrentGitBranchFromCommand().then(branchName => {\n        if (branchName) {\n            const workspaceFolder = vscode.workspace.workspaceFolders?.[0]\n            const repository = workspaceFolder?.uri.fsPath || 'unknown'\n            \n            const newState = { branch: branchName, repository }\n            logToOutput(`Initial git state: ${newState}`, 'info')\n            \n            lastKnownBranch = branchName\n            currentGitState = newState\n        }\n    }).catch(error => {\n        logToOutput(`Error getting initial git branch: ${error}`, 'error')\n    })\n}\n\n/**\n * Gets the current git branch using git command\n */\nasync function getCurrentGitBranchFromCommand(): Promise<string | null> {\n    try {\n        const workspaceFolder = vscode.workspace.workspaceFolders?.[0]\n        if (!workspaceFolder) {\n            return null\n        }\n        \n        const execAsync = util.promisify(child_process.exec)\n        const { stdout } = await execAsync('git branch --show-current', { \n            cwd: workspaceFolder.uri.fsPath \n        })\n        return stdout.trim()\n    } catch (error) {\n        logToOutput(`Error executing git command: ${error}`, 'error')\n        return null\n    }\n}\n\n/**\n * Handles branch checkout events\n */\nfunction handleBranchCheckout(newState: LocalGitState, previousState: LocalGitState): void {\n    if (!recording.isRecording) {\n        logToOutput('Not recording, skipping git checkout event', 'info')\n        return\n    }\n\n    if (isCurrentFileExported()) {\n        logToOutput('Current file is exported, skipping git checkout event', 'info')\n        return\n    }\n\n    recording.sequence++\n    const checkoutMessage = `Switched from branch '${previousState.branch}' to '${newState.branch}'`\n    \n    logToOutput(`Recording git checkout: ${checkoutMessage}`, 'info')\n    \n    addToFileQueue(\n        buildCsvRow({\n            sequence: recording.sequence,\n            rangeOffset: 0,\n            rangeLength: 0,\n            text: checkoutMessage,\n            type: ChangeType.GIT_BRANCH_CHECKOUT,\n        })\n    )\n    appendToFile()\n    \n    // Reset the file cache since files might have different content on the new branch\n    logToOutput('Resetting file cache due to branch checkout', 'info')\n    if (recording.activatedFiles) {\n        recording.activatedFiles.clear()\n    }\n}\n\n/**\n * Cleanup function to stop the interval\n */\nexport function cleanupGitProvider(): void {\n    if (gitStateCheckInterval) {\n        clearInterval(gitStateCheckInterval)\n        gitStateCheckInterval = undefined\n    }\n\n    if (gitHeadWatcher) {\n        gitHeadWatcher.dispose()\n        gitHeadWatcher = undefined\n    }\n} ",typescript,tab
+2,1714,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:29:15 AM [info] Activating crowd-code\n10:29:15 AM [info] Recording started\n10:29:15 AM [info] Initializing git provider using file system watchers...\n10:29:15 AM [info] Git repository found\n10:29:15 AM [info] Git provider initialized successfully\n10:29:15 AM [info] Initial git state: [object Object]\n",Log,tab
+3,2141,"src/gitProvider.ts",0,0,"",typescript,tab
+4,7177,"src/gitProvider.ts",0,7268,"/**\n * Git Provider for crowd-code 2.0\n * Detects git operations to annotate filesystem changes\n */\n\nimport * as vscode from 'vscode'\nimport { logToOutput } from './utilities'\n\n// Track recent git operations for filesystem change attribution\nlet sawHeadChange = false\nlet lastGitOperationTime = 0\nconst GIT_OPERATION_WINDOW_MS = 500\n\n// File system watchers\nlet gitHeadWatcher: vscode.FileSystemWatcher | undefined\nlet gitRefsWatcher: vscode.FileSystemWatcher | undefined\n\n/**\n * Check if there was a recent git operation\n * Returns 'git_checkout' if HEAD changed, 'git' for other operations, null otherwise\n */\nexport function getRecentGitOperation(): 'git' | 'git_checkout' | null {\n\tif (Date.now() - lastGitOperationTime > GIT_OPERATION_WINDOW_MS) {\n\t\treturn null\n\t}\n\tconst result = sawHeadChange ? 'git_checkout' : 'git'\n\tsawHeadChange = false\n\treturn result\n}\n\n/**\n * Setup git file watchers for the current workspace\n */\nfunction setupGitWatchers(): void {\n\t// Cleanup any existing watchers first\n\tdisposeWatchers()\n\n\tconst workspaceFolder = vscode.workspace.workspaceFolders?.[0]\n\tif (!workspaceFolder) {\n\t\tlogToOutput('No workspace folder found', 'info')\n\t\treturn\n\t}\n\n\tconst gitDir = vscode.Uri.joinPath(workspaceFolder.uri, '.git')\n\tvscode.workspace.fs.stat(gitDir).then(\n\t\t() => {\n\t\t\tlogToOutput('Git repository found', 'info')\n\n\t\t\t// Watch .git/HEAD for branch changes\n\t\t\tgitHeadWatcher = vscode.workspace.createFileSystemWatcher(\n\t\t\t\tnew vscode.RelativePattern(workspaceFolder, '.git/HEAD')\n\t\t\t)\n\t\t\tgitHeadWatcher.onDidChange(() => {\n\t\t\t\tlogToOutput('Git checkout detected', 'info')\n\t\t\t\tsawHeadChange = true\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\n\t\t\t// Watch .git/refs for other git operations (pull, stash, etc.)\n\t\t\tgitRefsWatcher = vscode.workspace.createFileSystemWatcher(\n\t\t\t\tnew vscode.RelativePattern(workspaceFolder, '.git/refs/**/*')\n\t\t\t)\n\t\t\tgitRefsWatcher.onDidChange(() => {\n\t\t\t\tlogToOutput('Git refs changed', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\t\t\tgitRefsWatcher.onDidCreate(() => {\n\t\t\t\tlogToOutput('Git refs created', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\t\t\tgitRefsWatcher.onDidDelete(() => {\n\t\t\t\tlogToOutput('Git refs deleted', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\n\t\t\tlogToOutput('Git provider initialized', 'info')\n\t\t},\n\t\t() => {\n\t\t\tlogToOutput('Not a git repository', 'info')\n\t\t}\n\t)\n}\n\nfunction disposeWatchers(): void {\n\tgitHeadWatcher?.dispose()\n\tgitHeadWatcher = undefined\n\tgitRefsWatcher?.dispose()\n\tgitRefsWatcher = undefined\n}\n\n/**\n * Initialize the git provider\n */\nexport function initializeGitProvider(context: vscode.ExtensionContext): void {\n\tlogToOutput('Initializing git provider...', 'info')\n\n\tsetupGitWatchers()\n\n\t// Reinitialize on workspace changes\n\tcontext.subscriptions.push(\n\t\tvscode.workspace.onDidChangeWorkspaceFolders(() => {\n\t\t\tlogToOutput('Workspace changed, reinitializing git provider...', 'info')\n\t\t\tsetupGitWatchers()\n\t\t})\n\t)\n}\n\n/**\n * Cleanup the git provider\n */\nexport function cleanupGitProvider(): void {\n\tdisposeWatchers()\n\tsawHeadChange = false\n\tlastGitOperationTime = 0\n}\n",typescript,content
+5,10117,"src/gitProvider.ts",0,0,"Switched from branch 'main' to 'crowd-code-2-0'",typescript,git_branch_checkout
+6,12193,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:29:15 AM [info] Activating crowd-code\n10:29:15 AM [info] Recording started\n10:29:15 AM [info] Initializing git provider using file system watchers...\n10:29:15 AM [info] Git repository found\n10:29:15 AM [info] Git provider initialized successfully\n10:29:15 AM [info] Initial git state: [object Object]\n10:29:25 AM [info] Branch checkout detected: main -> crowd-code-2-0\n10:29:25 AM [info] Recording git checkout: Switched from branch 'main' to 'crowd-code-2-0'\n10:29:25 AM [info] Resetting file cache due to branch checkout\n",Log,tab
+7,13668,"src/gitProvider.ts",0,0,"/**\n * Git Provider for crowd-code 2.0\n * Detects git operations to annotate filesystem changes\n */\n\nimport * as vscode from 'vscode'\nimport { logToOutput } from './utilities'\n\n// Track recent git operations for filesystem change attribution\nlet sawHeadChange = false\nlet lastGitOperationTime = 0\nconst GIT_OPERATION_WINDOW_MS = 500\n\n// File system watchers\nlet gitHeadWatcher: vscode.FileSystemWatcher | undefined\nlet gitRefsWatcher: vscode.FileSystemWatcher | undefined\n\n/**\n * Check if there was a recent git operation\n * Returns 'git_checkout' if HEAD changed, 'git' for other operations, null otherwise\n */\nexport function getRecentGitOperation(): 'git' | 'git_checkout' | null {\n\tif (Date.now() - lastGitOperationTime > GIT_OPERATION_WINDOW_MS) {\n\t\treturn null\n\t}\n\tconst result = sawHeadChange ? 'git_checkout' : 'git'\n\tsawHeadChange = false\n\treturn result\n}\n\n/**\n * Setup git file watchers for the current workspace\n */\nfunction setupGitWatchers(): void {\n\t// Cleanup any existing watchers first\n\tdisposeWatchers()\n\n\tconst workspaceFolder = vscode.workspace.workspaceFolders?.[0]\n\tif (!workspaceFolder) {\n\t\tlogToOutput('No workspace folder found', 'info')\n\t\treturn\n\t}\n\n\tconst gitDir = vscode.Uri.joinPath(workspaceFolder.uri, '.git')\n\tvscode.workspace.fs.stat(gitDir).then(\n\t\t() => {\n\t\t\tlogToOutput('Git repository found', 'info')\n\n\t\t\t// Watch .git/HEAD for branch changes\n\t\t\tgitHeadWatcher = vscode.workspace.createFileSystemWatcher(\n\t\t\t\tnew vscode.RelativePattern(workspaceFolder, '.git/HEAD')\n\t\t\t)\n\t\t\tgitHeadWatcher.onDidChange(() => {\n\t\t\t\tlogToOutput('Git checkout detected', 'info')\n\t\t\t\tsawHeadChange = true\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\n\t\t\t// Watch .git/refs for other git operations (pull, stash, etc.)\n\t\t\tgitRefsWatcher = vscode.workspace.createFileSystemWatcher(\n\t\t\t\tnew vscode.RelativePattern(workspaceFolder, '.git/refs/**/*')\n\t\t\t)\n\t\t\tgitRefsWatcher.onDidChange(() => {\n\t\t\t\tlogToOutput('Git refs changed', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\t\t\tgitRefsWatcher.onDidCreate(() => {\n\t\t\t\tlogToOutput('Git refs created', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\t\t\tgitRefsWatcher.onDidDelete(() => {\n\t\t\t\tlogToOutput('Git refs deleted', 'info')\n\t\t\t\tlastGitOperationTime = Date.now()\n\t\t\t})\n\n\t\t\tlogToOutput('Git provider initialized', 'info')\n\t\t},\n\t\t() => {\n\t\t\tlogToOutput('Not a git repository', 'info')\n\t\t}\n\t)\n}\n\nfunction disposeWatchers(): void {\n\tgitHeadWatcher?.dispose()\n\tgitHeadWatcher = undefined\n\tgitRefsWatcher?.dispose()\n\tgitRefsWatcher = undefined\n}\n\n/**\n * Initialize the git provider\n */\nexport function initializeGitProvider(context: vscode.ExtensionContext): void {\n\tlogToOutput('Initializing git provider...', 'info')\n\n\tsetupGitWatchers()\n\n\t// Reinitialize on workspace changes\n\tcontext.subscriptions.push(\n\t\tvscode.workspace.onDidChangeWorkspaceFolders(() => {\n\t\t\tlogToOutput('Workspace changed, reinitializing git provider...', 'info')\n\t\t\tsetupGitWatchers()\n\t\t})\n\t)\n}\n\n/**\n * Cleanup the git provider\n */\nexport function cleanupGitProvider(): void {\n\tdisposeWatchers()\n\tsawHeadChange = false\n\tlastGitOperationTime = 0\n}\n",typescript,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-6829bcbf-f7fb-4481-92ea-521e9af7eabb1754058671446-2025_08_01-16.31.17.606/source.csv

@@ -0,0 +1,4 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"utils/nn.py",0,0,"import math\nfrom typing import Tuple, Callable, List\n\nfrom flax import nnx\nimport jax\nimport jax.numpy as jnp\nimport einops\n\n\nclass PositionalEncoding(nnx.Module):\n    """"""https://uvadlc-notebooks.readthedocs.io/en/latest/tutorial_notebooks/JAX/tutorial6/Transformers_and_MHAttention.html""""""\n\n    def __init__(self, d_model: int, max_len: int = 5000):\n        self.d_model = d_model\n        self.max_len = max_len\n\n        pe = jnp.zeros((self.max_len, self.d_model))\n        position = jnp.arange(0, self.max_len, dtype=jnp.float32)[:, None]\n        div_term = jnp.exp(\n            jnp.arange(0, self.d_model, 2) * (-math.log(10000.0) / self.d_model)\n        )\n        pe = pe.at[:, 0::2].set(jnp.sin(position * div_term))\n        pe = pe.at[:, 1::2].set(jnp.cos(position * div_term))\n        self.pe = nnx.Variable(pe)\n\n    def __call__(self, x: jax.Array) -> jax.Array:\n        x = x + self.pe[: x.shape[2]]\n        return x\n\n\nclass STBlock(nnx.Module):\n    def __init__(\n        self,\n        dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.dim = dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.spatial_pos_enc = PositionalEncoding(self.dim)\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=False\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.temporal_pos_enc = PositionalEncoding(self.dim)\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array) -> jax.Array:\n        # --- Spatial attention ---\n        z_BTNM = self.spatial_pos_enc(x_BTNM)\n        z_BTNM = self.spatial_norm(z_BTNM)\n        z_BTNM = self.spatial_attention(z_BTNM)\n        x_BTNM = x_BTNM + z_BTNM\n\n        # --- Temporal attention ---\n        x_BNTM = x_BTNM.swapaxes(1, 2)\n        z_BNTM = self.temporal_pos_enc(x_BNTM)\n        z_BNTM = self.temporal_norm(z_BNTM)\n        z_BNTM = self.temporal_attention(z_BNTM)\n        x_BNTM = x_BNTM + z_BNTM\n        x_BTNM = x_BNTM.swapaxes(1, 2)\n\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n\n        return x_BTNM\n\n\nclass STTransformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        O: number of output features\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n        self.blocks = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                STBlock(\n                    dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    rngs=rngs,\n                )\n            )\n\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM)\n\n        x_BTNO = self.output_dense(x_BTNM)\n        return x_BTNO\n\nclass TransformerBlock(nnx.Module):\n    def __init__(\n        self,\n        model_dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.decode = decode\n\n        self.temporal_pos_enc = PositionalEncoding(self.model_dim)\n        self.spatial_pos_enc = PositionalEncoding(self.model_dim)\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        # --- Spatial attention ---\n        B, T, N, M = x_BTNM.shape\n        z_FNM = einops.rearrange(x_BTNM, ""b t n m -> (b t) n m"")\n        z_FNM = self.spatial_norm(z_FNM)\n        if self.decode:\n            assert pos_index is not None\n            z_FM = z_FNM[:, pos_index[1]]\n            z_F1M = jnp.reshape(z_FM, (B * T, 1, M))\n            z_F1M = self.spatial_attention(z_F1M)\n            z_FM = jnp.reshape(z_F1M, (B * T, M))\n            z_FNM = z_FNM.at[:, pos_index[1], :].set(z_FM)\n        else:\n            z_FNM = self.spatial_attention(z_FNM)\n        z_BTNM = einops.rearrange(z_FNM, ""(b t) n m -> b t n m"", t=T)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Temporal attention ---\n        z_PTM = einops.rearrange(x_BTNM, ""b t n m -> (b n) t m"")\n        z_PTM = self.temporal_norm(z_PTM)\n        if self.decode:\n            assert pos_index is not None\n            z_PM = z_PTM[:, pos_index[0]]\n            z_P1M = jnp.reshape(z_PM, (B * N, 1, M))\n            z_P1M = self.temporal_attention(z_P1M)\n            z_PM = jnp.reshape(z_P1M, (B * N, M))\n            z_PTM = z_PTM.at[:, pos_index[0], :].set(z_PM)\n        else:\n            z_PTM = self.temporal_attention(z_PTM)\n        z_BTNM = einops.rearrange(z_PTM, ""(b n) t m -> b t n m"", n=N)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n\n        return x_BTNM\n\nclass Transformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        O: number of output features\n        F: number of frames in batch\n        P: number of patch positions in batch\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n\n        self.pos_enc = PositionalEncoding(self.model_dim)\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n        self.blocks: List[TransformerBlock] = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                TransformerBlock(\n                    model_dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    decode=decode,\n                    rngs=rngs,\n                )\n            )\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM, pos_index)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        return x_BTNV\n\ndef normalize(x: jax.Array) -> jax.Array:\n    return x / (jnp.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-8)\n\n\nclass VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n    def __init__(\n        self, latent_dim: int, num_latents: int, dropout: float, rngs: nnx.Rngs\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(self.codebook.value)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = self.codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_DL + jax.lax.stop_gradient(z_DL - x_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n\n    def get_codes(self, indices_E: jax.Array) -> jax.Array:\n        return self.codebook[indices_E]\n\n\ndef _create_flash_attention_fn(use_flash_attention: bool, is_causal: bool) -> Callable:\n    """"""\n    Create an attention function that uses flash attention if enabled.\n\n    flax.nnx.MultiHeadAttention provides tensors with shape (batch..., length, num_heads, head_dim),\n    but jax.nn.dot_product_attention expects (batch, length, num_heads, head_dim). We reshape to\n    ensure compatibility. cuDNN's flash attention additionally requires a sequence length that\n    is a multiple of 4. We pad the sequence length to the nearest multiple of 4 and mask\n    accordingly. Note that cuDNN requires the mask to be broadcast before calling the attention\n    function due to strict shape checking.\n    """"""\n\n    # FIXME (f.srambical): keys and values could have different dimensionalities\n    def attention_fn(query_BSHD, key_BSHD, value_BSHD, bias=None, mask_B111=None, **kwargs):\n        implementation = ""cudnn"" if use_flash_attention else None\n\n        def _merge_batch_dims(x):\n            return einops.rearrange(x, ""... l h k -> (...) l h k"")\n\n        def _pad(x):\n            return jnp.pad(x, ((0, 0), (0, pad_size), (0, 0), (0, 0)))\n\n        original_shape = query_BSHD.shape\n        original_seq_len = query_BSHD.shape[-3]\n\n        # Pad to nearest multiple of 4\n        T = ((original_seq_len + 3) // 4) * 4\n        pad_size = T - original_seq_len\n\n        query_BTHD = _pad(_merge_batch_dims(query_BSHD))\n        key_BTHD = _pad(_merge_batch_dims(key_BSHD))\n        value_BTHD = _pad(_merge_batch_dims(value_BSHD))\n        B = query_BTHD.shape[0]\n\n        attention_mask = jnp.ones((T, T), dtype=jnp.bool_)\n        attention_mask = attention_mask.at[original_seq_len:, :].set(False)\n        attention_mask = attention_mask.at[:, original_seq_len:].set(False)\n\n        # Handle causal mask for cached decoder self-attention (from nnx.MultiHeadAttention)\n        if mask_B111 is not None:\n            mask_B111 = _merge_batch_dims(mask_B111)\n            # We need to broadcast T and S dimensions to target_seq_len since cudnn attention strictly checks the mask shape\n            # https://github.com/jax-ml/jax/issues/28974\n            # https://github.com/jax-ml/jax/blob/08c7677393672ccb85c10f1ed0bd506905c3c994/jax/_src/cudnn/fused_attention_stablehlo.py#L1830\n            # https://github.com/jax-ml/jax/blob/08c7677393672ccb85c10f1ed0bd506905c3c994/jax/_src/cudnn/fused_attention_stablehlo.py#L337\n            mask_B1QK = einops.repeat(mask_B111, ""... 1 1 -> ... t s"", t=T, s=T)\n            mask_B1QK = mask_B111.astype(jnp.bool)\n        else:\n            mask_11QK = attention_mask[jnp.newaxis, jnp.newaxis, :, :]\n            mask_B1QK = jnp.broadcast_to(mask_11QK, (B, 1, T, T))\n\n        bias_4d = _pad(_merge_batch_dims(bias)) if bias is not None else None\n\n        # NOTE: jax.nn.dot_product_attention does not support dropout\n        output_4d = jax.nn.dot_product_attention(\n            query=query_BTHD,\n            key=key_BTHD,\n            value=value_BTHD,\n            bias=bias_4d,\n            mask=mask_B1QK,\n            implementation=implementation,\n            is_causal=is_causal,\n        )\n        return output_4d[..., :original_seq_len, :, :].reshape(original_shape)\n\n    return attention_fn\n",python,tab
+2,333,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"4:31:17 PM [info] Activating crowd-code\n4:31:17 PM [info] Recording started\n4:31:17 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,381,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"4:31:17 PM [info] Git repository found\n4:31:17 PM [info] Git provider initialized successfully\n4:31:17 PM [info] Initial git state: [object Object]\n",Log,content

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-78f661d8-ef12-4bcf-a19e-75a3edb4b9f11762537370277-2025_11_07-18.42.56.849/source.csv

@@ -0,0 +1,118 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,203,"Untitled-1",0,0,"",plaintext,tab
+2,426,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"6:42:56 PM [info] Activating crowd-code\n6:42:57 PM [info] Recording started\n6:42:57 PM [info] Initializing git provider using file system watchers...\n6:42:57 PM [info] No workspace folder found\n",Log,tab
+3,2277,"extension-output-pdoom-org.crowd-code-#1-crowd-code",194,0,"6:42:59 PM [info] Retrying git provider initialization...\n6:42:59 PM [info] No workspace folder found\n",Log,content
+4,12391,"Untitled-1",0,0,"",plaintext,tab
+5,20501,"Untitled-1",0,0,"Hello, this is a new text file. You can start typing here.\n",plaintext,content
+6,22260,"Untitled-1",59,0,"This is the first line of the text file.",plaintext,content
+7,23758,"TERMINAL",0,0,"undefined[franz.srambical@hai-login2.haicore.berlin:~] $ Hello, this is a new text file. You can start typing here.",,terminal_command
+8,23793,"TERMINAL",0,0,"This is the first line of the text file.",,terminal_command
+9,23795,"TERMINAL",0,0,"",,terminal_command
+10,23799,"TERMINAL",0,0,"Run tests",,terminal_command
+11,23800,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login2:~",,terminal_output
+12,28638,"Untitled-1",98,0,"",plaintext,selection_command
+13,28844,"Untitled-1",39,0,"",plaintext,selection_command
+14,30447,"Untitled-1",39,19,"",plaintext,content
+15,31056,"Untitled-1",38,0,"",plaintext,selection_command
+16,31377,"Untitled-1",0,39,"",plaintext,content
+17,32417,"Untitled-1",1,0,"",plaintext,selection_command
+18,32711,"Untitled-1",1,40,"",plaintext,content
+19,33431,"Untitled-1",0,1,"",plaintext,content
+20,51569,"Untitled-1",0,0,"{\n  ""name"": ""crowd-pilot"",\n  ""displayName"": ""crowd-pilot-extension"",\n  ""description"": ""Teaching language models to code like humans."",\n  ""publisher"": ""p(doom)"",\n  ""version"": ""0.0.1"",\n  ""engines"": {\n    ""vscode"": ""^1.99.3""\n  },\n  ""categories"": [\n    ""Other""\n  ],\n  ""activationEvents"": [\n    ""onStartupFinished""\n  ],\n  ""main"": ""./out/extension.js"",\n  ""contributes"": {\n    ""commands"": [\n      {\n        ""command"": ""crowd-pilot.testRun"",\n        ""title"": ""Test Run""\n      },\n      {\n        ""command"": ""crowd-pilot.hideUi"",\n        ""title"": ""crowd-pilot: Hide Preview""\n      },\n      {\n        ""command"": ""crowd-pilot.sglangTest"",\n        ""title"": ""crowd-pilot: Test SGLang""\n      },\n      {\n        ""command"": ""crowd-pilot.modelRun"",\n        ""title"": ""crowd-pilot: Model Plan & Run""\n      }\n    ],\n    ""keybindings"": [\n      {\n        ""command"": ""crowd-pilot.modelRun"",\n        ""key"": ""tab"",\n        ""mac"": ""tab"",\n        ""when"": ""editorTextFocus || terminalFocus""\n      },\n      {\n        ""command"": ""crowd-pilot.modelRun"",\n        ""key"": ""tab"",\n        ""mac"": ""tab"",\n        ""when"": ""inQuickOpen && crowdPilot.uiVisible""\n      },\n      {\n        ""command"": ""crowd-pilot.hideUi"",\n        ""key"": ""escape"",\n        ""mac"": ""escape"",\n        ""when"": ""crowdPilot.uiVisible""\n      }\n    ]\n  },\n  ""scripts"": {\n    ""vscode:prepublish"": ""npm run compile"",\n    ""compile"": ""tsc -p ./"",\n    ""watch"": ""tsc -watch -p ./"",\n    ""pretest"": ""npm run compile && npm run lint"",\n    ""lint"": ""eslint src"",\n    ""test"": ""vscode-test""\n  },\n  ""devDependencies"": {\n    ""@types/vscode"": ""^1.105.0"",\n    ""@types/mocha"": ""^10.0.10"",\n    ""@types/node"": ""22.x"",\n    ""@typescript-eslint/eslint-plugin"": ""^8.45.0"",\n    ""@typescript-eslint/parser"": ""^8.45.0"",\n    ""eslint"": ""^9.36.0"",\n    ""typescript"": ""^5.9.3"",\n    ""@vscode/test-cli"": ""^0.0.11"",\n    ""@vscode/test-electron"": ""^2.5.2""\n  }\n}\n",plaintext,content
+21,52335,"Untitled-1",0,0,"",json,selection_command
+22,52923,"Untitled-1",2,0,"",json,selection_command
+23,53056,"Untitled-1",27,0,"",json,selection_command
+24,53206,"Untitled-1",69,0,"",json,selection_command
+25,53356,"Untitled-1",135,0,"",json,selection_command
+26,54110,"Untitled-1",161,0,"",json,selection_command
+27,54406,"Untitled-1",183,0,"",json,selection_command
+28,54652,"Untitled-1",198,0,"",json,selection_command
+29,54685,"Untitled-1",222,0,"",json,selection_command
+30,54720,"Untitled-1",227,0,"",json,selection_command
+31,54757,"Untitled-1",245,0,"",json,selection_command
+32,54787,"Untitled-1",257,0,"",json,selection_command
+33,54911,"Untitled-1",245,0,"",json,selection_command
+34,55167,"Untitled-1",227,0,"",json,selection_command
+35,55210,"Untitled-1",222,0,"",json,selection_command
+36,55230,"Untitled-1",198,0,"",json,selection_command
+37,55262,"Untitled-1",183,0,"",json,selection_command
+38,55293,"Untitled-1",161,0,"",json,selection_command
+39,55492,"Untitled-1",135,0,"",json,selection_command
+40,55555,"Untitled-1",161,0,"",json,selection_command
+41,55810,"Untitled-1",135,0,"",json,selection_command
+42,55963,"Untitled-1",69,0,"",json,selection_command
+43,56110,"Untitled-1",27,0,"",json,selection_command
+44,56259,"Untitled-1",2,0,"",json,selection_command
+45,56498,"Untitled-1",4,0,"",json,selection_command
+46,56666,"Untitled-1",5,0,"",json,selection_command
+47,56814,"Untitled-1",9,0,"",json,selection_command
+48,57010,"Untitled-1",12,0,"",json,selection_command
+49,57439,"Untitled-1",13,0,"",json,selection_command
+50,57640,"Untitled-1",18,0,"",json,selection_command
+51,57804,"Untitled-1",19,0,"",json,selection_command
+52,59096,"Untitled-1",23,0,"",json,selection_command
+53,59636,"Untitled-1",24,0,"",json,selection_command
+54,59969,"Untitled-1",24,0,"-",json,content
+55,59972,"Untitled-1",25,0,"",json,selection_keyboard
+56,60135,"Untitled-1",25,0,"e",json,content
+57,60137,"Untitled-1",26,0,"",json,selection_keyboard
+58,60472,"Untitled-1",26,0,"x",json,content
+59,60475,"Untitled-1",27,0,"",json,selection_keyboard
+60,60585,"Untitled-1",27,0,"t",json,content
+61,60588,"Untitled-1",28,0,"",json,selection_keyboard
+62,60636,"Untitled-1",28,0,"e",json,content
+63,60640,"Untitled-1",29,0,"",json,selection_keyboard
+64,60731,"Untitled-1",29,0,"n",json,content
+65,60733,"Untitled-1",30,0,"",json,selection_keyboard
+66,60913,"Untitled-1",30,0,"s",json,content
+67,60915,"Untitled-1",31,0,"",json,selection_keyboard
+68,60929,"Untitled-1",31,0,"i",json,content
+69,60931,"Untitled-1",32,0,"",json,selection_keyboard
+70,60959,"Untitled-1",32,0,"o",json,content
+71,60961,"Untitled-1",33,0,"",json,selection_keyboard
+72,61036,"Untitled-1",33,0,"n",json,content
+73,61038,"Untitled-1",34,0,"",json,selection_keyboard
+74,61905,"Untitled-1",33,0,"",json,selection_command
+75,83621,"Untitled-1",68,0,"",json,selection_command
+76,83875,"Untitled-1",110,0,"",json,selection_command
+77,83912,"Untitled-1",169,0,"",json,selection_command
+78,83976,"Untitled-1",191,0,"",json,selection_command
+79,83979,"Untitled-1",206,0,"",json,selection_command
+80,84006,"Untitled-1",230,0,"",json,selection_command
+81,84038,"Untitled-1",235,0,"",json,selection_command
+82,84070,"Untitled-1",253,0,"",json,selection_command
+83,84107,"Untitled-1",265,0,"",json,selection_command
+84,84136,"Untitled-1",270,0,"",json,selection_command
+85,84251,"Untitled-1",265,0,"",json,selection_command
+86,84513,"Untitled-1",253,0,"",json,selection_command
+87,84542,"Untitled-1",235,0,"",json,selection_command
+88,84569,"Untitled-1",230,0,"",json,selection_command
+89,84602,"Untitled-1",206,0,"",json,selection_command
+90,84635,"Untitled-1",191,0,"",json,selection_command
+91,84667,"Untitled-1",169,0,"",json,selection_command
+92,84788,"Untitled-1",110,0,"",json,selection_command
+93,84964,"Untitled-1",68,0,"",json,selection_command
+94,86093,"Untitled-1",110,0,"",json,selection_command
+95,86512,"Untitled-1",68,0,"",json,selection_command
+96,86854,"Untitled-1",110,0,"",json,selection_command
+97,87752,"Untitled-1",169,0,"",json,selection_command
+98,88002,"Untitled-1",191,0,"",json,selection_command
+99,88040,"Untitled-1",206,0,"",json,selection_command
+100,88063,"Untitled-1",230,0,"",json,selection_command
+101,88094,"Untitled-1",235,0,"",json,selection_command
+102,88131,"Untitled-1",253,0,"",json,selection_command
+103,88167,"Untitled-1",265,0,"",json,selection_command
+104,88195,"Untitled-1",270,0,"",json,selection_command
+105,88233,"Untitled-1",294,0,"",json,selection_command
+106,88465,"Untitled-1",318,0,"",json,selection_command
+107,88718,"Untitled-1",323,0,"",json,selection_command
+108,88831,"Untitled-1",355,0,"",json,selection_command
+109,89077,"Untitled-1",374,0,"",json,selection_command
+110,89108,"Untitled-1",392,0,"",json,selection_command
+111,89166,"Untitled-1",400,0,"",json,selection_command
+112,89367,"Untitled-1",433,0,"",json,selection_command
+113,90762,"Untitled-1",433,0,"-extension",json,content
+114,94571,"Untitled-1",443,0,"```json\n",json,content
+115,94608,"Untitled-1",443,0,"",json,selection_command
+116,96981,"Untitled-1",433,18,"",json,content
+117,96988,"Untitled-1",442,0,"",json,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-7ec4cd27-1a6f-46f4-897a-33c52986bfd51755936233909-2025_08_23-10.04.05.387/source.csv

@@ -0,0 +1,698 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"utils/train_utils.py",0,0,"import optax\nfrom jax.tree_util import tree_map, tree_reduce\n\n\ndef get_lr_schedule(\n    lr_schedule: str,\n    init_lr: float,\n    max_lr: float,\n    decay_end: float,\n    total_steps: int,\n    warmup_steps: int,\n    wsd_decay_steps: int,\n) -> optax.Schedule:\n    supported_schedules = [""wsd"", ""cos""]\n    if lr_schedule == ""cos"":\n        assert (\n            warmup_steps <= total_steps\n        ), ""Warmup steps can't be greater than total steps.""\n        return optax.warmup_cosine_decay_schedule(\n            init_value=init_lr,\n            peak_value=max_lr,\n            warmup_steps=warmup_steps,\n            decay_steps=total_steps,  # Note: decay_steps includes the warmup steps, so we need to pass total value\n            end_value=decay_end,\n        )\n    elif lr_schedule == ""wsd"":\n        assert (\n            warmup_steps + wsd_decay_steps <= total_steps\n        ), ""Warmup and decay period is longer than total steps.""\n        schedules = [\n            optax.linear_schedule(\n                init_value=init_lr, end_value=max_lr, transition_steps=warmup_steps\n            ),\n            optax.constant_schedule(value=max_lr),\n            optax.linear_schedule(\n                init_value=max_lr, end_value=decay_end, transition_steps=wsd_decay_steps\n            ),\n        ]\n        boundaries = [warmup_steps, total_steps - wsd_decay_steps]\n        return optax.join_schedules(schedules, boundaries)\n    else:\n        raise ValueError(\n            f""Learning rate schedule not supported. Please use one of {supported_schedules}""\n        )\n\n\ndef _count_leaf(x):\n    """"""Count parameters in a single leaf node.""""""\n    if hasattr(x, ""size""):\n        return x.size\n    return 0\n\n\ndef _count_component(component_params):\n    """"""Count total parameters in a component.""""""\n    return tree_reduce(\n        lambda x, y: x + y, tree_map(_count_leaf, component_params), initializer=0\n    )\n\n\ndef count_parameters_by_component(params):\n    """"""Count parameters for each component of the model.\n\n    Args:\n        params: Model parameters from nnx.split(model, nnx.Param, ...)\n\n    Returns:\n        Dictionary with parameter counts for each component\n    """"""\n    component_names = list(params.keys())\n    print(f""Counting all components: {component_names}"")\n\n    counts = {}\n    total_params = 0\n\n    for name in component_names:\n        component_params = params[name]\n        count = _count_component(component_params)\n        counts[name] = count\n        total_params += count\n\n    counts[""total""] = total_params\n    return counts\n\n\ndef bytes_to_gb(num_bytes):\n    return num_bytes / (1024**3)\n\n\ndef print_compiled_memory_stats(compiled_stats):\n    """"""from: https://github.com/AI-Hypercomputer/maxtext/blob/b18829fbaa48aec7ac350a03e62248e24c6a76b2/MaxText/max_utils.py#L739""""""\n    output_gb = bytes_to_gb(compiled_stats.output_size_in_bytes)\n    temp_gb = bytes_to_gb(compiled_stats.temp_size_in_bytes)\n    argument_gb = bytes_to_gb(compiled_stats.argument_size_in_bytes)\n    alias_gb = bytes_to_gb(compiled_stats.alias_size_in_bytes)\n    host_temp_gb = bytes_to_gb(compiled_stats.host_temp_size_in_bytes)\n    total_gb = output_gb + temp_gb + argument_gb - alias_gb\n    print(\n        f""Total memory size: {total_gb:.1f} GB, Output size: {output_gb:.1f} GB, Temp size: {temp_gb:.1f} GB, ""\n        f""Argument size: {argument_gb:.1f} GB, Host temp size: {host_temp_gb:.1f} GB.""\n    )\n\n\ndef print_compiled_cost_analysis(cost_stats):\n    flops = float(cost_stats.get(""flops"", 0.0))\n    bytes_accessed = float(cost_stats.get(""bytes accessed"", 0.0))\n    gb = bytes_to_gb(bytes_accessed) if bytes_accessed else 0.0\n    intensity = (flops / bytes_accessed) if bytes_accessed else float(""nan"")\n    print(\n        f""FLOPs: {flops:.3e}, Bytes: {bytes_accessed:.3e} ({gb:.1f} GB), ""\n        f""Intensity: {intensity:.1f} FLOPs/byte""\n    )\n",python,tab
+2,369,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:04:05 AM [info] Activating crowd-code\n10:04:05 AM [info] Recording started\n10:04:05 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,518,"extension-output-pdoom-org.crowd-code-#1-crowd-code",153,0,"10:04:05 AM [info] Git repository found\n10:04:05 AM [info] Git provider initialized successfully\n10:04:05 AM [info] Initial git state: [object Object]\n",Log,content
+4,1047,"utils/train_utils.py",0,0,"",python,tab
+5,1346,"utils/train_utils.py",3825,0,"",python,selection_command
+6,2928,"utils/train_utils.py",3825,0,"\n",python,content
+7,3331,"utils/train_utils.py",3826,0,"\n",python,content
+8,3547,"utils/train_utils.py",3827,0,"d",python,content
+9,3547,"utils/train_utils.py",3828,0,"",python,selection_keyboard
+10,3563,"utils/train_utils.py",3828,0,"e",python,content
+11,3564,"utils/train_utils.py",3829,0,"",python,selection_keyboard
+12,3676,"utils/train_utils.py",3829,0,"f",python,content
+13,3677,"utils/train_utils.py",3830,0,"",python,selection_keyboard
+14,3785,"utils/train_utils.py",3830,0," ",python,content
+15,3785,"utils/train_utils.py",3831,0,"",python,selection_keyboard
+16,8217,"utils/train_utils.py",3830,0,"",python,selection_command
+17,19115,"utils/train_utils.py",3831,0,"",python,selection_command
+18,19455,"utils/train_utils.py",3830,1,"",python,content
+19,19655,"utils/train_utils.py",3829,1,"",python,content
+20,19809,"utils/train_utils.py",3828,1,"",python,content
+21,19984,"utils/train_utils.py",3827,1,"",python,content
+22,20150,"utils/train_utils.py",3827,0,"def print_mem_stats(label: str):\n  max_logging.log(f""\nMemstats: {label}:"")\n  try:\n    for d in jax.local_devices():\n      stats = d.memory_stats()\n      used = round(stats[""bytes_in_use""] / 2**30, 2)\n      limit = round(stats[""bytes_limit""] / 2**30, 2)\n      max_logging.log(f""\tUsing (GB) {used} / {limit} ({used/limit:%}) on {d}"")\n  except (RuntimeError, KeyError, TypeError) as ex:\n    max_logging.log(f""\tMemstats unavailable, error: {ex}"")",python,content
+23,20850,"utils/train_utils.py",4271,0,"",python,selection_command
+24,21022,"utils/train_utils.py",4211,0,"",python,selection_command
+25,21284,"utils/train_utils.py",4139,0,"",python,selection_command
+26,21317,"utils/train_utils.py",4079,0,"",python,selection_command
+27,21331,"utils/train_utils.py",4026,0,"",python,selection_command
+28,21361,"utils/train_utils.py",3973,0,"",python,selection_command
+29,21404,"utils/train_utils.py",3942,0,"",python,selection_command
+30,21430,"utils/train_utils.py",3908,0,"",python,selection_command
+31,21466,"utils/train_utils.py",3901,0,"",python,selection_command
+32,21496,"utils/train_utils.py",3858,0,"",python,selection_command
+33,21968,"utils/train_utils.py",3827,0,"",python,selection_command
+34,22551,"utils/train_utils.py",3860,0,"",python,selection_command
+35,22791,"utils/train_utils.py",3827,0,"",python,selection_command
+36,23421,"utils/train_utils.py",3826,0,"",python,selection_command
+37,23733,"utils/train_utils.py",3825,0,"",python,selection_command
+38,23733,"utils/train_utils.py",3819,0,"",python,selection_command
+39,23734,"utils/train_utils.py",3770,0,"",python,selection_command
+40,23741,"utils/train_utils.py",3695,0,"",python,selection_command
+41,23874,"utils/train_utils.py",3684,0,"",python,selection_command
+42,23875,"utils/train_utils.py",3607,0,"",python,selection_command
+43,23876,"utils/train_utils.py",3543,0,"",python,selection_command
+44,24019,"utils/train_utils.py",3477,0,"",python,selection_command
+45,24020,"utils/train_utils.py",3429,0,"",python,selection_command
+46,24021,"utils/train_utils.py",3383,0,"",python,selection_command
+47,24022,"utils/train_utils.py",3382,0,"",python,selection_command
+48,24244,"utils/train_utils.py",3381,0,"",python,selection_command
+49,24250,"utils/train_utils.py",3375,0,"",python,selection_command
+50,24250,"utils/train_utils.py",3288,0,"",python,selection_command
+51,24251,"utils/train_utils.py",3176,0,"",python,selection_command
+52,24251,"utils/train_utils.py",3165,0,"",python,selection_command
+53,24252,"utils/train_utils.py",3105,0,"",python,selection_command
+54,24372,"utils/train_utils.py",3034,0,"",python,selection_command
+55,24530,"utils/train_utils.py",2971,0,"",python,selection_command
+56,24626,"utils/train_utils.py",2902,0,"",python,selection_command
+57,24775,"utils/train_utils.py",2841,0,"",python,selection_command
+58,24965,"utils/train_utils.py",2776,0,"",python,selection_command
+59,25175,"utils/train_utils.py",2644,0,"",python,selection_command
+60,27038,"utils/train_utils.py",3831,0,"",python,selection_command
+61,27661,"utils/train_utils.py",3859,0,"\n    """"""from: https://github.com/AI-Hypercomputer/maxtext/blob/b18829fbaa48aec7ac350a03e62248e24c6a76b2/MaxText/max_utils.py#L739""""""",python,content
+62,27663,"utils/train_utils.py",3864,0,"",python,selection_command
+63,29213,"utils/train_utils.py",3860,4,"",python,content
+64,33658,"utils/train_utils.py",3859,1,"",python,content
+65,33729,"utils/train_utils.py",3859,0,"\n    ",python,content
+66,34455,"utils/train_utils.py",3863,0,"",python,selection_command
+67,34766,"utils/train_utils.py",3995,0,"",python,selection_command
+68,39172,"utils/train_utils.py",3992,0,"",python,selection_command
+69,39469,"utils/train_utils.py",3992,1," ",python,selection_command
+70,39654,"utils/train_utils.py",3992,1," ",python,selection_command
+71,39909,"utils/train_utils.py",3992,1," ",python,selection_command
+72,39924,"utils/train_utils.py",3992,1," ",python,selection_command
+73,39975,"utils/train_utils.py",3992,1," ",python,selection_command
+74,39980,"utils/train_utils.py",3992,1," ",python,selection_command
+75,40126,"utils/train_utils.py",3992,1," ",python,selection_command
+76,40126,"utils/train_utils.py",3992,1," ",python,selection_command
+77,40130,"utils/train_utils.py",3992,1," ",python,selection_command
+78,40598,"utils/train_utils.py",3992,0,"",python,selection_command
+79,40810,"utils/train_utils.py",4345,0," ",python,content
+80,40810,"utils/train_utils.py",4293,0," ",python,content
+81,40810,"utils/train_utils.py",4213,0," ",python,content
+82,40810,"utils/train_utils.py",4160,0," ",python,content
+83,40810,"utils/train_utils.py",4107,0," ",python,content
+84,40810,"utils/train_utils.py",4076,0," ",python,content
+85,40810,"utils/train_utils.py",4042,0," ",python,content
+86,40810,"utils/train_utils.py",4035,0," ",python,content
+87,40810,"utils/train_utils.py",3992,0," ",python,content
+88,40811,"utils/train_utils.py",3993,0,"",python,selection_keyboard
+89,41565,"utils/train_utils.py",4354,0," ",python,content
+90,41566,"utils/train_utils.py",4301,0," ",python,content
+91,41566,"utils/train_utils.py",4220,0," ",python,content
+92,41566,"utils/train_utils.py",4166,0," ",python,content
+93,41566,"utils/train_utils.py",4112,0," ",python,content
+94,41566,"utils/train_utils.py",4080,0," ",python,content
+95,41566,"utils/train_utils.py",4045,0," ",python,content
+96,41566,"utils/train_utils.py",4037,0," ",python,content
+97,41566,"utils/train_utils.py",3993,0," ",python,content
+98,41566,"utils/train_utils.py",3994,0,"",python,selection_keyboard
+99,41942,"utils/train_utils.py",3993,0,"",python,selection_command
+100,42604,"utils/train_utils.py",3996,0,"",python,selection_command
+101,44787,"utils/train_utils.py",3996,1,"m",python,selection_command
+102,44847,"utils/train_utils.py",3996,11,"max_logging",python,selection_command
+103,44998,"utils/train_utils.py",3996,12,"max_logging.",python,selection_command
+104,45356,"utils/train_utils.py",3996,15,"max_logging.log",python,selection_command
+105,46120,"utils/train_utils.py",3996,15,"",python,content
+106,46415,"utils/train_utils.py",3996,0,"p",python,content
+107,46415,"utils/train_utils.py",3997,0,"",python,selection_keyboard
+108,46511,"utils/train_utils.py",3997,0,"r",python,content
+109,46511,"utils/train_utils.py",3998,0,"",python,selection_keyboard
+110,46673,"utils/train_utils.py",3998,0,"i",python,content
+111,46674,"utils/train_utils.py",3999,0,"",python,selection_keyboard
+112,46744,"utils/train_utils.py",3999,0,"n",python,content
+113,46744,"utils/train_utils.py",4000,0,"",python,selection_keyboard
+114,46800,"utils/train_utils.py",4000,0,"t",python,content
+115,46800,"utils/train_utils.py",4001,0,"",python,selection_keyboard
+116,47106,"utils/train_utils.py",4000,0,"",python,selection_command
+117,47651,"utils/train_utils.py",4034,0,"",python,selection_command
+118,47865,"utils/train_utils.py",4044,0,"",python,selection_command
+119,47914,"utils/train_utils.py",4080,0,"",python,selection_command
+120,47939,"utils/train_utils.py",4113,0,"",python,selection_command
+121,47957,"utils/train_utils.py",4168,0,"",python,selection_command
+122,47990,"utils/train_utils.py",4223,0,"",python,selection_command
+123,48210,"utils/train_utils.py",4305,0,"",python,selection_command
+124,48481,"utils/train_utils.py",4223,0,"",python,selection_command
+125,49417,"utils/train_utils.py",4372,0,"t",python,content
+126,49417,"utils/train_utils.py",4367,5,"",python,content
+127,49417,"utils/train_utils.py",4365,0,"  pr",python,content
+128,49417,"utils/train_utils.py",4357,8,"",python,content
+129,49417,"utils/train_utils.py",4238,0,"t",python,content
+130,49417,"utils/train_utils.py",4233,5,"",python,content
+131,49417,"utils/train_utils.py",4231,0,"pr",python,content
+132,49417,"utils/train_utils.py",4223,8,"",python,content
+133,53218,"utils/train_utils.py",0,0,"",python,selection_command
+134,54689,"utils/train_utils.py",12,0,"\nimport jax",python,content
+135,55629,"utils/train_utils.py",13,0,"",python,selection_command
+136,56205,"utils/train_utils.py",13,11,"",python,content
+137,56299,"utils/train_utils.py",0,0,"",python,selection_command
+138,56686,"utils/train_utils.py",0,0,"import jax\n",python,content
+139,56688,"utils/train_utils.py",0,0,"",python,selection_command
+140,57999,"utils/train_utils.py",4360,0,"",python,selection_command
+141,58919,"utils/train_utils.py",4306,0,"",python,selection_command
+142,59163,"utils/train_utils.py",4234,0,"",python,selection_command
+143,59213,"utils/train_utils.py",4179,0,"",python,selection_command
+144,59218,"utils/train_utils.py",4124,0,"",python,selection_command
+145,59252,"utils/train_utils.py",4091,0,"",python,selection_command
+146,59288,"utils/train_utils.py",4055,0,"",python,selection_command
+147,59328,"utils/train_utils.py",4045,0,"",python,selection_command
+148,59401,"utils/train_utils.py",4011,0,"",python,selection_command
+149,59730,"utils/train_utils.py",3879,0,"",python,selection_command
+150,59832,"utils/train_utils.py",3846,0,"",python,selection_command
+151,76578,"utils/train_utils.py",3838,0,"",python,selection_command
+152,100328,"utils/train_utils.py",3837,0,"",python,selection_command
+153,100527,"utils/train_utils.py",3836,0,"",python,selection_command
+154,100731,"utils/train_utils.py",3837,0,"",python,selection_command
+155,100866,"utils/train_utils.py",3838,0,"",python,selection_command
+156,103918,"train_dynamics.py",0,0,"from dataclasses import dataclass, field\nimport itertools\nimport os\nfrom typing import cast, Optional\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_type: str = ""maskgit""  # supported options: maskgit, causal\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[Genie, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        # Dynamics\n        dyna_type=args.dyna_type,\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        decode=False,\n        rngs=rngs,\n    )\n    del genie.lam.decoder\n    return genie, rng\n\n\ndef build_optimizer(genie: Genie, args: Args) -> tuple[nnx.Optimizer, optax.Schedule]:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.Optimizer(genie, tx)\n    return optimizer, lr_schedule\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.Optimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> ocp.CheckpointManager:\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    return checkpoint_manager\n\n\ndef restore_or_initialize_components(\n    args: Args,\n    checkpoint_manager: ocp.CheckpointManager,\n    optimizer: nnx.Optimizer,\n    grain_iterator: grain.DataLoaderIterator,\n    rng: jax.Array,\n    replicated_sharding: NamedSharding,\n    restore_step: Optional[int] = None,\n) -> tuple[int, nnx.Optimizer, grain.DataLoaderIterator, jax.Array]:\n    step = 0\n    if restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            restore_step,\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = restore_step or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        rng, _rng = jax.random.split(rng)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, _rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n    return step, optimizer, grain_iterator, rng\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    genie, rng = build_model(args, rng)\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer, lr_schedule = build_optimizer(genie, args)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    mesh, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    grain_iterator = build_dataloader(args)\n\n    # --- Restore checkpoint ---\n    step, optimizer, grain_iterator, rng = restore_or_initialize_components(\n        args, checkpoint_manager, optimizer, grain_iterator, rng, replicated_sharding\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def dynamics_loss_fn(\n        model: Genie, inputs: dict\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        model.train()\n        outputs = model(inputs, training=True)\n        mask = outputs[""mask""]\n        outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n        ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n            outputs[""token_logits""], outputs[""video_tokens""]\n        )\n        ce_loss = (mask * ce_loss).sum() / mask.sum()\n        acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n        acc = (mask * acc).sum() / mask.sum()\n        select_probs = jax.nn.softmax(outputs[""token_logits""])\n        gt_val = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n        recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n        psnr = jnp.asarray(pix.psnr(gt_val, recon)).mean()\n        ssim = jnp.asarray(pix.ssim(gt_val, recon)).mean()\n        _, index_counts_lam = jnp.unique_counts(\n            jnp.ravel(outputs[""lam_indices""]),\n            size=args.num_latent_actions,\n            fill_value=0,\n        )\n        _, index_counts_tokenizer = jnp.unique_counts(\n            jnp.ravel(outputs[""video_tokens""]),\n            size=args.num_patch_latents,\n            fill_value=0,\n        )\n        codebook_usage_lam = (index_counts_lam != 0).mean()\n        codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n        metrics = dict(\n            cross_entropy_loss=ce_loss,\n            masked_token_accuracy=acc,\n            select_logit=outputs[""token_logits""].max(-1).mean(),\n            select_p=select_probs.max(-1).mean(),\n            entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n            psnr=psnr,\n            ssim=ssim,\n            codebook_usage_lam=codebook_usage_lam,\n            codebook_usage_tokenizer=codebook_usage_tokenizer,\n        )\n        return ce_loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.Optimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            return dynamics_loss_fn(model, inputs)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""dynamics""]\n            )\n        return loss, recon, metrics\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    if jax.process_index() == 0:\n        # TODO (f.srambical): use maxtext's get_shaped_batch instead of loading from the dataloader\n        first_videos = next(dataloader)\n        sample_inputs = dict(videos=first_videos, mask_rng=rng)\n        compiled = train_step.lower(optimizer, sample_inputs).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader = itertools.chain([first_videos], dataloader)\n    print(f""Starting training from step {step}..."")\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(videos=videos, mask_rng=_rng_mask)\n            loss, recon, metrics = train_step(optimizer, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                            recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+157,105632,"utils/train_utils.py",0,0,"",python,tab
+158,112240,"train_tokenizer.py",0,0,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast, Optional\n\nimport einops\nimport itertools\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    init_lr: float = 0.0\n    max_lr: float = 3e-4\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        20000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_tokenizer""\n    tags: list[str] = field(default_factory=lambda: [""tokenizer""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[TokenizerVQVAE, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    return (\n        TokenizerVQVAE(\n            in_dim=args.image_channels,\n            model_dim=args.model_dim,\n            ffn_dim=args.ffn_dim,\n            latent_dim=args.latent_dim,\n            num_latents=args.num_latents,\n            patch_size=args.patch_size,\n            num_blocks=args.num_blocks,\n            num_heads=args.num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.codebook_dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        ),\n        rng,\n    )\n\n\ndef build_optimizer(\n    model: TokenizerVQVAE, args: Args\n) -> tuple[nnx.Optimizer, optax.Schedule]:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.Optimizer(model, tx)\n    return optimizer, lr_schedule\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.Optimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> ocp.CheckpointManager:\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    return checkpoint_manager\n\n\ndef restore_checkpoint_if_needed(\n    args: Args,\n    checkpoint_manager: ocp.CheckpointManager,\n    optimizer: nnx.Optimizer,\n    grain_iterator: grain.DataLoaderIterator,\n    restore_step: Optional[int] = None,\n) -> tuple[int, nnx.Optimizer, grain.DataLoaderIterator]:\n    step = 0\n    if restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            restore_step,\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = restore_step or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    return step, optimizer, grain_iterator\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    tokenizer, rng = build_model(args, rng)\n\n    _, params, _ = nnx.split(tokenizer, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer, lr_schedule = build_optimizer(tokenizer, args)\n    del tokenizer\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    mesh, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    grain_iterator = build_dataloader(args)\n\n    # --- Restore checkpoint ---\n    step, optimizer, grain_iterator = restore_checkpoint_if_needed(\n        args, checkpoint_manager, optimizer, grain_iterator\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def tokenizer_loss_fn(\n        model: TokenizerVQVAE, inputs: dict\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        model.train()\n        outputs = model(inputs, training=True)\n        outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n        mse = jnp.square(gt - outputs[""recon""]).mean()\n        q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n        commitment_loss = jnp.square(\n            outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n        ).mean()\n        loss = mse + q_loss + args.vq_beta * commitment_loss\n\n        gt_clipped = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n        recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n        psnr = jnp.asarray(pix.psnr(gt_clipped, recon)).mean()\n        ssim = jnp.asarray(pix.ssim(gt_clipped, recon)).mean()\n        _, index_counts = jnp.unique_counts(\n            jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n        )\n        codebook_usage = (index_counts != 0).mean()\n        metrics = dict(\n            loss=loss,\n            mse=mse,\n            q_loss=q_loss,\n            commitment_loss=commitment_loss,\n            psnr=psnr,\n            ssim=ssim,\n            codebook_usage=codebook_usage,\n        )\n        return loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.Optimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: TokenizerVQVAE) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            return tokenizer_loss_fn(model, inputs)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""encoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""encoder""]\n            )\n            metrics[""vq_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""vq""]\n            )\n            metrics[""decoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""decoder""]\n            )\n        return loss, recon, metrics\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    if jax.process_index() == 0:\n        # TODO (f.srambical): use maxtext's get_shaped_batch instead of loading from the dataloader\n        first_videos = next(dataloader)\n        sample_inputs = dict(videos=first_videos)\n        compiled = train_step.lower(optimizer, sample_inputs).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader = itertools.chain([first_videos], dataloader)\n    print(f""Starting training from step {step}..."")\n    jax.config.update(""jax_transfer_guard"", ""disallow"")\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            inputs = dict(videos=videos)\n            loss, recon, metrics = train_step(optimizer, inputs)\n            # metrics[""lr""] = lr_schedule(step)\n            with jax.default_device(""cpu""):\n                step_on_cpu = jnp.array(step)\n                metrics[""lr""] = float(lr_schedule(step_on_cpu))\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+159,113683,"utils/train_utils.py",0,0,"",python,tab
+160,117799,"utils/train_utils.py",3842,0,"",python,selection_command
+161,117935,"utils/train_utils.py",3857,0,"",python,selection_command
+162,118135,"utils/train_utils.py",3858,0,"",python,selection_command
+163,123617,"utils/train_utils.py",3838,0,"",python,selection_command
+164,133991,"utils/train_utils.py",3842,0,"",python,selection_command
+165,134134,"utils/train_utils.py",3857,0,"",python,selection_command
+166,134653,"utils/train_utils.py",3842,0,"",python,selection_command
+167,136734,"train_tokenizer.py",0,0,"",python,tab
+168,138123,"train_tokenizer.py",11515,0,"",python,selection_command
+169,138737,"train_tokenizer.py",10920,0,"",python,selection_command
+170,139177,"train_tokenizer.py",10215,0,"",python,selection_command
+171,195279,"train_tokenizer.py",10920,0,"",python,selection_command
+172,195805,"train_tokenizer.py",11515,0,"",python,selection_command
+173,196459,"train_tokenizer.py",12140,0,"",python,selection_command
+174,197171,"train_tokenizer.py",12107,0,"",python,selection_command
+175,197417,"train_tokenizer.py",12097,0,"",python,selection_command
+176,197433,"train_tokenizer.py",12066,0,"",python,selection_command
+177,197458,"train_tokenizer.py",11996,0,"",python,selection_command
+178,197498,"train_tokenizer.py",11977,0,"",python,selection_command
+179,197528,"train_tokenizer.py",11952,0,"",python,selection_command
+180,197979,"train_tokenizer.py",11944,0,"",python,selection_command
+181,278923,"train_tokenizer.py",11969,0,"",python,selection_command
+182,279209,"train_tokenizer.py",11988,0,"",python,selection_command
+183,279216,"train_tokenizer.py",12058,0,"",python,selection_command
+184,279242,"train_tokenizer.py",12093,0,"",python,selection_command
+185,279429,"train_tokenizer.py",12099,0,"",python,selection_command
+186,279675,"train_tokenizer.py",12132,0,"",python,selection_command
+187,279706,"train_tokenizer.py",12232,0,"",python,selection_command
+188,279739,"train_tokenizer.py",12272,0,"",python,selection_command
+189,279867,"train_tokenizer.py",12322,0,"",python,selection_command
+190,279868,"train_tokenizer.py",12394,0,"",python,selection_command
+191,279869,"train_tokenizer.py",12458,0,"",python,selection_command
+192,279869,"train_tokenizer.py",12521,0,"",python,selection_command
+193,279901,"train_tokenizer.py",12575,0,"",python,selection_command
+194,279937,"train_tokenizer.py",12640,0,"",python,selection_command
+195,279970,"train_tokenizer.py",12692,0,"",python,selection_command
+196,280000,"train_tokenizer.py",12748,0,"",python,selection_command
+197,280048,"train_tokenizer.py",12781,0,"",python,selection_command
+198,280079,"train_tokenizer.py",12815,0,"",python,selection_command
+199,280100,"train_tokenizer.py",12848,0,"",python,selection_command
+200,280141,"train_tokenizer.py",12889,0,"",python,selection_command
+201,280172,"train_tokenizer.py",12954,0,"",python,selection_command
+202,280218,"train_tokenizer.py",13002,0,"",python,selection_command
+203,280314,"train_tokenizer.py",12954,0,"",python,selection_command
+204,280560,"train_tokenizer.py",12889,0,"",python,selection_command
+205,280594,"train_tokenizer.py",12848,0,"",python,selection_command
+206,280622,"train_tokenizer.py",12815,0,"",python,selection_command
+207,280653,"train_tokenizer.py",12781,0,"",python,selection_command
+208,280679,"train_tokenizer.py",12748,0,"",python,selection_command
+209,280712,"train_tokenizer.py",12692,0,"",python,selection_command
+210,281020,"train_tokenizer.py",12640,0,"",python,selection_command
+211,281349,"train_tokenizer.py",12575,0,"",python,selection_command
+212,283539,"train_tokenizer.py",12639,0,"\n        ",python,content
+213,283973,"train_tokenizer.py",12644,4,"",python,content
+214,290202,"train_tokenizer.py",12644,0,"print_mem_stats(""before training"")",python,content
+215,290501,"train_tokenizer.py",12677,0,"",python,selection_command
+216,290724,"train_tokenizer.py",12676,0,"",python,selection_command
+217,290874,"train_tokenizer.py",12668,0,"",python,selection_command
+218,291013,"train_tokenizer.py",12661,0,"",python,selection_command
+219,305410,"train_tokenizer.py",12668,0,"",python,selection_command
+220,305844,"train_tokenizer.py",12661,0,"",python,selection_command
+221,306527,"train_tokenizer.py",12661,1,"b",python,selection_command
+222,306606,"train_tokenizer.py",12661,6,"before",python,selection_command
+223,306813,"train_tokenizer.py",12661,15,"before training",python,selection_command
+224,308739,"train_tokenizer.py",12661,15,"",python,content
+225,309066,"train_tokenizer.py",12661,0,"A",python,content
+226,309066,"train_tokenizer.py",12662,0,"",python,selection_keyboard
+227,309418,"train_tokenizer.py",12662,0,"f",python,content
+228,309418,"train_tokenizer.py",12663,0,"",python,selection_keyboard
+229,309683,"train_tokenizer.py",12663,0,"t",python,content
+230,309684,"train_tokenizer.py",12664,0,"",python,selection_keyboard
+231,309738,"train_tokenizer.py",12664,0,"e",python,content
+232,309738,"train_tokenizer.py",12665,0,"",python,selection_keyboard
+233,309874,"train_tokenizer.py",12665,0,"r",python,content
+234,309875,"train_tokenizer.py",12666,0,"",python,selection_keyboard
+235,309977,"train_tokenizer.py",12666,0," ",python,content
+236,309977,"train_tokenizer.py",12667,0,"",python,selection_keyboard
+237,310218,"train_tokenizer.py",12667,0,"p",python,content
+238,310219,"train_tokenizer.py",12668,0,"",python,selection_keyboard
+239,310315,"train_tokenizer.py",12668,0,"a",python,content
+240,310316,"train_tokenizer.py",12669,0,"",python,selection_keyboard
+241,310382,"train_tokenizer.py",12669,0,"r",python,content
+242,310383,"train_tokenizer.py",12670,0,"",python,selection_keyboard
+243,310516,"train_tokenizer.py",12670,0,"a",python,content
+244,310516,"train_tokenizer.py",12671,0,"",python,selection_keyboard
+245,310604,"train_tokenizer.py",12671,0,"m",python,content
+246,310605,"train_tokenizer.py",12672,0,"",python,selection_keyboard
+247,310715,"train_tokenizer.py",12672,0,"s",python,content
+248,310716,"train_tokenizer.py",12673,0,"",python,selection_keyboard
+249,310811,"train_tokenizer.py",12673,0," ",python,content
+250,310811,"train_tokenizer.py",12674,0,"",python,selection_keyboard
+251,311402,"train_tokenizer.py",12674,0,"i",python,content
+252,311402,"train_tokenizer.py",12675,0,"",python,selection_keyboard
+253,311446,"train_tokenizer.py",12675,0,"n",python,content
+254,311446,"train_tokenizer.py",12676,0,"",python,selection_keyboard
+255,311579,"train_tokenizer.py",12676,0,"i",python,content
+256,311580,"train_tokenizer.py",12677,0,"",python,selection_keyboard
+257,311676,"train_tokenizer.py",12677,0,"t",python,content
+258,311676,"train_tokenizer.py",12678,0,"",python,selection_keyboard
+259,311783,"train_tokenizer.py",12678,0,"i",python,content
+260,311784,"train_tokenizer.py",12679,0,"",python,selection_keyboard
+261,311988,"train_tokenizer.py",12679,0,"a",python,content
+262,311988,"train_tokenizer.py",12680,0,"",python,selection_keyboard
+263,312287,"train_tokenizer.py",12680,0,"l",python,content
+264,312287,"train_tokenizer.py",12681,0,"",python,selection_keyboard
+265,312342,"train_tokenizer.py",12681,0,"i",python,content
+266,312342,"train_tokenizer.py",12682,0,"",python,selection_keyboard
+267,312430,"train_tokenizer.py",12682,0,"z",python,content
+268,312430,"train_tokenizer.py",12683,0,"",python,selection_keyboard
+269,312554,"train_tokenizer.py",12683,0,"e",python,content
+270,312554,"train_tokenizer.py",12684,0,"",python,selection_keyboard
+271,312716,"train_tokenizer.py",12684,0,"d",python,content
+272,312716,"train_tokenizer.py",12685,0,"",python,selection_keyboard
+273,312979,"train_tokenizer.py",12684,0,"",python,selection_command
+274,316460,"train_tokenizer.py",12640,0,"",python,selection_command
+275,316895,"train_tokenizer.py",12644,0,"",python,selection_command
+276,317982,"train_tokenizer.py",0,0,"",python,selection_command
+277,320130,"train_tokenizer.py",41,0,"",python,selection_command
+278,320378,"train_tokenizer.py",51,0,"",python,selection_command
+279,320408,"train_tokenizer.py",85,0,"",python,selection_command
+280,320443,"train_tokenizer.py",86,0,"",python,selection_command
+281,320693,"train_tokenizer.py",100,0,"",python,selection_command
+282,320831,"train_tokenizer.py",117,0,"",python,selection_command
+283,323546,"train_tokenizer.py",592,0,"\n    print_mem_stats,",python,content
+284,324557,"train_tokenizer.py",12423,0,"",python,selection_command
+285,330415,"train_tokenizer.py",12487,0,"",python,selection_command
+286,330647,"train_tokenizer.py",12550,0,"",python,selection_command
+287,330672,"train_tokenizer.py",12604,0,"",python,selection_command
+288,331194,"train_tokenizer.py",12669,0,"",python,selection_command
+289,349311,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+290,351497,"train_tokenizer.py",0,0,"",python,tab
+291,351500,"TERMINAL",0,0,"",,terminal_focus
+292,352508,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+293,353522,"TERMINAL",0,0,"squeue",,terminal_command
+294,353522,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  22981   xiao.liu interacti     1    32  R 2025-08-22T14:55:05 2025-08-22T14:55:05 19:14:53   23:59:00 hai001\r\n  22980   xiao.liu interacti     1    32  R 2025-08-22T14:54:50 2025-08-22T14:54:50 19:15:08   23:59:00 hai002\r\n  23178 alfred.ngu  standard     1    16  R 2025-08-23T01:26:22 2025-08-23T01:26:39  8:43:19 1-00:00:00 hai007\r\n  23177 alfred.ngu  standard     1    16  R 2025-08-23T01:26:18 2025-08-23T01:26:19  8:43:39 1-00:00:00 hai006\r\n  22675 nishant.ku  standard     3   192  R 2025-08-22T11:17:07 2025-08-22T11:17:07 22:52:51 1-00:00:00 hai[003-005]\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+295,362786,"TERMINAL",0,0,"salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G",,terminal_command
+296,362835,"TERMINAL",0,0,"]633;Csalloc: Granted job allocation 23180\r\n",,terminal_output
+297,362943,"TERMINAL",0,0,"salloc: Nodes hai008 are ready for job\r\n",,terminal_output
+298,363301,"TERMINAL",0,0,"Running inside SLURM, Job ID 23180.\r\n",,terminal_output
+299,363377,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/jafar[?2004h[franz.srambical@hai008.haicore.berlin:~/jafar] $ ",,terminal_output
+300,366299,"TERMINAL",0,0,"\r(reverse-i-search)`': ",,terminal_output
+301,366700,"TERMINAL",0,0,"b': source /home/franz.srambical/maxtext/.venv/bin/activatesall\ra': python3 -m MaxText.train MaxText/configs/base.yml   run_name=h100_mfu_340m   hardware=gpu   dataset_type=synthetic   steps=60   log_period=1   enable_checkpointing=False   gcs_metrics=False   metrics_file=/tmp/h100_mfu_metrics.jsonl   base_output_directory=/tmp/maxtext   attention=autoselected   per_device_batch_size=4   base_emb_dim=1536   base_num_query_heads=12   base_num_kv_heads=12   base_mlp_dim=4096   base_num_decoder_layers=10   head_dim=128   logits_via_embedding=True",,terminal_output
+302,366771,"TERMINAL",0,0,"\rs': python3 -m MaxText.train MaxText/configs/base.yml   run_name=h100_mfu_340m   hardware=gpu   dataset_type=synthetic   steps=60   log_period=1   enable_checkpointing=False   gcs_metrics=False   metrics_file=/tmp/h100_mfu_metrics.jsonl   base_output_directory=/tmp/maxtext   attention=autoselected   per_device_batch_size=4   base_emb_dim=1536   base_num_query_heads=12   base_num_kv_heads=12   base_mlp_dim=4096   base_num_decoder_layers=10   head_dim=128   logits_via_embedding=True",,terminal_output
+303,366850,"TERMINAL",0,0,"\rh': . ""/fast/home/franz.srambical/.cursor-server/bin/af58d92614edb1f72bdd756615d131bf8dfa5290/out/vs/workbench/contrib/terminal/common/scripts/shellIntegration-bash.sh""\r\n\r\r\n\r",,terminal_output
+304,366966,"TERMINAL",0,0," ': bash slurm/utils/create_dev_dir.sh . ../sbatch-runner/coinrun-lam-nan-investigation-to-107k\r\n\r",,terminal_output
+305,368082,"TERMINAL",0,0,"experiments/dynamics_grain_full_restore.sh \r",,terminal_output
+306,368804,"TERMINAL",0,0,"\r[23@[franz.srambical@hai008.haicore.berlin:~/jafar] $ bash ",,terminal_output
+307,369098,"TERMINAL",0,0,"",,terminal_output
+308,369420,"TERMINAL",0,0,"",,terminal_output
+309,369506,"TERMINAL",0,0,"",,terminal_output
+310,370269,"TERMINAL",0,0,"",,terminal_output
+311,370308,"TERMINAL",0,0,"",,terminal_output
+312,370448,"TERMINAL",0,0,"",,terminal_output
+313,370585,"TERMINAL",0,0,"",,terminal_output
+314,370816,"TERMINAL",0,0,"t",,terminal_output
+315,370912,"TERMINAL",0,0,"ok",,terminal_output
+316,371313,"TERMINAL",0,0,"",,terminal_output
+317,371443,"TERMINAL",0,0,"",,terminal_output
+318,371523,"TERMINAL",0,0,"",,terminal_output
+319,371914,"TERMINAL",0,0,"",,terminal_output
+320,372041,"TERMINAL",0,0,"/",,terminal_output
+321,372337,"TERMINAL",0,0,"t",,terminal_output
+322,372443,"TERMINAL",0,0,"ok",,terminal_output
+323,372580,"TERMINAL",0,0,"enizer_",,terminal_output
+324,373011,"TERMINAL",0,0,"",,terminal_output
+325,373489,"TERMINAL",0,0,"g",,terminal_output
+326,373693,"TERMINAL",0,0,"r",,terminal_output
+327,373880,"TERMINAL",0,0,"ain_checkpointing.sh ",,terminal_output
+328,377933,"utils/train_utils.py",0,0,"",python,tab
+329,377933,"utils/train_utils.py",3842,0,"",python,selection_command
+330,379871,"train_tokenizer.py",0,0,"",python,tab
+331,379872,"train_tokenizer.py",12669,0,"",python,selection_command
+332,380832,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+333,387862,"TERMINAL",0,0,"Running on 1 devices.\r\n",,terminal_output
+334,391879,"TERMINAL",0,0,"Counting all components: ['decoder', 'encoder', 'vq']\r\nParameter counts:\r\n{'decoder': 16858736, 'encoder': 16858752, 'vq': 32768, 'total': 33750256}\r\n",,terminal_output
+335,419970,"TERMINAL",0,0,"Total memory size: 8.8 GB, Output size: 0.4 GB, Temp size: 8.3 GB, Argument size: 0.4 GB, Host temp size: 0.0 GB.\r\nFLOPs: 4.481e+13, Bytes: 3.947e+11 (367.6 GB), Intensity: 113.5 FLOPs/byte\r\n\r\nMemstats: After params initialized:\r\n\tUsing (GB) 0.41 / 59.39 (0.690352%) on cuda:0\r\nStarting training from step 0...\r\n",,terminal_output
+336,420371,"TERMINAL",0,0,"jax.errors.SimplifiedTraceback: For simplicity, JAX has removed its internal frames from the traceback of the following exception. Set JAX_TRACEBACK_FILTERING=off to include these.\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n  File ""/fast/home/franz.srambical/jafar/train_tokenizer.py"", line 440, in <module>\r\n    main(args)\r\n  File ""/fast/home/franz.srambical/jafar/train_tokenizer.py"", line 385, in main\r\n    step_on_cpu = jnp.array(step)\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/numpy/array.py"", line 268, in array\r\n    out_array: Array = lax._convert_element_type(\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/lax/lax.py"", line 1726, in _convert_element_type\r\n    return convert_element_type_p.bind(\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/core.py"", line 536, in bind\r\n    return self._true_bind(*args, **params)\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/core.py"", line 552, in _true_bind\r\n    return self.bind_with_trace(prev_trace, args, params)\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/lax/lax.py"", line 4902, in _convert_element_type_bind_with_trace\r\n    operand = core.Primitive.bind_with_trace(convert_element_type_p, trace, args, params)\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/core.py"", line 562, in bind_with_trace\r\n    return trace.process_primitive(self, args, params)\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/core.py"", line 1066, in process_primitive\r\n    return primitive.impl(*args, **params)\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/jax/_src/dispatch.py"", line 91, in apply_primitive\r\n    outs = fun(*args)\r\njaxlib._jax.XlaRuntimeError: INVALID_ARGUMENT: Disallowed host-to-device transfer: aval=ShapedArray(int32[]), dst_sharding=SingleDeviceSharding(device=CpuDevice(id=0), memory_kind=unpinned_host)\r\n",,terminal_output
+337,423510,"TERMINAL",0,0,"/home/franz.srambical/.local/share/uv/python/cpython-3.10.18-linux-x86_64-gnu/lib/python3.10/multiprocessing/resource_tracker.py:224: UserWarning: resource_tracker: There appear to be 10 leaked shared_memory objects to clean up at shutdown\r\n  warnings.warn('resource_tracker: There appear to be %d '\r\n",,terminal_output
+338,423822,"TERMINAL",0,0,"srun: error: hai008: task 0: Exited with exit code 1\r\n]0;franz.srambical@hai-login1:~/jafar[?2004h[franz.srambical@hai008.haicore.berlin:~/jafar] $ ",,terminal_output
+339,463313,"train_tokenizer.py",12661,48,"",python,content
+340,463338,"train_tokenizer.py",12665,0,"",python,selection_command
+341,463627,"train_tokenizer.py",12717,0,"",python,selection_command
+342,463893,"train_tokenizer.py",12773,0,"",python,selection_command
+343,463893,"train_tokenizer.py",12806,0,"",python,selection_command
+344,463946,"train_tokenizer.py",12840,0,"",python,selection_command
+345,463963,"train_tokenizer.py",12873,0,"",python,selection_command
+346,463992,"train_tokenizer.py",12914,0,"",python,selection_command
+347,464042,"train_tokenizer.py",12979,0,"",python,selection_command
+348,464075,"train_tokenizer.py",13027,0,"",python,selection_command
+349,464092,"train_tokenizer.py",13071,0,"",python,selection_command
+350,464138,"train_tokenizer.py",13117,0,"",python,selection_command
+351,464152,"train_tokenizer.py",13181,0,"",python,selection_command
+352,464192,"train_tokenizer.py",13229,0,"",python,selection_command
+353,464250,"train_tokenizer.py",13247,0,"",python,selection_command
+354,464266,"train_tokenizer.py",13252,0,"",python,selection_command
+355,464315,"train_tokenizer.py",13282,0,"",python,selection_command
+356,464672,"train_tokenizer.py",13252,0,"",python,selection_command
+357,464932,"train_tokenizer.py",13247,0,"",python,selection_command
+358,464972,"train_tokenizer.py",13229,0,"",python,selection_command
+359,464986,"train_tokenizer.py",13181,0,"",python,selection_command
+360,465009,"train_tokenizer.py",13117,0,"",python,selection_command
+361,465052,"train_tokenizer.py",13071,0,"",python,selection_command
+362,465072,"train_tokenizer.py",13027,0,"",python,selection_command
+363,465112,"train_tokenizer.py",12979,0,"",python,selection_command
+364,465162,"train_tokenizer.py",12914,0,"",python,selection_command
+365,465294,"train_tokenizer.py",12873,0,"",python,selection_command
+366,465512,"train_tokenizer.py",12840,0,"",python,selection_command
+367,465626,"train_tokenizer.py",12806,0,"",python,selection_command
+368,466012,"train_tokenizer.py",12773,0,"",python,selection_command
+369,471272,"train_tokenizer.py",12773,0,"",python,selection_command
+370,471552,"train_tokenizer.py",12717,0,"",python,selection_command
+371,471802,"train_tokenizer.py",12768,0,"\n    ",python,content
+372,472090,"train_tokenizer.py",12773,0,"f",python,content
+373,472090,"train_tokenizer.py",12774,0,"",python,selection_keyboard
+374,472233,"train_tokenizer.py",12774,0,"i",python,content
+375,472233,"train_tokenizer.py",12775,0,"",python,selection_keyboard
+376,472313,"train_tokenizer.py",12775,0,"s",python,content
+377,472313,"train_tokenizer.py",12776,0,"",python,selection_keyboard
+378,472396,"train_tokenizer.py",12776,0,"r",python,content
+379,472396,"train_tokenizer.py",12777,0,"",python,selection_keyboard
+380,472954,"train_tokenizer.py",12776,1,"",python,content
+381,473072,"train_tokenizer.py",12775,1,"",python,content
+382,473162,"train_tokenizer.py",12775,0,"r",python,content
+383,473163,"train_tokenizer.py",12776,0,"",python,selection_keyboard
+384,473308,"train_tokenizer.py",12776,0,"s",python,content
+385,473308,"train_tokenizer.py",12777,0,"",python,selection_keyboard
+386,473373,"train_tokenizer.py",12777,0,"t",python,content
+387,473373,"train_tokenizer.py",12778,0,"",python,selection_keyboard
+388,474633,"train_tokenizer.py",12778,0,"_",python,content
+389,474634,"train_tokenizer.py",12779,0,"",python,selection_keyboard
+390,474906,"train_tokenizer.py",12779,0,"s",python,content
+391,474907,"train_tokenizer.py",12780,0,"",python,selection_keyboard
+392,475050,"train_tokenizer.py",12780,0,"t",python,content
+393,475051,"train_tokenizer.py",12781,0,"",python,selection_keyboard
+394,475055,"train_tokenizer.py",12781,0,"e",python,content
+395,475055,"train_tokenizer.py",12782,0,"",python,selection_keyboard
+396,475194,"train_tokenizer.py",12782,0,"p",python,content
+397,475194,"train_tokenizer.py",12783,0,"",python,selection_keyboard
+398,475273,"train_tokenizer.py",12783,0," ",python,content
+399,475274,"train_tokenizer.py",12784,0,"",python,selection_keyboard
+400,475930,"train_tokenizer.py",12784,0,"= step",python,content
+401,476138,"train_tokenizer.py",12789,0,"",python,selection_command
+402,477073,"train_tokenizer.py",12811,0,"",python,selection_command
+403,477313,"train_tokenizer.py",12844,0,"",python,selection_command
+404,477353,"train_tokenizer.py",12878,0,"",python,selection_command
+405,477354,"train_tokenizer.py",12911,0,"",python,selection_command
+406,477410,"train_tokenizer.py",12952,0,"",python,selection_command
+407,477433,"train_tokenizer.py",13017,0,"",python,selection_command
+408,477451,"train_tokenizer.py",13065,0,"",python,selection_command
+409,477513,"train_tokenizer.py",13109,0,"",python,selection_command
+410,477553,"train_tokenizer.py",13155,0,"",python,selection_command
+411,477562,"train_tokenizer.py",13219,0,"",python,selection_command
+412,477593,"train_tokenizer.py",13267,0,"",python,selection_command
+413,477633,"train_tokenizer.py",13269,0,"",python,selection_command
+414,477691,"train_tokenizer.py",13290,0,"",python,selection_command
+415,477706,"train_tokenizer.py",13320,0,"",python,selection_command
+416,477738,"train_tokenizer.py",13345,0,"",python,selection_command
+417,477793,"train_tokenizer.py",13424,0,"",python,selection_command
+418,477802,"train_tokenizer.py",13455,0,"",python,selection_command
+419,477833,"train_tokenizer.py",13481,0,"",python,selection_command
+420,477883,"train_tokenizer.py",13523,0,"",python,selection_command
+421,477914,"train_tokenizer.py",13565,0,"",python,selection_command
+422,477953,"train_tokenizer.py",13604,0,"",python,selection_command
+423,477994,"train_tokenizer.py",13630,0,"",python,selection_command
+424,477994,"train_tokenizer.py",13652,0,"",python,selection_command
+425,478033,"train_tokenizer.py",13708,0,"",python,selection_command
+426,478073,"train_tokenizer.py",13785,0,"",python,selection_command
+427,478106,"train_tokenizer.py",13837,0,"",python,selection_command
+428,478171,"train_tokenizer.py",13919,0,"",python,selection_command
+429,478172,"train_tokenizer.py",13974,0,"",python,selection_command
+430,478210,"train_tokenizer.py",14043,0,"",python,selection_command
+431,478250,"train_tokenizer.py",14065,0,"",python,selection_command
+432,478282,"train_tokenizer.py",14145,0,"",python,selection_command
+433,478313,"train_tokenizer.py",14223,0,"",python,selection_command
+434,478362,"train_tokenizer.py",14300,0,"",python,selection_command
+435,478378,"train_tokenizer.py",14349,0,"",python,selection_command
+436,478435,"train_tokenizer.py",14392,0,"",python,selection_command
+437,478443,"train_tokenizer.py",14462,0,"",python,selection_command
+438,478473,"train_tokenizer.py",14535,0,"",python,selection_command
+439,478513,"train_tokenizer.py",14590,0,"",python,selection_command
+440,478554,"train_tokenizer.py",14666,0,"",python,selection_command
+441,478586,"train_tokenizer.py",14697,0,"",python,selection_command
+442,478633,"train_tokenizer.py",14723,0,"",python,selection_command
+443,478634,"train_tokenizer.py",14769,0,"",python,selection_command
+444,478691,"train_tokenizer.py",14805,0,"",python,selection_command
+445,478713,"train_tokenizer.py",14881,0,"",python,selection_command
+446,478753,"train_tokenizer.py",14936,0,"",python,selection_command
+447,478794,"train_tokenizer.py",14977,0,"",python,selection_command
+448,478810,"train_tokenizer.py",15003,0,"",python,selection_command
+449,479273,"train_tokenizer.py",15048,0,"",python,selection_command
+450,479473,"train_tokenizer.py",15003,0,"",python,selection_command
+451,479713,"train_tokenizer.py",14977,0,"",python,selection_command
+452,479778,"train_tokenizer.py",14936,0,"",python,selection_command
+453,479802,"train_tokenizer.py",14881,0,"",python,selection_command
+454,479834,"train_tokenizer.py",14805,0,"",python,selection_command
+455,479893,"train_tokenizer.py",14769,0,"",python,selection_command
+456,479893,"train_tokenizer.py",14723,0,"",python,selection_command
+457,479914,"train_tokenizer.py",14697,0,"",python,selection_command
+458,479953,"train_tokenizer.py",14666,0,"",python,selection_command
+459,480010,"train_tokenizer.py",14590,0,"",python,selection_command
+460,480011,"train_tokenizer.py",14535,0,"",python,selection_command
+461,480074,"train_tokenizer.py",14462,0,"",python,selection_command
+462,480113,"train_tokenizer.py",14392,0,"",python,selection_command
+463,480114,"train_tokenizer.py",14349,0,"",python,selection_command
+464,480153,"train_tokenizer.py",14300,0,"",python,selection_command
+465,480193,"train_tokenizer.py",14223,0,"",python,selection_command
+466,480234,"train_tokenizer.py",14145,0,"",python,selection_command
+467,480266,"train_tokenizer.py",14065,0,"",python,selection_command
+468,480282,"train_tokenizer.py",14043,0,"",python,selection_command
+469,480313,"train_tokenizer.py",13974,0,"",python,selection_command
+470,480362,"train_tokenizer.py",13919,0,"",python,selection_command
+471,480401,"train_tokenizer.py",13837,0,"",python,selection_command
+472,480433,"train_tokenizer.py",13785,0,"",python,selection_command
+473,480473,"train_tokenizer.py",13708,0,"",python,selection_command
+474,480538,"train_tokenizer.py",13652,0,"",python,selection_command
+475,480553,"train_tokenizer.py",13630,0,"",python,selection_command
+476,480554,"train_tokenizer.py",13604,0,"",python,selection_command
+477,480593,"train_tokenizer.py",13565,0,"",python,selection_command
+478,480650,"train_tokenizer.py",13523,0,"",python,selection_command
+479,480651,"train_tokenizer.py",13481,0,"",python,selection_command
+480,480713,"train_tokenizer.py",13455,0,"",python,selection_command
+481,480714,"train_tokenizer.py",13424,0,"",python,selection_command
+482,480763,"train_tokenizer.py",13345,0,"",python,selection_command
+483,480795,"train_tokenizer.py",13320,0,"",python,selection_command
+484,480858,"train_tokenizer.py",13290,0,"",python,selection_command
+485,480874,"train_tokenizer.py",13269,0,"",python,selection_command
+486,480897,"train_tokenizer.py",13267,0,"",python,selection_command
+487,480913,"train_tokenizer.py",13219,0,"",python,selection_command
+488,480970,"train_tokenizer.py",13155,0,"",python,selection_command
+489,481002,"train_tokenizer.py",13109,0,"",python,selection_command
+490,481041,"train_tokenizer.py",13065,0,"",python,selection_command
+491,481073,"train_tokenizer.py",13017,0,"",python,selection_command
+492,481082,"train_tokenizer.py",12952,0,"",python,selection_command
+493,481130,"train_tokenizer.py",12911,0,"",python,selection_command
+494,481153,"train_tokenizer.py",12878,0,"",python,selection_command
+495,481194,"train_tokenizer.py",12844,0,"",python,selection_command
+496,481275,"train_tokenizer.py",12811,0,"",python,selection_command
+497,481290,"train_tokenizer.py",12789,0,"",python,selection_command
+498,481291,"train_tokenizer.py",12733,0,"",python,selection_command
+499,481313,"train_tokenizer.py",12681,0,"",python,selection_command
+500,481429,"train_tokenizer.py",12733,0,"",python,selection_command
+501,481673,"train_tokenizer.py",12789,0,"",python,selection_command
+502,481793,"train_tokenizer.py",12733,0,"",python,selection_command
+503,482075,"train_tokenizer.py",12616,0,"",python,selection_command
+504,482113,"train_tokenizer.py",12562,0,"",python,selection_command
+505,482153,"train_tokenizer.py",12499,0,"",python,selection_command
+506,482233,"train_tokenizer.py",12562,0,"",python,selection_command
+507,482673,"train_tokenizer.py",12616,0,"",python,selection_command
+508,483513,"train_tokenizer.py",12733,0,"",python,selection_command
+509,483553,"train_tokenizer.py",12789,0,"",python,selection_command
+510,483553,"train_tokenizer.py",12811,0,"",python,selection_command
+511,483753,"train_tokenizer.py",12844,0,"",python,selection_command
+512,484010,"train_tokenizer.py",12878,0,"",python,selection_command
+513,484153,"train_tokenizer.py",12911,0,"",python,selection_command
+514,484363,"train_tokenizer.py",12952,0,"",python,selection_command
+515,484521,"train_tokenizer.py",13017,0,"",python,selection_command
+516,484651,"train_tokenizer.py",13065,0,"",python,selection_command
+517,484881,"train_tokenizer.py",13109,0,"",python,selection_command
+518,485433,"train_tokenizer.py",13065,0,"",python,selection_command
+519,495313,"train_tokenizer.py",13017,0,"",python,selection_command
+520,495434,"train_tokenizer.py",12952,0,"",python,selection_command
+521,497651,"train_tokenizer.py",12996,0,"\n            ",python,content
+522,497914,"train_tokenizer.py",13009,0,"i",python,content
+523,497914,"train_tokenizer.py",13010,0,"",python,selection_keyboard
+524,498041,"train_tokenizer.py",13010,0,"f",python,content
+525,498042,"train_tokenizer.py",13011,0,"",python,selection_keyboard
+526,498123,"train_tokenizer.py",13011,0," ",python,content
+527,498124,"train_tokenizer.py",13012,0,"",python,selection_keyboard
+528,499562,"train_tokenizer.py",13012,0,"step == first_step:",python,content
+529,500333,"train_tokenizer.py",13031,0,"\n                print_mem_stats(""After first train step"")",python,content
+530,500673,"train_tokenizer.py",13088,0,"",python,selection_command
+531,501610,"train_tokenizer.py",13087,0,"",python,selection_command
+532,501738,"train_tokenizer.py",13083,0,"",python,selection_command
+533,502442,"train_tokenizer.py",13065,22,"",python,content
+534,503322,"train_tokenizer.py",13064,0,"",python,selection_command
+535,503433,"train_tokenizer.py",13067,0,"\n                ",python,content
+536,503633,"train_tokenizer.py",13084,0,"def print_mem_stats(label: str):\n  max_logging.log(f""\nMemstats: {label}:"")\n  try:\n    for d in jax.local_devices():\n      stats = d.memory_stats()\n      used = round(stats[""bytes_in_use""] / 2**30, 2)\n      limit = round(stats[""bytes_limit""] / 2**30, 2)\n      max_logging.log(f""\tUsing (GB) {used} / {limit} ({used/limit:%}) on {d}"")\n  except (RuntimeError, KeyError, TypeError) as ex:\n    max_logging.log(f""\tMemstats unavailable, error: {ex}"")",python,content
+537,503930,"train_tokenizer.py",13528,0,"",python,selection_command
+538,504713,"train_tokenizer.py",13080,462,"",python,content
+539,504730,"train_tokenizer.py",13064,0,"",python,selection_command
+540,508234,"train_tokenizer.py",13065,0,"After first train step",python,content
+541,508234,"train_tokenizer.py",13086,0,"",python,selection_command
+542,509833,"train_tokenizer.py",13065,22,"",python,content
+543,509850,"train_tokenizer.py",13064,0,"",python,selection_command
+544,527753,"train_tokenizer.py",13065,0,"",python,selection_command
+545,527953,"train_tokenizer.py",13065,0,"After params initialized",python,content
+546,527954,"train_tokenizer.py",13089,0,"",python,selection_keyboard
+547,528234,"train_tokenizer.py",13088,0,"",python,selection_command
+548,531393,"TERMINAL",0,0,"^C[?2004l\r[?2004h[?2004l\r\r\n]0;franz.srambical@hai-login1:~/jafar[?2004h[franz.srambical@hai008.haicore.berlin:~/jafar] $ [franz.srambical@hai008.haicore.berlin:~/jafar] $ ",,terminal_output
+549,531473,"TERMINAL",0,0,"bash experiments/tokenizer_grain_checkpointing.sh ",,terminal_output
+550,531961,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+551,533362,"TERMINAL",0,0,"^Csrun: interrupt (one more within 1 sec to abort)\r\nsrun: StepId=23180.1 task 0: running\r\n",,terminal_output
+552,535402,"train_tokenizer.py",13030,0,"",python,selection_command
+553,536793,"train_tokenizer.py",0,0,"",python,selection_command
+554,536802,"TERMINAL",0,0,"Running on 1 devices.\r\n",,terminal_output
+555,537202,"train_tokenizer.py",41,0,"",python,selection_command
+556,537433,"train_tokenizer.py",51,0,"",python,selection_command
+557,537498,"train_tokenizer.py",85,0,"",python,selection_command
+558,537507,"train_tokenizer.py",86,0,"",python,selection_command
+559,537535,"train_tokenizer.py",100,0,"",python,selection_command
+560,537563,"train_tokenizer.py",117,0,"",python,selection_command
+561,537603,"train_tokenizer.py",177,0,"",python,selection_command
+562,537833,"train_tokenizer.py",597,0,"",python,selection_command
+563,538674,"train_tokenizer.py",931,0,"",python,selection_command
+564,539793,"train_tokenizer.py",1435,0,"",python,selection_command
+565,540002,"train_tokenizer.py",1925,0,"",python,selection_command
+566,540753,"TERMINAL",0,0,"Counting all components: ['decoder', 'encoder', 'vq']\r\nParameter counts:\r\n{'decoder': 16858736, 'encoder': 16858752, 'vq': 32768, 'total': 33750256}\r\n",,terminal_output
+567,542593,"train_tokenizer.py",2487,0,"",python,selection_command
+568,542594,"train_tokenizer.py",2969,0,"",python,selection_command
+569,543313,"train_tokenizer.py",3402,0,"",python,selection_command
+570,543603,"train_tokenizer.py",4048,0,"",python,selection_command
+571,545833,"train_tokenizer.py",12740,0,"",python,selection_command
+572,547073,"train_tokenizer.py",12754,0,"",python,selection_command
+573,547233,"train_tokenizer.py",12757,0,"",python,selection_command
+574,547713,"train_tokenizer.py",12758,0,"",python,selection_command
+575,548027,"train_tokenizer.py",12758,8,"",python,content
+576,548159,"train_tokenizer.py",12758,0,"a",python,content
+577,548159,"train_tokenizer.py",12759,0,"",python,selection_keyboard
+578,548234,"train_tokenizer.py",12759,0,"l",python,content
+579,548234,"train_tokenizer.py",12760,0,"",python,selection_keyboard
+580,548546,"train_tokenizer.py",12760,0,"o",python,content
+581,548546,"train_tokenizer.py",12761,0,"",python,selection_keyboard
+582,549100,"train_tokenizer.py",12760,1,"",python,content
+583,549281,"train_tokenizer.py",12760,0,"l",python,content
+584,549282,"train_tokenizer.py",12761,0,"",python,selection_keyboard
+585,549420,"train_tokenizer.py",12761,0,"o",python,content
+586,549420,"train_tokenizer.py",12762,0,"",python,selection_keyboard
+587,549513,"train_tokenizer.py",12762,0,"w",python,content
+588,549514,"train_tokenizer.py",12763,0,"",python,selection_keyboard
+589,549818,"train_tokenizer.py",12762,0,"",python,selection_command
+590,551538,"TERMINAL",0,0,"^Csrun: interrupt (one more within 1 sec to abort)\r\nsrun: StepId=23180.1 task 0: running\r\n",,terminal_output
+591,551689,"TERMINAL",0,0,"^Csrun: sending Ctrl-C to StepId=23180.1\r\nsrun: forcing job termination\r\nsrun: Job step aborted: Waiting up to 32 seconds for job step to finish.\r\n[2025-08-23T10:13:16.933] error: *** STEP 23180.1 ON hai008 CANCELLED AT 2025-08-23T10:13:16 DUE to SIGNAL Killed ***\r\n",,terminal_output
+592,551873,"TERMINAL",0,0,"^L",,terminal_output
+593,551953,"TERMINAL",0,0,"^P",,terminal_output
+594,552059,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/jafar[?2004h[franz.srambical@hai008.haicore.berlin:~/jafar] $ [franz.srambical@hai008.haicore.berlin:~/jafar] $ bash experiments/tokenizer_grain_checkpointing.sh ",,terminal_output
+595,552713,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+596,553713,"train_tokenizer.py",12717,0,"",python,selection_command
+597,556953,"TERMINAL",0,0,"Running on 1 devices.\r\n",,terminal_output
+598,560993,"TERMINAL",0,0,"Counting all components: ['decoder', 'encoder', 'vq']\r\nParameter counts:\r\n{'decoder': 16858736, 'encoder': 16858752, 'vq': 32768, 'total': 33750256}\r\n",,terminal_output
+599,594170,"TERMINAL",0,0,"Total memory size: 8.8 GB, Output size: 0.4 GB, Temp size: 8.3 GB, Argument size: 0.4 GB, Host temp size: 0.0 GB.\r\nFLOPs: 4.481e+13, Bytes: 3.947e+11 (367.6 GB), Intensity: 113.5 FLOPs/byte\r\nStarting training from step 0...\r\n",,terminal_output
+600,594513,"TERMINAL",0,0,"\r\nMemstats: After params initialized:\r\n\tUsing (GB) 0.44 / 59.39 (0.740865%) on cuda:0\r\n",,terminal_output
+601,595393,"TERMINAL",0,0,"Step 0, loss: 0.25430572032928467\r\n",,terminal_output
+602,604010,"TERMINAL",0,0,"^Csrun: interrupt (one more within 1 sec to abort)\r\nsrun: StepId=23180.2 task 0: running\r\n",,terminal_output
+603,604530,"TERMINAL",0,0,"^Csrun: sending Ctrl-C to StepId=23180.2\r\nsrun: forcing job termination\r\nsrun: Job step aborted: Waiting up to 32 seconds for job step to finish.\r\n[2025-08-23T10:14:09.728] error: *** STEP 23180.2 ON hai008 CANCELLED AT 2025-08-23T10:14:09 DUE to SIGNAL Killed ***\r\n",,terminal_output
+604,604913,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/jafar[?2004h[franz.srambical@hai008.haicore.berlin:~/jafar] $ ",,terminal_output
+605,674737,"train_tokenizer.py",12665,0,"",python,selection_command
+606,674996,"train_tokenizer.py",12600,0,"",python,selection_command
+607,675037,"train_tokenizer.py",12546,0,"",python,selection_command
+608,675049,"train_tokenizer.py",12483,0,"",python,selection_command
+609,675076,"train_tokenizer.py",12419,0,"",python,selection_command
+610,675357,"train_tokenizer.py",12347,0,"",python,selection_command
+611,675677,"train_tokenizer.py",12419,0,"",python,selection_command
+612,675837,"train_tokenizer.py",12483,0,"",python,selection_command
+613,676236,"train_tokenizer.py",12546,0,"",python,selection_command
+614,676597,"train_tokenizer.py",12600,0,"",python,selection_command
+615,677036,"train_tokenizer.py",12665,0,"",python,selection_command
+616,677196,"train_tokenizer.py",12717,0,"",python,selection_command
+617,677396,"train_tokenizer.py",12665,0,"",python,selection_command
+618,677636,"train_tokenizer.py",12600,0,"",python,selection_command
+619,677657,"train_tokenizer.py",12546,0,"",python,selection_command
+620,677716,"train_tokenizer.py",12483,0,"",python,selection_command
+621,677729,"train_tokenizer.py",12419,0,"",python,selection_command
+622,677764,"train_tokenizer.py",12347,0,"",python,selection_command
+623,677809,"train_tokenizer.py",12297,0,"",python,selection_command
+624,677929,"train_tokenizer.py",12347,0,"",python,selection_command
+625,678196,"train_tokenizer.py",12419,0,"",python,selection_command
+626,678197,"train_tokenizer.py",12483,0,"",python,selection_command
+627,678249,"train_tokenizer.py",12546,0,"",python,selection_command
+628,678276,"train_tokenizer.py",12600,0,"",python,selection_command
+629,679004,"train_tokenizer.py",12665,0,"",python,selection_command
+630,679116,"train_tokenizer.py",12717,0,"",python,selection_command
+631,705890,"train_tokenizer.py",12665,0,"",python,selection_command
+632,705993,"train_tokenizer.py",12600,0,"",python,selection_command
+633,706273,"train_tokenizer.py",12546,0,"",python,selection_command
+634,706274,"train_tokenizer.py",12483,0,"",python,selection_command
+635,706314,"train_tokenizer.py",12419,0,"",python,selection_command
+636,706353,"train_tokenizer.py",12347,0,"",python,selection_command
+637,706374,"train_tokenizer.py",12297,0,"",python,selection_command
+638,706401,"train_tokenizer.py",12257,0,"",python,selection_command
+639,706441,"train_tokenizer.py",12157,0,"",python,selection_command
+640,706473,"train_tokenizer.py",12124,0,"",python,selection_command
+641,706513,"train_tokenizer.py",12118,0,"",python,selection_command
+642,706673,"train_tokenizer.py",12124,0,"",python,selection_command
+643,706934,"train_tokenizer.py",12157,0,"",python,selection_command
+644,706945,"train_tokenizer.py",12257,0,"",python,selection_command
+645,706974,"train_tokenizer.py",12297,0,"",python,selection_command
+646,707033,"train_tokenizer.py",12347,0,"",python,selection_command
+647,707048,"train_tokenizer.py",12419,0,"",python,selection_command
+648,707073,"train_tokenizer.py",12483,0,"",python,selection_command
+649,707294,"train_tokenizer.py",12546,0,"",python,selection_command
+650,707473,"train_tokenizer.py",12600,0,"",python,selection_command
+651,707713,"train_tokenizer.py",12665,0,"",python,selection_command
+652,707761,"train_tokenizer.py",12717,0,"",python,selection_command
+653,707768,"train_tokenizer.py",12770,0,"",python,selection_command
+654,707974,"train_tokenizer.py",12792,0,"",python,selection_command
+655,708214,"train_tokenizer.py",12825,0,"",python,selection_command
+656,708233,"train_tokenizer.py",12859,0,"",python,selection_command
+657,708273,"train_tokenizer.py",12892,0,"",python,selection_command
+658,708414,"train_tokenizer.py",12933,0,"",python,selection_command
+659,708673,"train_tokenizer.py",12998,0,"",python,selection_command
+660,708694,"train_tokenizer.py",13033,0,"",python,selection_command
+661,708854,"train_tokenizer.py",13093,0,"",python,selection_command
+662,709233,"train_tokenizer.py",13033,0,"",python,selection_command
+663,721862,"train_tokenizer.py",13093,0,"",python,selection_command
+664,722081,"train_tokenizer.py",13141,0,"",python,selection_command
+665,722095,"train_tokenizer.py",13185,0,"",python,selection_command
+666,722134,"train_tokenizer.py",13231,0,"",python,selection_command
+667,722169,"train_tokenizer.py",13295,0,"",python,selection_command
+668,722193,"train_tokenizer.py",13343,0,"",python,selection_command
+669,722214,"train_tokenizer.py",13361,0,"",python,selection_command
+670,722257,"train_tokenizer.py",13366,0,"",python,selection_command
+671,722294,"train_tokenizer.py",13396,0,"",python,selection_command
+672,722334,"train_tokenizer.py",13421,0,"",python,selection_command
+673,722345,"train_tokenizer.py",13500,0,"",python,selection_command
+674,722394,"train_tokenizer.py",13531,0,"",python,selection_command
+675,722425,"train_tokenizer.py",13557,0,"",python,selection_command
+676,722454,"train_tokenizer.py",13599,0,"",python,selection_command
+677,722481,"train_tokenizer.py",13641,0,"",python,selection_command
+678,722614,"train_tokenizer.py",13599,0,"",python,selection_command
+679,722894,"train_tokenizer.py",13557,0,"",python,selection_command
+680,722894,"train_tokenizer.py",13531,0,"",python,selection_command
+681,722934,"train_tokenizer.py",13500,0,"",python,selection_command
+682,722953,"train_tokenizer.py",13421,0,"",python,selection_command
+683,722985,"train_tokenizer.py",13396,0,"",python,selection_command
+684,723025,"train_tokenizer.py",13366,0,"",python,selection_command
+685,723054,"train_tokenizer.py",13361,0,"",python,selection_command
+686,723094,"train_tokenizer.py",13343,0,"",python,selection_command
+687,723254,"train_tokenizer.py",13295,0,"",python,selection_command
+688,723494,"train_tokenizer.py",13231,0,"",python,selection_command
+689,723713,"train_tokenizer.py",13185,0,"",python,selection_command
+690,723753,"train_tokenizer.py",13141,0,"",python,selection_command
+691,723953,"train_tokenizer.py",13093,0,"",python,selection_command
+692,724121,"train_tokenizer.py",13033,0,"",python,selection_command
+693,933467,"train_tokenizer.py",13093,0,"",python,selection_command
+694,933627,"train_tokenizer.py",13141,0,"",python,selection_command
+695,936379,"train_tokenizer.py",13185,0,"",python,selection_command
+696,937659,"train_tokenizer.py",13141,0,"",python,selection_command
+697,999836,"TERMINAL",0,0,"\r[franz.srambical@hai008.haicore.berlin:~/jafar] $ ",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-83398cd5-40f9-4801-bd37-06200c9680561758650419267-2025_09_23-20.00.24.831/source.csv

@@ -0,0 +1,719 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,4,"experiments/dynamics_grain_tok_restore.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=2\n#SBATCH --ntasks-per-node=8\n#SBATCH --time=12:00:00\n#SBATCH --cpus-per-task=1\n#SBATCH --gres=gpu:8\n#SBATCH --mem=100GB\n\nsource .venv/bin/activate\n\ndata_dir=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/coinrun/array_records_10m""\nckpt_dir=""${PWD}/checkpoints/dynamics_openai_grain_tok_restore""\ntokenizer_ckpt_dir=""${PWD}/checkpoints/tokenizer_openai_grain_checkpointing""\n\nexport PYTHONUNBUFFERED=1\nsrun python train_dynamics.py \\n    --dyna_type 'causal' \\n    --batch_size 120 \\n    --image_height=64 \\n    --image_width=64 \\n    --patch_size=16 \\n    --tokenizer_checkpoint $tokenizer_ckpt_dir \\n    --log_checkpoint_interval 5 \\n    --ckpt_dir $ckpt_dir \\n    --num_steps 300000 \\n    --warmup_steps 10000 \\n    --seed 0 \\n    --init_lr=0.0000866 \\n    --max_lr=0.0000866 \\n    --data_dir $data_dir",shellscript,tab
+2,150,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"8:00:24 PM [info] Activating crowd-code\n8:00:24 PM [info] Recording started\n8:00:24 PM [info] Initializing git provider using file system watchers...\n8:00:24 PM [info] Git repository found\n8:00:24 PM [info] Git provider initialized successfully\n8:00:24 PM [info] Initial git state: [object Object]\n",Log,tab
+3,1038,"TERMINAL",0,0,"",,terminal_focus
+4,1181,"experiments/dynamics_grain_tok_restore.sh",0,0,"",shellscript,tab
+5,1485,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+6,5893,"TERMINAL",0,0,"",,terminal_command
+7,12655,"TERMINAL",0,0,"",,terminal_command
+8,115088,"experiments/dynamics_grain_tok_restore.sh",0,0,"Switched from branch 'new-nnx-optimizer-api' to 'main'",shellscript,git_branch_checkout
+9,120844,"TERMINAL",0,0,"",,terminal_command
+10,260094,"experiments/dynamics_grain_tok_restore.sh",0,0,"Switched from branch 'main' to 'split-requirements'",shellscript,git_branch_checkout
+11,263190,"requirements.txt",0,0,"# core requirements\ndm_pix>=0.4.3\neinops>=0.8.0\nflax>=0.10.7\njax[cuda12]>=0.6.2\noptax>=0.2.3\ntyro>=0.8.5\nwandb>=0.17.4\ngrain>=0.2.10\narray-record>=0.7.2\n\n# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0\nprocgen>=0.10.7\ntqdm>=4.67.1",pip-requirements,tab
+12,265314,"requirements.txt",0,0,"",pip-requirements,selection_command
+13,266099,"requirements.txt",20,0,"",pip-requirements,selection_command
+14,266176,"requirements.txt",34,0,"",pip-requirements,selection_command
+15,266432,"requirements.txt",48,0,"",pip-requirements,selection_command
+16,266459,"requirements.txt",61,0,"",pip-requirements,selection_command
+17,266508,"requirements.txt",80,0,"",pip-requirements,selection_command
+18,266526,"requirements.txt",93,0,"",pip-requirements,selection_command
+19,266553,"requirements.txt",105,0,"",pip-requirements,selection_command
+20,266588,"requirements.txt",119,0,"",pip-requirements,selection_command
+21,266621,"requirements.txt",133,0,"",pip-requirements,selection_command
+22,266655,"requirements.txt",153,0,"",pip-requirements,selection_command
+23,266688,"requirements.txt",154,0,"",pip-requirements,selection_command
+24,266721,"requirements.txt",171,0,"",pip-requirements,selection_command
+25,266754,"requirements.txt",192,0,"",pip-requirements,selection_command
+26,266961,"requirements.txt",171,0,"",pip-requirements,selection_command
+27,267362,"requirements.txt",192,0,"",pip-requirements,selection_command
+28,267447,"requirements.txt",211,0,"",pip-requirements,selection_command
+29,267558,"requirements.txt",240,0,"",pip-requirements,selection_command
+30,267820,"requirements.txt",258,0,"",pip-requirements,selection_command
+31,267842,"requirements.txt",274,0,"",pip-requirements,selection_command
+32,268436,"requirements.txt",258,0,"",pip-requirements,selection_command
+33,268490,"requirements.txt",240,0,"",pip-requirements,selection_command
+34,268748,"requirements.txt",211,0,"",pip-requirements,selection_command
+35,268781,"requirements.txt",192,0,"",pip-requirements,selection_command
+36,268816,"requirements.txt",171,0,"",pip-requirements,selection_command
+37,268846,"requirements.txt",154,0,"",pip-requirements,selection_command
+38,268878,"requirements.txt",153,0,"",pip-requirements,selection_command
+39,268911,"requirements.txt",133,0,"",pip-requirements,selection_command
+40,269458,"requirements.txt",119,0,"",pip-requirements,selection_command
+41,269509,"requirements.txt",105,0,"",pip-requirements,selection_command
+42,269632,"requirements.txt",93,0,"",pip-requirements,selection_command
+43,269774,"requirements.txt",80,0,"",pip-requirements,selection_command
+44,269907,"requirements.txt",61,0,"",pip-requirements,selection_command
+45,310804,"requirements-coinrun.py",0,0,"",python,tab
+46,321869,"requirements-minecraft.py",0,0,"",python,tab
+47,323139,"requirements-coinrun.py",0,0,"",python,tab
+48,323505,"requirements.txt",0,0,"",pip-requirements,tab
+49,323979,"requirements-coinrun.py",0,0,"",python,tab
+50,324639,"requirements-minecraft.py",0,0,"",python,tab
+51,325739,"requirements.txt",0,0,"",pip-requirements,tab
+52,326606,"requirements.txt",166,0,"",pip-requirements,selection_mouse
+53,328385,"requirements.txt",154,16,"# input pipeline",pip-requirements,selection_command
+54,330144,"requirements.txt",154,37,"# input pipeline\nffmpeg-python==0.2.0",pip-requirements,selection_command
+55,330214,"requirements.txt",154,56,"# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9",pip-requirements,selection_command
+56,330361,"requirements.txt",154,85,"# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3",pip-requirements,selection_command
+57,330480,"requirements.txt",154,103,"# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0",pip-requirements,selection_command
+58,330731,"requirements.txt",154,119,"# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0\nprocgen>=0.10.7",pip-requirements,selection_command
+59,330762,"requirements.txt",154,132,"# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0\nprocgen>=0.10.7\ntqdm>=4.67.1",pip-requirements,selection_command
+60,333536,"requirements.txt",153,133,"",pip-requirements,content
+61,335660,"requirements-coinrun.py",0,0,"",python,tab
+62,336502,"requirements-coinrun.py",0,0,"\n# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0\nprocgen>=0.10.7\ntqdm>=4.67.1",python,content
+63,336507,"requirements-coinrun.py",1,0,"",python,selection_command
+64,337098,"requirements-coinrun.py",0,0,"",python,selection_command
+65,337277,"requirements-coinrun.py",0,1,"",python,content
+66,338610,"requirements.txt",0,0,"",pip-requirements,tab
+67,339582,"requirements.txt",152,1,"",pip-requirements,content
+68,339589,"requirements.txt",133,0,"",pip-requirements,selection_command
+69,341663,"requirements.txt",0,152,"# core requirements\ndm_pix>=0.4.3\neinops>=0.8.0\nflax>=0.10.7\njax[cuda12]>=0.6.2\noptax>=0.2.3\ntyro>=0.8.5\nwandb>=0.17.4\ngrain>=0.2.10\narray-record>=0.7.2",pip-requirements,selection_command
+70,343016,"requirements.txt",133,0,"",pip-requirements,selection_command
+71,360994,"requirements-coinrun.py",0,0,"",python,tab
+72,364126,"requirements.txt",0,0,"",pip-requirements,tab
+73,386818,"requirements.txt",0,0,"",pip-requirements,selection_command
+74,403354,"requirements-minecraft.txt",0,0,"",pip-requirements,tab
+75,405540,"requirements-coinrun.txt",0,0,"# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0\nprocgen>=0.10.7\ntqdm>=4.67.1",pip-requirements,tab
+76,406758,"requirements-coinrun.txt",132,0,"",pip-requirements,selection_mouse
+77,406758,"requirements-coinrun.txt",131,0,"",pip-requirements,selection_command
+78,408883,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+79,408941,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+80,409191,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+81,409221,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+82,409256,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+83,409287,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_command
+84,409590,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+85,409695,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+86,409946,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+87,409982,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+88,410005,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+89,410038,"requirements-coinrun.txt",131,0,"",pip-requirements,selection_command
+90,410247,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+91,410356,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+92,410606,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+93,410631,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+94,410663,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+95,410696,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_command
+96,411016,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+97,411098,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+98,411348,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+99,411377,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+100,411411,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+101,411444,"requirements-coinrun.txt",131,0,"",pip-requirements,selection_command
+102,411683,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+103,411791,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+104,412046,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+105,412076,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+106,412106,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+107,412140,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_command
+108,412315,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+109,412408,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+110,412667,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+111,412694,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+112,412720,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+113,412755,"requirements-coinrun.txt",131,0,"",pip-requirements,selection_command
+114,413054,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+115,413124,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+116,413238,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+117,413485,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+118,413514,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+119,413547,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_command
+120,418464,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+121,418571,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+122,418818,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+123,418847,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+124,418882,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+125,418912,"requirements-coinrun.txt",131,0,"",pip-requirements,selection_command
+126,419120,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+127,419221,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+128,419340,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+129,419603,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+130,419632,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+131,419657,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_command
+132,423268,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+133,423359,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+134,423604,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+135,423635,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+136,423670,"requirements-coinrun.txt",115,0,"",pip-requirements,selection_command
+137,423857,"requirements-coinrun.txt",97,0,"",pip-requirements,selection_command
+138,423943,"requirements-coinrun.txt",68,0,"",pip-requirements,selection_command
+139,424073,"requirements-coinrun.txt",49,0,"",pip-requirements,selection_command
+140,424572,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+141,424734,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_command
+142,426390,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+143,430172,"requirements-coinrun.txt",17,21,"",pip-requirements,content
+144,431512,"requirements-minecraft.txt",0,0,"",pip-requirements,tab
+145,431784,"requirements-minecraft.txt",0,0,"\nffmpeg-python==0.2.0",pip-requirements,content
+146,431791,"requirements-minecraft.txt",1,0,"",pip-requirements,selection_command
+147,433989,"requirements-coinrun.txt",0,0,"",pip-requirements,tab
+148,434440,"requirements-coinrun.txt",0,0,"",pip-requirements,selection_command
+149,435050,"requirements-coinrun.txt",2,0,"",pip-requirements,selection_command
+150,439128,"requirements-coinrun.txt",2,14,"",pip-requirements,content
+151,439611,"requirements-coinrun.txt",2,0,"c",pip-requirements,content
+152,439611,"requirements-coinrun.txt",3,0,"",pip-requirements,selection_keyboard
+153,439681,"requirements-coinrun.txt",3,0,"o",pip-requirements,content
+154,439682,"requirements-coinrun.txt",4,0,"",pip-requirements,selection_keyboard
+155,439712,"requirements-coinrun.txt",4,0,"i",pip-requirements,content
+156,439712,"requirements-coinrun.txt",5,0,"",pip-requirements,selection_keyboard
+157,439781,"requirements-coinrun.txt",5,0,"n",pip-requirements,content
+158,439781,"requirements-coinrun.txt",6,0,"",pip-requirements,selection_keyboard
+159,439942,"requirements-coinrun.txt",6,0,"r",pip-requirements,content
+160,439943,"requirements-coinrun.txt",7,0,"",pip-requirements,selection_keyboard
+161,440041,"requirements-coinrun.txt",7,0,"u",pip-requirements,content
+162,440041,"requirements-coinrun.txt",8,0,"",pip-requirements,selection_keyboard
+163,440195,"requirements-coinrun.txt",8,0,"n",pip-requirements,content
+164,440196,"requirements-coinrun.txt",9,0,"",pip-requirements,selection_keyboard
+165,440223,"requirements-coinrun.txt",9,0," ",pip-requirements,content
+166,440224,"requirements-coinrun.txt",10,0,"",pip-requirements,selection_keyboard
+167,440713,"requirements-coinrun.txt",10,0,"p",pip-requirements,content
+168,440713,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_keyboard
+169,440812,"requirements-coinrun.txt",11,0,"r",pip-requirements,content
+170,440812,"requirements-coinrun.txt",12,0,"",pip-requirements,selection_keyboard
+171,440890,"requirements-coinrun.txt",12,0,"e",pip-requirements,content
+172,440890,"requirements-coinrun.txt",13,0,"",pip-requirements,selection_keyboard
+173,441052,"requirements-coinrun.txt",13,0,"p",pip-requirements,content
+174,441053,"requirements-coinrun.txt",14,0,"",pip-requirements,selection_keyboard
+175,441137,"requirements-coinrun.txt",14,0,"r",pip-requirements,content
+176,441138,"requirements-coinrun.txt",15,0,"",pip-requirements,selection_keyboard
+177,441265,"requirements-coinrun.txt",15,0,"o",pip-requirements,content
+178,441265,"requirements-coinrun.txt",16,0,"",pip-requirements,selection_keyboard
+179,441730,"requirements-coinrun.txt",16,0,"x",pip-requirements,content
+180,441730,"requirements-coinrun.txt",17,0,"",pip-requirements,selection_keyboard
+181,441836,"requirements-coinrun.txt",17,0,"e",pip-requirements,content
+182,441836,"requirements-coinrun.txt",18,0,"",pip-requirements,selection_keyboard
+183,442007,"requirements-coinrun.txt",18,0,"s",pip-requirements,content
+184,442007,"requirements-coinrun.txt",19,0,"",pip-requirements,selection_keyboard
+185,442322,"requirements-coinrun.txt",18,1,"",pip-requirements,content
+186,442475,"requirements-coinrun.txt",17,1,"",pip-requirements,content
+187,442640,"requirements-coinrun.txt",16,1,"",pip-requirements,content
+188,442659,"requirements-coinrun.txt",16,0,"c",pip-requirements,content
+189,442659,"requirements-coinrun.txt",17,0,"",pip-requirements,selection_keyboard
+190,442856,"requirements-coinrun.txt",17,0,"e",pip-requirements,content
+191,442856,"requirements-coinrun.txt",18,0,"",pip-requirements,selection_keyboard
+192,442974,"requirements-coinrun.txt",18,0,"s",pip-requirements,content
+193,442975,"requirements-coinrun.txt",19,0,"",pip-requirements,selection_keyboard
+194,443039,"requirements-coinrun.txt",19,0,"s",pip-requirements,content
+195,443039,"requirements-coinrun.txt",20,0,"",pip-requirements,selection_keyboard
+196,443094,"requirements-coinrun.txt",20,0,"i",pip-requirements,content
+197,443094,"requirements-coinrun.txt",21,0,"",pip-requirements,selection_keyboard
+198,443148,"requirements-coinrun.txt",21,0,"n",pip-requirements,content
+199,443148,"requirements-coinrun.txt",22,0,"",pip-requirements,selection_keyboard
+200,443254,"requirements-coinrun.txt",22,0,"g",pip-requirements,content
+201,443254,"requirements-coinrun.txt",23,0,"",pip-requirements,selection_keyboard
+202,444695,"requirements-coinrun.txt",22,1,"",pip-requirements,content
+203,444938,"requirements-coinrun.txt",21,1,"",pip-requirements,content
+204,444966,"requirements-coinrun.txt",20,1,"",pip-requirements,content
+205,445000,"requirements-coinrun.txt",19,1,"",pip-requirements,content
+206,445033,"requirements-coinrun.txt",18,1,"",pip-requirements,content
+207,445062,"requirements-coinrun.txt",17,1,"",pip-requirements,content
+208,445099,"requirements-coinrun.txt",16,1,"",pip-requirements,content
+209,445129,"requirements-coinrun.txt",15,1,"",pip-requirements,content
+210,445167,"requirements-coinrun.txt",14,1,"",pip-requirements,content
+211,445287,"requirements-coinrun.txt",13,1,"",pip-requirements,content
+212,445448,"requirements-coinrun.txt",12,1,"",pip-requirements,content
+213,445602,"requirements-coinrun.txt",11,1,"",pip-requirements,content
+214,445736,"requirements-coinrun.txt",10,1,"",pip-requirements,content
+215,446003,"requirements-coinrun.txt",10,0,"a",pip-requirements,content
+216,446004,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_keyboard
+217,446016,"requirements-coinrun.txt",11,0,"t",pip-requirements,content
+218,446016,"requirements-coinrun.txt",12,0,"",pip-requirements,selection_keyboard
+219,446178,"requirements-coinrun.txt",12,0,"a",pip-requirements,content
+220,446178,"requirements-coinrun.txt",13,0,"",pip-requirements,selection_keyboard
+221,446179,"requirements-coinrun.txt",13,0," ",pip-requirements,content
+222,446180,"requirements-coinrun.txt",14,0,"",pip-requirements,selection_keyboard
+223,446337,"requirements-coinrun.txt",14,0,"g",pip-requirements,content
+224,446337,"requirements-coinrun.txt",15,0,"",pip-requirements,selection_keyboard
+225,446542,"requirements-coinrun.txt",14,1,"",pip-requirements,content
+226,446707,"requirements-coinrun.txt",13,1,"",pip-requirements,content
+227,446832,"requirements-coinrun.txt",12,1,"",pip-requirements,content
+228,446983,"requirements-coinrun.txt",11,1,"",pip-requirements,content
+229,447125,"requirements-coinrun.txt",10,1,"",pip-requirements,content
+230,447273,"requirements-coinrun.txt",10,0,"d",pip-requirements,content
+231,447274,"requirements-coinrun.txt",11,0,"",pip-requirements,selection_keyboard
+232,447366,"requirements-coinrun.txt",11,0,"a",pip-requirements,content
+233,447366,"requirements-coinrun.txt",12,0,"",pip-requirements,selection_keyboard
+234,447371,"requirements-coinrun.txt",12,0,"t",pip-requirements,content
+235,447372,"requirements-coinrun.txt",13,0,"",pip-requirements,selection_keyboard
+236,447589,"requirements-coinrun.txt",13,0,"a",pip-requirements,content
+237,447589,"requirements-coinrun.txt",14,0,"",pip-requirements,selection_keyboard
+238,447591,"requirements-coinrun.txt",14,0," ",pip-requirements,content
+239,447591,"requirements-coinrun.txt",15,0,"",pip-requirements,selection_keyboard
+240,447819,"requirements-coinrun.txt",15,0,"g",pip-requirements,content
+241,447819,"requirements-coinrun.txt",16,0,"",pip-requirements,selection_keyboard
+242,447946,"requirements-coinrun.txt",16,0,"e",pip-requirements,content
+243,447946,"requirements-coinrun.txt",17,0,"",pip-requirements,selection_keyboard
+244,448080,"requirements-coinrun.txt",17,0,"n",pip-requirements,content
+245,448081,"requirements-coinrun.txt",18,0,"",pip-requirements,selection_keyboard
+246,448143,"requirements-coinrun.txt",18,0,"e",pip-requirements,content
+247,448143,"requirements-coinrun.txt",19,0,"",pip-requirements,selection_keyboard
+248,448165,"requirements-coinrun.txt",19,0,"r",pip-requirements,content
+249,448165,"requirements-coinrun.txt",20,0,"",pip-requirements,selection_keyboard
+250,448337,"requirements-coinrun.txt",20,0,"a",pip-requirements,content
+251,448337,"requirements-coinrun.txt",21,0,"",pip-requirements,selection_keyboard
+252,448351,"requirements-coinrun.txt",21,0,"t",pip-requirements,content
+253,448351,"requirements-coinrun.txt",22,0,"",pip-requirements,selection_keyboard
+254,448364,"requirements-coinrun.txt",22,0,"i",pip-requirements,content
+255,448364,"requirements-coinrun.txt",23,0,"",pip-requirements,selection_keyboard
+256,448434,"requirements-coinrun.txt",23,0,"o",pip-requirements,content
+257,448434,"requirements-coinrun.txt",24,0,"",pip-requirements,selection_keyboard
+258,448476,"requirements-coinrun.txt",24,0,"n",pip-requirements,content
+259,448477,"requirements-coinrun.txt",25,0,"",pip-requirements,selection_keyboard
+260,448782,"requirements-coinrun.txt",24,0,"",pip-requirements,selection_command
+261,449789,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+262,453098,"requirements-coinrun.txt",43,1,"",pip-requirements,content
+263,453402,"requirements-coinrun.txt",43,0,"j",pip-requirements,content
+264,453403,"requirements-coinrun.txt",44,0,"",pip-requirements,selection_keyboard
+265,453922,"requirements-coinrun.txt",43,1,"",pip-requirements,content
+266,454546,"requirements-coinrun.txt",42,0,"",pip-requirements,selection_command
+267,454714,"requirements-coinrun.txt",43,0,"9",pip-requirements,content
+268,454718,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+269,455413,"requirements-coinrun.txt",43,1,"",pip-requirements,content
+270,456358,"requirements-coinrun.txt",42,0,"",pip-requirements,selection_command
+271,456515,"requirements-coinrun.txt",43,0,"9",pip-requirements,content
+272,456520,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+273,456894,"requirements-coinrun.txt",26,18,"hf-transfer==0.1.9",pip-requirements,selection_command
+274,457798,"requirements-coinrun.txt",26,47,"hf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3",pip-requirements,selection_command
+275,457888,"requirements-coinrun.txt",26,65,"hf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0",pip-requirements,selection_command
+276,458061,"requirements-coinrun.txt",26,81,"hf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0\nprocgen>=0.10.7",pip-requirements,selection_command
+277,458388,"requirements-coinrun.txt",26,65,"hf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0",pip-requirements,selection_command
+278,458450,"requirements-coinrun.txt",26,47,"hf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3",pip-requirements,selection_command
+279,464120,"requirements-coinrun.txt",26,48,"",pip-requirements,content
+280,465094,"requirements-minecraft.txt",0,0,"",pip-requirements,tab
+281,465814,"requirements-minecraft.txt",0,0,"",pip-requirements,selection_command
+282,466335,"requirements-minecraft.txt",0,0,"\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3",pip-requirements,content
+283,466337,"requirements-minecraft.txt",1,0,"",pip-requirements,selection_command
+284,466813,"requirements-minecraft.txt",0,0,"",pip-requirements,selection_command
+285,467030,"requirements-minecraft.txt",0,1,"",pip-requirements,content
+286,467886,"requirements-coinrun.txt",0,0,"",pip-requirements,tab
+287,470019,"requirements-coinrun.txt",44,0,"",pip-requirements,selection_command
+288,470112,"requirements-coinrun.txt",60,0,"",pip-requirements,selection_command
+289,470588,"requirements-coinrun.txt",44,0,"",pip-requirements,selection_command
+290,470680,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+291,470851,"requirements-coinrun.txt",0,0,"",pip-requirements,selection_command
+292,471380,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+293,471454,"requirements-coinrun.txt",44,0,"",pip-requirements,selection_command
+294,471725,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+295,471978,"requirements-coinrun.txt",26,18,"",pip-requirements,content
+296,472980,"requirements.txt",0,0,"",pip-requirements,tab
+297,473979,"requirements.txt",20,0,"",pip-requirements,selection_command
+298,474048,"requirements.txt",34,0,"",pip-requirements,selection_command
+299,474155,"requirements.txt",48,0,"",pip-requirements,selection_command
+300,474406,"requirements.txt",61,0,"",pip-requirements,selection_command
+301,474435,"requirements.txt",80,0,"",pip-requirements,selection_command
+302,474468,"requirements.txt",93,0,"",pip-requirements,selection_command
+303,474500,"requirements.txt",105,0,"",pip-requirements,selection_command
+304,474536,"requirements.txt",119,0,"",pip-requirements,selection_command
+305,474568,"requirements.txt",133,0,"",pip-requirements,selection_command
+306,474837,"requirements.txt",152,0,"\npre-commit>=4.2.0",pip-requirements,content
+307,474862,"requirements.txt",153,0,"",pip-requirements,selection_command
+308,477802,"requirements-coinrun.txt",0,0,"",pip-requirements,tab
+309,478397,"requirements-coinrun.txt",42,0,"",pip-requirements,selection_command
+310,534105,"requirements-coinrun.txt",41,13,"",pip-requirements,content
+311,534111,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+312,535229,"requirements.txt",0,0,"",pip-requirements,tab
+313,536066,"requirements-coinrun.txt",0,0,"",pip-requirements,tab
+314,536753,"requirements-minecraft.txt",0,0,"",pip-requirements,tab
+315,536970,"requirements-minecraft.txt",18,0,"\ntqdm>=4.67.1",pip-requirements,content
+316,536972,"requirements-minecraft.txt",19,0,"",pip-requirements,selection_command
+317,538021,"requirements-minecraft.txt",19,13,"",pip-requirements,content
+318,538184,"requirements-minecraft.txt",48,0,"",pip-requirements,selection_command
+319,538409,"requirements-minecraft.txt",68,0,"\ntqdm>=4.67.1",pip-requirements,content
+320,538415,"requirements-minecraft.txt",69,0,"",pip-requirements,selection_command
+321,544520,"requirements-coinrun.txt",0,0,"",pip-requirements,tab
+322,545165,"requirements-coinrun.txt",0,0,"",pip-requirements,selection_command
+323,545552,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+324,546213,"requirements-coinrun.txt",0,0,"",pip-requirements,selection_command
+325,546428,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+326,549027,"requirements-coinrun.txt",0,0,"",pip-requirements,selection_command
+327,549246,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+328,569618,"requirements-coinrun.txt",26,0,"tqdm>=4.67.1\n",pip-requirements,content
+329,569625,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+330,570584,"requirements-coinrun.txt",39,15,"",pip-requirements,content
+331,570585,"requirements-coinrun.txt",39,0,"",pip-requirements,selection_command
+332,575666,"requirements-coinrun.txt",39,0,"\n",pip-requirements,content
+333,575673,"requirements-coinrun.txt",39,0,"",pip-requirements,selection_command
+334,576102,"requirements-coinrun.txt",40,0,"",pip-requirements,selection_command
+335,576424,"requirements-coinrun.txt",40,0,"\n",pip-requirements,content
+336,576433,"requirements-coinrun.txt",40,0,"",pip-requirements,selection_command
+337,576727,"requirements-coinrun.txt",41,0,"",pip-requirements,selection_command
+338,576978,"requirements-coinrun.txt",41,0,"\n",pip-requirements,content
+339,576980,"requirements-coinrun.txt",41,0,"",pip-requirements,selection_command
+340,577707,"requirements-coinrun.txt",42,0,"\n",pip-requirements,content
+341,577709,"requirements-coinrun.txt",42,0,"",pip-requirements,selection_command
+342,577971,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+343,578130,"requirements-coinrun.txt",43,0,"\n",pip-requirements,content
+344,578133,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+345,578253,"requirements-coinrun.txt",44,0,"",pip-requirements,selection_command
+346,581767,"requirements-coinrun.txt",44,0,"\n",pip-requirements,content
+347,581772,"requirements-coinrun.txt",44,0,"",pip-requirements,selection_command
+348,582354,"requirements-coinrun.txt",45,0,"",pip-requirements,selection_command
+349,583201,"requirements-coinrun.txt",45,0,"\n",pip-requirements,content
+350,583205,"requirements-coinrun.txt",45,0,"",pip-requirements,selection_command
+351,583587,"requirements-coinrun.txt",44,1,"",pip-requirements,content
+352,583827,"requirements-coinrun.txt",44,0,"\n",pip-requirements,content
+353,583834,"requirements-coinrun.txt",44,0,"",pip-requirements,selection_command
+354,583981,"requirements-coinrun.txt",43,1,"",pip-requirements,content
+355,584123,"requirements-coinrun.txt",43,0,"\n",pip-requirements,content
+356,584129,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+357,584286,"requirements-coinrun.txt",42,1,"",pip-requirements,content
+358,584406,"requirements-coinrun.txt",42,0,"\n",pip-requirements,content
+359,584422,"requirements-coinrun.txt",42,0,"",pip-requirements,selection_command
+360,584569,"requirements-coinrun.txt",41,1,"",pip-requirements,content
+361,584679,"requirements-coinrun.txt",41,0,"\n",pip-requirements,content
+362,584680,"requirements-coinrun.txt",41,0,"",pip-requirements,selection_command
+363,584835,"requirements-coinrun.txt",40,1,"",pip-requirements,content
+364,584973,"requirements-coinrun.txt",40,0,"\n",pip-requirements,content
+365,584974,"requirements-coinrun.txt",40,0,"",pip-requirements,selection_command
+366,585114,"requirements-coinrun.txt",39,1,"",pip-requirements,content
+367,585447,"requirements-coinrun.txt",39,0,"\n",pip-requirements,content
+368,585455,"requirements-coinrun.txt",39,0,"",pip-requirements,selection_command
+369,585808,"requirements-coinrun.txt",26,13,"",pip-requirements,content
+370,586273,"requirements-coinrun.txt",26,0,"\ntqdm>=4.67.1",pip-requirements,content
+371,586275,"requirements-coinrun.txt",27,0,"",pip-requirements,selection_command
+372,586722,"requirements-coinrun.txt",26,0,"e-commit>=4.2.0\npr",pip-requirements,content
+373,586724,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+374,588324,"requirements-coinrun.txt",26,0,"hf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\n",pip-requirements,content
+375,588326,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+376,589327,"requirements-coinrun.txt",62,0,"",pip-requirements,selection_command
+377,589420,"requirements-coinrun.txt",88,0,"",pip-requirements,selection_command
+378,589542,"requirements-coinrun.txt",91,0,"",pip-requirements,selection_command
+379,589682,"requirements-coinrun.txt",104,0,"",pip-requirements,selection_command
+380,590519,"requirements-coinrun.txt",91,0,"",pip-requirements,selection_command
+381,590620,"requirements-coinrun.txt",88,0,"",pip-requirements,selection_command
+382,590787,"requirements-coinrun.txt",62,0,"",pip-requirements,selection_command
+383,590924,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+384,591088,"requirements-coinrun.txt",17,0,"",pip-requirements,selection_command
+385,591432,"requirements-coinrun.txt",43,0,"",pip-requirements,selection_command
+386,591727,"requirements-coinrun.txt",26,19,"",pip-requirements,content
+387,592036,"requirements-coinrun.txt",26,29,"",pip-requirements,content
+388,592356,"requirements-coinrun.txt",26,16,"",pip-requirements,content
+389,592651,"requirements-coinrun.txt",26,3,"",pip-requirements,content
+390,593368,"requirements-coinrun.txt",26,13,"",pip-requirements,content
+391,595612,"requirements-coinrun.txt",15,0,"",pip-requirements,selection_command
+392,599311,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+393,599479,"requirements-coinrun.txt",27,0,"",pip-requirements,selection_command
+394,599741,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+395,599761,"requirements-coinrun.txt",29,0,"",pip-requirements,selection_command
+396,599792,"requirements-coinrun.txt",30,0,"",pip-requirements,selection_command
+397,599826,"requirements-coinrun.txt",31,0,"",pip-requirements,selection_command
+398,599862,"requirements-coinrun.txt",32,0,"",pip-requirements,selection_command
+399,601894,"requirements-coinrun.txt",31,1,"\n",pip-requirements,selection_command
+400,602176,"requirements-coinrun.txt",31,0,"",pip-requirements,selection_command
+401,603026,"requirements-coinrun.txt",32,0,"",pip-requirements,selection_command
+402,604424,"requirements-coinrun.txt",31,1,"\n",pip-requirements,selection_command
+403,604490,"requirements-coinrun.txt",30,2,"\n\n",pip-requirements,selection_command
+404,604630,"requirements-coinrun.txt",29,3,"\n\n\n",pip-requirements,selection_command
+405,604759,"requirements-coinrun.txt",28,4,"\n\n\n\n",pip-requirements,selection_command
+406,604902,"requirements-coinrun.txt",27,5,"\n\n\n\n\n",pip-requirements,selection_command
+407,605029,"requirements-coinrun.txt",26,6,"\n\n\n\n\n\n",pip-requirements,selection_command
+408,605162,"requirements-coinrun.txt",0,32,"# coinrun data generation\n\n\n\n\n\n\n",pip-requirements,selection_command
+409,605301,"requirements-coinrun.txt",0,32,"",pip-requirements,content
+410,605906,"requirements-coinrun.txt",0,0,"# coinrun data generation\n\n\n\n\n\n\n",pip-requirements,content
+411,605908,"requirements-coinrun.txt",0,0,"",pip-requirements,selection_command
+412,606249,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+413,606397,"requirements-coinrun.txt",27,0,"",pip-requirements,selection_command
+414,606649,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+415,606678,"requirements-coinrun.txt",29,0,"",pip-requirements,selection_command
+416,606912,"requirements-coinrun.txt",28,0,"",pip-requirements,selection_command
+417,606972,"requirements-coinrun.txt",27,0,"",pip-requirements,selection_command
+418,607139,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+419,608231,"requirements-coinrun.txt",26,6,"",pip-requirements,content
+420,613996,"requirements.txt",0,0,"",pip-requirements,tab
+421,613997,"requirements.txt",153,0,"",pip-requirements,selection_command
+422,616887,"requirements.txt",0,0,"# core requirements\ndm_pix>=0.4.3\neinops>=0.8.0\nflax>=0.10.7\njax[cuda12]>=0.6.2\noptax>=0.2.3\ntyro>=0.8.5\nwandb>=0.17.4\ngrain>=0.2.10\narray-record>=0.7.2\n\n# input pipeline\nffmpeg-python==0.2.0\nhf-transfer==0.1.9\nhuggingface-hub[cli]>=0.34.3\npre-commit>=4.2.0\nprocgen>=0.10.7\ntqdm>=4.67.1",pip-requirements,tab
+423,616888,"requirements.txt",265,0,"",pip-requirements,selection_mouse
+424,617004,"requirements.txt",258,7,"procgen",pip-requirements,selection_mouse
+425,617140,"requirements.txt",258,16,"procgen>=0.10.7\n",pip-requirements,selection_mouse
+426,619281,"requirements.txt",0,0,"",pip-requirements,tab
+427,619281,"requirements-coinrun.txt",0,0,"",pip-requirements,tab
+428,620132,"requirements-coinrun.txt",26,0,"procgen>=0.10.7\n",pip-requirements,content
+429,620971,"requirements-coinrun.txt",41,1,"",pip-requirements,content
+430,620978,"requirements-coinrun.txt",26,0,"",pip-requirements,selection_command
+431,630211,"requirements.txt",0,0,"",pip-requirements,tab
+432,1899365,"requirements.txt",170,0,"",pip-requirements,selection_mouse
+433,1899365,"requirements.txt",169,0,"",pip-requirements,selection_command
+434,1900780,"requirements-coinrun.txt",0,0,"",pip-requirements,tab
+435,1937923,"input_pipeline/coinrun",0,0,"",plaintext,tab
+436,1949072,"requirements-coinrun.txt",0,0,"",pip-requirements,tab
+437,2770779,"requirements-coinrun.txt",41,0,"",pip-requirements,selection_mouse
+438,2770784,"requirements-coinrun.txt",40,0,"",pip-requirements,selection_command
+439,2771647,"requirements-coinrun.txt",41,0,"",pip-requirements,selection_mouse
+440,2771647,"requirements-coinrun.txt",40,0,"",pip-requirements,selection_command
+441,2771688,"requirements-coinrun.txt",40,1,"7",pip-requirements,selection_mouse
+442,2771689,"requirements-coinrun.txt",41,0,"",pip-requirements,selection_command
+443,2771998,"requirements-coinrun.txt",40,0,"",pip-requirements,selection_command
+444,2862196,"README.md",0,0,"<h1 align=""center"">🧞‍♀️ Jasmine: A simple, performant and scalable JAX-based world modeling codebase 🧞‍♀️</h1>\n\n<p align=""center"">\n    <a href= ""https://github.com/FLAIROx/jafar/blob/main/LICENSE"">\n        <img src=""https://img.shields.io/badge/license-Apache2.0-blue.svg"" /></a>\n    <a href= ""https://github.com/psf/black"">\n        <img src=""https://img.shields.io/badge/code%20style-black-000000.svg"" /></a>\n</p>\n\nJasmine is a production-ready JAX-based world modeling codebase. It currently implements the high-level architecture of [Genie: Generative Interactive Environments](https://arxiv.org/abs/2402.15391) (Bruce et al., 2024) with [MaskGIT](https://arxiv.org/abs/2202.04200) (Chang et al., 2022), as well as an autoregressive (causal) baseline. A diffusion baseline is coming soon.\n\nJasmine scales from single hosts to hundreds of xPUs thanks to XLA and strives to be an easily hackable, batteries-included foundation for world modeling research.\n\n<h2 name=""overview"" id=""overview"">Overview</h2>\n\n- Asynchronous & distributed checkpointing thanks to [orbax.checkpoint](https://github.com/google/orbax)\n    - Jasmine also supports mixing and matching hardware topologies (e.g. train on four nodes, load the checkpoint on a single node)\n- Optimized dataloading thanks to [Grain](https://github.com/google/grain)\n    - Dataloading scales with the number of processes (i.e. nodes/xPUs)\n- Checkpointing of model weights, optimizer and dataloader states\n- Full reproducibility with **identical** training curves (thanks to seeded dataloading and training, and [JAX' approach to pseudo random numbers](https://docs.jax.dev/en/latest/random-numbers.html))\n- Automatic checkpoint deletion/retention according to specified retention policy thanks to `orbax.checkpoint.CheckpointManager`\n- Mixed precision training using `bfloat16`\n    - `int8` training is on the roadmap via [aqt](https://github.com/google/aqt)\n- FlashAttention thanks to [cuDNN SDPA](https://github.com/jax-ml/jax/blob/a155c5a9997924170e0067d552351a9833c12c11/jax/_src/cudnn/fused_attention_stablehlo.py#L842)\n- Frame-level KV cache resets for accelerated spatiotemporal attention in causal baseline (still in PR)\n- Activation checkpointing (even onto host memory if desired)\n- DDP (changing to FSDP requires changing **a single line of code**)\n- WSD learning rate schedule\n    -  No need to retrain from scratch if you want to train for longer\n- Index-shuffling during dataloading\n- Google-native stack\n    - https://github.com/google/orbax for checkpointing\n    - https://github.com/google/grain for dataloading\n    - https://github.com/google-deepmind/dm_pix for image manipulation\n    - https://github.com/google/array_record as the data format\n- Easy model inspection thanks to [treescope](https://github.com/google-deepmind/treescope)\n- Modularized training script for easy inspection using notebooks ([demo notebook](https://colab.research.google.com/drive/1zHkciFIZxXloJgue9F5LtFlA0m00rJIf?usp=sharing))\n- Easy model surgery thanks to the new [flax.nnx](https://flax.readthedocs.io/en/latest/migrating/linen_to_nnx.html) API\n- [Shape suffixes](https://medium.com/@NoamShazeer/shape-suffixes-good-coding-style-f836e72e24fd) throughout the repository\n\n<h2 name=""start"" id=""start"">Setup 🧗</h2>\n\nJasmine requires `python 3.10`, `jax 0.6.2`, and `flax 0.10.7`. To install the requirements, run:\n\n```bash\npip install -r requirements.txt\npre-commit install\n```\n\n---\n\n<h2 name=""dataset"" id=""dataset"">Dataset 📂</h2>\n\nYou can either download our preprocessed dataset from [Hugging Face](https://huggingface.co/datasets/p-doom/open_ai_minecraft_arrayrecords_chunked) or preprocess [OpenAI's VPT dataset](https://github.com/openai/Video-Pre-Training) manually.\n\n### Option 1: Use Preprocessed Dataset (Recommended)\n\nThe easiest way to get started is to download our preprocessed dataset from Hugging Face. This script will handle downloading and extracting it:\n\n```bash\nbash input_pipeline/download/huggingface/download_openai_array_records.sh\n```\n\n---\n\n### Option 2: Manual Download & Preprocessing of OpenAI's VPT Dataset\n\nIf you prefer to use the raw VPT dataset from OpenAI and preprocess it yourself, follow these steps:\n\n1. **Download index files:**\n   This will download the initial index file:\n\n   ```bash\n   bash input_pipeline/download/openai/download_index_files.sh\n   ```\n\n2. **Download from all index files:**\n   This may take a long time depending on your bandwidth:\n\n   ```bash\n   python input_pipeline/download/openai/download_videos.py --index_file_path data/open_ai_index_files/all_7xx_Apr_6.json\n   python input_pipeline/download/openai/download_videos.py --index_file_path data/open_ai_index_files/all_8xx_Jun_29.json\n   python input_pipeline/download/openai/download_videos.py --index_file_path data/open_ai_index_files/all_9xx_Jun_29.json\n   python input_pipeline/download/openai/download_videos.py --index_file_path data/open_ai_index_files/all_10xx_Jun_29.json\n   ```\n\n3. **Preprocess videos into ArrayRecords:**\n   For efficient distributed training, convert the raw videos into the arrayrecord format (make sure to have [ffmpeg](https://github.com/FFmpeg/FFmpeg) installed on your machine):\n\n   ```bash\n   python input_pipeline/preprocess/video_to_array_records.py\n   ```\n\n> **Note:** This is a large dataset and may take considerable time and storage to download and process.\n\n\n<h2 name=""train"" id=""train"">Quick Start 🚀 </h2>\n\nGenie has three components: a [video tokenizer](models/tokenizer.py), a [latent action model](models/lam.py), and a [dynamics model](models/dynamics.py). Each of these components are trained separately, however, the dynamics model requires a pre-trained video tokenizer (and latent action model).\n\nTo train the video tokenizer, run:\n\n```bash\npython train_tokenizer.py --ckpt_dir <path>\n```\n\nTo train the latent action model, run:\n\n```bash\npython train_lam.py --ckpt_dir <path>\n```\n\nOnce the tokenizer and LAM are trained, the dynamics model can be trained with:\n\n```bash\npython train_dynamics.py --tokenizer_checkpoint <path> --lam_checkpoint <path>\n```\n\nLogging with `wandb` is supported. To enable logging, set the `WANDB_API_KEY` environment variable or run:\n\n```bash\nwandb login\n```\n\nTraining can then be logged by setting the `--log` flag:\n\n```bash\npython train_tokenizer.py --log --entity <wandb-entity> --project <wandb-project>\n```\n\n<h2 name=""cite"" id=""cite"">Citing 📜 </h2>\n\nJasmine was built by [Mihir Mahajan](https://maharajamihir.github.io/), [Alfred Nguyen](https://avocadoali.github.io/) and [Franz Srambical](https://srambical.fr/), but started as a fork of [Jafar](https://github.com/flairox/jafar), built by [Matthew Jackson](https://matthewtjackson.com) and [Timon Willi](https://www.timonwilli.com).\n\nIf you use Jasmine in your work, please cite us, Jafar, and the original Genie paper as follows:\n\n```\n@article{\n    mahajan2025jasmine,\n    title={Jasmine: A simple, performant and scalable JAX-based world modeling codebase},\n    author={Mihir Mahajan and Alfred Nguyen and Franz Srambical and Stefan Bauer},\n    journal = {p(doom) blog},\n    year={2025},\n    url={https://pdoom.org/jasmine.html},\n    note = {https://pdoom.org/blog.html}\n}\n```\n```\n@inproceedings{\n    willi2024jafar,\n    title={Jafar: An Open-Source Genie Reimplemention in Jax},\n    author={Timon Willi and Matthew Thomas Jackson and Jakob Nicolaus Foerster},\n    booktitle={First Workshop on Controllable Video Generation @ ICML 2024},\n    year={2024},\n    url={https://openreview.net/forum?id=ZZGaQHs9Jb}\n}\n```\n```\n@inproceedings{\n    bruce2024genie,\n    title={Genie: Generative Interactive Environments},\n    author={Jake Bruce and Michael D Dennis and Ashley Edwards and Jack Parker-Holder and Yuge Shi and Edward Hughes and Matthew Lai and Aditi Mavalankar and Richie Steigerwald and Chris Apps and Yusuf Aytar and Sarah Maria Elisabeth Bechtle and Feryal Behbahani and Stephanie C.Y. Chan and Nicolas Heess and Lucy Gonzalez and Simon Osindero and Sherjil Ozair and Scott Reed and Jingwei Zhang and Konrad Zolna and Jeff Clune and Nando de Freitas and Satinder Singh and Tim Rockt{\""a}schel},\n    booktitle={Forty-first International Conference on Machine Learning},\n    year={2024},\n    url={https://openreview.net/forum?id=bJbSbJskOS}\n}\n```\n",markdown,tab
+445,2862197,"README.md",3943,14,"input_pipeline",markdown,selection_command
+446,2868173,"README.md",3956,0,"",markdown,selection_command
+447,2869447,"README.md",3943,0,"",markdown,selection_command
+448,2870118,"README.md",3943,1,"i",markdown,selection_command
+449,2870207,"README.md",3943,14,"input_pipeline",markdown,selection_command
+450,2870727,"README.md",3943,14,"",markdown,content
+451,2871002,"README.md",3943,0,"d",markdown,content
+452,2871003,"README.md",3944,0,"",markdown,selection_keyboard
+453,2871134,"README.md",3944,0,"a",markdown,content
+454,2871134,"README.md",3945,0,"",markdown,selection_keyboard
+455,2871137,"README.md",3945,0,"t",markdown,content
+456,2871137,"README.md",3946,0,"",markdown,selection_keyboard
+457,2871305,"README.md",3946,0,"a",markdown,content
+458,2871305,"README.md",3947,0,"",markdown,selection_keyboard
+459,2871411,"README.md",3947,0,"1",markdown,content
+460,2871411,"README.md",3948,0,"",markdown,selection_keyboard
+461,2871925,"README.md",3947,1,"",markdown,content
+462,2871957,"README.md",3946,0,"",markdown,selection_command
+463,2872212,"README.md",3947,0,"",markdown,selection_command
+464,2872537,"README.md",3948,0,"",markdown,selection_command
+465,2876892,"README.md",3956,0,"",markdown,selection_command
+466,2878196,"README.md",3948,0,"",markdown,selection_command
+467,2879101,"README.md",3948,8,"",markdown,content
+468,2879472,"README.md",3948,0,"i",markdown,content
+469,2879472,"README.md",3949,0,"",markdown,selection_keyboard
+470,2879574,"README.md",3949,0,"n",markdown,content
+471,2879574,"README.md",3950,0,"",markdown,selection_keyboard
+472,2879680,"README.md",3950,0,"e",markdown,content
+473,2879681,"README.md",3951,0,"",markdown,selection_keyboard
+474,2880081,"README.md",3950,1,"",markdown,content
+475,2880239,"README.md",3949,1,"",markdown,content
+476,2880348,"README.md",3948,1,"",markdown,content
+477,2880532,"README.md",3948,0,"m",markdown,content
+478,2880532,"README.md",3949,0,"",markdown,selection_keyboard
+479,2880554,"README.md",3949,0,"i",markdown,content
+480,2880554,"README.md",3950,0,"",markdown,selection_keyboard
+481,2880689,"README.md",3950,0,"n",markdown,content
+482,2880689,"README.md",3951,0,"",markdown,selection_keyboard
+483,2880910,"README.md",3951,0,"e",markdown,content
+484,2880910,"README.md",3952,0,"",markdown,selection_keyboard
+485,2881169,"README.md",3952,0,"c",markdown,content
+486,2881169,"README.md",3953,0,"",markdown,selection_keyboard
+487,2881840,"README.md",3953,0,"r",markdown,content
+488,2881840,"README.md",3954,0,"",markdown,selection_keyboard
+489,2882028,"README.md",3954,0,"a",markdown,content
+490,2882028,"README.md",3955,0,"",markdown,selection_keyboard
+491,2882063,"README.md",3955,0,"f",markdown,content
+492,2882063,"README.md",3956,0,"",markdown,selection_keyboard
+493,2882217,"README.md",3956,0,"t",markdown,content
+494,2882217,"README.md",3957,0,"",markdown,selection_keyboard
+495,2882537,"README.md",3956,0,"",markdown,selection_command
+496,2888555,"README.md",4005,0,"",markdown,selection_command
+497,2888711,"README.md",3956,0,"",markdown,selection_command
+498,2893786,"README.md",4281,0,"",markdown,selection_command
+499,2894026,"README.md",4462,0,"",markdown,selection_command
+500,2895102,"README.md",4281,0,"",markdown,selection_command
+501,2899628,"README.md",4281,1,"i",markdown,selection_command
+502,2899671,"README.md",4281,14,"input_pipeline",markdown,selection_command
+503,2899839,"README.md",4281,15,"input_pipeline/",markdown,selection_command
+504,2900008,"README.md",4281,23,"input_pipeline/download",markdown,selection_command
+505,2900344,"README.md",4281,23,"",markdown,content
+506,2901084,"README.md",4281,0,"i",markdown,content
+507,2901084,"README.md",4282,0,"",markdown,selection_keyboard
+508,2901111,"README.md",4282,0,"n",markdown,content
+509,2901111,"README.md",4283,0,"",markdown,selection_keyboard
+510,2901213,"README.md",4283,0,"p",markdown,content
+511,2901214,"README.md",4284,0,"",markdown,selection_keyboard
+512,2901324,"README.md",4284,0,"u",markdown,content
+513,2901325,"README.md",4285,0,"",markdown,selection_keyboard
+514,2901377,"README.md",4285,0,"t",markdown,content
+515,2901377,"README.md",4286,0,"",markdown,selection_keyboard
+516,2901547,"README.md",4286,0,"/",markdown,content
+517,2901548,"README.md",4287,0,"",markdown,selection_keyboard
+518,2901653,"README.md",4287,0,"m",markdown,content
+519,2901653,"README.md",4288,0,"",markdown,selection_keyboard
+520,2901804,"README.md",4288,0,"i",markdown,content
+521,2901804,"README.md",4289,0,"",markdown,selection_keyboard
+522,2901880,"README.md",4289,0,"n",markdown,content
+523,2901880,"README.md",4290,0,"",markdown,selection_keyboard
+524,2901991,"README.md",4290,0,"e",markdown,content
+525,2901991,"README.md",4291,0,"",markdown,selection_keyboard
+526,2902227,"README.md",4291,0,"c",markdown,content
+527,2902227,"README.md",4292,0,"",markdown,selection_keyboard
+528,2902386,"README.md",4292,0,"r",markdown,content
+529,2902387,"README.md",4293,0,"",markdown,selection_keyboard
+530,2902624,"README.md",4293,0,"a",markdown,content
+531,2902624,"README.md",4294,0,"",markdown,selection_keyboard
+532,2902706,"README.md",4294,0,"f",markdown,content
+533,2902707,"README.md",4295,0,"",markdown,selection_keyboard
+534,2902824,"README.md",4295,0,"t",markdown,content
+535,2902825,"README.md",4296,0,"",markdown,selection_keyboard
+536,2903320,"README.md",4295,0,"",markdown,selection_command
+537,2904539,"README.md",4287,0,"",markdown,selection_command
+538,2904689,"README.md",4286,0,"",markdown,selection_command
+539,2904975,"README.md",4281,0,"",markdown,selection_command
+540,2905477,"README.md",4281,5,"",markdown,content
+541,2906260,"README.md",4281,0,"d",markdown,content
+542,2906260,"README.md",4282,0,"",markdown,selection_keyboard
+543,2906349,"README.md",4282,0,"a",markdown,content
+544,2906349,"README.md",4283,0,"",markdown,selection_keyboard
+545,2906358,"README.md",4283,0,"t",markdown,content
+546,2906358,"README.md",4284,0,"",markdown,selection_keyboard
+547,2906520,"README.md",4284,0,"a",markdown,content
+548,2906520,"README.md",4285,0,"",markdown,selection_keyboard
+549,2906630,"README.md",4284,0,"",markdown,selection_command
+550,2907697,"README.md",4332,0,"",markdown,selection_command
+551,2907839,"README.md",4453,0,"",markdown,selection_command
+552,2912699,"README.md",4453,1,"i",markdown,selection_command
+553,2912973,"README.md",4453,14,"input_pipeline",markdown,selection_command
+554,2913186,"README.md",4453,15,"input_pipeline/",markdown,selection_command
+555,2913358,"README.md",4453,23,"input_pipeline/download",markdown,selection_command
+556,2914088,"README.md",4453,24,"input_pipeline/download/",markdown,selection_command
+557,2914433,"README.md",4453,24,"input_pipeline/download/",markdown,selection_command
+558,2914502,"README.md",4453,24,"input_pipeline/download/",markdown,selection_command
+559,2914640,"README.md",4453,24,"input_pipeline/download/",markdown,selection_command
+560,2922165,"README.md",4821,24,"",markdown,content
+561,2922165,"README.md",4698,24,"",markdown,content
+562,2922165,"README.md",4575,24,"",markdown,content
+563,2922165,"README.md",4453,24,"",markdown,content
+564,2922570,"README.md",4749,0,"d",markdown,content
+565,2922570,"README.md",4650,0,"d",markdown,content
+566,2922570,"README.md",4551,0,"d",markdown,content
+567,2922570,"README.md",4453,0,"d",markdown,content
+568,2922570,"README.md",4454,0,"",markdown,selection_keyboard
+569,2922800,"README.md",4753,0,"a",markdown,content
+570,2922800,"README.md",4653,0,"a",markdown,content
+571,2922800,"README.md",4553,0,"a",markdown,content
+572,2922800,"README.md",4454,0,"a",markdown,content
+573,2922800,"README.md",4455,0,"",markdown,selection_keyboard
+574,2922812,"README.md",4757,0,"t",markdown,content
+575,2922812,"README.md",4656,0,"t",markdown,content
+576,2922812,"README.md",4555,0,"t",markdown,content
+577,2922812,"README.md",4455,0,"t",markdown,content
+578,2922812,"README.md",4456,0,"",markdown,selection_keyboard
+579,2922999,"README.md",4761,0,"a",markdown,content
+580,2922999,"README.md",4659,0,"a",markdown,content
+581,2922999,"README.md",4557,0,"a",markdown,content
+582,2922999,"README.md",4456,0,"a",markdown,content
+583,2922999,"README.md",4457,0,"",markdown,selection_keyboard
+584,2923119,"README.md",4765,0,"/",markdown,content
+585,2923119,"README.md",4662,0,"/",markdown,content
+586,2923119,"README.md",4559,0,"/",markdown,content
+587,2923119,"README.md",4457,0,"/",markdown,content
+588,2923119,"README.md",4458,0,"",markdown,selection_keyboard
+589,2923381,"README.md",4769,0,"m",markdown,content
+590,2923381,"README.md",4665,0,"m",markdown,content
+591,2923381,"README.md",4561,0,"m",markdown,content
+592,2923381,"README.md",4458,0,"m",markdown,content
+593,2923382,"README.md",4459,0,"",markdown,selection_keyboard
+594,2923495,"README.md",4773,0,"i",markdown,content
+595,2923495,"README.md",4668,0,"i",markdown,content
+596,2923495,"README.md",4563,0,"i",markdown,content
+597,2923495,"README.md",4459,0,"i",markdown,content
+598,2923495,"README.md",4460,0,"",markdown,selection_keyboard
+599,2923552,"README.md",4777,0,"n",markdown,content
+600,2923552,"README.md",4671,0,"n",markdown,content
+601,2923552,"README.md",4565,0,"n",markdown,content
+602,2923552,"README.md",4460,0,"n",markdown,content
+603,2923552,"README.md",4461,0,"",markdown,selection_keyboard
+604,2923653,"README.md",4781,0,"e",markdown,content
+605,2923653,"README.md",4674,0,"e",markdown,content
+606,2923653,"README.md",4567,0,"e",markdown,content
+607,2923653,"README.md",4461,0,"e",markdown,content
+608,2923653,"README.md",4462,0,"",markdown,selection_keyboard
+609,2923853,"README.md",4785,0,"c",markdown,content
+610,2923853,"README.md",4677,0,"c",markdown,content
+611,2923853,"README.md",4569,0,"c",markdown,content
+612,2923853,"README.md",4462,0,"c",markdown,content
+613,2923853,"README.md",4463,0,"",markdown,selection_keyboard
+614,2924016,"README.md",4789,0,"r",markdown,content
+615,2924016,"README.md",4680,0,"r",markdown,content
+616,2924016,"README.md",4571,0,"r",markdown,content
+617,2924016,"README.md",4463,0,"r",markdown,content
+618,2924016,"README.md",4464,0,"",markdown,selection_keyboard
+619,2924259,"README.md",4793,0,"a",markdown,content
+620,2924259,"README.md",4683,0,"a",markdown,content
+621,2924259,"README.md",4573,0,"a",markdown,content
+622,2924259,"README.md",4464,0,"a",markdown,content
+623,2924259,"README.md",4465,0,"",markdown,selection_keyboard
+624,2924290,"README.md",4797,0,"f",markdown,content
+625,2924290,"README.md",4686,0,"f",markdown,content
+626,2924290,"README.md",4575,0,"f",markdown,content
+627,2924290,"README.md",4465,0,"f",markdown,content
+628,2924290,"README.md",4466,0,"",markdown,selection_keyboard
+629,2924429,"README.md",4801,0,"t",markdown,content
+630,2924429,"README.md",4689,0,"t",markdown,content
+631,2924429,"README.md",4577,0,"t",markdown,content
+632,2924429,"README.md",4466,0,"t",markdown,content
+633,2924429,"README.md",4467,0,"",markdown,selection_keyboard
+634,2924564,"README.md",4805,0,"/",markdown,content
+635,2924564,"README.md",4692,0,"/",markdown,content
+636,2924564,"README.md",4579,0,"/",markdown,content
+637,2924564,"README.md",4467,0,"/",markdown,content
+638,2924564,"README.md",4468,0,"",markdown,selection_keyboard
+639,2924710,"README.md",4467,0,"",markdown,selection_command
+640,2927852,"README.md",5153,0,"",markdown,selection_command
+641,2928973,"README.md",5153,1,"i",markdown,selection_command
+642,2929103,"README.md",5153,14,"input_pipeline",markdown,selection_command
+643,2929301,"README.md",5153,15,"input_pipeline/",markdown,selection_command
+644,2929498,"README.md",5153,25,"input_pipeline/preprocess",markdown,selection_command
+645,2929719,"README.md",5153,26,"input_pipeline/preprocess/",markdown,selection_command
+646,2929988,"README.md",5153,25,"input_pipeline/preprocess",markdown,selection_command
+647,2930171,"README.md",5153,25,"",markdown,content
+648,2930621,"README.md",5153,0,"d",markdown,content
+649,2930622,"README.md",5154,0,"",markdown,selection_keyboard
+650,2930726,"README.md",5154,0,"a",markdown,content
+651,2930726,"README.md",5155,0,"",markdown,selection_keyboard
+652,2930729,"README.md",5155,0,"t",markdown,content
+653,2930729,"README.md",5156,0,"",markdown,selection_keyboard
+654,2930955,"README.md",5156,0,"a",markdown,content
+655,2930956,"README.md",5157,0,"",markdown,selection_keyboard
+656,2931535,"README.md",5156,1,"",markdown,content
+657,2931708,"README.md",5155,1,"",markdown,content
+658,2931925,"README.md",5154,0,"",markdown,selection_command
+659,2932728,"README.md",5153,0,"",markdown,selection_command
+660,2935681,"README.md",5154,0,"",markdown,selection_command
+661,2935734,"README.md",5155,0,"",markdown,selection_command
+662,2935795,"README.md",5155,0,"t",markdown,content
+663,2935795,"README.md",5156,0,"",markdown,selection_keyboard
+664,2935993,"README.md",5156,0,"a",markdown,content
+665,2935993,"README.md",5157,0,"",markdown,selection_keyboard
+666,2936394,"README.md",5156,0,"",markdown,selection_command
+667,4780376,"pyproject.toml",0,0,"[project]\nname = ""jafar""\nversion = ""0.1.0""\nrequires-python = "">=3.10""\ndependencies = []\n\n[dependency-groups]\ncore = [\n  ""dm-pix>=0.4.3"",\n  ""einops>=0.8.0"",\n  ""flax>=0.10.7"",\n  ""jax[cuda12]>=0.6.2"",\n  ""optax>=0.2.3"",\n  ""tyro>=0.8.5"",\n  ""wandb>=0.17.4"",\n  ""grain>=0.2.10"",\n  ""array-record>=0.7.2"",\n  ""pre-commit>=4.2.0"",\n]\ncoinrun = [\n  ""procgen>=0.10.7"",\n]\nminecraft = [\n  ""hf-transfer==0.1.9"",\n  ""huggingface-hub[cli]>=0.34.3"",\n  ""ffmpeg-python==0.2.0"",\n  ""tqdm>=4.67.1"",\n]\n\n[build-system]\nrequires = [""uv-build>=0.8.21,<0.9.0""]\nbuild-backend = ""uv_build""\n\n\n",plaintext,tab
+668,4783725,"pyproject.toml",0,557,"",plaintext,content
+669,4783758,"pyproject.toml",0,1," ",plaintext,content
+670,4783998,"README.md",0,0,"",markdown,tab
+671,4869408,"data/coinrun/generate_coinrun_dataset.py",0,0,"""""""\nGenerates a dataset of random-action CoinRun episodes.\nEpisodes are saved individually as memory-mapped files for efficient loading.\n""""""\n\nfrom dataclasses import dataclass\n\nfrom gym3 import types_np\nimport numpy as np\nfrom procgen import ProcgenGym3Env\nimport tyro\nimport json\nimport os\nfrom utils import save_chunks\n\n\n@dataclass\nclass Args:\n    num_episodes_train: int = 10000\n    num_episodes_val: int = 500\n    num_episodes_test: int = 500\n    output_dir: str = ""data/coinrun_episodes""\n    min_episode_length: int = 1000\n    max_episode_length: int = 1000\n    chunk_size: int = 160\n    chunks_per_file: int = 100\n    seed: int = 0\n\n\nargs = tyro.cli(Args)\nassert (\n    args.max_episode_length >= args.min_episode_length\n), ""Maximum episode length must be greater than or equal to minimum episode length.""\n\nif args.min_episode_length < args.chunk_size:\n    print(\n        ""Warning: Minimum episode length is smaller than chunk size. Note that episodes shorter than the chunk size will be discarded.""\n    )\n\n\n# --- Generate episodes ---\ndef generate_episodes(num_episodes, split):\n    episode_idx = 0\n    episode_metadata = []\n    obs_chunks = []\n    act_chunks = []\n    file_idx = 0\n    output_dir_split = os.path.join(args.output_dir, split)\n    while episode_idx < num_episodes:\n        seed = np.random.randint(0, 10000)\n        env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=seed)\n\n        observations_seq = []\n        actions_seq = []\n        episode_obs_chunks = []\n        episode_act_chunks = []\n\n        # --- Run episode ---\n        step_t = 0\n        first_obs = True\n        for step_t in range(args.max_episode_length):\n            _, obs, first = env.observe()\n            action = types_np.sample(env.ac_space, bshape=(env.num,))\n            env.act(action)\n            observations_seq.append(obs[""rgb""])\n            actions_seq.append(action)\n            if len(observations_seq) == args.chunk_size:\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n                observations_seq = []\n                actions_seq = []\n            if first and not first_obs:\n                break\n            first_obs = False\n\n        # --- Save episode ---\n        if step_t + 1 >= args.min_episode_length:\n            if observations_seq:\n                if len(observations_seq) < args.chunk_size:\n                    print(\n                        f""Warning: Inconsistent chunk_sizes. Episode has {len(observations_seq)} frames, ""\n                        f""which is smaller than the requested chunk_size: {args.chunk_size}. ""\n                        ""This might lead to performance degradation during training.""\n                    )\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n\n            obs_chunks_data = [\n                np.concatenate(seq, axis=0).astype(np.uint8)\n                for seq in episode_obs_chunks\n            ]\n            act_chunks_data = [\n                np.concatenate(act, axis=0) for act in episode_act_chunks\n            ]\n            obs_chunks.extend(obs_chunks_data)\n            act_chunks.extend(act_chunks_data)\n\n            ep_metadata, file_idx, obs_chunks, act_chunks = save_chunks(\n                file_idx, args.chunks_per_file, output_dir_split, obs_chunks, act_chunks\n            )\n            episode_metadata.extend(ep_metadata)\n\n            print(f""Episode {episode_idx} completed, length: {step_t + 1}."")\n            episode_idx += 1\n        else:\n            print(f""Episode too short ({step_t + 1}), resampling..."")\n\n    if len(obs_chunks) > 0:\n        print(\n            f""Warning: Dropping {len(obs_chunks)} chunks for consistent number of chunks per file."",\n            ""Consider changing the chunk_size and chunks_per_file parameters to prevent data-loss."",\n        )\n\n    print(f""Done generating {split} split"")\n    return episode_metadata\n\n\ndef get_action_space():\n    env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=0)\n    return env.ac_space.eltype.n\n\n\ndef main():\n    # Set random seed and create dataset directories\n    np.random.seed(args.seed)\n    # --- Generate episodes ---\n    train_episode_metadata = generate_episodes(args.num_episodes_train, ""train"")\n    val_episode_metadata = generate_episodes(args.num_episodes_val, ""val"")\n    test_episode_metadata = generate_episodes(args.num_episodes_test, ""test"")\n\n    # --- Save metadata ---\n    metadata = {\n        ""env"": ""coinrun"",\n        ""num_actions"": get_action_space(),\n        ""num_episodes_train"": args.num_episodes_train,\n        ""num_episodes_val"": args.num_episodes_val,\n        ""num_episodes_test"": args.num_episodes_test,\n        ""avg_episode_len_train"": np.mean(\n            [ep[""avg_seq_len""] for ep in train_episode_metadata]\n        ),\n        ""avg_episode_len_val"": np.mean(\n            [ep[""avg_seq_len""] for ep in val_episode_metadata]\n        ),\n        ""avg_episode_len_test"": np.mean(\n            [ep[""avg_seq_len""] for ep in test_episode_metadata]\n        ),\n        ""episode_metadata_train"": train_episode_metadata,\n        ""episode_metadata_val"": val_episode_metadata,\n        ""episode_metadata_test"": test_episode_metadata,\n    }\n    with open(os.path.join(args.output_dir, ""metadata.json""), ""w"") as f:\n        json.dump(metadata, f)\n\n    print(f""Done generating dataset."")\n\n\nif __name__ == ""__main__"":\n    main()\n",python,tab
+672,4870140,"data/coinrun/generate_coinrun_dataset.py",321,0,"",python,selection_mouse
+673,4871817,"data/coinrun/generate_coinrun_dataset.py",1499,0,"",python,selection_command
+674,4872559,"data/coinrun/generate_coinrun_dataset.py",0,0,"",python,selection_command
+675,4876799,"data/coinrun/generate_coinrun_dataset.py",4,0,"",python,selection_command
+676,4876848,"data/coinrun/generate_coinrun_dataset.py",59,0,"",python,selection_command
+677,4877101,"data/coinrun/generate_coinrun_dataset.py",137,0,"",python,selection_command
+678,4877125,"data/coinrun/generate_coinrun_dataset.py",141,0,"",python,selection_command
+679,4877165,"data/coinrun/generate_coinrun_dataset.py",142,0,"",python,selection_command
+680,4877197,"data/coinrun/generate_coinrun_dataset.py",176,0,"",python,selection_command
+681,4877231,"data/coinrun/generate_coinrun_dataset.py",177,0,"",python,selection_command
+682,4877263,"data/coinrun/generate_coinrun_dataset.py",203,0,"",python,selection_command
+683,4877298,"data/coinrun/generate_coinrun_dataset.py",222,0,"",python,selection_command
+684,4877330,"data/coinrun/generate_coinrun_dataset.py",257,0,"",python,selection_command
+685,4877363,"data/coinrun/generate_coinrun_dataset.py",269,0,"",python,selection_command
+686,4877396,"data/coinrun/generate_coinrun_dataset.py",281,0,"",python,selection_command
+687,4877699,"data/coinrun/generate_coinrun_dataset.py",0,0,"",python,selection_command
+688,4881224,"data/coinrun/generate_coinrun_dataset.py",4,0,"",python,selection_command
+689,4881292,"data/coinrun/generate_coinrun_dataset.py",59,0,"",python,selection_command
+690,4881409,"data/coinrun/generate_coinrun_dataset.py",137,0,"",python,selection_command
+691,4881674,"data/coinrun/generate_coinrun_dataset.py",141,0,"",python,selection_command
+692,4881699,"data/coinrun/generate_coinrun_dataset.py",142,0,"",python,selection_command
+693,4881722,"data/coinrun/generate_coinrun_dataset.py",176,0,"",python,selection_command
+694,4881755,"data/coinrun/generate_coinrun_dataset.py",177,0,"",python,selection_command
+695,4881787,"data/coinrun/generate_coinrun_dataset.py",203,0,"",python,selection_command
+696,4881821,"data/coinrun/generate_coinrun_dataset.py",222,0,"",python,selection_command
+697,4881854,"data/coinrun/generate_coinrun_dataset.py",257,0,"",python,selection_command
+698,4881888,"data/coinrun/generate_coinrun_dataset.py",269,0,"",python,selection_command
+699,4881920,"data/coinrun/generate_coinrun_dataset.py",281,0,"",python,selection_command
+700,4882335,"data/coinrun/generate_coinrun_dataset.py",291,0,"",python,selection_command
+701,4882441,"data/coinrun/generate_coinrun_dataset.py",321,0,"",python,selection_command
+702,4882949,"data/coinrun/generate_coinrun_dataset.py",291,0,"",python,selection_command
+703,4905982,"README.md",0,0,"",markdown,tab
+704,4911629,"data/coinrun/generate_coinrun_dataset.py",0,0,"",python,tab
+705,4913944,"README.md",0,0,"",markdown,tab
+706,4917140,"data/minecraft/huggingface/download_openai_array_records.sh",0,0,"#!/bin/bash\n\n# Download and extract array records from Hugging Face\n# \n# This script performs a two-step process:\n# 1. Downloads compressed array records from a Hugging Face dataset repository\n# 2. Extracts the compressed tar files in parallel for better performance\n#\n# Usage:\n#   ./download_openai_array_records.sh [hf_download_dir] [final_dataset_dir]\n#\n# Arguments:\n#   hf_download_dir    - Directory to store compressed downloads (default: data/minecraft_arrayrecords_compressed)\n#   final_dataset_dir  - Directory for extracted array records (default: data/minecraft_arrayrecords)\n\n# Set default directories if not provided as arguments\nhf_download_dir=""${1:-data/minecraft_arrayrecords_compressed}"" \nfinal_dataset_dir=""${2:-data/minecraft_arrayrecords}""          \n\nmkdir -p $hf_download_dir\nmkdir -p $final_dataset_dir\n\n# Step 1: Download compressed dataset from Hugging Face\necho ""Starting download from Hugging Face...""\nrepo_id=p-doom/open_ai_minecraft_arrayrecords_chunked\nstart_time_hf_download=$(date +%s)\n\nHF_HUB_ENABLE_HF_TRANSFER=1 HF_HUB_DISABLE_SYMLINKS=1 \\nhuggingface-cli download --repo-type dataset $repo_id --local-dir $hf_download_dir\n\nend_time_hf_download=$(date +%s)\necho ""Download completed. Time taken: $((end_time_hf_download - start_time_hf_download)) seconds""\n\n# Step 2: Extract compressed array records in parallel\necho ""Starting parallel extraction of tar files...""\nnum_workers=64  # Number of parallel extraction processes\nstart_time_uncompress=$(date +%s)\n\n# Find all shard tar files and extract them in parallel:\nxargs -0 -P $num_workers -I {} bash -c 'echo ""Extracting {}""; tar -xf ""{}"" -C ""'$final_dataset_dir'""'\n\nend_time_uncompress=$(date +%s)\n\n# Display timing summary\necho ""================================""\necho ""Extraction completed successfully!""\necho ""Uncompress time: $((end_time_uncompress - start_time_uncompress)) seconds""\necho ""Download time: $((end_time_hf_download - start_time_hf_download)) seconds""\necho ""Total time: $((end_time_uncompress - start_time_hf_download)) seconds""\necho ""Final dataset location: $final_dataset_dir""\n",shellscript,tab
+707,4920228,"README.md",0,0,"",markdown,tab
+708,4924670,"data/minecraft/openai/download_actions_files.py",0,0,"import subprocess\nimport json\nimport tyro\nfrom dataclasses import dataclass\nimport os\nfrom multiprocessing import Pool, cpu_count\nfrom tqdm import tqdm\n\n\n@dataclass\nclass Args:\n    index_file: str = ""data/open_ai_index_files/all_6xx_Jun_29.json""\n    output_dir: str = ""data/open_ai_minecraft_actions_files""\n    num_workers: int = -1  # -1 means use all available cores\n\n\ndef flatten_path(relpath):\n    """"""Convert nested path to flattened filename with subdirectory as prefix\n    e.g. data/6.10/filename.mp4 -> 6.10_filename.mp4\n    """"""\n\n    parts = relpath.split(""/"")\n\n    if len(parts) >= 3:\n        subdir = parts[1]\n        filename = parts[2]\n        return f""{subdir}_{filename}""\n    else:\n        return relpath.replace(""/"", ""_"")\n\n\ndef download_file(args):\n    try:\n        url, base_dir, output_dir = args\n        jsonl_url = url.rsplit(""."", 1)[0] + "".jsonl""\n        filename = flatten_path(jsonl_url)\n        output_file = os.path.join(output_dir, filename)\n        subprocess.run(\n            [""wget"", ""-q"", base_dir + jsonl_url, ""-O"", output_file], check=True\n        )\n        return {""file"": jsonl_url, ""success"": True}\n    except subprocess.CalledProcessError as e:\n        # delete file if it exists\n        if os.path.exists(output_file):\n            os.remove(output_file)\n        return {""file"": jsonl_url, ""success"": False, ""error"": str(e)}\n\n\ndef download_actions_files(index_file: str, output_dir: str, num_workers: int):\n    # load json file\n    with open(index_file, ""r"") as f:\n        data = json.load(f)\n\n    base_dir = data[""basedir""]\n    urls = data[""relpaths""]\n\n    # Prepare arguments for each process\n    args_list = [(url, base_dir, output_dir) for url in urls]\n\n    results = []\n    with tqdm(total=len(args_list), desc=""Downloading actions files"") as pbar:\n        with Pool(processes=num_workers) as pool:\n            for result in pool.imap_unordered(download_file, args_list):\n                results.append(result)\n                pbar.update(1)\n\n    # save results to json\n    meta_data_file_name = index_file.split(""/"")[-1].split(""."")[0] + ""_metadata.json""\n    with open(os.path.join(output_dir, meta_data_file_name), ""w"") as f:\n        json.dump(results, f)\n\n    # print number of failed downloads\n    failed_downloads = [result for result in results if not result[""success""]]\n    print(f""Number of failed downloads: {len(failed_downloads)}"")\n\n    # print number of successful downloads\n    successful_downloads = [result for result in results if result[""success""]]\n    print(f""Number of successful downloads: {len(successful_downloads)}"")\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    if not os.path.exists(args.output_dir):\n        os.makedirs(args.output_dir)\n\n    if args.num_workers == -1:\n        args.num_workers = cpu_count()\n\n    print(f""Index file: {args.index_file}"")\n    print(f""Output directory: {args.output_dir}"")\n    print(f""Number of workers: {args.num_workers}"")\n\n    download_actions_files(args.index_file, args.output_dir, args.num_workers)\n",python,tab
+709,4930223,"README.md",0,0,"",markdown,tab
+710,4933542,"data/minecraft/openai/download_videos.py",0,0,"import json\nimport requests\nimport os\nimport tyro\nimport logging\nfrom urllib.parse import urljoin\nfrom dataclasses import dataclass\nfrom pathlib import Path\nfrom tqdm import tqdm\nfrom multiprocessing import Pool, cpu_count\nimport time\n\n\n@dataclass\nclass DownloadVideos:\n    index_file_path: str = ""data/open_ai_index_files/all_6xx_Jun_29.json""\n    num_workers: int = -1  # -1 means use all available cores\n    output_dir: str = ""data/minecraft_videos/""\n\n\ndef download_single_file(args):\n    """"""Download a single file - designed to be used with multiprocessing""""""\n    relpath, url, output_path = args\n\n    if os.path.exists(output_path):\n        return f""Skipped {relpath} (already exists)""\n\n    # No need to create parent directories since we're flattening the structure\n    try:\n        response = requests.get(url, stream=True, timeout=30)\n        if response.status_code == 200:\n            file_size = 0\n            with open(output_path, ""wb"") as f:\n                for chunk in response.iter_content(chunk_size=8192):\n                    if chunk:\n                        f.write(chunk)\n                        file_size += len(chunk)\n\n            # Convert to MB for logging\n            file_size_mb = file_size / (1024 * 1024)\n            return f""Downloaded {relpath} ({file_size_mb:.2f} MB)""\n        else:\n            return f""Failed to download {relpath}: HTTP {response.status_code}""\n    except requests.exceptions.RequestException as e:\n        return f""Request failed for {relpath}: {e}""\n    except Exception as e:\n        return f""Unexpected error downloading {relpath}: {e}""\n\n\ndef flatten_path(relpath):\n    """"""Convert nested path to flattened filename with subdirectory as prefix\n    e.g. data/6.10/filename.mp4 -> 6.10_filename.mp4\n    """"""\n\n    parts = relpath.split(""/"")\n\n    if len(parts) >= 3:\n        subdir = parts[1]\n        filename = parts[2]\n        return f""{subdir}_{filename}""\n    else:\n        return relpath.replace(""/"", ""_"")\n\n\ndef download_dataset(index_file_path, output_dir, num_workers=64):\n    # Load the index file\n    with open(index_file_path, ""r"") as f:\n        index_data = json.load(f)\n\n    basedir = index_data[""basedir""]\n    relpaths = index_data[""relpaths""]\n\n    # Filter for mp4 files only and flatten the path structure\n    mp4_files = []\n    for relpath in relpaths:\n        if relpath.endswith("".mp4""):\n            url = urljoin(basedir, relpath)\n            flattened_filename = flatten_path(relpath)\n            output_path = os.path.join(output_dir, flattened_filename)\n            mp4_files.append((relpath, url, output_path))\n\n    print(f""Found {len(mp4_files)} MP4 files to download"")\n    print(f""Using {num_workers} workers for parallel downloads"")\n\n    start_time = time.time()\n\n    if num_workers > len(mp4_files):\n        num_workers = len(mp4_files)\n\n    with tqdm(\n        total=len(mp4_files), desc=""Overall Download Progress"", unit=""files""\n    ) as pbar:\n        with Pool(processes=num_workers) as pool:\n            results = []\n            for result in pool.imap_unordered(\n                download_single_file,\n                [\n                    (relpath, url, output_path)\n                    for relpath, url, output_path in mp4_files\n                ],\n            ):\n                results.append(result)\n                pbar.update(1)\n    # Print final results summary\n    successful_downloads = sum(1 for r in results if ""Downloaded"" in r)\n    skipped_files = sum(1 for r in results if ""Skipped"" in r)\n    failed_downloads = len(results) - successful_downloads - skipped_files\n\n    print(f""\nDownload Summary:"")\n    print(f""  Successful downloads: {successful_downloads}"")\n    print(f""  Skipped files: {skipped_files}"")\n    print(f""  Failed downloads: {failed_downloads}"")\n\n    end_time = time.time()\n    total_time = end_time - start_time\n    print(f""Download completed in {total_time:.2f} seconds"")\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(DownloadVideos)\n    os.makedirs(args.output_dir, exist_ok=True)\n\n    if args.num_workers == -1:\n        args.num_workers = cpu_count()\n\n    print(f""Index file path: {args.index_file_path}"")\n    print(f""Output directory: {args.output_dir}"")\n    print(f""Number of workers: {args.num_workers}"")\n\n    download_dataset(args.index_file_path, args.output_dir, args.num_workers)\n",python,tab
+711,4935141,"data/minecraft/openai/download_videos.py",2315,0,"",python,selection_mouse
+712,4935141,"data/minecraft/openai/download_videos.py",2314,0,"",python,selection_command
+713,4935394,"data/minecraft/openai/download_videos.py",0,0,"",python,selection_command
+714,4937027,"data/minecraft/openai/download_videos.py",235,0,"",python,selection_mouse
+715,4939290,"README.md",0,0,"",markdown,tab
+716,4951060,"data/coinrun/generate_coinrun_dataset.py",0,0,"",python,tab
+717,4951272,"data/coinrun/generate_coinrun_dataset.py",320,0,"",python,selection_mouse
+718,4951279,"data/coinrun/generate_coinrun_dataset.py",319,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-83ab1212-b02e-4773-9400-fd9fd124ffa11763043835189-2025_11_13-15.24.15.09/source.csv

@@ -0,0 +1,209 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"#!/usr/bin/env python3\n""""""\nCommon utilities for dataset serialization scripts.\n""""""\n\nfrom __future__ import annotations\n\nfrom dataclasses import dataclass\nfrom pathlib import Path\nfrom typing import List, Optional, Tuple, Dict\n\nimport difflib\nimport re\nimport pandas as pd\nfrom datasets import Dataset, load_dataset\n\n\n_ANSI_CSI_RE = re.compile(r""\x1b\[[0-9;?]*[ -/]*[@-~]"")\n_ANSI_OSC_TERMINATED_RE = re.compile(r""\x1b\][\s\S]*?(?:\x07|\x1b\\)"")\n_ANSI_OSC_LINE_FALLBACK_RE = re.compile(r""\x1b\][^\n]*$"")\n_BRACKETED_PASTE_ENABLE = ""\x1b[?2004h""\n_BRACKETED_PASTE_DISABLE = ""\x1b[?2004l""\n_OSC_633 = ""\x1b]633;""\n_OSC_0 = ""\x1b]0;""\n\n\n@dataclass\nclass SerializeConfig:\n    output_dir: str\n    shard_size: int\n    target_chars: int\n    overlap_chars: int\n    min_session_chars: int\n    max_docs: Optional[int]\n    long_pause_threshold_ms: int\n    csv_root: Optional[str]\n    val_ratio: float\n    arrayrecord_group_size: Optional[int] = None\n\n\ndef _clean_text(text: str) -> str:\n    # Normalize line endings and strip trailing spaces; preserve tabs/newlines.\n    return text.replace(""\r\n"", ""\n"").replace(""\r"", ""\n"").rstrip()\n\n\ndef _fenced_block(path: str, language: Optional[str], content: str) -> str:\n    lang = (language or """").lower()\n    return f""```{lang}\n{content}\n```\n""\n\n\ndef _apply_change(content: str, offset: int, length: int, new_text: str) -> str:\n    # Mirrors crowd_code_player.replay_file.apply_change\n    base = str(content)\n    text = str(new_text) if pd.notna(new_text) else """"\n    text = text.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n    if offset > len(base):\n        base = base + ("" "" * (offset - len(base)))\n    return base[:offset] + text + base[offset + length:]\n\n\ndef _apply_backspaces(text: str) -> str:\n    out: List[str] = []\n    for ch in text:\n        if ch == ""\b"":  # \x08\n            if out:\n                out.pop()\n        else:\n            out.append(ch)\n    return """".join(out)\n\n\ndef _normalize_terminal_output(raw: str) -> str:\n    """"""\n    Normalize PTY/terminal output for training:\n      - Apply backspaces (\x08)\n      - Strip OSC (window title/shell integration) first, keeping BEL/ST terminators intact\n      - Resolve carriage returns (\r) by keeping the last rewrite per line\n      - Strip CSI (coloring etc.)\n      - Finally drop any remaining BEL (\x07)\n    """"""\n    if not raw:\n        return raw\n    s = _apply_backspaces(raw)\n    # Remove OSC sequences that are properly terminated (BEL or ST)\n    s = _ANSI_OSC_TERMINATED_RE.sub("""", s)\n    # Fallback: drop any unterminated OSC up to end-of-line only\n    s = ""\n"".join(_ANSI_OSC_LINE_FALLBACK_RE.sub("""", line) for line in s.split(""\n""))\n    # Resolve carriage returns per line:\n    # - If there are multiple rewrites, keep the last non-empty chunk\n    # - If it's CRLF (ending with '\r' before '\n'), keep the content before '\r'\n    resolved_lines: List[str] = []\n    for seg in s.split(""\n""):\n        parts = seg.split(""\r"")\n        chosen = """"\n        # pick last non-empty part if available; else last part\n        for p in reversed(parts):\n            if p != """":\n                chosen = p\n                break\n        if chosen == """" and parts:\n            chosen = parts[-1]\n        resolved_lines.append(chosen)\n    s = ""\n"".join(resolved_lines)\n    # Strip ANSI escape sequences\n    s = _ANSI_CSI_RE.sub("""", s)\n    # Remove any remaining BEL beeps\n    s = s.replace(""\x07"", """")\n    return s\n\n\ndef _line_numbered_output(content: str, start_line: Optional[int] = None, end_line: Optional[int] = None) -> str:\n    # TODO (f.srambical): check whether this corresponds **exactly** to the output of cat -n {file_path} | sed -n '{vstart},{vend}p'\n    lines = content.splitlines()\n    total = len(lines)\n    if total == 0:\n        return """"\n    s = 1 if start_line is None else max(1, min(start_line, total))\n    e = total if end_line is None else max(1, min(end_line, total))\n    if e < s:\n        # FIXME (f.srambical): If this does not happen, remove the condition\n        raise ValueError(""This should never happen!"")\n        e = s\n    buf: List[str] = []\n    for idx in range(s, e + 1):\n        buf.append(f""{idx:6}\t{lines[idx - 1]}"")\n    return ""\n"".join(buf)\n\n\ndef _compute_viewport(total_lines: int, center_line: int, radius: int) -> Tuple[int, int]:\n    if total_lines <= 0:\n        return (1, 0)\n    start = max(1, center_line - radius)\n    end = min(total_lines, center_line + radius)\n    if end < start:\n        # FIXME (f.srambical): If this does not happen, remove the condition\n        raise ValueError(""This should never happen!"")\n    return (start, end)\n\n\ndef _escape_single_quotes_for_sed(text: str) -> str:\n    # Close quote, add an escaped single quote, reopen quote: '""'""'\n    return text.replace(""'"", ""'\""'\""'"")\n\n\ndef _compute_changed_block_lines(before: str, after: str) -> Tuple[int, int, List[str]]:\n    """"""\n    Return 1-based start and end line numbers in 'before' that should be replaced,\n    and the replacement lines from 'after'.\n    For pure deletions, the replacement list may be empty.\n    """"""\n    before_lines = before.splitlines()\n    after_lines = after.splitlines()\n    sm = difflib.SequenceMatcher(a=before_lines, b=after_lines, autojunk=False)\n    opcodes = [op for op in sm.get_opcodes() if op[0] != ""equal""]\n    if not opcodes:\n        # FIXME (f.srambical): clean this up\n        raise ValueError(""No diff opcodes found for content change"")\n                # No visible change; choose a safe single-line replace at end of file\n        start_line = max(1, len(before_lines))\n        end_line = start_line\n        repl = after_lines[start_line - 1:start_line] if after_lines else [""""]\n        return (start_line, end_line, repl)\n\n    first = opcodes[0]\n    last = opcodes[-1]\n    # i1/i2 refer to 'before' indices, j1/j2 to 'after'\n    start_line = (first[1] + 1) if (first[1] + 1) > 0 else 1\n    end_line = last[2] # no increment since we go from 'exclusive' to 'inclusive' indexing\n    replacement_lines = after_lines[first[3]:last[4]]\n    return (start_line, end_line, replacement_lines)\n\n\ndef _session_to_transcript(\n    df: pd.DataFrame,\n    long_pause_threshold_ms: int,\n) -> str:\n\n    file_states: Dict[str, str] = {}\n    terminal_state: str = """"\n    per_file_event_counts: Dict[str, int] = {}\n    per_file_cursor_positions: Dict[str, Tuple[int, int]] = {}  # (offset, length) for each file\n    last_time_ms: Optional[int] = None\n\n    parts: List[str] = []\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_time: int = row[""Time""]\n        language: Optional[str] = row[""Language""]\n\n        # Long pause detection\n        if last_time_ms is not None:\n            delta = event_time - last_time_ms\n            if delta > long_pause_threshold_ms:\n                # TODO (f.srambical): think about whether we want to emit this as an observation or not\n                parts.append(f""<obs long_pause ms=\""{delta}\"" />"")\n        last_time_ms = event_time\n\n        event_type = row[""Type""]\n\n        match event_type:\n            case ""tab"":\n                # File switch event\n                parts.append(f""<act focus file=\""{file_path}\"" />"")\n                \n                # If Text is present, this is the first time opening the file\n                # and the entire file content is captured\n                text = row[""Text""]\n                if pd.notna(text):\n                    file_content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = file_content\n                    parts.append(f""// observation: file={file_path}"")\n                    parts.append(_fenced_block(file_path, language, _clean_text(file_content)))\n\n            case ""terminal_command"":\n                # Terminal command execution\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<act terminal_command />"")\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(command_str)))\n\n            case ""terminal_output"":\n                # Terminal output capture\n                output = row[""Text""]\n                output_str = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<obs terminal_output />"")\n                parts.append(_fenced_block(file_path, None, _clean_text(output_str)))\n\n            case ""terminal_focus"":\n                # Terminal focus event\n                parts.append(f""<act focus target=\""terminal\"" />"")\n\n            case ""git_branch_checkout"":\n                # Git branch checkout event\n                branch_info = row[""Text""]\n                branch_str = str(branch_info).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<act git_branch_checkout />"")\n                parts.append(f""// git: {_clean_text(branch_str)}"")\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                # Handle cursor movement\n                offset = row[""RangeOffset""]\n                length = row[""RangeLength""]\n                old_cursor = per_file_cursor_positions.get(file_path, (0, 0))\n                new_cursor = (offset, length)\n                per_file_cursor_positions[file_path] = new_cursor\n                \n                # Emit cursor movement observation if position changed\n                if old_cursor != new_cursor:\n                    parts.append(f""<act cursor file=\""{file_path}\"" offset=\""{offset}\"" len=\""{length}\"" />"")\n\n            case ""content"":\n                # Handle file edit events\n                offset = row[""RangeOffset""]\n                length = row[""RangeLength""]\n                new_text = row[""Text""]\n                new_text_str = str(new_text) if pd.notna(new_text) else """"\n\n                operation = ""noop""\n                if length == 0 and new_text_str:\n                    operation = ""insert""\n                elif length > 0 and not new_text_str:\n                    operation = ""delete""\n                elif length > 0 and new_text_str:\n                    operation = ""replace""\n\n                parts.append(f""<act {operation} file=\""{file_path}\"" offset=\""{offset}\"" len=\""{length}\"" />"")\n\n                if new_text_str and (operation == ""insert"" or operation == ""replace""):\n                    parts.append(_fenced_block(file_path, language, _clean_text(new_text_str)))\n\n                before = file_states.get(file_path, """")\n                after = _apply_change(before, offset, length, new_text)\n                file_states[file_path] = after\n                per_file_event_counts[file_path] = per_file_event_counts.get(file_path, 0) + 1\n\n                # Update cursor position after edit (cursor moves to end of inserted/replaced text)\n                per_file_cursor_positions[file_path] = (offset + len(new_text_str), 0)\n\n            case _:\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    return ""\n"".join(parts).strip()\n\n\ndef session_to_bash_formatted_transcript(\n    df: pd.DataFrame,\n    viewport_radius: int = 10,\n    normalize_terminal_output: bool = True,\n) -> str:\n    r""""""\n    Serialize a session to a bash-like transcript comprised of:\n      - Commands (bash fenced blocks): cat -n, sed -i 'S,Ec\...' && cat -n | sed -n 'VSTART,VENDp'\n      - Outputs (<stdout>...</stdout>) that reflect the file state after each action\n    Tracks per-file state and a per-file viewport. Viewport only shifts when selection moves out of bounds\n    or when first initialized.\n    """"""\n    file_states: Dict[str, str] = {}\n    per_file_viewport: Dict[str, Optional[Tuple[int, int]]] = {}\n\n    parts: List[str] = []\n    terminal_output_buffer: List[str] = []\n    pending_edits_before: Dict[str, Optional[str]] = {}\n\n    def _flush_terminal_output_buffer() -> None:\n        if not terminal_output_buffer:\n            return\n        aggregated = """".join(terminal_output_buffer)\n        out = aggregated\n        if normalize_terminal_output:\n            out = _normalize_terminal_output(out)\n        cleaned = _clean_text(out)\n        if cleaned.strip():\n            parts.append(f""<stdout>\n{cleaned}\n</stdout>"")\n        terminal_output_buffer.clear()\n\n    def _flush_pending_edit_for_file(target_file: str) -> None:\n        before_snapshot = pending_edits_before.get(target_file)\n        if before_snapshot is None:\n            return\n        after_state = file_states.get(target_file, """")\n        try:\n            start_line, end_line, repl_lines = _compute_changed_block_lines(before_snapshot, after_state)\n        except ValueError:\n            pending_edits_before[target_file] = None\n            return\n        before_total_lines = len(before_snapshot.splitlines())\n        if end_line < start_line:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            if start_line <= max(1, before_total_lines):\n                sed_cmd = f""sed -i '{start_line}i\\\n{sed_payload}' {target_file}""\n            else:\n                sed_cmd = f""sed -i '$a\\\n{sed_payload}' {target_file}""\n        elif not repl_lines:\n            sed_cmd = f""sed -i '{start_line},{end_line}d' {target_file}""\n        else:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            sed_cmd = f""sed -i '{start_line},{end_line}c\\\n{sed_payload}' {target_file}""\n        total_lines = len(after_state.splitlines())\n        center = (start_line + end_line) // 2\n        vp = _compute_viewport(total_lines, center, viewport_radius)\n        per_file_viewport[target_file] = vp\n        vstart, vend = vp\n        chained_cmd = f""{sed_cmd} && cat -n {target_file} | sed -n '{vstart},{vend}p'""\n        parts.append(_fenced_block(target_file, ""bash"", _clean_text(chained_cmd)))\n        viewport_output = _line_numbered_output(after_state, vstart, vend)\n        parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n        pending_edits_before[target_file] = None\n\n    def _flush_all_pending_edits() -> None:\n        for fname in list(pending_edits_before.keys()):\n            _flush_pending_edit_for_file(fname)\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_type = row[""Type""]\n        \n        match event_type:\n            case ""tab"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                text = row[""Text""]\n                if pd.notna(text):\n                    content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = content\n                    # First open with full file capture\n                    cmd = f""cat -n {file_path}""\n                    parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                    output = _line_numbered_output(content)\n                    parts.append(f""<stdout>\n{output}\n</stdout>"")\n                else:\n                    # File switch without content snapshot: show current viewport only\n                    content = file_states.get(file_path, """")\n                    total_lines = len(content.splitlines())\n                    vp = per_file_viewport.get(file_path)\n                    if not vp or vp[1] == 0:\n                        vp = _compute_viewport(total_lines, 1, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                    if vp:\n                        vstart, vend = vp\n                        cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                        parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                        viewport_output = _line_numbered_output(content, vstart, vend)\n                        parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n\n            case ""content"":\n                _flush_terminal_output_buffer()\n                offset = int(row[""RangeOffset""])\n                length = int(row[""RangeLength""])\n                new_text = row[""Text""]\n                before = file_states.get(file_path, """")\n                after = _apply_change(before, offset, length, new_text)\n                if pending_edits_before.get(file_path) is None:\n                    pending_edits_before[file_path] = before\n                file_states[file_path] = after\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                offset = int(row[""RangeOffset""])\n                content = file_states.get(file_path, """")\n                total_lines = len(content.splitlines())\n                target_line = content[:offset].count(""\n"") + 1\n                vp = per_file_viewport.get(file_path)\n                should_emit = False\n                if not vp or vp[1] == 0:\n                    vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                    per_file_viewport[file_path] = vp\n                    should_emit = True\n                else:\n                    vstart, vend = vp\n                    if target_line < vstart or target_line > vend:\n                        vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                        should_emit = True\n                if should_emit and vp:\n                    vstart, vend = vp\n                    cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                    parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                    viewport_output = _line_numbered_output(content, vstart, vend)\n                    parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n\n            case ""terminal_command"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(command_str)))\n\n            case ""terminal_output"":\n                output = row[""Text""]\n                raw_output = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                terminal_output_buffer.append(raw_output)\n\n            case ""terminal_focus"" | ""git_branch_checkout"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                # FIXME (f.srambical): handle these events \n                pass\n\n            case _:\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    _flush_all_pending_edits()\n    _flush_terminal_output_buffer()\n    return ""\n"".join(parts).strip()\n\ndef load_hf_csv(hf_path: str, split: str) -> Dataset:\n    loaded = load_dataset(hf_path, split=split)\n\n    assert isinstance(loaded, Dataset), ""Expected a Dataset from load_dataset""\n    return loaded\n\n\ndef _discover_local_sessions(root: Path) -> List[Path]:\n    # Recursively find all CSV files\n    paths: List[Path] = []\n    for p in root.rglob(""*.csv""):\n        if p.is_file():\n            paths.append(p)\n    paths.sort()\n    return paths\n\n\ndef _chunk_text(text: str, target_chars: int, overlap_chars: int) -> List[str]:\n    """"""Split a long text into overlapping chunks near target length.""""""\n    if target_chars <= 0:\n        return [text]\n    n = len(text)\n    if n <= target_chars:\n        return [text]\n\n    chunks: List[str] = []\n    start = 0\n    # Ensure sane overlap\n    overlap = max(0, min(overlap_chars, target_chars // 2))\n    while start < n:\n        end_target = min(start + target_chars, n)\n        if end_target < n:\n            end = end_target\n        else:\n            end = n\n        chunk = text[start:end].strip()\n        chunks.append(chunk)\n        if end == n:\n            break\n        # advance with overlap\n        start = max(0, end - overlap)\n        if start >= n:\n            break\n    return chunks\n\n\n",python,tab
+2,1550,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"3:24:14 PM [info] Activating crowd-code\n3:24:15 PM [info] Recording started\n3:24:15 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,1856,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"3:24:15 PM [info] Git repository found\n3:24:15 PM [info] Git provider initialized successfully\n3:24:15 PM [info] Initial git state: [object Object]\n",Log,content
+4,2014,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"",python,tab
+5,7234,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+6,7378,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"",python,tab
+7,69985,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset copy.sh",0,0,"#!/bin/bash\n\nset -uex\n\nOUTPUT_DIR=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/array_record/""\nCSV_ROOT=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/csv/""\n\nuv run crowd-pilot/serialize_dataset_array_record.py --csv_root=$CSV_ROOT --output_dir=$OUTPUT_DIR",shellscript,tab
+8,82832,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",0,0,"#!/bin/bash\n\nset -uex\n\nOUTPUT_DIR=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/array_record/""\nCSV_ROOT=""/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/csv/""\n\nuv run crowd-pilot/serialize_dataset_array_record.py --csv_root=$CSV_ROOT --output_dir=$OUTPUT_DIR",shellscript,tab
+9,84098,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",309,0,"",shellscript,selection_mouse
+10,84109,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",308,0,"",shellscript,selection_command
+11,84295,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",210,0,"",shellscript,selection_command
+12,84353,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",208,0,"",shellscript,selection_command
+13,84527,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",120,0,"",shellscript,selection_command
+14,86460,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+15,89601,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",0,0,"",shellscript,tab
+16,90435,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",23,0,"",shellscript,selection_command
+17,90561,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",33,0,"",shellscript,selection_command
+18,90755,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",36,0,"",shellscript,selection_command
+19,90915,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",40,0,"",shellscript,selection_command
+20,91336,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",121,0,"",shellscript,selection_command
+21,91489,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",120,0,"",shellscript,selection_command
+22,92041,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",119,0,"",shellscript,selection_command
+23,93224,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",119,0,"_",shellscript,content
+24,93225,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",120,0,"",shellscript,selection_keyboard
+25,93583,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",120,0,"b",shellscript,content
+26,93584,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",121,0,"",shellscript,selection_keyboard
+27,93852,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",121,0,"a",shellscript,content
+28,93852,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",122,0,"",shellscript,selection_keyboard
+29,94011,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",122,0,"s",shellscript,content
+30,94011,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",123,0,"",shellscript,selection_keyboard
+31,94030,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",123,0,"h",shellscript,content
+32,94030,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",124,0,"",shellscript,selection_keyboard
+33,94332,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",124,0,"_",shellscript,content
+34,94332,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",125,0,"",shellscript,selection_keyboard
+35,94358,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",125,0,"v",shellscript,content
+36,94358,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",126,0,"",shellscript,selection_keyboard
+37,94541,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",126,0,"e",shellscript,content
+38,94541,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",127,0,"",shellscript,selection_keyboard
+39,94591,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",127,0,"r",shellscript,content
+40,94591,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",128,0,"",shellscript,selection_keyboard
+41,94667,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",128,0,"s",shellscript,content
+42,94667,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",129,0,"",shellscript,selection_keyboard
+43,94668,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",129,0,"i",shellscript,content
+44,94668,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",130,0,"",shellscript,selection_keyboard
+45,94669,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",130,0,"o",shellscript,content
+46,94670,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",131,0,"",shellscript,selection_keyboard
+47,94767,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",131,0,"n",shellscript,content
+48,94768,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",132,0,"",shellscript,selection_keyboard
+49,94932,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",131,0,"",shellscript,selection_command
+50,95874,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",221,0,"",shellscript,selection_command
+51,96027,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",223,0,"",shellscript,selection_command
+52,97605,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",223,0,"j",shellscript,content
+53,97605,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",224,0,"",shellscript,selection_keyboard
+54,97897,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",223,0,"",shellscript,selection_command
+55,99948,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",223,1,"",shellscript,content
+56,100925,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",224,0,"",shellscript,selection_command
+57,101055,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",223,0,"",shellscript,selection_command
+58,101200,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",135,0,"",shellscript,selection_command
+59,101334,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",23,0,"",shellscript,selection_command
+60,101532,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",33,0,"",shellscript,selection_command
+61,101713,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",36,0,"",shellscript,selection_command
+62,102117,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",35,0,"",shellscript,selection_command
+63,103528,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",34,0,"",shellscript,selection_command
+64,103901,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",35,0,"",shellscript,selection_command
+65,103978,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",35,1,"/",shellscript,selection_command
+66,105022,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",35,99,"/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/array_record_bash_version/""",shellscript,selection_command
+67,105481,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",35,98,"/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/array_record_bash_version/",shellscript,selection_command
+68,105929,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",132,0,"",shellscript,selection_command
+69,106370,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+70,112319,"TERMINAL",0,0,"",,terminal_focus
+71,112319,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",0,0,"",shellscript,tab
+72,113248,"TERMINAL",0,0,"source /home/franz.srambical/crowd-pilot/.venv/bin/activate",,terminal_command
+73,113283,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+74,114872,"TERMINAL",0,0,"ls /fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/array_record_bash_version/",,terminal_command
+75,114874,"TERMINAL",0,0,"]633;Cls: cannot access '/fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/array_record_bash_version/': No such file or directory\r\n]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+76,121081,"TERMINAL",0,0,"ls /fast/project/HFMI_SynergyUnit/jafar_ws/data/crowd-pilot/crowd-code-0.1/",,terminal_command
+77,121081,"TERMINAL",0,0,"]633;Carray_record  bash_format_array_record  csv  parquet\r\n]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+78,125800,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",107,0,"",shellscript,selection_command
+79,126984,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",107,25,"",shellscript,content
+80,127314,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",107,0,"b",shellscript,content
+81,127314,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",108,0,"",shellscript,selection_keyboard
+82,127559,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",108,0,"a",shellscript,content
+83,127559,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",109,0,"",shellscript,selection_keyboard
+84,127608,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",109,0,"s",shellscript,content
+85,127608,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",110,0,"",shellscript,selection_keyboard
+86,127619,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",110,0,"h",shellscript,content
+87,127619,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",111,0,"",shellscript,selection_keyboard
+88,128009,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",111,0,"\n",shellscript,content
+89,128688,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",111,1,"",shellscript,content
+90,129078,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",111,0,"_",shellscript,content
+91,129078,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",112,0,"",shellscript,selection_keyboard
+92,129305,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",112,0,"v",shellscript,content
+93,129305,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",113,0,"",shellscript,selection_keyboard
+94,129608,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",112,1,"",shellscript,content
+95,129624,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",112,0,"f",shellscript,content
+96,129625,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",113,0,"",shellscript,selection_keyboard
+97,129632,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",113,0,"o",shellscript,content
+98,129632,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",114,0,"",shellscript,selection_keyboard
+99,129733,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",114,0,"r",shellscript,content
+100,129733,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",115,0,"",shellscript,selection_keyboard
+101,129803,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",115,0,"m",shellscript,content
+102,129803,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",116,0,"",shellscript,selection_keyboard
+103,130015,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",116,0,"a",shellscript,content
+104,130016,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",117,0,"",shellscript,selection_keyboard
+105,130016,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",117,0,"t",shellscript,content
+106,130016,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",118,0,"",shellscript,selection_keyboard
+107,130320,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",118,0,"_",shellscript,content
+108,130320,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",119,0,"",shellscript,selection_keyboard
+109,130452,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",119,0,"a",shellscript,content
+110,130453,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",120,0,"",shellscript,selection_keyboard
+111,130453,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",120,0,"r",shellscript,content
+112,130453,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",121,0,"",shellscript,selection_keyboard
+113,130585,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",121,0,"r",shellscript,content
+114,130585,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",122,0,"",shellscript,selection_keyboard
+115,130756,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",122,0,"a",shellscript,content
+116,130756,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",123,0,"",shellscript,selection_keyboard
+117,130756,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",123,0,"y",shellscript,content
+118,130756,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",124,0,"",shellscript,selection_keyboard
+119,131039,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",124,0,"_",shellscript,content
+120,131039,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",125,0,"",shellscript,selection_keyboard
+121,131129,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",125,0,"r",shellscript,content
+122,131129,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",126,0,"",shellscript,selection_keyboard
+123,131194,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",126,0,"e",shellscript,content
+124,131194,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",127,0,"",shellscript,selection_keyboard
+125,131368,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",127,0,"c",shellscript,content
+126,131368,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",128,0,"",shellscript,selection_keyboard
+127,131435,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",128,0,"o",shellscript,content
+128,131435,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",129,0,"",shellscript,selection_keyboard
+129,131536,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",129,0,"r",shellscript,content
+130,131536,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",130,0,"",shellscript,selection_keyboard
+131,131705,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",130,0,"d",shellscript,content
+132,131706,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",131,0,"",shellscript,selection_keyboard
+133,132112,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",130,0,"",shellscript,selection_command
+134,132345,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",134,0,"",shellscript,selection_command
+135,132417,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",135,0,"",shellscript,selection_command
+136,132960,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",134,0,"",shellscript,selection_command
+137,136592,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",135,0,"",shellscript,selection_command
+138,137296,"slurm/dev/franz/berlin/crowd-pilot/generate_array_record_dataset_bash_version.sh",134,0,"",shellscript,selection_command
+139,167938,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"",python,tab
+140,169399,"crowd-pilot/crowd-pilot/serialization_utils.py",14370,0,"",python,selection_command
+141,169571,"crowd-pilot/crowd-pilot/serialization_utils.py",14344,0,"",python,selection_command
+142,172766,"crowd-pilot/crowd-pilot/serialization_utils.py",14334,0,"",python,selection_command
+143,172931,"crowd-pilot/crowd-pilot/serialization_utils.py",14302,0,"",python,selection_command
+144,173139,"crowd-pilot/crowd-pilot/serialization_utils.py",14326,0,"\n        ",python,content
+145,173254,"crowd-pilot/crowd-pilot/serialization_utils.py",14335,0,"\n        ",python,content
+146,173254,"crowd-pilot/crowd-pilot/serialization_utils.py",14327,8,"",python,content
+147,173931,"crowd-pilot/crowd-pilot/serialization_utils.py",14336,0,"i",python,content
+148,173932,"crowd-pilot/crowd-pilot/serialization_utils.py",14337,0,"",python,selection_keyboard
+149,174127,"crowd-pilot/crowd-pilot/serialization_utils.py",14337,0,"f",python,content
+150,174128,"crowd-pilot/crowd-pilot/serialization_utils.py",14338,0,"",python,selection_keyboard
+151,174129,"crowd-pilot/crowd-pilot/serialization_utils.py",14338,0," ",python,content
+152,174129,"crowd-pilot/crowd-pilot/serialization_utils.py",14339,0,"",python,selection_keyboard
+153,176988,"crowd-pilot/crowd-pilot/serialization_utils.py",14339,0,"i",python,content
+154,176988,"crowd-pilot/crowd-pilot/serialization_utils.py",14340,0,"",python,selection_keyboard
+155,177224,"crowd-pilot/crowd-pilot/serialization_utils.py",14340,0," ",python,content
+156,177225,"crowd-pilot/crowd-pilot/serialization_utils.py",14341,0,"",python,selection_keyboard
+157,177794,"crowd-pilot/crowd-pilot/serialization_utils.py",14341,0,"%",python,content
+158,177794,"crowd-pilot/crowd-pilot/serialization_utils.py",14342,0,"",python,selection_keyboard
+159,178081,"crowd-pilot/crowd-pilot/serialization_utils.py",14342,0," ",python,content
+160,178082,"crowd-pilot/crowd-pilot/serialization_utils.py",14343,0,"",python,selection_keyboard
+161,179224,"crowd-pilot/crowd-pilot/serialization_utils.py",14343,0,"1",python,content
+162,179225,"crowd-pilot/crowd-pilot/serialization_utils.py",14344,0,"",python,selection_keyboard
+163,179441,"crowd-pilot/crowd-pilot/serialization_utils.py",14344,0,"0",python,content
+164,179441,"crowd-pilot/crowd-pilot/serialization_utils.py",14345,0,"",python,selection_keyboard
+165,179615,"crowd-pilot/crowd-pilot/serialization_utils.py",14345,0,"0",python,content
+166,179615,"crowd-pilot/crowd-pilot/serialization_utils.py",14346,0,"",python,selection_keyboard
+167,180088,"crowd-pilot/crowd-pilot/serialization_utils.py",14346,0," ",python,content
+168,180089,"crowd-pilot/crowd-pilot/serialization_utils.py",14347,0,"",python,selection_keyboard
+169,180366,"crowd-pilot/crowd-pilot/serialization_utils.py",14347,0,"=",python,content
+170,180367,"crowd-pilot/crowd-pilot/serialization_utils.py",14348,0,"",python,selection_keyboard
+171,180441,"crowd-pilot/crowd-pilot/serialization_utils.py",14348,0,"=",python,content
+172,180442,"crowd-pilot/crowd-pilot/serialization_utils.py",14349,0,"",python,selection_keyboard
+173,180729,"crowd-pilot/crowd-pilot/serialization_utils.py",14349,0," ",python,content
+174,180730,"crowd-pilot/crowd-pilot/serialization_utils.py",14350,0,"",python,selection_keyboard
+175,181059,"crowd-pilot/crowd-pilot/serialization_utils.py",14350,0,"0",python,content
+176,181059,"crowd-pilot/crowd-pilot/serialization_utils.py",14351,0,"",python,selection_keyboard
+177,181404,"crowd-pilot/crowd-pilot/serialization_utils.py",14351,0,":",python,content
+178,181404,"crowd-pilot/crowd-pilot/serialization_utils.py",14352,0,"",python,selection_keyboard
+179,181681,"crowd-pilot/crowd-pilot/serialization_utils.py",14352,0,"\n            ",python,content
+180,221617,"crowd-pilot/crowd-pilot/serialization_utils.py",14365,0,"b",python,content
+181,221618,"crowd-pilot/crowd-pilot/serialization_utils.py",14366,0,"",python,selection_keyboard
+182,221810,"crowd-pilot/crowd-pilot/serialization_utils.py",14366,0,"r",python,content
+183,221811,"crowd-pilot/crowd-pilot/serialization_utils.py",14367,0,"",python,selection_keyboard
+184,221857,"crowd-pilot/crowd-pilot/serialization_utils.py",14367,0,"e",python,content
+185,221857,"crowd-pilot/crowd-pilot/serialization_utils.py",14368,0,"",python,selection_keyboard
+186,222047,"crowd-pilot/crowd-pilot/serialization_utils.py",14368,0,"a",python,content
+187,222047,"crowd-pilot/crowd-pilot/serialization_utils.py",14369,0,"",python,selection_keyboard
+188,222088,"crowd-pilot/crowd-pilot/serialization_utils.py",14369,0,"k",python,content
+189,222088,"crowd-pilot/crowd-pilot/serialization_utils.py",14370,0,"",python,selection_keyboard
+190,222192,"crowd-pilot/crowd-pilot/serialization_utils.py",14370,0,"p",python,content
+191,222192,"crowd-pilot/crowd-pilot/serialization_utils.py",14371,0,"",python,selection_keyboard
+192,222268,"crowd-pilot/crowd-pilot/serialization_utils.py",14371,0,"o",python,content
+193,222269,"crowd-pilot/crowd-pilot/serialization_utils.py",14372,0,"",python,selection_keyboard
+194,222270,"crowd-pilot/crowd-pilot/serialization_utils.py",14372,0,"i",python,content
+195,222270,"crowd-pilot/crowd-pilot/serialization_utils.py",14373,0,"",python,selection_keyboard
+196,222351,"crowd-pilot/crowd-pilot/serialization_utils.py",14373,0,"n",python,content
+197,222352,"crowd-pilot/crowd-pilot/serialization_utils.py",14374,0,"",python,selection_keyboard
+198,222515,"crowd-pilot/crowd-pilot/serialization_utils.py",14374,0,"t",python,content
+199,222516,"crowd-pilot/crowd-pilot/serialization_utils.py",14375,0,"",python,selection_keyboard
+200,222952,"crowd-pilot/crowd-pilot/serialization_utils.py",14375,0,"()",python,content
+201,222952,"crowd-pilot/crowd-pilot/serialization_utils.py",14376,0,"",python,selection_keyboard
+202,223008,"crowd-pilot/crowd-pilot/serialization_utils.py",14376,1,")",python,content
+203,223009,"crowd-pilot/crowd-pilot/serialization_utils.py",14377,0,"",python,selection_keyboard
+204,223280,"crowd-pilot/crowd-pilot/serialization_utils.py",14376,0,"",python,selection_command
+205,226101,"crowd-pilot/crowd-pilot/serialize_dataset_array_record.py",0,0,"#!/usr/bin/env python3\n""""""\nCSV sessions -> ArrayRecord shards for MaxText Grain pretraining.\n""""""\n\nfrom __future__ import annotations\n\nimport argparse\nimport os\nfrom pathlib import Path\nfrom typing import List, Tuple, cast\nimport random\n\nimport pandas as pd\n\nfrom array_record.python import array_record_module as arm # type: ignore\n\nimport tensorflow as tf \nfrom serialization_utils import (\n    SerializeConfig,\n    session_to_bash_formatted_transcript,\n    _discover_local_sessions,\n    _chunk_text,\n)\n\n\ndef to_array_record(\n    cfg: SerializeConfig,\n) -> None:\n    os.makedirs(cfg.output_dir, exist_ok=True)\n\n    required_cols = [""Sequence"", ""Time"", ""File"", ""RangeOffset"", ""RangeLength"", ""Text"", ""Language"", ""Type""]\n\n    session_dataframes: List[Tuple[pd.DataFrame, str]] = []\n    root = Path(cast(str, cfg.csv_root)).expanduser().resolve()\n    csv_files = _discover_local_sessions(root)\n    assert csv_files, f""No CSV files found under {root}""\n    for csv_file in csv_files:\n        df = pd.read_csv(csv_file)\n        missing_local = [c for c in required_cols if c not in df.columns]\n        assert not missing_local, f""Missing required CSV columns in {csv_file}: {missing_local}""\n        session_dataframes.append((df, str(csv_file)))\n\n    random.seed(42)\n    session_dataframes = [(df, path) for df, path in session_dataframes]\n    random.shuffle(session_dataframes)\n    \n    total_sessions = len(session_dataframes)\n    val_count = int(total_sessions * cfg.val_ratio)\n    train_count = total_sessions - val_count\n\n    train_rows = 0\n    val_rows = 0\n    train_shard_idx = 0\n    val_shard_idx = 0\n    docs_written = 0\n\n    def write_shard(chunks: List[str], split: str, shard_idx: int) -> int:\n        if not chunks:\n            return 0\n        out_path = Path(cfg.output_dir) / f""{split}_{shard_idx:05d}.array_record""\n        group_size = cfg.arrayrecord_group_size\n        options = f""group_size:{group_size}""\n        writer = arm.ArrayRecordWriter(str(out_path), options)\n        try:\n            for chunk in chunks:\n                example = tf.train.Example(\n                    features=tf.train.Features(\n                        feature={\n                            ""text"": tf.train.Feature(\n                                bytes_list=tf.train.BytesList(value=[chunk.encode(""utf-8"")])\n                            )\n                        }\n                    )\n                )\n                writer.write(example.SerializeToString())\n        finally:\n            writer.close()\n        return len(chunks)\n\n    for i, (session_df, session_path) in enumerate(session_dataframes):\n        session_df = pd.DataFrame(session_df.copy())\n        transcript = session_to_bash_formatted_transcript(\n            session_df,\n        )\n        if len(transcript.strip()) < cfg.min_session_chars:\n            print(f""Skipping session {session_path} because it's too short ({len(transcript.strip())} chars)"")\n            continue\n        chunks = _chunk_text(transcript, cfg.target_chars, cfg.overlap_chars)\n        if not chunks:\n            continue\n        docs_written += len(chunks)\n        \n        if i < train_count:\n            rows_written = write_shard(chunks, ""train"", train_shard_idx)\n            train_rows += rows_written\n            train_shard_idx += 1\n        else:\n            rows_written = write_shard(chunks, ""val"", val_shard_idx)\n            val_rows += rows_written\n            val_shard_idx += 1\n            \n        if cfg.max_docs and docs_written >= cfg.max_docs:\n            break\n\n    print(f""Wrote {train_rows} train and {val_rows} val documents to {cfg.output_dir}"")\n\n\ndef parse_args() -> SerializeConfig:\n    p = argparse.ArgumentParser(description=""Serialize HF CSV sessions to ArrayRecord for MaxText Grain"")\n    p.add_argument(""--csv_root"", type=str, required=True, help=""Root directory containing per-session CSV files"")\n    p.add_argument(""--output_dir"", type=str, required=True, help=""Output directory for ArrayRecord shards"")\n    p.add_argument(""--shard_size"", type=int, default=20000, help=""Rows per shard (currently one session per shard)"")\n    # FIXME(f.srambical): It is awkward that the target number is in character-space instead of in token-space.\n    p.add_argument(""--target_chars"", type=int, default=8192, help=""Target characters per document chunk. This should be ~3-4x the max token length of the model you are using."")\n    p.add_argument(""--overlap_chars"", type=int, default=128, help=""Character overlap between chunks"")\n    p.add_argument(""--min_session_chars"", type=int, default=1024, help=""Minimum characters to keep a session"")\n    p.add_argument(""--max_docs"", type=int, default=None, help=""Stop after writing this many unique docs"")\n    p.add_argument(""--long_pause_threshold_ms"", type=int, default=120000, help=""Threshold (ms) to annotate long pauses and emit a keyframe"")\n    p.add_argument(""--val_ratio"", type=float, default=0.10, help=""Fraction of sessions to route to validation [0,1)"")\n    p.add_argument(""--arrayrecord_group_size"", type=int, default=1, help=""ArrayRecord group_size option controlling index granularity and compression grouping"")\n    args = p.parse_args()\n    return SerializeConfig(\n        output_dir=args.output_dir,\n        shard_size=args.shard_size,\n        target_chars=args.target_chars,\n        overlap_chars=args.overlap_chars,\n        min_session_chars=args.min_session_chars,\n        max_docs=args.max_docs,\n        long_pause_threshold_ms=args.long_pause_threshold_ms,\n        csv_root=(args.csv_root if args.csv_root else None),\n        val_ratio=args.val_ratio,\n        arrayrecord_group_size=args.arrayrecord_group_size,\n    )\n\n\ndef main() -> None:\n    cfg = parse_args()\n    to_array_record(cfg)\n\n\nif __name__ == ""__main__"":\n    main()",python,tab
+206,226435,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"",python,tab
+207,228553,"crowd-pilot/crowd-pilot/serialization_utils.py",14377,0,"",python,selection_mouse
+208,228561,"crowd-pilot/crowd-pilot/serialization_utils.py",14376,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-a313d008-5546-415a-a27c-b4bbbd49fb041754912780018-2025_08_11-13.46.25.836/source.csv

@@ -0,0 +1,69 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,313,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"1:46:25 PM [info] Activating crowd-code\n1:46:25 PM [info] Recording started\n1:46:25 PM [info] Initializing git provider using file system watchers...\n1:46:25 PM [info] Git repository found\n1:46:25 PM [info] Git provider initialized successfully\n1:46:25 PM [info] Initial git state: [object Object]\n",Log,tab
+3,11854,"sz.py",0,0,"#!/usr/bin/env python3\nimport os, sys\nimport token\nimport tokenize\nimport itertools\nfrom tabulate import tabulate\n\nTOKEN_WHITELIST = [token.OP, token.NAME, token.NUMBER, token.STRING]\n\ndef is_docstring(t):\n  return t.type == token.STRING and t.string.startswith('""""""') and t.line.strip().startswith('""""""')\n\ndef is_js_token(s): return len(s) and not s.startswith('//')\n\ndef gen_stats(base_path="".""):\n  table = []\n  for path, _, files in os.walk(os.path.join(base_path, ""tinygrad"")):\n    for name in files:\n      if not (name.endswith("".py"") or name.endswith("".js"")): continue\n      if any(s in path.replace('\\', '/') for s in ['tinygrad/runtime/autogen', 'tinygrad/viz/assets']): continue\n      filepath = os.path.join(path, name)\n      relfilepath = os.path.relpath(filepath, base_path).replace('\\', '/')\n      if name.endswith("".js""):\n        with open(filepath) as file_: lines = [line.strip() for line in file_.readlines()]\n        token_count, line_count = sum(len(line.split()) for line in lines if is_js_token(line)), sum(1 for line in lines if is_js_token(line))\n      else:\n        with tokenize.open(filepath) as file_:\n          tokens = [t for t in tokenize.generate_tokens(file_.readline) if t.type in TOKEN_WHITELIST and not is_docstring(t)]\n          token_count, line_count = len(tokens), len(set([x for t in tokens for x in range(t.start[0], t.end[0]+1)]))\n      if line_count > 0: table.append([relfilepath, line_count, token_count/line_count])\n  return table\n\ndef gen_diff(table_old, table_new):\n  table = []\n  files_new = set([x[0] for x in table_new])\n  files_old = set([x[0] for x in table_old])\n  added, deleted, unchanged = files_new - files_old, files_old - files_new, files_new & files_old\n  if added:\n    for file in added:\n      file_stat = [stats for stats in table_new if file in stats]\n      table.append([file_stat[0][0], file_stat[0][1], file_stat[0][1]-0, file_stat[0][2], file_stat[0][2]-0])\n  if deleted:\n    for file in deleted:\n      file_stat = [stats for stats in table_old if file in stats]\n      table.append([file_stat[0][0], 0, 0 - file_stat[0][1], 0, 0-file_stat[0][2]])\n  if unchanged:\n    for file in unchanged:\n      file_stat_old = [stats for stats in table_old if file in stats]\n      file_stat_new = [stats for stats in table_new if file in stats]\n      if file_stat_new[0][1]-file_stat_old[0][1] != 0 or file_stat_new[0][2]-file_stat_old[0][2] != 0:\n        table.append([file_stat_new[0][0], file_stat_new[0][1], file_stat_new[0][1]-file_stat_old[0][1], file_stat_new[0][2],\n                      file_stat_new[0][2]-file_stat_old[0][2]])\n  return table\n\ndef display_diff(diff): return ""+""+str(diff) if diff > 0 else str(diff)\n\nif __name__ == ""__main__"":\n  if len(sys.argv) == 3:\n    headers = [""Name"", ""Lines"", ""Diff"", ""Tokens/Line"", ""Diff""]\n    table = gen_diff(gen_stats(sys.argv[1]), gen_stats(sys.argv[2]))\n  elif len(sys.argv) == 2:\n    headers = [""Name"", ""Lines"", ""Tokens/Line""]\n    table = gen_stats(sys.argv[1])\n  else:\n    headers = [""Name"", ""Lines"", ""Tokens/Line""]\n    table = gen_stats(""."")\n\n  if table:\n    if len(sys.argv) == 3:\n      print(""### Changes"")\n      print(""```"")\n      print(tabulate([headers] + sorted(table, key=lambda x: -x[1]), headers=""firstrow"", intfmt=(..., ""d"", ""+d""),\n                     floatfmt=(..., ..., ..., "".1f"", ""+.1f""))+""\n"")\n      print(f""\ntotal lines changes: {display_diff(sum([x[2] for x in table]))}"")\n      print(""```"")\n    else:\n      print(tabulate([headers] + sorted(table, key=lambda x: -x[1]), headers=""firstrow"", floatfmt="".1f"")+""\n"")\n      groups = sorted([('/'.join(x[0].rsplit(""/"", 1)[0].split(""/"")[0:2]), x[1], x[2]) for x in table])\n      for dir_name, group in itertools.groupby(groups, key=lambda x:x[0]):\n        print(f""{dir_name:30s} : {sum([x[1] for x in group]):6d}"")\n      total_lines = sum([x[1] for x in table])\n      print(f""\ntotal line count: {total_lines}"")\n      max_line_count = int(os.getenv(""MAX_LINE_COUNT"", ""-1""))\n      assert max_line_count == -1 or total_lines <= max_line_count, f""OVER {max_line_count} LINES""\n",python,tab
+4,17972,"test_driven_development.sh",0,0,"#!/bin/bash\npython3 test/external/process_replay/reset.py\nCAPTURE_PROCESS_REPLAY=1 pytest -n auto test/test_tiny.py test/test_uop_graph.py test/test_ops.py test/test_linearizer.py\nwhile true; do\n  if python3 test/test_tiny.py; then\n    PYTHONPATH=""."" python3 test/external/process_replay/process_replay.py\n  fi\ndone\n",shellscript,tab
+5,80622,"examples/stable_diffusion.py",0,0,"# https://arxiv.org/pdf/2112.10752.pdf\n# https://github.com/ekagra-ranjan/huggingface-blog/blob/main/stable_diffusion.md\nimport tempfile\nfrom pathlib import Path\nimport argparse\nfrom collections import namedtuple\nfrom typing import Dict, Any\n\nfrom PIL import Image\nimport numpy as np\nfrom tinygrad import Device, GlobalCounters, dtypes, Tensor, TinyJit\nfrom tinygrad.helpers import Timing, Context, getenv, fetch, colored, tqdm\nfrom tinygrad.nn import Conv2d, GroupNorm\nfrom tinygrad.nn.state import torch_load, load_state_dict, get_state_dict\nfrom extra.models.clip import Closed, Tokenizer\nfrom extra.models.unet import UNetModel\nfrom extra.bench_log import BenchEvent, WallTimeEvent\n\nclass AttnBlock:\n  def __init__(self, in_channels):\n    self.norm = GroupNorm(32, in_channels)\n    self.q = Conv2d(in_channels, in_channels, 1)\n    self.k = Conv2d(in_channels, in_channels, 1)\n    self.v = Conv2d(in_channels, in_channels, 1)\n    self.proj_out = Conv2d(in_channels, in_channels, 1)\n\n  # copied from AttnBlock in ldm repo\n  def __call__(self, x):\n    h_ = self.norm(x)\n    q,k,v = self.q(h_), self.k(h_), self.v(h_)\n\n    # compute attention\n    b,c,h,w = q.shape\n    q,k,v = [x.reshape(b,c,h*w).transpose(1,2) for x in (q,k,v)]\n    h_ = Tensor.scaled_dot_product_attention(q,k,v).transpose(1,2).reshape(b,c,h,w)\n    return x + self.proj_out(h_)\n\nclass ResnetBlock:\n  def __init__(self, in_channels, out_channels=None):\n    self.norm1 = GroupNorm(32, in_channels)\n    self.conv1 = Conv2d(in_channels, out_channels, 3, padding=1)\n    self.norm2 = GroupNorm(32, out_channels)\n    self.conv2 = Conv2d(out_channels, out_channels, 3, padding=1)\n    self.nin_shortcut = Conv2d(in_channels, out_channels, 1) if in_channels != out_channels else lambda x: x\n\n  def __call__(self, x):\n    h = self.conv1(self.norm1(x).swish())\n    h = self.conv2(self.norm2(h).swish())\n    return self.nin_shortcut(x) + h\n\nclass Mid:\n  def __init__(self, block_in):\n    self.block_1 = ResnetBlock(block_in, block_in)\n    self.attn_1 = AttnBlock(block_in)\n    self.block_2 = ResnetBlock(block_in, block_in)\n\n  def __call__(self, x):\n    return x.sequential([self.block_1, self.attn_1, self.block_2])\n\nclass Decoder:\n  def __init__(self):\n    sz = [(128, 256), (256, 512), (512, 512), (512, 512)]\n    self.conv_in = Conv2d(4,512,3, padding=1)\n    self.mid = Mid(512)\n\n    arr = []\n    for i,s in enumerate(sz):\n      arr.append({""block"":\n        [ResnetBlock(s[1], s[0]),\n         ResnetBlock(s[0], s[0]),\n         ResnetBlock(s[0], s[0])]})\n      if i != 0: arr[-1]['upsample'] = {""conv"": Conv2d(s[0], s[0], 3, padding=1)}\n    self.up = arr\n\n    self.norm_out = GroupNorm(32, 128)\n    self.conv_out = Conv2d(128, 3, 3, padding=1)\n\n  def __call__(self, x):\n    x = self.conv_in(x)\n    x = self.mid(x)\n\n    for l in self.up[::-1]:\n      print(""decode"", x.shape)\n      for b in l['block']: x = b(x)\n      if 'upsample' in l:\n        # https://pytorch.org/docs/stable/generated/torch.nn.functional.interpolate.html ?\n        bs,c,py,px = x.shape\n        x = x.reshape(bs, c, py, 1, px, 1).expand(bs, c, py, 2, px, 2).reshape(bs, c, py*2, px*2)\n        x = l['upsample']['conv'](x)\n      x.realize()\n\n    return self.conv_out(self.norm_out(x).swish())\n\nclass Encoder:\n  def __init__(self):\n    sz = [(128, 128), (128, 256), (256, 512), (512, 512)]\n    self.conv_in = Conv2d(3,128,3, padding=1)\n\n    arr = []\n    for i,s in enumerate(sz):\n      arr.append({""block"":\n        [ResnetBlock(s[0], s[1]),\n         ResnetBlock(s[1], s[1])]})\n      if i != 3: arr[-1]['downsample'] = {""conv"": Conv2d(s[1], s[1], 3, stride=2, padding=(0,1,0,1))}\n    self.down = arr\n\n    self.mid = Mid(512)\n    self.norm_out = GroupNorm(32, 512)\n    self.conv_out = Conv2d(512, 8, 3, padding=1)\n\n  def __call__(self, x):\n    x = self.conv_in(x)\n\n    for l in self.down:\n      print(""encode"", x.shape)\n      for b in l['block']: x = b(x)\n      if 'downsample' in l: x = l['downsample']['conv'](x)\n\n    x = self.mid(x)\n    return self.conv_out(self.norm_out(x).swish())\n\nclass AutoencoderKL:\n  def __init__(self):\n    self.encoder = Encoder()\n    self.decoder = Decoder()\n    self.quant_conv = Conv2d(8, 8, 1)\n    self.post_quant_conv = Conv2d(4, 4, 1)\n\n  def __call__(self, x):\n    latent = self.encoder(x)\n    latent = self.quant_conv(latent)\n    latent = latent[:, 0:4]  # only the means\n    print(""latent"", latent.shape)\n    latent = self.post_quant_conv(latent)\n    return self.decoder(latent)\n\ndef get_alphas_cumprod(beta_start=0.00085, beta_end=0.0120, n_training_steps=1000):\n  betas = np.linspace(beta_start ** 0.5, beta_end ** 0.5, n_training_steps, dtype=np.float32) ** 2\n  alphas = 1.0 - betas\n  alphas_cumprod = np.cumprod(alphas, axis=0)\n  return Tensor(alphas_cumprod)\n\nunet_params: Dict[str,Any] = {\n  ""adm_in_ch"": None,\n  ""in_ch"": 4,\n  ""out_ch"": 4,\n  ""model_ch"": 320,\n  ""attention_resolutions"": [4, 2, 1],\n  ""num_res_blocks"": 2,\n  ""channel_mult"": [1, 2, 4, 4],\n  ""n_heads"": 8,\n  ""transformer_depth"": [1, 1, 1, 1],\n  ""ctx_dim"": 768,\n  ""use_linear"": False,\n}\n\nclass StableDiffusion:\n  def __init__(self):\n    self.alphas_cumprod = get_alphas_cumprod()\n    self.model = namedtuple(""DiffusionModel"", [""diffusion_model""])(diffusion_model = UNetModel(**unet_params))\n    self.first_stage_model = AutoencoderKL()\n    self.cond_stage_model = namedtuple(""CondStageModel"", [""transformer""])(transformer = namedtuple(""Transformer"", [""text_model""])(text_model = Closed.ClipTextTransformer()))\n\n  def get_x_prev_and_pred_x0(self, x, e_t, a_t, a_prev):\n    temperature = 1\n    sigma_t = 0\n    sqrt_one_minus_at = (1-a_t).sqrt()\n    #print(a_t, a_prev, sigma_t, sqrt_one_minus_at)\n\n    pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()\n\n    # direction pointing to x_t\n    dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t\n\n    x_prev = a_prev.sqrt() * pred_x0 + dir_xt\n    return x_prev, pred_x0\n\n  def get_model_output(self, unconditional_context, context, latent, timestep, unconditional_guidance_scale):\n    # put into diffuser\n    latents = self.model.diffusion_model(latent.expand(2, *latent.shape[1:]), timestep, unconditional_context.cat(context, dim=0))\n    unconditional_latent, latent = latents[0:1], latents[1:2]\n\n    e_t = unconditional_latent + unconditional_guidance_scale * (latent - unconditional_latent)\n    return e_t\n\n  def decode(self, x):\n    x = self.first_stage_model.post_quant_conv(1/0.18215 * x)\n    x = self.first_stage_model.decoder(x)\n\n    # make image correct size and scale\n    x = (x + 1.0) / 2.0\n    x = x.reshape(3,512,512).permute(1,2,0).clip(0,1)*255\n    return x.cast(dtypes.uint8)\n\n  def __call__(self, unconditional_context, context, latent, timestep, alphas, alphas_prev, guidance):\n    e_t = self.get_model_output(unconditional_context, context, latent, timestep, guidance)\n    x_prev, _ = self.get_x_prev_and_pred_x0(latent, e_t, alphas, alphas_prev)\n    #e_t_next = get_model_output(x_prev)\n    #e_t_prime = (e_t + e_t_next) / 2\n    #x_prev, pred_x0 = get_x_prev_and_pred_x0(latent, e_t_prime, index)\n    return x_prev.realize()\n\n# ** ldm.models.autoencoder.AutoencoderKL (done!)\n# 3x512x512 <--> 4x64x64 (16384)\n# decode torch.Size([1, 4, 64, 64]) torch.Size([1, 3, 512, 512])\n# section 4.3 of paper\n# first_stage_model.encoder, first_stage_model.decoder\n\n# ** ldm.modules.diffusionmodules.openaimodel.UNetModel\n# this is what runs each time to sample. is this the LDM?\n# input:  4x64x64\n# output: 4x64x64\n# model.diffusion_model\n# it has attention?\n\n# ** ldm.modules.encoders.modules.FrozenCLIPEmbedder\n# cond_stage_model.transformer.text_model\n\nif __name__ == ""__main__"":\n  default_prompt = ""a horse sized cat eating a bagel""\n  parser = argparse.ArgumentParser(description='Run Stable Diffusion', formatter_class=argparse.ArgumentDefaultsHelpFormatter)\n  parser.add_argument('--steps', type=int, default=6, help=""Number of steps in diffusion"")\n  parser.add_argument('--prompt', type=str, default=default_prompt, help=""Phrase to render"")\n  parser.add_argument('--out', type=str, default=Path(tempfile.gettempdir()) / ""rendered.png"", help=""Output filename"")\n  parser.add_argument('--noshow', action='store_true', help=""Don't show the image"")\n  parser.add_argument('--fp16', action='store_true', help=""Cast the weights to float16"")\n  parser.add_argument('--timing', action='store_true', help=""Print timing per step"")\n  parser.add_argument('--seed', type=int, help=""Set the random latent seed"")\n  parser.add_argument('--guidance', type=float, default=7.5, help=""Prompt strength"")\n  args = parser.parse_args()\n\n  model = StableDiffusion()\n\n  # load in weights\n  with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):\n    load_state_dict(model, torch_load(fetch('https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt', 'sd-v1-4.ckpt'))['state_dict'], strict=False)\n\n    if args.fp16:\n      for k,v in get_state_dict(model).items():\n        if k.startswith(""model""):\n          v.replace(v.cast(dtypes.float16).realize())\n\n  # run through CLIP to get context\n  tokenizer = Tokenizer.ClipTokenizer()\n  prompt = Tensor([tokenizer.encode(args.prompt)])\n  context = model.cond_stage_model.transformer.text_model(prompt).realize()\n  print(""got CLIP context"", context.shape)\n\n  prompt = Tensor([tokenizer.encode("""")])\n  unconditional_context = model.cond_stage_model.transformer.text_model(prompt).realize()\n  print(""got unconditional CLIP context"", unconditional_context.shape)\n\n  # done with clip model\n  del model.cond_stage_model\n\n  timesteps = list(range(1, 1000, 1000//args.steps))\n  print(f""running for {timesteps} timesteps"")\n  alphas = model.alphas_cumprod[Tensor(timesteps)]\n  alphas_prev = Tensor([1.0]).cat(alphas[:-1])\n\n  # start with random noise\n  if args.seed is not None: Tensor.manual_seed(args.seed)\n  latent = Tensor.randn(1,4,64,64)\n\n  @TinyJit\n  def run(model, *x): return model(*x).realize()\n\n  # this is diffusion\n  with Context(BEAM=getenv(""LATEBEAM"")):\n    for index, timestep in (t:=tqdm(list(enumerate(timesteps))[::-1])):\n      GlobalCounters.reset()\n      t.set_description(""%3d %3d"" % (index, timestep))\n      with Timing(""step in "", enabled=args.timing, on_exit=lambda _: f"", using {GlobalCounters.mem_used/1e9:.2f} GB""):\n        with WallTimeEvent(BenchEvent.STEP):\n          tid = Tensor([index])\n          latent = run(model, unconditional_context, context, latent, Tensor([timestep]), alphas[tid], alphas_prev[tid], Tensor([args.guidance]))\n          if args.timing: Device[Device.DEFAULT].synchronize()\n    del run\n\n  # upsample latent space to image with autoencoder\n  x = model.decode(latent)\n  print(x.shape)\n\n  # save image\n  im = Image.fromarray(x.numpy())\n  print(f""saving {args.out}"")\n  im.save(args.out)\n  # Open image.\n  if not args.noshow: im.show()\n\n  # validation!\n  if args.prompt == default_prompt and args.steps == 6 and args.seed == 0 and args.guidance == 7.5:\n    ref_image = Tensor(np.array(Image.open(Path(__file__).parent / ""stable_diffusion_seed0.png"")))\n    distance = (((x.cast(dtypes.float) - ref_image.cast(dtypes.float)) / ref_image.max())**2).mean().item()\n    assert distance < 3e-3, colored(f""validation failed with {distance=}"", ""red"")  # higher distance with WINO\n    print(colored(f""output validated with {distance=}"", ""green""))\n",python,tab
+6,120509,"examples/gpt2.py",0,0,"#!/usr/bin/env python3\nimport os, argparse, contextlib\nfrom typing import Optional, Union\nwith contextlib.suppress(ImportError): import tiktoken\nfrom tinygrad import Tensor, TinyJit, Device, GlobalCounters, Variable, dtypes\nfrom tinygrad.uop.ops import UOp\nfrom tinygrad.helpers import Timing, DEBUG, JIT, getenv, fetch, colored, trange\nfrom tinygrad.nn import Embedding, Linear, LayerNorm\nfrom tinygrad.nn.state import gguf_load, torch_load, load_state_dict, get_state_dict\nfrom extra.bench_log import BenchEvent, WallTimeEvent\n\nMAX_CONTEXT = getenv(""MAX_CONTEXT"", 128)\nHALF = getenv(""HALF"")\n\nclass Attention:\n  def __init__(self, dim, n_heads):\n    self.c_attn = Linear(dim, 3*dim, bias=True)\n    self.c_proj = Linear(dim, dim, bias=True)\n    self.n_heads = n_heads\n    self.dim = dim\n    self.head_dim = dim // n_heads\n\n  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]) -> Tensor:\n    if mask is not None or start_pos.val == 0:\n      # no symbolic shape qkv when consuming prompts\n      start_pos = start_pos.val\n\n    if HALF: x = x.half()\n    xqkv = self.c_attn(x)\n    xq, xk, xv = [xqkv.shrink((None, None, (i*self.dim, (i+1)*self.dim))).reshape(None, None, self.n_heads, self.head_dim) for i in range(3)]\n    bsz, seqlen, _, _ = xq.shape\n\n    # create kv cache\n    if not hasattr(self, ""cache_kv""):\n      self.cache_kv = Tensor.zeros(2, bsz, MAX_CONTEXT, self.n_heads, self.head_dim, dtype=x.dtype).contiguous().realize()\n\n    # update the cache\n    self.cache_kv.shrink((None, None,(start_pos,start_pos+seqlen),None,None)).assign(Tensor.stack(xk, xv)).realize()\n\n    if start_pos > 0:\n      keys = self.cache_kv[0].shrink((None, (0, start_pos+seqlen), None, None))\n      values = self.cache_kv[1].shrink((None, (0, start_pos+seqlen), None, None))\n    else:\n      keys = xk\n      values = xv\n\n    xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)\n    return self.c_proj(xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2).reshape(bsz, seqlen, self.dim))\n\nclass FeedForward:\n  def __init__(self, dim, hidden_dim):\n    self.c_fc = Linear(dim, hidden_dim, bias=True)\n    self.c_proj = Linear(hidden_dim, dim, bias=True)\n\n  def __call__(self, x:Tensor) -> Tensor:\n    return self.c_proj(self.c_fc(x).gelu())\n\nclass TransformerBlock:\n  def __init__(self, dim, n_heads, norm_eps):\n    self.attn = Attention(dim, n_heads)\n    self.mlp = FeedForward(dim, 4*dim)\n    self.ln_1 = LayerNorm(dim, norm_eps)\n    self.ln_2 = LayerNorm(dim, norm_eps)\n\n  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]):\n    h = x + self.attn(self.ln_1(x), start_pos, mask).float()\n    return (h + self.mlp(self.ln_2(h)))\n\nclass Transformer:\n  def __init__(self, dim, n_heads, n_layers, norm_eps, vocab_size, max_seq_len=1024):\n    self.vocab_size = vocab_size\n    self.wte = Embedding(vocab_size, dim)\n    self.wpe = Embedding(max_seq_len, dim)\n    self.h = [TransformerBlock(dim, n_heads, norm_eps) for _ in range(n_layers)]\n    self.ln_f = LayerNorm(dim, norm_eps)\n    self.lm_head = Linear(dim, vocab_size, bias=False)\n    self.forward_jit = TinyJit(self.forward)\n\n  def forward(self, tokens:Union[Tensor,UOp], start_pos:Variable, temperature:float=0.0):\n    if not hasattr(self, 'allpos'): self.allpos = Tensor.arange(0, MAX_CONTEXT).reshape(1, -1).realize()\n    if isinstance(tokens, UOp):\n      seqlen = 1\n      tok_emb = self.wte.weight.shrink(((tokens, tokens+1), None))\n    else:\n      seqlen = tokens.shape[1]\n      tok_emb = self.wte(tokens)\n\n    # not symbolic when consuming the prompt\n    selected_pos = (0, seqlen) if start_pos.val == 0 else (start_pos, start_pos+1)\n    pos_emb = self.wpe(self.allpos.shrink((None, selected_pos)))\n\n    h = tok_emb + pos_emb\n\n    if HALF: h = h.half()\n\n    mask = Tensor.full((1, 1, seqlen, start_pos.val+seqlen), float(""-inf""), dtype=h.dtype).triu(start_pos.val+1) if seqlen > 1 else None\n\n    for hi in self.h: h = hi(h, start_pos, mask)\n\n    logits = self.lm_head(self.ln_f(h))\n\n    if logits.shape[1] == 0:\n      # special case for empty prompt\n      logits = Tensor.ones((logits.shape[0], self.vocab_size), dtype=logits.dtype, device=logits.device)\n    else:\n      logits = logits[:, -1, :]\n\n    if temperature < 1e-6:\n      ret = logits.argmax(-1)\n    else:\n      ret = (logits / temperature).softmax().multinomial()\n    return ret.flatten().realize()\n\n  def __call__(self, tokens:Union[Tensor,UOp], start_pos:Variable, temperature:float=0.0) -> Tensor:\n    forward = (self.forward_jit if JIT and (isinstance(tokens, UOp) or tokens.shape[1] == 1) else self.forward)\n    return forward(tokens, start_pos, temperature)\n\nVOCAB_SIZE = 50257\nMODEL_PARAMS = {\n  'gpt2':         dict(n_layers=12, n_heads=12, dim=768, norm_eps=1e-5, vocab_size=VOCAB_SIZE),   # 124M params\n  'gpt2-medium':  dict(n_layers=24, n_heads=16, dim=1024, norm_eps=1e-5, vocab_size=VOCAB_SIZE),  # 350M params\n  'gpt2-large':   dict(n_layers=36, n_heads=20, dim=1280, norm_eps=1e-5, vocab_size=VOCAB_SIZE),  # 774M params\n  'gpt2-xl':      dict(n_layers=48, n_heads=25, dim=1600, norm_eps=1e-5, vocab_size=VOCAB_SIZE),  # 1558M params\n}\n\nclass GPT2:\n  @staticmethod\n  def build(model_size=""gpt2""):\n    tokenizer = tiktoken.get_encoding(""gpt2"")\n\n    model = Transformer(**MODEL_PARAMS[model_size])\n    weights = torch_load(fetch(f'https://huggingface.co/{model_size}/resolve/main/pytorch_model.bin'))\n    # special treatment for the Conv1D weights we need to transpose\n    transposed = ('attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight')\n    for k in weights:\n      if k.endswith(transposed):\n        weights[k] = weights[k].T\n    # lm head and wte are tied\n    weights['lm_head.weight'] = weights['wte.weight']\n\n    with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):\n      load_state_dict(model, weights)\n\n      if HALF:\n        for l in get_state_dict(model).values():\n          l.replace(l.half().realize())\n\n    return GPT2(model, tokenizer)\n\n  @staticmethod\n  def build_gguf(model_size: str):\n    q_type = model_size[len(""gpt2_gguf_""):].upper()\n    fn = fetch(f""https://huggingface.co/PrunaAI/gpt2-GGUF-smashed/resolve/main/gpt2.{q_type}.gguf?download=true"")\n    gguf_tensor = Tensor.empty(os.stat(fn).st_size, dtype=dtypes.uint8, device=f""disk:{fn}"").to(Device.DEFAULT)\n    kv_data, state_dict = gguf_load(gguf_tensor)\n\n    gpt2_params = {\n      ""dim"": kv_data[""gpt2.embedding_length""], ""n_heads"": kv_data[""gpt2.attention.head_count""],\n      ""n_layers"": kv_data[""gpt2.block_count""], ""norm_eps"": kv_data[""gpt2.attention.layer_norm_epsilon""],\n      ""vocab_size"": VOCAB_SIZE, ""max_seq_len"": kv_data[""gpt2.context_length""],\n    }\n    def _remap_gguf_key(key: str):\n      replaces = [\n        (""blk."", ""h.""), ("".attn_qkv.bias"", "".attn.c_attn.bias""), ("".attn_qkv.weight"", "".attn.c_attn.weight""),\n        ("".ffn_norm.bias"", "".ln_2.bias""), ("".ffn_norm.weight"", "".ln_2.weight""), ("".attn_norm.bias"", "".ln_1.bias""),\n        ("".attn_norm.weight"", "".ln_1.weight""), ("".attn_output.bias"", "".attn.c_proj.bias""), ("".attn_output.weight"", "".attn.c_proj.weight""),\n        ("".ffn_up.bias"", "".mlp.c_fc.bias""), ("".ffn_up.weight"", "".mlp.c_fc.weight""), ("".ffn_down.bias"", "".mlp.c_proj.bias""),\n        ("".ffn_down.weight"", "".mlp.c_proj.weight""), (""token_embd.weight"", ""wte.weight""), (""output.weight"", ""lm_head.weight""),\n        (""output_norm.bias"", ""ln_f.bias""), (""output_norm.weight"", ""ln_f.weight""), (""position_embd.weight"", ""wpe.weight""),\n      ]\n      for ostr, ns in replaces: key = key.replace(ostr, ns)\n      return key\n    state_dict = { _remap_gguf_key(k): v for k, v in state_dict.items() }\n    model = Transformer(**gpt2_params)\n    with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):\n      load_state_dict(model, state_dict)\n    return GPT2(model, tiktoken.get_encoding(""gpt2""))\n\n  def __init__(self, model, tokenizer):\n    self.model = model\n    self.tokenizer = tokenizer\n\n  def generate(self, prompt:str, max_length:int, temperature:float, timing:bool=False, batch_size:int=1):\n    prompt_tokens = self.tokenizer.encode(prompt, allowed_special={""<|endoftext|>""})\n    toks = [prompt_tokens[:] for _ in range(batch_size)]\n    start_pos = 0\n    for _ in trange(max_length, disable=(timing==True)):\n      GlobalCounters.reset()\n      if timing: print("""")\n      st = GlobalCounters.time_sum_s\n      with Timing(""ran model in "", on_exit=(lambda et: (f"", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU"" if DEBUG>=2 else """")+\n                  f"", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB""+\n                  (f"", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s"" if DEBUG>=2 else """")) if DEBUG else None, enabled=timing):\n        with WallTimeEvent(BenchEvent.STEP):\n          if batch_size == 1 and len(toks[0][start_pos:]) == 1:\n            tokens = Variable(""tokens"", 0, VOCAB_SIZE-1).bind(toks[0][start_pos])\n          else:\n            tokens = Tensor([x[start_pos:] for x in toks])\n          tok = self.model(tokens, Variable(""start_pos"", 1 if start_pos else 0, MAX_CONTEXT-1).bind(start_pos), temperature).tolist()\n      start_pos = len(toks[0])\n      for i,t in enumerate(tok): toks[i].append(t)\n    return [self.tokenizer.decode(x) for x in toks]\n\n# **** main code ****\n\nif __name__ == ""__main__"":\n  print(f""using {Device.DEFAULT} backend"")\n  default_prompt = ""What is the answer to life, the universe, and everything?""\n\n  parser = argparse.ArgumentParser(description='Run GPT2 in tinygrad', formatter_class=argparse.ArgumentDefaultsHelpFormatter)\n  parser.add_argument('--prompt', type=str, default=default_prompt, help=""Phrase to start with"")\n  parser.add_argument('--count', type=int, default=100, help=""Max number of tokens to generate"")\n  parser.add_argument('--temperature', type=float, default=0.8, help=""Temperature in the softmax"")\n  parser.add_argument('--model_size', type=str, default=""gpt2-medium"", help=""Size of model to use [gpt2, gpt2-medium, gpt2-large, gpt2-xl]"")\n  parser.add_argument('--timing', action='store_true', help=""Print timing per token"")\n  parser.add_argument('--seed', type=int, help=""Set the random seed"")\n  parser.add_argument('--batch_size', type=int, default=1, help=""Set the input batch size"")\n  parser.add_argument('--benchmark', type=int, default=-1, help=""Benchmark GPT with the given number of tokens"")\n  parser.add_argument('--noshow', action='store_true', help=""Don't show the output"")\n  args = parser.parse_args()\n\n  if args.seed is not None:\n    Tensor.manual_seed(args.seed)\n\n  print(f""using {args.model_size}"")\n  gpt2 = GPT2.build_gguf(args.model_size) if args.model_size.startswith(""gpt2_gguf_"") else GPT2.build(args.model_size)\n\n  if args.benchmark != -1:\n    gpt2.model(Tensor.rand(args.batch_size, args.benchmark), Variable(""a"", 0, MAX_CONTEXT).bind(0)).realize()\n  else:\n    texts = gpt2.generate(args.prompt, args.count, args.temperature, timing=args.timing, batch_size=args.batch_size)\n    if not args.noshow:\n      print('Generating text...')\n      if len(texts) == 1: print(texts[0])\n      else:\n        for i,text in enumerate(texts): print(colored(f""Response {i}:"", ""green""), text)\n\n    # validate output!\n    if args.temperature == 0 and args.model_size == ""gpt2-medium"" and args.count == 10:\n      expected = {\n        default_prompt: ""What is the answer to life, the universe, and everything?\n\nThe answer is that we are all one"",\n        ""Hello."": ""Hello. I'm a little late to the party, but"",\n      }\n      try:\n        assert texts[0] == expected[args.prompt]\n        print(colored(""output validated"", ""green""))\n      except KeyError:\n        pass\n",python,tab
+7,121560,"examples/gpt2.py",593,0,"",python,selection_mouse
+8,482010,"examples/gpt2.py",864,0,"",python,selection_command
+9,482406,"tinygrad/__init__.py",0,0,"import os\nif int(os.getenv(""TYPED"", ""0"")):\n  from typeguard import install_import_hook\n  install_import_hook(__name__)\nfrom tinygrad.tensor import Tensor                                    # noqa: F401\nfrom tinygrad.engine.jit import TinyJit                               # noqa: F401\nfrom tinygrad.uop.ops import UOp\nVariable = UOp.variable\nfrom tinygrad.dtype import dtypes                                     # noqa: F401\nfrom tinygrad.helpers import GlobalCounters, fetch, Context, getenv   # noqa: F401\nfrom tinygrad.device import Device                                    # noqa: F401\n",python,tab
+10,482416,"tinygrad/__init__.py",318,0,"",python,selection_command
+11,484052,"tinygrad/__init__.py",327,0,"",python,selection_command
+12,484164,"tinygrad/__init__.py",329,0,"",python,selection_command
+13,484524,"tinygrad/__init__.py",332,0,"",python,selection_command
+14,484722,"tinygrad/__init__.py",333,0,"",python,selection_command
+15,484999,"tinygrad/uop/ops.py",0,0,"from __future__ import annotations\nfrom typing import Any, Callable, cast, TYPE_CHECKING, Type, Sequence\nimport sys, time, functools, itertools, math, operator, hashlib, os, types, pickle, pathlib, inspect, weakref\nfrom dataclasses import dataclass, field\nfrom enum import Enum, auto\nfrom tinygrad.uop import Ops, GroupOp\nfrom tinygrad.uop.mathtraits import MathTrait\nfrom tinygrad.dtype import ConstType, ImageDType, dtypes, DType, truncate, PtrDType\nfrom tinygrad.helpers import ContextVar, all_int, prod, getenv, all_same, Context, partition, temp, unwrap, T, argfix, Metadata, flatten\nfrom tinygrad.helpers import PICKLE_BUFFERS, PROFILE, dedup, cdiv, cmod, diskcache_put, to_function_name, cpu_profile, TracingKey\nif TYPE_CHECKING:\n  from tinygrad.shape.shapetracker import ShapeTracker\n  from tinygrad.device import Buffer, MultiBuffer\n\n# https://en.wikipedia.org/wiki/Identity_element\ndef identity_element(op:Ops, dt:DType) -> ConstType: return dtypes.as_const({Ops.ADD:0, Ops.MUL:1, Ops.MAX:dtypes.min(dt)}[op], dt)\n\ndef can_pad(root:UOp, edges:dict[UOp, None]) -> bool:\n  return all(u.op not in GroupOp.UnsafePad for u in root.toposort(gate=lambda x:x not in edges))\n\n# With True as the default, this matches the old symbolic behavior\ndef resolve(x:UOp|bool, default:bool=True):\n  if isinstance(x, bool): return x\n  assert x.dtype == dtypes.bool, ""UOp in resolve must be bool""\n  # NOTE: generating the text for the exception is expensive, so we do this\n  return bool(sx.vmin) if (sx:=x.simplify()).vmin == sx.vmax else default\n\n# smax/smin are replacements for max/min that preserve symbolic\ndef _suop(lst, uop_fxn, python_fxn):\n  uops, nums = partition(lst, lambda x: isinstance(x, UOp))\n  return ssimplify(functools.reduce(uop_fxn, uops + ([python_fxn(nums)] if nums else [])))\ndef smax(*lst): return _suop(argfix(*lst), UOp.maximum, max)\ndef smin(*lst): return _suop(argfix(*lst), UOp.minimum, min)\ndef srender(x) -> str: return x.render() if isinstance(x, UOp) else str(x)\n\ndef ssimplify(uop): return uop.ssimplify() if isinstance(uop, UOp) else uop\ndef sym_infer(uop: UOp|int, var_vals: dict[UOp, int]) -> int: return uop.sym_infer(var_vals) if isinstance(uop, UOp) else uop\n\n# used for UOp and UPat\ndef pretty_print(x:Any, rep:Callable, srcfn=lambda x: x.src, cache=None, d=0)->str:\n  def dfs(x:Any, cache:dict):\n    for s in srcfn(x) or []:\n      cache.setdefault(s, [len(cache), 0, False])[1] += 1\n      if cache[s][1] == 1: dfs(s, cache)\n  if cache is None: dfs(x, cache:={})\n  if (cx:=cache.setdefault(x, [0,0,False]))[2]: return f""{' '*d} x{cx[0]}""\n  cx[2], srcs = True, ('None' if srcfn(x) is None else ''.join(f'\n{pretty_print(s, rep, srcfn, cache, d+2)},' for s in srcfn(x)))\n  return f""{' '*d}{f'x{cx[0]}:=' * (cx[1]>1)}{rep(x)}"" % srcs\n\nclass UOpMetaClass(type):\n  ucache:dict[tuple, weakref.ReferenceType[UOp]] = {}\n  def __call__(cls, op:Ops, dtype:DType=dtypes.void, src:tuple[UOp,...]=tuple(), arg:Any=None, tag:Any=None,\n               metadata:tuple[Metadata,...]|None=None, _buffer:Buffer|None=None):\n    if (wret:=UOpMetaClass.ucache.get(key:=(op, dtype, src, arg, tag), None)) is not None and (ret:=wret()) is not None: return ret\n    UOpMetaClass.ucache[key] = ref = weakref.ref(created:=super().__call__(*key))\n    for s in src: s.children.add(ref)\n    if metadata is not None: all_metadata[created] = metadata\n    # NOTE: this value is set by pickle when pickling a realized tensor\n    if _buffer is not None:\n      assert op is Ops.BUFFER, f""trying to set Buffer {_buffer} for {op}""\n      buffers[created] = _buffer\n    return created\n\n# some uops map to other stuff\nbuffers:weakref.WeakKeyDictionary[UOp, Buffer|MultiBuffer] = weakref.WeakKeyDictionary() # this maps BUFFER uops to their device Buffers\nall_metadata:weakref.WeakKeyDictionary[UOp, tuple[Metadata, ...]] = weakref.WeakKeyDictionary() # TODO: should this be here?\n\n# NOTE: this should be frozen, but frozen is slower\n@dataclass(eq=False, slots=True)\nclass UOp(MathTrait, metaclass=UOpMetaClass):\n  op:Ops\n  dtype:DType = dtypes.void\n  src:tuple[UOp, ...] = tuple()\n  arg:Any = None\n  tag:Any = None\n  children:set[weakref.ref[UOp]] = field(default_factory=set)\n  def __del__(self):\n    if Ops is not None and self.op is Ops.BUFFER and (buffer:=buffers.get(self)) is not None: buffer.ref(-1)\n    try:\n      if (ref:=UOpMetaClass.ucache.get(k:=(self.op, self.dtype, self.src, self.arg, self.tag))) is not None:\n        for s in self.src: s.children.discard(ref)\n        del UOpMetaClass.ucache[k]\n    except AttributeError: pass\n  def __reduce__(self):\n    args = [self.op, self.dtype, self.src, self.arg, self.tag, self.metadata]\n    if self.op is Ops.BUFFER and self.realized is not None and PICKLE_BUFFERS: args.append(self.realized)\n    return UOp, tuple(args)\n  def replace(self, **kwargs) -> UOp:\n    new_args = (kwargs.pop(""op"", self.op), kwargs.pop(""dtype"", self.dtype), kwargs.pop(""src"", self.src),\n                kwargs.pop(""arg"", self.arg), kwargs.pop(""tag"", self.tag))\n    assert len(kwargs) == 0, f""unused kwargs in replace {list(kwargs)}""\n    if (self.op, self.dtype, self.src, self.arg, self.tag) == new_args: return self\n    return UOp(*new_args)\n  def rtag(self, tag=True): return self.replace(tag=tag)\n  @functools.cached_property\n  def key(self) -> bytes:\n    return hashlib.sha256(str((self.op, self.dtype, self.arg)).encode() + b"""".join([s.key for s in self.src])).digest()\n  def __repr__(self): return pretty_print(self, lambda x: f""{type(self).__name__}({x.op}, {x.dtype}, arg={x.argstr()}{x.tagstr()}, src=(%s))"")\n  def argstr(self): return f'({"", "".join(map(str, self.arg))})' if self.op is Ops.REDUCE_AXIS else repr(self.arg)\n  def tagstr(self): return f"", tag={self.tag}"" if self.tag is not None else """"\n\n  @functools.cached_property\n  def parents(self:UOp) -> dict[UOp, None]:\n    ret = {s:None for s in self.src}\n    for s in self.src: ret.update(s.parents)\n    return ret\n  @property\n  def sparents(self:UOp) -> dict[UOp, None]: return {self:None, **self.parents}\n\n  def toposort(self, gate:Callable|None=None) -> dict[UOp, None]:\n    ret: dict[UOp, None] = {}\n    stack: list[tuple[UOp, bool]] = [(self, False)] # each stack entry is (node, visited_flag)\n    while stack:\n      node, visited = stack.pop()\n      if node in ret: continue\n      if not visited:\n        if gate is None or gate(node):\n          stack.append((node, True))  # push node back on stack to process after its parents\n          for parent in reversed(node.src): stack.append((parent, False)) # push parents on the stack\n      else: ret[node] = None # second time i'm seeing this node, add it to returned toposort\n    return ret\n\n  # returns map of UOps to their children in the graph rooted by self\n  def get_children_map(self) -> dict[UOp, dict[UOp, None]]:\n    ret: dict[UOp, dict[UOp, None]] = {}\n    for u in self.toposort():\n      ret[u] = {}\n      for s in u.src: ret[s][u] = None\n    return ret\n\n  @functools.cached_property\n  def tuplize(self:UOp) -> tuple:\n    return (self.op.value, self.arg, self.dtype,)+tuple([x.tuplize for x in self.src])\n\n  # *** uop shape stuff ***\n\n  @functools.cached_property\n  def st(self) -> ShapeTracker|None:\n    if self.op in GroupOp.Block or self.op is Ops.INDEX: return None\n    from tinygrad.shape.shapetracker import ShapeTracker\n    # VIEW and MovementOps define a new ShapeTracker from the arg\n    if self.op is Ops.VIEW: return self.arg\n    if self.op in GroupOp.Movement: return unwrap(self.src[0].st).mop(self.op, self.arg)\n    # CONST with a DEVICE has a shape of ()\n    if self.op is Ops.CONST and len(self.src) and self.src[0].op is Ops.DEVICE: return ShapeTracker.from_shape(())\n    # BufferOps and ASSIGN flow ShapeTracker from a direct edge\n    if self.op in {Ops.STORE, Ops.ASSIGN, Ops.LOAD}: return self.src[0].st\n    if self.op in GroupOp.Buffer: return views[0] if (views:=[x.st for x in self.src if x.op is Ops.VIEW]) else None\n\n    # BUFFER/BUFFER_VIEW and KERNEL only have a size\n    if self.op in {Ops.BUFFER, Ops.BUFFER_VIEW}: return ShapeTracker.from_shape((self.size,))\n    if self.op is Ops.KERNEL: return ShapeTracker.from_shape((self.arg.ast.size,))\n    if self.op in {Ops.DEFINE_GLOBAL, Ops.DEFINE_LOCAL, Ops.DEFINE_REG}:\n      sz = cast(PtrDType, self.dtype).size\n      return ShapeTracker.from_shape((sz,)) if sz > 0 else None\n\n    # hack for PTX, CASTing the ptr loses the shape\n    if self.op is Ops.CAST and self.src[0].op is Ops.DEFINE_GLOBAL: return None\n\n    # otherwise we get the shape from sources\n    if not (src_sts := [x.st for x in self.src if x.st is not None]): return None\n    assert all_same([x.shape for x in src_sts]), f""UOp sources must have the same shape {self} {[x.shape for x in src_sts]}""\n    match self.op:\n      case Ops.MULTI: shape = tuple(self.src[0].shape[a]*len(self.device) if a == self.axis else s for a,s in enumerate(self.src[0].shape))\n      case Ops.BITCAST:\n        shape = src_sts[0].shape\n        if self.dtype.itemsize != (input_sz:=self.src[0].dtype.itemsize): shape = shape[:-1]+((shape[-1]*input_sz) // self.dtype.itemsize,)\n      case Ops.REDUCE_AXIS | Ops.WMMA: shape = src_sts[0].reduce(self.axis_arg)\n      case _: shape = src_sts[0].shape\n    return ShapeTracker.from_shape(shape)\n\n  @functools.cached_property\n  def full_shape(self) -> tuple[sint, ...]:\n    if self.op is Ops.VIEW: return self.shape\n    # NOTE: if a parent doesn't have st its full_shape is empty\n    parent_shapes = [x.full_shape for x in self.src]\n    return tuple(smax(x) for x in itertools.zip_longest(*parent_shapes, fillvalue=1))\n  @property\n  def shape(self) -> tuple[sint, ...]:\n    assert self.st is not None, f""{self.op} doesn't have a shape""\n    return unwrap(self.st).shape\n  @property\n  def size(self) -> int: return self.arg[0] if self.op is Ops.BUFFER_VIEW else self.arg if self.op is Ops.BUFFER else unwrap(self.st).size\n\n  # *** uop evaluation ***\n\n  def simplify(self):\n    # late import!\n    from tinygrad.uop.symbolic import symbolic\n    with Context(TRACK_MATCH_STATS=0):\n      return graph_rewrite(self, symbolic)\n  def ssimplify(self) -> UOp|ConstType: return ret.arg if (ret:=self.simplify()).op is Ops.CONST else ret\n  def _eval(self, dtype, expected_type:Type[T]) -> T:\n    assert self.dtype in dtype, f""eval with wrong dtype {self}""\n    vmin, vmax = (simple_self:=self.simplify())._min_max\n    if vmin != vmax: raise ValueError(f""eval failed to be a single number, range is {vmin} to {vmax} in {simple_self.render()}"")\n    assert isinstance(vmin, expected_type), f""vmin is wrong dtype {type(vmin)} != {expected_type}""\n    return vmin\n  def __bool__(self): return self._eval((dtypes.bool,), bool)\n  def __int__(self): return self._eval(dtypes.ints, int)\n  def __float__(self): return self._eval(dtypes.floats, float)\n  def substitute(self, dvars:dict[UOp, UOp], name:str|None=None):\n    dvars = {k:v for k,v in dvars.items() if k is not v}\n    if len(dvars) == 0: return self\n    with Context(TRACK_MATCH_STATS=(0 if name is None else TRACK_MATCH_STATS.value)):\n      return graph_rewrite(self, _substitute, dvars, bottom_up=True, name=name)\n\n  # *** uop syntactic sugar ***\n\n  @property\n  def st_arg(self) -> ShapeTracker:\n    assert self.op in GroupOp.Buffer, f""st_arg called on {self.op}""\n    return unwrap(self.st)\n  @property\n  def axis_arg(self) -> tuple[int, ...]:\n    assert self.op in {Ops.REDUCE_AXIS, Ops.WMMA}, f""axis_arg called on {self.op}""\n    ret = self.arg[1] if self.op is Ops.REDUCE_AXIS else self.arg[7]\n    assert isinstance(ret, tuple) and all(isinstance(x, int) for x in ret), f""axis_arg trying to return {ret}""\n    return ret\n  def sink(self, *srcs:UOp|None, **kwargs): return UOp(Ops.SINK, dtypes.void, (self,)+tuple([x for x in srcs if x is not None]), **kwargs)\n  def detach(self): return UOp(Ops.DETACH, self.dtype, (self,))\n  def index(self, idx:UOp, valid:UOp|None=None): return UOp(Ops.INDEX, self.dtype, (self,idx,valid) if valid is not None else (self,idx))\n  def __getitem__(self, idx): return self.index(idx)\n  def const_like(self, b:ConstLike):\n    # constants can optionally have a DEVICE source\n    return UOp.const(self.dtype, b, device=self._device, shape=self.shape if self.st is not None else None)\n  def broadcast(self, count:int):\n    assert self.dtype.count == 1\n    if count == 1: return self\n    return UOp(Ops.VECTORIZE, self.dtype.vec(count), (self,)*count)\n  def cast(self, dtype:DType):\n    if self.dtype == dtype: return self\n    return UOp(Ops.CAST, dtype, (self,))\n  def cast_vec(self, dtype:DType): return UOp(Ops.CAST, dtype.vec(self.dtype.count), (self,))\n  def bitcast(self, dtype:DType): return UOp(Ops.BITCAST, dtype, (self,))\n  def gep(self, i:tuple[int, ...]|int):\n    if isinstance(i, tuple) and len(i) == 1: return self.gep(i[0])\n    if isinstance(i, int):\n      # NOTE: these are just shortcuts to not have to create and fold later\n      if self.op is Ops.VECTORIZE: return self.src[i]\n      if self.op is Ops.VCONST: return UOp.const(self.dtype.scalar(), self.arg[i])\n      if self.op is Ops.CONST: return UOp.const(self.dtype.scalar(), self.arg)\n      i = (i,)\n    return UOp(Ops.GEP, self.dtype.scalar().vec(len(i)) if len(i) > 1 else self.dtype.scalar(), (self,), i)\n  def load(self, *src:UOp, **kwargs): return UOp(Ops.LOAD, dtype=kwargs.pop(""dtype"", self.dtype.base), src=(self,)+src, **kwargs)\n  def store(self, *src:UOp, **kwargs): return UOp(Ops.STORE, dtypes.void, (self,)+src, **kwargs)\n  def assign(self, x:UOp): return UOp(Ops.ASSIGN, self.dtype, (self, x))\n  def barrier(self, *src:UOp): return UOp(Ops.BARRIER, src=(self,)+src)\n  def alu(self, op, *src:UOp, **kwargs):\n    out_dtype = (self, *src)[-1].dtype\n    if op in {Ops.CMPLT, Ops.CMPNE, Ops.CMPEQ}: out_dtype = dtypes.bool.vec(out_dtype.count) if out_dtype.count > 1 else dtypes.bool\n    return UOp(op, out_dtype, (self,)+src, **kwargs)\n  @staticmethod\n  def const(dtype:DType, b:ConstLike, device:str|tuple[str, ...]|None=None, shape:tuple[sint, ...]|None=None):\n    if isinstance(b, UOp): return b.unbind()[0] if b.op is Ops.BIND else b\n    if isinstance(b, tuple) and all_same(b): b = b[0]  # doesn't have to be a VCONST if they are all the same\n    ret = UOp(Ops.VCONST if isinstance(b, tuple) else Ops.CONST, dtype, arg=dtypes.as_const(b, dtype))\n    if shape is not None:\n      from tinygrad.shape.shapetracker import ShapeTracker\n      ret = ret.replace(src=(UOp(Ops.VIEW, dtypes.void, (), ShapeTracker.from_shape(shape, (0,)*len(shape))),))\n    if device is not None:\n      ret = ret.replace(src=(UOp(Ops.DEVICE, arg=device).view(unwrap(ret.st)),))\n    return ret\n  @staticmethod\n  def range(dtype:DType, end:sint, idx:int): return UOp(Ops.RANGE, dtype=dtype, src=(sint_to_uop(end),), arg=idx)\n  def r(self, op:Ops, axis:tuple[int, ...]):\n    axis = tuple(sorted([x for x in axis if resolve(self.shape[x] != 1)]))\n    if len(axis) == 0: return self\n    # move any non reduce axis before the first reduce axis\n    move_early, rest = partition(range(axis[0], len(self.shape)), lambda i: i not in axis and resolve(self.shape[i] != 1))\n    permaxis = tuple(range(axis[0])) + tuple(move_early) + tuple(rest)\n    ret = self.permute(permaxis)\n    new_axis = tuple([x for x in range(axis[0]+len(move_early), len(self.shape)) if resolve(ret.shape[x] != 1)])\n    assert len(axis) == len(new_axis)\n    ret = UOp(Ops.REDUCE_AXIS, self.dtype, (ret,), (op, new_axis))\n    return ret.reshape(tuple([x if i not in axis else 1 for i,x in enumerate(self.shape)]))\n  def reduce(self, *src:UOp, **kwargs): return UOp(Ops.REDUCE, kwargs.pop('dtype', self.dtype), src=(self,)+src, **kwargs)\n  def contiguous(self): return self.alu(Ops.CONTIGUOUS)\n  def contiguous_backward(self): return self.alu(Ops.CONTIGUOUS_BACKWARD)\n  def fuse(self): return self.alu(Ops.FUSE)\n  def allreduce(self, op, device:str|tuple[str, ...]|UOp):\n    assert isinstance(self.device, tuple), f""allreduce must be on tuple {self.device} isn't""\n    return UOp(Ops.ALLREDUCE, self.dtype, (self, UOp(Ops.DEVICE, arg=device) if not isinstance(device, UOp) else device), op)\n\n  # *** from MultiLazyBuffer ***\n\n  def multi(self, axis:int|None):\n    assert isinstance(self.device, tuple), f""multi device must be tuple, {self.device} isn't""\n    assert axis is not None, ""multi None is no longer supported""\n    return UOp(Ops.MULTI, self.dtype, (self,), axis)\n\n  @property\n  def bounds(self):\n    if self.axis is None: raise RuntimeError(""bounds is not defined when axis is None"")\n    return tuple(itertools.pairwise(itertools.accumulate([self.src[0].shape[self.axis] for _ in self.device], initial=0)))\n\n  @functools.cached_property\n  def axis(self) -> int|None:\n    if self.op is Ops.MULTI: return self.arg\n    # NOTE: they all have to share an axis, we always choose [-1]\n    if self.op in GroupOp.ALU: return axes[-1] if (axes := dedup([x.axis for x in self.src if x.axis is not None])) else None\n    if len(self.src) == 0: return None\n    src_axis = self.src[0].axis\n    if self.op is Ops.REDUCE_AXIS: return None if src_axis is not None and src_axis in self.arg[1] else src_axis\n    if self.op is Ops.RESHAPE:\n      if src_axis is None: return None\n      arg_acc:list[sint] = list(itertools.accumulate(self.arg, operator.mul, initial=1))\n      # new_axis is the last one that preserves prod(prior to new_axis) and must not move items between shards\n      # TODO: what to do about shrinking to self.shape[self.axis]==1 len(self.real_lbs)==1?\n      return len(arg_acc) - arg_acc[::-1].index(prod(self.src[0].shape[:src_axis])) - 1\n    if self.op is Ops.PERMUTE: return self.arg.index(src_axis) if src_axis is not None else None\n    return src_axis\n\n  def _unshard(self, axis:int) -> UOp:\n    bsz, dcount = self.shape[axis], len(self.device)\n    dnum = UOp.variable(""_device_num"", 0, dcount-1)\n    return self.pad(tuple((0,0) if a != axis else (bsz*dnum, bsz*(dcount-1) - bsz*dnum) for a in range(len(self.shape))))\n\n  def _shard(self, axis:int) -> UOp:\n    dcount = len(self.device)\n    dnum = UOp.variable(""_device_num"", 0, dcount-1)\n    if self.shape[axis] % dcount != 0: raise RuntimeError(f""multi axis uneven: {self.shape[axis]=} {axis=} {dcount=}"")\n    sz = self.shape[axis] // dcount\n    return self.shrink(tuple((0,s) if i != axis else (dnum*sz,dnum*sz+sz) for i,s in enumerate(self.shape)))\n  def shard(self, devices:tuple[str, ...], axis:int) -> UOp: return self.copy_to_device(devices)._shard(axis).multi(axis)\n\n  # *** from LazyBuffer ***\n\n  def copy_to_device(self, device:str|tuple[str, ...]|UOp, arg=None):\n    assert arg is None or isinstance(self.device, tuple)\n    inp = self if arg is None else UOp(Ops.MSELECT, self.dtype, src=(self,), arg=arg)\n    return UOp(Ops.COPY, self.dtype, (inp, UOp(Ops.DEVICE, arg=device) if not isinstance(device, UOp) else device))\n  def mselect(self, arg:int) -> UOp: return UOp(Ops.MSELECT, self.dtype, (self,), arg)\n  @property\n  def metadata(self) -> tuple[Metadata, ...]|None: return all_metadata.get(self, None)\n\n  # *** uop movement ops ***\n\n  @property\n  def base(self) -> UOp:\n    if (self.op is Ops.VIEW and len(self.src) != 0) or self.op in GroupOp.Movement: return self.src[0].base\n    if self.op is Ops.MULTI: return self.src[0].base  # MULTI is really a VIEW\n    return self\n  def view(self, new_st:ShapeTracker) -> UOp: return UOp(Ops.VIEW, self.dtype, (self,), new_st)\n\n  def _mop(self, op:Ops, arg) -> UOp:\n    ret = UOp(op, self.dtype, (self,), arg)\n    if self.st == ret.st: return self  # ignore NOOPs, also check ret.st\n    return ret\n\n  def reshape(self, arg:tuple[sint, ...]): return self._mop(Ops.RESHAPE, arg)\n  def pad(self, arg:tuple[tuple[sint, sint], ...]): return self._mop(Ops.PAD, arg)\n  def expand(self, arg:tuple[sint, ...]): return self._mop(Ops.EXPAND, arg)\n  def permute(self, arg:tuple[sint, ...]): return self._mop(Ops.PERMUTE, arg)\n  def shrink(self, arg:tuple[tuple[sint, sint], ...]): return self._mop(Ops.SHRINK, arg)\n  def flip(self, arg:tuple[bool, ...]): return self._mop(Ops.FLIP, arg)\n\n  # *** uop UNIQUE ***\n\n  # TODO: use this in Buffer\n  unique_num = itertools.count(0)\n  @staticmethod\n  def unique(): return UOp(Ops.UNIQUE, arg=next(UOp.unique_num))\n\n  # *** uop Buffer stuff ***\n\n  @staticmethod\n  def new_buffer(device:str|tuple[str, ...], size:int, dtype:DType): return UOp(Ops.BUFFER, dtype, (UOp.unique(), UOp(Ops.DEVICE, arg=device)), size)\n  @property\n  def device(self) -> str|tuple[str, ...]: return cast(str|tuple[str, ...], unwrap(self._device))\n  @functools.cached_property\n  def _device(self) -> str|tuple[str, ...]|None:\n    if self.op is Ops.DEVICE: return self.arg\n    if self.op is Ops.MSELECT:\n      assert isinstance(self.src[0].device, tuple), ""mselect must be on tuple device""\n      return self.src[0].device[self.arg]\n    if self.op is Ops.MSTACK: return tuple(cast(str, x.device) for x in self.src)\n    if self.op in {Ops.COPY, Ops.BUFFER, Ops.ALLREDUCE}: return self.src[1].device\n    return next((x._device for x in self.src if x._device is not None), None)\n  @property\n  def buf_uop(self) -> UOp:\n    if self.op is Ops.BUFFER: return self\n    if self.op is Ops.MSELECT: return self.src[0].buf_uop.mselect(self.arg)\n    if self.op is Ops.MSTACK: return UOp(Ops.MSTACK, self.dtype, src=tuple(x.buf_uop for x in self.src))\n    assert self.op is Ops.ASSIGN, f""must be ASSIGN {self.op}""\n    return self.src[0].base\n  @property\n  def buffer(self) -> Buffer|MultiBuffer:\n    from tinygrad.device import Buffer, MultiBuffer\n    if self is not self.base:\n      assert unwrap(self.st).contiguous, ""VIEW only works here if it's contiguous""\n      return self.src[0].buffer\n    if self.op is Ops.MSELECT:\n      ret = self.src[0].buffer\n      assert isinstance(ret, MultiBuffer)\n      return ret.bufs[self.arg]\n    if self.op is Ops.MSTACK:\n      ret = MultiBuffer.__new__(MultiBuffer)\n      ret.bufs = [cast(Buffer, x.buffer) for x in self.src]\n      assert all_same([x.size for x in ret.bufs]) and all_same([x.dtype for x in ret.bufs]), ""multibuffers mismatch buffers""\n      return ret\n    assert self.op is Ops.BUFFER, f""must be BUFFER {self.op}""\n    if (cret:=buffers.get(self)) is not None: return cret\n    rdtype = self.dtype if isinstance(self.dtype, ImageDType) else self.dtype.base\n    if isinstance(self.device, tuple): ret = MultiBuffer(self.device, self.size, rdtype).ref(1)\n    else: ret = Buffer(self.device, self.size, rdtype).ref(1)\n    buffers[self] = ret\n    return ret\n  @property\n  def realized(self) -> Buffer|MultiBuffer|None:\n    # NOTE: this is used by the JIT to determine which inputs we capture\n    return self.buffer if self.op in {Ops.BUFFER, Ops.MSTACK} and self.buffer.is_allocated() else None\n  @property\n  def is_realized(self) -> bool:\n    return all(x.base.realized is not None for x in self.base.src) if self.base.op is Ops.MULTI else self.base.realized is not None\n\n  # *** uop Variable stuff ***\n\n  @staticmethod\n  def variable(name:str, min_val:ConstType, max_val:ConstType, dtype:DType=dtypes.int) -> UOp:\n    assert not isinstance(min_val, UOp) and not isinstance(max_val, UOp), f""can't create Variable {name} with {min_val}/{max_val}""\n    return UOp(Ops.DEFINE_VAR, dtype, arg=(name, min_val, max_val))\n  @property\n  def expr(self):\n    assert self.op is Ops.DEFINE_VAR, f""op is {self.op}, need DEFINE_VAR""\n    return self.arg[0]\n  def bind(self, val:int|UOp):\n    assert self.op is Ops.DEFINE_VAR, f""op is {self.op}, need DEFINE_VAR""\n    uval = self.const_like(val) if isinstance(val, int) else val\n    assert self.arg[1] <= uval.vmin and uval.vmax <= self.arg[2], f""bind {val} not in range [{self.arg[1]}, {self.arg[2]}]""\n    return UOp(Ops.BIND, self.dtype, (self, uval))\n  def unbind(self) -> tuple[Variable, int]:\n    assert self.op is Ops.BIND and self.src[0].op is Ops.DEFINE_VAR and self.src[1].op is Ops.CONST, f""can't unbind {self}""\n    return self.src[0], self.src[1].arg\n  @property\n  def val(self) -> int: return self.unbind()[1]\n  def vars(self) -> set[UOp]:\n    bound_vars = set([x for x in self.toposort() if x.op is Ops.BIND and x.src[0].op is Ops.DEFINE_VAR])\n    bound_var_base = set(x.src[0] for x in bound_vars)\n    all_vars = set([x for x in self.toposort() if x.op is Ops.DEFINE_VAR])\n    return bound_vars.union(set([x for x in all_vars if x not in bound_var_base]))\n  def variables(self) -> list[Variable]:\n    st_vars: list[set[Variable]] = [x.arg.vars() for x in self.toposort() if x.op is Ops.VIEW]\n    return sorted(set.union(*st_vars, set([x.unbind()[0] if x.op is not Ops.DEFINE_VAR else x for x in self.vars()])), key=lambda v: v.arg)\n\n  # *** uop symbolic stuff ***\n\n  def is_increasing(self:UOp) -> bool:\n    # is f a monotonically increasing function regards its input\n    if self.op in GroupOp.Irreducible: return True\n    if self.op is Ops.ADD: return self.src[0].is_increasing() and self.src[1].is_increasing()\n    if self.op in (Ops.MUL, Ops.IDIV) and self.src[1].op is Ops.CONST and self.src[1].arg >= 0: return self.src[0].is_increasing()\n    return False  # False if not sure\n  def const_factor(self) -> int:\n    """"""largest known int that divides self""""""\n    # TODO: for negatives it's not the largest\n    if self.op is Ops.CONST: return self.arg\n    if self.op is Ops.VCONST: return math.gcd(*self.arg)\n    if self.op is Ops.ADD: return math.gcd(self.src[0].const_factor(), self.src[1].const_factor())\n    if self.op is Ops.MUL: return self.src[0].arg if self.src[0].op is Ops.CONST else self.src[1].arg if self.src[1].op is Ops.CONST else 1\n    return 1\n  def divides(self, v:int) -> UOp|None:\n    if v==1: return self\n    if self.op is Ops.CONST: return self.const_like(self.arg//v) if self.arg%v == 0 else None\n    if self.op is Ops.VCONST: return self.const_like(tuple(x//v for x in self.arg)) if all(x%v == 0 for x in self.arg) else None\n    if self.op is Ops.ADD: return d0+d1 if (d0:=self.src[0].divides(v)) is not None and (d1:=self.src[1].divides(v)) is not None else None\n    if self.op is Ops.MUL:\n      if (d0:=self.src[0].divides(v)) is not None: return d0 * self.src[1]\n      if (d1:=self.src[1].divides(v)) is not None: return self.src[0] * d1\n    return None # generic None if we aren't sure\n  @property\n  def vmin(self) -> ConstType: return self._min_max[0]\n  @property\n  def vmax(self) -> ConstType: return self._min_max[1]\n  @functools.cached_property\n  def _min_max(self) -> tuple[ConstType, ConstType]:\n    if self.op in GroupOp.Binary and not dtypes.is_float(self.dtype):\n      (s0_vmin, s0_vmax), (s1_vmin, s1_vmax) = self.src[0]._min_max, self.src[1]._min_max\n      if self.op is Ops.ADD: return s0_vmin+s1_vmin, s0_vmax+s1_vmax\n      if self.op is Ops.SUB: return s0_vmin-s1_vmax, s0_vmax-s1_vmin\n      if self.op is Ops.AND and s1_vmin == s1_vmax and s0_vmin >= 0 and s1_vmin >= 0: return min(0, s0_vmin), min(s0_vmax, s1_vmax)\n      if self.op is Ops.MUL: return min(vals:=(s0_vmin*s1_vmin, s0_vmin*s1_vmax, s0_vmax*s1_vmin, s0_vmax*s1_vmax)), max(vals)\n      # SHL/SHR on consts only\n      if self.op is Ops.SHL and s1_vmin == s1_vmax and all_int(t:=(s0_vmin, s0_vmax, s1_vmin)): return t[0] << t[2], t[1] << t[2]\n      if self.op is Ops.SHR and s1_vmin == s1_vmax and all_int(t:=(s0_vmin, s0_vmax, s1_vmin)): return t[0] >> t[2], t[1] >> t[2]\n      if self.op is Ops.MOD:\n        if s1_vmin > 0: return (0, s1_vmax-1) if s0_vmin >= 0 else (-(s1_vmax-1), 0) if s0_vmax <= 0 else (-(s1_vmax-1), s1_vmax-1)\n        if s1_vmax < 0: return (0, -s1_vmin-1) if s0_vmin >= 0 else (-(-s1_vmin-1), 0) if s0_vmax <= 0 else (-(-s1_vmin-1), -s1_vmin-1)\n      if self.op is Ops.IDIV:\n        assert isinstance(s0_vmin, int) and isinstance(s0_vmax, int) and isinstance(s1_vmin, int) and isinstance(s1_vmax, int)\n        if (c:=s1_vmin) == s1_vmax:  # s1 is a const\n          if c > 0: return cdiv(s0_vmin, c), cdiv(s0_vmax, c)\n          if c < 0: return cdiv(s0_vmax, c), cdiv(s0_vmin, c)\n        if (s0_vmax <= 0 and s1_vmax < 0): return cdiv(s0_vmax, s1_vmin), cdiv(s0_vmin, s1_vmax)\n        if (s0_vmin >= 0 and s1_vmin > 0): return cdiv(s0_vmin, s1_vmax), cdiv(s0_vmax, s1_vmin)\n        if (s0_vmax <= 0 and s1_vmin > 0): return cdiv(s0_vmin, s1_vmin), cdiv(s0_vmax, s1_vmax)\n        if (s0_vmin >= 0 and s1_vmax < 0): return cdiv(s0_vmax, s1_vmax), cdiv(s0_vmin, s1_vmin)\n      if self.op is Ops.MAX: return max(s0_vmin, s1_vmin), max(s0_vmax, s1_vmax)\n      if self.op is Ops.CMPLT: return (s0_vmax<s1_vmin, s0_vmin<s1_vmax)\n      if self.op is Ops.CMPNE: return ((s0_vmax < s1_vmin) or (s1_vmax < s0_vmin), not (s0_vmin == s0_vmax == s1_vmin == s1_vmax))\n      if self.dtype == dtypes.bool:\n        if self.op is Ops.OR: return s0_vmin or s1_vmin, s0_vmax or s1_vmax\n        if self.op is Ops.AND: return s0_vmin and s1_vmin, s0_vmax and s1_vmax\n    # float has NAN issue and we use explicit NAN in transcendental\n    if self.op is Ops.WHERE and dtypes.is_int(self.dtype): return min(self.src[1].vmin, self.src[2].vmin), max(self.src[1].vmax, self.src[2].vmax)\n    # NOTE: returned UOp is assumed to be CONST\n    if self.op is Ops.DEFINE_VAR and self.arg: return self.arg[1], self.arg[2]\n    if self.op is Ops.RANGE: return 0, (self.src[0]-1).vmax\n    if self.op is Ops.BIND: return self.src[0]._min_max # ignore the bound value\n    if self.op in {Ops.UNROLL, Ops.VECTORIZE}: return min(x.vmin for x in self.src), max(x.vmax for x in self.src)\n    # TODO: Ops.SPECIAL is Ops.DEFINE_VAR\n    if self.op is Ops.SPECIAL: return 0, self.arg[1]-1 if isinstance(self.arg[1], int) else self.arg[1].vmax\n    if self.op is Ops.CONST: return self.arg, self.arg\n    if self.op is Ops.VCONST: return (min(self.arg), max(self.arg))\n    # TODO: CAST to bool/unsigned is not monotone, still some case can be simplified\n    if self.op is Ops.CAST and self.dtype in (dtypes.floats+dtypes.sints):\n      return max(dtypes.min(self.dtype), self.src[0].vmin), min(self.src[0].vmax, dtypes.max(self.dtype))\n    return dtypes.min(self.dtype), dtypes.max(self.dtype)\n\n  @functools.cached_property\n  def _sym_fxn(self):\n    sself = self.simplify()\n    varnames = tuple(x.arg[0] for x in sself.toposort() if x.op is Ops.DEFINE_VAR)\n    # TODO: sanitize varnames, or don't use naked eval while staying fast\n    return eval(""lambda ""+','.join(varnames)+"": ""+sself.render(pm=renderer_infer)), varnames  # pylint: disable=eval-used\n\n  def sym_infer(self, var_vals:dict[UOp, int]):\n    fxn, varnames = self._sym_fxn\n    return fxn(**{k.arg[0]:v for k,v in var_vals.items() if k.arg[0] in varnames})\n\n  def render(self, simplify=True, pm:PatternMatcher|None=None) -> str:\n    ret = graph_rewrite(self.simplify() if simplify else self, renderer if pm is None else pm)\n    return ret.arg if ret.op is Ops.NOOP else str(ret)\n\nclass AxisType(Enum):\n  GLOBAL = auto(); LOCAL = auto(); LOOP = auto(); GROUP_REDUCE = auto(); REDUCE = auto(); UPCAST = auto(); UNROLL = auto()  # noqa: E702\n\n@dataclass(frozen=True)\nclass KernelInfo:\n  name: str = ""test""            # name of the kernel\n  axis_types: tuple[AxisType, ...] = tuple()\n  dont_use_locals: bool = False # don't use local indexing\n  applied_opts: tuple = tuple()\n  opts_to_apply: tuple|None = None\n  @property\n  def function_name(self): return to_function_name(self.name)\n\n# ******** ops in python ********\n\ndef safe_exp2(x):\n  try: return 2 ** x\n  except OverflowError: return math.inf\n\ndef safe_pow(x, y):\n  try: return math.nan if isinstance(p:=pow(x, y), complex) else p\n  except ZeroDivisionError: return math.inf\n  except ValueError: return math.inf if x > 0 else -math.inf\n\npython_alu: dict[Ops, Callable]  = {\n  Ops.LOG2: lambda x: math.log2(x) if x > 0 else -math.inf if x == 0 else math.nan, Ops.EXP2: safe_exp2,\n  Ops.SQRT: lambda x: math.sqrt(x) if x >= 0 else math.nan, Ops.RECIP: lambda x: 1/x if x != 0 else math.copysign(math.inf, x),\n  Ops.SIN: lambda x: math.sin(x) if not math.isinf(x) else math.nan, Ops.POW: safe_pow,\n  Ops.NEG: operator.neg, Ops.ADD: operator.add, Ops.SUB: operator.sub, Ops.MUL: operator.mul, Ops.CMPNE: operator.ne, Ops.CMPLT: operator.lt,\n  Ops.XOR: operator.xor, Ops.OR: operator.or_, Ops.AND: operator.and_, Ops.SHR: operator.rshift, Ops.SHL: operator.lshift, Ops.MAX: max,\n  Ops.MOD: cmod, Ops.IDIV: cdiv, Ops.MULACC: lambda x,y,z: (x*y)+z, Ops.WHERE: lambda x,y,z: y if x else z, Ops.CMPEQ: operator.eq}\n\ndef exec_alu(op:Ops, dtype:DType, operands, truncate_output=True):\n  if dtype.count > 1:\n    return tuple([exec_alu(op, dtype.scalar(), [x[i] if isinstance(x, tuple) else x for x in operands]) for i in range(dtype.count)])\n  alu = python_alu[op](*operands)\n  return truncate.get(dtype, lambda x: x)(alu) if truncate_output else alu\n\n# ***** uop helpers *****\n\ndef print_uops(uops:list[UOp]):\n  for i,u in enumerate(uops):\n    formatted_parents = [(uops.index(x) if x.op is not Ops.CONST else f""{x.arg}"") if x in uops else ""--"" for x in u.src]\n    print(f""{i:4d} {str(u.op):20s}: {str(u.dtype):30s} "" f""{str(formatted_parents):32s} {u.arg}"")\n\n# ***** pattern matcher *****\n\ndef get_location() -> tuple[str, int]:\n  frm = sys._getframe(1)\n  # skip over ops.py/mathtraits.py (unless there's nothing but ops.py/mathtraits.py)\n  while pathlib.Path(frm.f_code.co_filename).name in (""ops.py"", ""mathtraits.py"") and frm.f_back is not None and \\n      not frm.f_back.f_code.co_filename.startswith(""<frozen""):\n    frm = frm.f_back\n  return frm.f_code.co_filename, frm.f_lineno\n\n@functools.cache\ndef lines(fn) -> list[str]:\n  with open(fn) as f: return f.readlines()\n\ndef printable(loc:tuple[str, int]) -> str:\n  try: return lines(loc[0])[loc[1]-1].strip()\n  except FileNotFoundError: return ""<missing>""\n\nclass UPat(MathTrait):\n  __slots__ = (""op"", ""dtype"", ""arg"", ""name"", ""src"")\n  def __init__(self, op:Ops|tuple[Ops, ...]|set[Ops]|None=None, dtype:DType|tuple[DType, ...]|None=None,\n               src:tuple[UPat, ...]|list[UPat]|UPat|None=None, arg:Any=None,\n               name:str|None=None, allow_any_len:bool=False, custom_early_reject:set[Ops]|None=None, location=None):\n    assert op is None or isinstance(op, (Ops, tuple, set)), ""op must be Ops or tuple of Ops""\n    self.op: tuple[Ops, ...]|None = (op,) if isinstance(op, Ops) else (tuple(op) if isinstance(op, set) else op)\n    self.dtype: tuple[DType, ...]|None = (dtype,) if isinstance(dtype, DType) else dtype\n    self.arg, self.name, self._in_src, self.custom_early_reject = arg, name, src, custom_early_reject\n    self.src: Any = None\n    assert self.name != ""ctx"", ""UPat can't be named ctx""\n    assert dtype is None or isinstance(dtype, DType) or all(isinstance(x, DType) for x in dtype), f""invalid dtype {dtype}""\n\n    # try all permutations if it's a list\n    if isinstance(src, list): self.src = list(itertools.permutations(src)) if not all_same(src) else [tuple(src)]\n    # only one if it's a tuple\n    elif isinstance(src, tuple): self.src = [src]\n    # repeat if it's a UPat\n    elif isinstance(src, UPat): self.src = [itertools.repeat(src)]\n\n    self.strict_length = not (allow_any_len or isinstance(src, UPat) or src is None)\n    self.required_len: int = 0 if isinstance(src, UPat) or src is None else len(src)\n    self.location = location or get_location()\n\n    if custom_early_reject is not None: self.early_reject = custom_early_reject\n    else:\n      upat_match = [src] if isinstance(src, UPat) else ([] if src is None else self.src[0])\n      self.early_reject = {pp.op[0] for pp in upat_match if pp.op is not None and len(pp.op) == 1}\n\n  def __reduce__(self):\n    return UPat, (self.op, self.dtype, self._in_src, self.arg, self.name, not self.strict_length, self.custom_early_reject, self.location)\n  def named(self, name:str): return UPat(self.op, self.dtype, self._in_src, self.arg, name, not self.strict_length, self.custom_early_reject)\n\n  @staticmethod\n  def any(*src): return UPatAny(src=src)\n  def or_casted(self, name:str|None=None): return UPat.any(self if name is None else self.named(name), UPat(Ops.CAST, name=name, src=(self,)))\n\n  @staticmethod\n  @functools.cache\n  def var(name:str|None=None, dtype:DType|tuple[DType, ...]|None=None): return UPat(dtype=dtype, name=name)\n  @staticmethod\n  @functools.cache\n  def cvar(name:str|None=None, dtype:DType|None=None, vec=True): return UPat((Ops.CONST,Ops.VCONST) if vec else Ops.CONST, dtype, name=name)\n  @staticmethod\n  def const(dtype:DType|tuple[DType, ...]|None, b:ConstType): return UPat(Ops.CONST, dtype=dtype, arg=b)\n\n  # copied from UOp\n  def sink(self, *srcs:UPat|None, **kwargs): return UPat(Ops.SINK, dtypes.void, (self,)+tuple([x for x in srcs if x is not None]), **kwargs)\n  def index(self, idx:UPat, valid:UPat|None=None): return UPat(Ops.INDEX, self.dtype, (self,idx,valid) if valid is not None else (self,idx))\n  def view(self, st=None, **kwargs): return UPat(Ops.VIEW, self.dtype, (self,), st, **kwargs)\n  def cast(self, dtype=None, **kwargs): return UPat(Ops.CAST, dtype, (self,), **kwargs)\n  def bitcast(self, dtype=None): return UPat(Ops.BITCAST, dtype, (self,))\n  def gep(self, i:int|None=None, **kwargs): return UPat(Ops.GEP, None, (self,), (i,) if i is not None else None, **kwargs)\n  def load(self, *src:UPat, **kwargs): return UPat(Ops.LOAD, src=(self,)+src, **kwargs)\n  def store(self, *src:UPat, **kwargs): return UPat(Ops.STORE, self.dtype, (self,)+src, **kwargs)\n  def assign(self, x:UPat, **kwargs): return UPat(Ops.ASSIGN, self.dtype, (self,x), **kwargs)\n  def reduce(self, *src:UPat, **kwargs): return UPat(Ops.REDUCE, self.dtype, src=(self,)+src, **kwargs)\n  def fuse(self): return self.alu(Ops.FUSE)\n  def or_broadcasted(self, **kwargs): return UPat.any(self, UPat(Ops.VECTORIZE, self.dtype, src=self, **kwargs))\n\n  def const_like(self, b:ConstLike): return UPat.const(self.dtype, cast(ConstType, b))\n  def alu(self, op:Ops, *src:UPat):\n    asrc = (self,)+src\n    return UPat(op, dtypes.bool if op in {Ops.CMPLT, Ops.CMPNE} else asrc[-1].dtype, list(asrc) if op in GroupOp.Commutative else asrc)\n\n  def __repr__(self):\n    def rep(x):\n      form = ""UPat(%s, %s, name=%s, dtype=%s, allow_any_len=%s, src=%s)""\n      return form % (None if x.op is None else ('(%s)'%', '.join(map(str, x.op))), x.arg, repr(x.name),\n        set(x.dtype) if x.dtype else None, not x.strict_length, ""[%s]"" if x.src and len(x.src)>1 else (""(%s)"" if x.src else ""%s""))\n    return pretty_print(self, rep, srcfn=lambda x:None if x.src is None else [next(x.src[0])] if isinstance(x.src[0], itertools.repeat) else x.src[0])\n\n  def match(self:UPat, uop:UOp, store:dict[str, UOp]) -> list[dict[str, UOp]]:\n    if (self.op is not None and uop.op not in self.op) or \\n       (self.name is not None and store.setdefault(self.name, uop) is not uop) or \\n       (self.dtype is not None and uop.dtype not in self.dtype and uop.dtype.scalar() not in self.dtype) or \\n       (self.arg is not None and self.arg != uop.arg) or \\n       (len(uop.src) < self.required_len) or \\n       (self.strict_length and len(uop.src) != self.required_len): return []\n    if self.src is None: return [store]\n    res: list[dict[str, UOp]] = []\n    for vp in self.src:\n      stores, new_stores = [store.copy()], []\n      for uu, vv in zip(uop.src, vp):\n        for s in stores: new_stores.extend(vv.match(uu, s))\n        stores, new_stores = new_stores, []\n      res.extend(stores)\n    return res\n\nclass UPatAny(UPat):\n  def match(self:UPat, uop:UOp, store:dict[str, UOp]) -> list[dict[str, UOp]]:\n    matches = [x.match(uop, store.copy()) for x in self.src[0]]\n    return flatten([x for x in matches if x is not None])\n\ndef deconstruct_function(fxn:Callable) -> tuple:\n  new_globals = {k:v for k,v in fxn.__globals__.items() if k in fxn.__code__.co_names}\n  for co in fxn.__code__.co_consts:\n    if isinstance(co, types.CodeType): new_globals.update({k:v for k,v in fxn.__globals__.items() if k in co.co_names})\n  # NOTE: optional round trip through pickle!\n  assert fxn.__closure__ is None, ""closures are not supported in pattern matchers""\n  ret = fxn.__code__, new_globals, fxn.__name__, fxn.__defaults__\n  return pickle.loads(pickle.dumps(ret)) if getenv(""TEST_PICKLE"") else ret\n\n@functools.cache\ndef upat_interpret(p:UPat, fxn:Callable) -> Callable:\n  real_fxn = types.FunctionType(*deconstruct_function(fxn))\n  if 'ctx' in inspect.signature(real_fxn).parameters:\n    def universal_match(uop, ctx):\n      for match in p.match(uop, {}):\n        if (ret:=real_fxn(ctx=ctx, **match)) is not None: return ret  # pylint: disable=not-callable\n      return None\n  else:\n    def universal_match(uop, _):\n      for match in p.match(uop, {}):\n        if (ret:=real_fxn(**match)) is not None: return ret  # pylint: disable=not-callable\n      return None\n  return universal_match\n\nclass PatternMatcher:\n  def __init__(self, patterns:Sequence[tuple[UPat, Callable|tuple]], compiled=bool(getenv(""UPAT_COMPILE"", 1))):\n    if compiled: from tinygrad.uop.upat import upat_compile\n    # if this comes from a pickle, we reconstruct the lambda functions here\n    self.patterns:list[tuple[UPat, Callable]] = [(p,types.FunctionType(*fxn) if isinstance(fxn, tuple) else fxn) for p,fxn in patterns]\n    # NOTE: use of DefaultDict here is very dangerous! all keys will live for the lifetime of the PatternMatcher!\n    self.pdict: dict[Ops, list[tuple[UPat, Callable, set]]] = {}\n    # uop is required, arg is optional\n    for p,fxn in self.patterns:\n      assert p.op is not None\n      if compiled and (match:=upat_compile(p, fxn)) is not None: pass # pylint: disable=E0606\n      else: match = upat_interpret(p, fxn)\n      for uop in p.op: self.pdict.setdefault(uop, []).append((p, match, p.early_reject))\n\n  def __reduce__(self): return PatternMatcher, ([(x,deconstruct_function(fxn) if fxn.__name__ == ""<lambda>"" else fxn) for x,fxn in self.patterns],)\n\n  @functools.cache  # pylint: disable=method-cache-max-size-none\n  def __add__(self, more:PatternMatcher): return PatternMatcher(self.patterns+more.patterns)\n\n  def rewrite(self, uop:UOp, ctx=None) -> UOp|None:\n    ler = {u.op for u in uop.src}\n    for _,match,early_reject in self.pdict.get(uop.op, []):\n      if not early_reject.issubset(ler): continue\n      if (ret:=match(uop, ctx)) is not None and ret is not uop: return ret\n    return None\n\n  def fixed_point_rewrite(self, uop:UOp, ctx=None) -> UOp:\n    # apply rewrite rules until a fixed point is reached. may return `uop` itself if PatternMatcher doesn't match\n    new_n: UOp|None = uop\n    seen = set()\n    while new_n is not None:\n      if new_n in seen: raise RuntimeError(""infinite loop in fixed_point_rewrite"")\n      seen.add(new_n)\n      last_n, new_n = new_n, self.rewrite(new_n, ctx)\n    return last_n\n\n# *** non-blocking UOp tracker ***\n\nucount = itertools.count()\nuop_number:weakref.WeakKeyDictionary[UOp, int] = weakref.WeakKeyDictionary()\nuop_fields:dict[int, tuple] = {}\ndef track_uop(u:UOp):\n  if (cret:=uop_number.get(u)) is not None: return cret\n  uop_number[u] = num = next(ucount)\n  # KERNEL also has a UOp in the arg\n  arg = type(u.arg)(track_uop(u.arg.ast), u.arg.metadata) if u.op is Ops.KERNEL else u.arg\n  uop_fields[num] = (u.op, u.dtype, tuple(track_uop(s) for s in u.src), arg, u.tag)\n  return num\n\n# *** tracking pattern matcher ***\n\nVIZ = ContextVar(""VIZ"", 0)\nTRACK_MATCH_STATS = ContextVar(""TRACK_MATCH_STATS"", 2 if VIZ else 0)\nmatch_stats:dict[UPat, list[int|float]] = dict()\n\n@dataclass(frozen=True)\nclass TrackedGraphRewrite:\n  loc:tuple[str, int]                    # location that called graph_rewrite\n  sink:int                               # the sink input to graph_rewrite\n  matches:list[tuple[int, int, tuple]]   # before/after UOp, UPat location\n  name:str|None                          # optional name of the rewrite\n  depth:int                              # depth if it's a subrewrite\n  bottom_up:bool\n\ntracked_keys:list[TracingKey] = []\ntracked_ctxs:list[list[TrackedGraphRewrite]] = []\n_name_cnt:dict[str, itertools.count] = {}\n\nif getenv(""CAPTURE_PROCESS_REPLAY""):\n  replay_capture: dict[str, bytes] = {}\n  import atexit\n  @atexit.register\n  def save_to_diskcache():\n    for k,v in replay_capture.items(): diskcache_put(""process_replay"", k, v, prepickled=True)\n\ndef track_rewrites(name:Callable[..., str|TracingKey]|bool=True):\n  def _decorator(func):\n    def __wrapper(*args, **kwargs):\n      fn = key = func.__name__\n      if TRACK_MATCH_STATS >= 2:\n        tracked_keys.append(key:=TracingKey(n:=f""{fn} n{next(_name_cnt.setdefault(fn, itertools.count(1)))}"", (n,), cat=fn))\n        tracked_ctxs.append([])\n      with cpu_profile(key, ""TINY"") as e:\n        ret = func(*args, **kwargs)\n      if TRACK_MATCH_STATS >= 2 and callable(name):\n        name_ret = name(*args, **kwargs, ret=ret)\n        assert isinstance(name_ret, (TracingKey, str)), f""name function returned {type(name_ret)}""\n        tracked_keys[-1] = k = TracingKey(n:=tracked_keys[-1].display_name.replace(fn, name_ret), (n,)) if isinstance(name_ret, str) else name_ret\n        e.name = TracingKey(k.display_name if isinstance(name_ret, str) else f""{fn} for {k.display_name}"", k.keys, cat=fn)\n      if getenv(""CAPTURE_PROCESS_REPLAY""):\n        # find the unittest frame we're capturing in\n        frm = sys._getframe(1)\n        while (f_back:=frm.f_back) is not None and ""unittest"" not in f_back.f_code.co_filename: frm = f_back\n        loc = f""{frm.f_code.co_filename.split('/')[-1]}:{frm.f_lineno} {frm.f_code.co_name}""\n        # capture global context vars and all the args passed in\n        with Context(PICKLE_BUFFERS=0):\n          inputs = (fn, args, kwargs, ContextVar._cache)\n          replay_capture[hashlib.sha256(pickle.dumps(inputs)).hexdigest()] = pickle.dumps(inputs+(loc, ret))\n      return ret\n    return __wrapper\n  return _decorator\n\nactive_rewrites:list[TrackedGraphRewrite] = []\ndef track_matches(func):\n  def _track_func(*args, **kwargs):\n    if tracking:=(TRACK_MATCH_STATS >= 2 and tracked_ctxs):\n      loc = ((frm:=sys._getframe(1)).f_code.co_filename, frm.f_lineno)\n      depth = len(active_rewrites)\n      tracked_ctxs[-1].append(ctx:=TrackedGraphRewrite(loc, track_uop(args[0]), [], kwargs.get(""name"", None), depth, kwargs.get(""bottom_up"", False)))\n      active_rewrites.append(ctx)\n    with cpu_profile(kwargs.get(""name"", ""<unnamed>""), ""TINY"", display=tracking):\n      ret = func(*args, **kwargs)\n    if tracking: active_rewrites.pop()\n    return ret\n  return _track_func\n\nclass TrackedPatternMatcher(PatternMatcher):\n  def rewrite(self, uop:UOp, ctx=None) -> UOp|None:\n    ret = None\n    ler = {u.op for u in uop.src}\n    for p,match,early_reject in self.pdict.get(uop.op, []):\n      if p not in match_stats: match_stats[p] = [0,0,0.0,0.0]\n      st = time.perf_counter()\n      if not early_reject.issubset(ler):\n        match_stats[p][2] += time.perf_counter()-st\n        continue\n      match_stats[p][1] += 1\n      if (ret:=match(uop, ctx)) is not None and ret is not uop:\n        match_stats[p][0] += 1\n        match_stats[p][3] += (et:=time.perf_counter()-st)\n        if TRACK_MATCH_STATS >= 3: print(f""{et*1e6:7.2f} us -- "", printable(p.location))\n        if TRACK_MATCH_STATS >= 2 and isinstance(ret, UOp) and active_rewrites:\n          active_rewrites[-1].matches.append((track_uop(uop), track_uop(ret), p.location))\n        return ret\n      match_stats[p][2] += time.perf_counter()-st\n    return None\n\nif TRACK_MATCH_STATS or PROFILE:\n  PatternMatcher = TrackedPatternMatcher  # type: ignore\n  import atexit\n  @atexit.register\n  def print_match_stats():\n    if TRACK_MATCH_STATS >= 2:\n      with open(fn:=temp(""rewrites.pkl"", append_user=True), ""wb"") as f:\n        print(f""rewrote {len(tracked_ctxs)} graphs and matched {sum(len(r.matches) for x in tracked_ctxs for r in x)} times, saved to {fn}"")\n        pickle.dump((tracked_keys, tracked_ctxs, uop_fields), f)\n    if VIZ: launch_viz(VIZ, temp(""rewrites.pkl"", append_user=True))\n    if getenv(""PRINT_MATCH_STATS"", TRACK_MATCH_STATS.value):\n      ret = [0,0,0.0,0.0]\n      for k,v in sorted(list(match_stats.items()), key=lambda x: x[1][2]+x[1][3]):\n        loc_str = f""{k.location[0].split('/')[-1]}:{k.location[1]}""\n        if v[1] != 0: print(f""{v[0]:6d} / {v[1]:7d} -- {v[3]*1000.:9.2f} / {(v[2]+v[3])*1000.:9.2f} ms -- {loc_str:20s}"", printable(k.location))\n        ret = [x+y for x,y in zip(ret, v)]\n      print(f""{ret[0]:6d} / {ret[1]:7d} -- {ret[3]*1000.:9.2f} / {(ret[2]+ret[3])*1000.:9.2f} ms -- TOTAL"")\n      print(f""{len(match_stats)} rules, {sum(v[0] > 0 for v in match_stats.values())} matched once"")\n\n  def launch_viz(var:ContextVar, data:str):\n    os.environ[(env_str:=var.key)] = ""0""\n    os.environ[f""{env_str}_DATA""] = data\n    os.environ[f""{env_str}_VALUE""] = str(var.value)\n    if not int(os.getenv(""VIZ"", ""0"")) and not int(os.getenv(""PROFILE"", ""0"")):\n      args = ['--kernels', getenv(""VIZ_DATA"", """")] if getenv(""VIZ_DATA"", """") else []\n      args += ['--profile', getenv(""PROFILE_DATA"", """")] if getenv(""PROFILE_DATA"", """") else []\n      os.execv(sys.executable, [sys.executable] + [os.path.join(os.path.dirname(__file__), ""../"", ""viz"", ""serve.py"")] + args)\n\n# *** simple graph rewrite engine ***\n\nclass RewriteNotReady(Exception): pass\nclass RewriteContext:\n  def __init__(self, pm, bpm, ctx=None):\n    self.pm: PatternMatcher|None = pm\n    self.bpm: PatternMatcher|None = bpm\n    self.ctx = ctx\n    self.replace: dict[UOp, UOp] = {}\n\n  def unified_rewrite(self, root:UOp) -> UOp:\n    stack: list[tuple[UOp, int, UOp]] = [(root, 0, root)]\n    while stack:\n      if len(stack) >= 200000: raise RuntimeError(""infinite loop in graph_rewrite"")\n      n, stage, new_n = stack.pop()\n      if n in self.replace: continue  # skip any nodes we have seen\n      try:\n        if stage == 0:\n          # if bottom up, we rewrite this node early. in both cases, we add its parents to the stack\n          if self.bpm is not None: new_n = self.bpm.fixed_point_rewrite(new_n, self.ctx)\n          stack.append((n, 1, new_n))\n          for x in reversed(new_n.src): stack.append((x, 0, x))\n        elif stage == 1:\n          try: new_src = tuple([self.replace[x] for x in new_n.src])\n          except KeyError: raise RewriteNotReady  # pylint: disable=raise-missing-from\n          if new_src == new_n.src:\n            # if top down, do the rewrite. if no rewrite or bottom up, we are done rewriting this node so we add it to the dict\n            if self.pm is None or (new_src_n:=self.pm.rewrite(new_n, self.ctx)) is None:\n              self.replace[n] = new_n\n              continue\n          else:\n            # if srcs changed from rewrites, construct a new UOp with the new srcs\n            new_src_n = UOp(new_n.op, new_n.dtype, new_src, new_n.arg, new_n.tag)\n          # trigger a rewrite of new_src_n, then after that rewrite is done, link it back to n\n          stack.append((n, 2, new_src_n))\n          stack.append((new_src_n, 0, new_src_n))\n        else:\n          # in stage 2, we link the result of new_n to the result of n\n          try: self.replace[n] = self.replace[new_n]\n          except KeyError: raise RewriteNotReady  # pylint: disable=raise-missing-from\n      except RewriteNotReady:\n        # retry this later\n        stack.insert(0, (n, stage, new_n))\n    return self.replace[root]\n\n@track_matches\ndef graph_rewrite(sink:UOp, pm:PatternMatcher, ctx=None, bottom_up=False, name=None, bpm=None) -> UOp:\n  rewrite_ctx = RewriteContext(pm if not bottom_up else None, pm if bottom_up else bpm, ctx)\n  return rewrite_ctx.unified_rewrite(sink)\n\n@track_matches\ndef graph_rewrite_map(sink:UOp, pm:PatternMatcher, ctx=None, bottom_up=False, name=None, bpm=None,\n                      input_map:dict[UOp, UOp]|None=None, ) -> dict[UOp, UOp]:\n  rewrite_ctx = RewriteContext(pm if not bottom_up else None, pm if bottom_up else bpm, ctx)\n  new_map: dict[UOp, UOp] = {}\n  for k in sink.toposort():\n    new_map[k] = v = rewrite_ctx.unified_rewrite(k)\n    if k is not v and k.metadata is not None: all_metadata[v] = tuple(dedup(all_metadata.get(v, ())))+k.metadata\n  if input_map is not None:\n    for k,v in input_map.items(): new_map[k] = new_map.get(v,v)\n  return new_map\n\ndef sint_to_uop(x:sint, dtype:DType=dtypes.int) -> UOp: return UOp.const(dtype, x) if isinstance(x, int) else x\n\n_substitute = PatternMatcher([(UPat(tuple(Ops), name=""x""), lambda ctx,x: ctx.get(x,None))])\n\n# for debug\nsyms = { Ops.ADD: ""+"", Ops.SUB: ""-"", Ops.IDIV: ""//"", Ops.MOD: ""%"", Ops.SHL: ""<<"", Ops.SHR: "">>"",\n         Ops.MUL: ""*"", Ops.CMPLT: ""<"", Ops.CMPNE: ""!="", Ops.AND: ""&"", Ops.OR: ""|"", Ops.XOR: ""^""}\nrenderer = PatternMatcher([\n  (UPat((Ops.DEFINE_VAR, Ops.SPECIAL), name=""x""), lambda x: UOp(Ops.NOOP, arg=x.arg[0])),\n  (UPat(Ops.RANGE, name=""x""), lambda x: UOp(Ops.NOOP, arg=f""ridx{x.arg}"")),\n  (UPat((Ops.CONST, Ops.VCONST), name=""x""), lambda x: UOp(Ops.NOOP, arg=str(x.arg))),\n  (UPat(Ops.UNROLL, name=""x""), lambda x: UOp(Ops.NOOP, arg=f""UNROLL({x.src[0].arg}, {x.arg})"")),\n  (UPat(Ops.CAST, name=""x""), lambda x: UOp(Ops.NOOP, arg=f""({str(x.dtype)[7:]})({x.src[0].arg})"")),\n  (UPat(Ops.LOAD), lambda: UOp(Ops.NOOP, arg=""load"")),\n  (UPat(Ops.BIND, src=UPat(Ops.NOOP), name=""x""), lambda x: x.src[0]),\n  #(UPat(Ops.BIND, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""{x.src[0].arg}[={x.src[1].arg}]"")),\n  (UPat(Ops.NEG, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""(-{x.src[0].arg})"")),\n  (UPat(Ops.RECIP, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""(1/{x.src[0].arg})"")),\n  (UPat(Ops.MAX, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""max({x.src[0].arg}, {x.src[1].arg})"")),\n  (UPat(Ops.MULACC, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""({x.src[0].arg}*{x.src[1].arg}+{x.src[2].arg})"")),\n  (UPat(Ops.WHERE, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""({x.src[1].arg} if {x.src[0].arg} else {x.src[2].arg})"")),\n  (UPat(GroupOp.ALU, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""({x.src[0].arg}{syms[x.op]}{x.src[1].arg})"")),\n])\nrenderer_infer = PatternMatcher([\n  (UPat(Ops.MOD, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""cmod({x.src[0].arg}, {x.src[1].arg})"")),\n  (UPat(Ops.IDIV, src=UPat(Ops.NOOP), name=""x""), lambda x: UOp(Ops.NOOP, arg=f""cdiv({x.src[0].arg}, {x.src[1].arg})"")),\n  *renderer.patterns\n])\n\n# *** what was symbolic.py ***\n\nsint = int|UOp\nVariable = UOp\n\nConstLike = ConstType|Variable|tuple[ConstType, ...]\n",python,tab
+16,485019,"tinygrad/uop/ops.py",22784,0,"",python,selection_command
+17,486621,"tinygrad/__init__.py",0,0,"",python,tab
+18,486633,"tinygrad/__init__.py",333,0,"",python,selection_command
+19,486760,"examples/gpt2.py",0,0,"",python,tab
+20,486769,"examples/gpt2.py",864,0,"",python,selection_command
+21,490868,"examples/gpt2.py",823,0,"",python,selection_command
+22,495277,"examples/gpt2.py",908,0,"",python,selection_command
+23,495533,"examples/gpt2.py",955,0,"",python,selection_command
+24,495562,"examples/gpt2.py",1008,0,"",python,selection_command
+25,495589,"examples/gpt2.py",1040,0,"",python,selection_command
+26,495625,"examples/gpt2.py",1041,0,"",python,selection_command
+27,495658,"examples/gpt2.py",1067,0,"",python,selection_command
+28,495689,"examples/gpt2.py",1093,0,"",python,selection_command
+29,495830,"examples/gpt2.py",1235,0,"",python,selection_command
+30,496088,"examples/gpt2.py",1268,0,"",python,selection_command
+31,496123,"examples/gpt2.py",1269,0,"",python,selection_command
+32,496153,"examples/gpt2.py",1291,0,"",python,selection_command
+33,496190,"examples/gpt2.py",1329,0,"",python,selection_command
+34,498903,"examples/gpt2.py",1335,0,"",python,selection_command
+35,499152,"examples/gpt2.py",1339,0,"",python,selection_command
+36,499527,"examples/gpt2.py",1340,0,"",python,selection_command
+37,500851,"examples/gpt2.py",1349,0,"",python,selection_command
+38,501076,"examples/gpt2.py",1351,0,"",python,selection_command
+39,501567,"examples/gpt2.py",1357,0,"",python,selection_command
+40,501785,"examples/gpt2.py",1358,0,"",python,selection_command
+41,520618,"examples/gpt2.py",1363,0,"",python,selection_command
+42,521029,"examples/gpt2.py",1364,0,"",python,selection_command
+43,521897,"examples/gpt2.py",1365,0,"",python,selection_command
+44,522144,"examples/gpt2.py",1367,0,"",python,selection_command
+45,522600,"examples/gpt2.py",1370,0,"",python,selection_command
+46,522817,"examples/gpt2.py",1372,0,"",python,selection_command
+47,523065,"examples/gpt2.py",1383,0,"",python,selection_command
+48,523095,"examples/gpt2.py",1385,0,"",python,selection_command
+49,523258,"examples/gpt2.py",1389,0,"",python,selection_command
+50,523516,"examples/gpt2.py",1390,0,"",python,selection_command
+51,523546,"examples/gpt2.py",1397,0,"",python,selection_command
+52,523572,"examples/gpt2.py",1399,0,"",python,selection_command
+53,523800,"examples/gpt2.py",1403,0,"",python,selection_command
+54,524051,"examples/gpt2.py",1404,0,"",python,selection_command
+55,524078,"examples/gpt2.py",1412,0,"",python,selection_command
+56,524115,"examples/gpt2.py",1414,0,"",python,selection_command
+57,524307,"examples/gpt2.py",1419,0,"",python,selection_command
+58,524561,"examples/gpt2.py",1420,0,"",python,selection_command
+59,524596,"examples/gpt2.py",1421,0,"",python,selection_command
+60,524623,"examples/gpt2.py",1422,0,"",python,selection_command
+61,524663,"examples/gpt2.py",1427,0,"",python,selection_command
+62,524689,"examples/gpt2.py",1429,0,"",python,selection_command
+63,524728,"examples/gpt2.py",1439,0,"",python,selection_command
+64,524763,"examples/gpt2.py",1442,0,"",python,selection_command
+65,524792,"examples/gpt2.py",1449,0,"",python,selection_command
+66,524949,"examples/gpt2.py",1452,0,"",python,selection_command
+67,525953,"examples/gpt2.py",1329,0,"",python,selection_command
+68,526084,"examples/gpt2.py",1335,0,"",python,selection_command
+69,607176,"examples/gpt2.py",1329,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-be01f682-9bec-4d8c-a265-7f60592d40e81756061513793-2025_08_24-20.51.59.750/source.csv

@@ -0,0 +1,6 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"MaxText/train.py",0,0,"# Copyright 2023–2025 Google LLC\n#\n# Licensed under the Apache License, Version 2.0 (the ""License"");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n#    https://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an ""AS IS"" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\n# pylint: disable=g-bad-todo, abstract-method, consider-using-with\n""""""Training loop and Decoding of the model.""""""\n\n# Calling jax.device_count here prevents a ""TPU platform already registered"" error.\n# See github.com/google/maxtext/issues/20 for more\n\nfrom typing import Any, Sequence\nimport datetime\nimport functools\nimport os\n\nfrom absl import app\n\nimport numpy as np\n\nimport pathwaysutils  # pylint: disable=unused-import\n\nimport tensorflow as tf\n\nimport jax\nimport jax.numpy as jnp\n\nfrom flax import linen as nn\nfrom flax.linen import partitioning as nn_partitioning\n\nfrom cloud_tpu_diagnostics import diagnostic\nfrom cloud_tpu_diagnostics.configuration import debug_configuration\nfrom cloud_tpu_diagnostics.configuration import diagnostic_configuration\nfrom cloud_tpu_diagnostics.configuration import stack_trace_configuration\n\nfrom MaxText import checkpointing\nfrom MaxText import exceptions\nfrom MaxText import max_logging\nfrom MaxText import max_utils\nfrom MaxText import maxtext_utils\nfrom MaxText import train_utils\nfrom MaxText import profiler\nfrom MaxText import pyconfig\nfrom MaxText.layers.multi_token_prediction import calculate_mtp_acceptance_rate, calculate_mtp_loss\nfrom MaxText.data_loader import DataLoader\nfrom MaxText.input_pipeline.input_pipeline_interface import create_data_iterator\nfrom MaxText.globals import EPS\nfrom MaxText.metric_logger import MetricLogger\nfrom MaxText.utils import gcs_utils\nfrom MaxText.utils.goodput_utils import (\n    GoodputEvent,\n    create_goodput_recorder,\n    maybe_monitor_goodput,\n    maybe_record_goodput,\n)\nfrom MaxText.vertex_tensorboard import VertexTensorboardManager\n# Placeholder: internal\n\nimport MaxText as mt\n# pylint: disable=too-many-positional-arguments\n\n\ndef validate_train_config(config):\n  """"""Validates the configuration is set correctly for 'train.py'.""""""\n\n  assert config.run_name, ""Erroring out, need a real run_name""\n  if config.dataset_path and not config.dataset_path.startswith(""gs://""):\n    max_logging.log(""WARNING: 'dataset_path' might be pointing your local file system"")\n  if not config.base_output_directory.startswith(""gs://""):\n    max_logging.log(""WARNING: 'base_output_directory' might be pointing your local file system"")\n  assert config.steps > 0, ""You must set steps or learning_rate_schedule_steps to a positive integer.""\n\n  if config.quantization in (""fp8"", ""nanoo_fp8""):\n    # pylint: disable=line-too-long\n    assert config.gradient_accumulation_steps == 1, (\n        ""fp8 can't be used with gradient_accumulation_steps right now. Please use other quantization or set ""\n        ""gradient_accumulation_steps to 1""\n    )\n\n  # Check if GPU Flash Attention is being used with sequence packing\n  if config.attention == ""cudnn_flash_te"" and config.packing and config.dataset_type != ""synthetic"":\n    raise ValueError(\n        ""cudnn_flash_te only supports BSHD format. The THD (seq packing) support is going to be available in ""\n        ""Transformer Engine 2.0 release. ""\n        ""Please disable sequence packing (set packing=False) or use a different attention mechanism. ""\n        ""With synthetic data, the format is not important as packing is not applied.""\n    )\n\n\ndef get_first_step(state):\n  return int(state.step)\n\n\n# -----------------------------------------------------------------------------\n# Top-level Functions\n# -----------------------------------------------------------------------------\n\n\ndef record_activation_metrics(output_metrics, intermediate_outputs, config):\n  """"""Adds the activation metrics to the metrics dict""""""\n\n  if config.scan_layers:\n    metrics_dict = intermediate_outputs[""intermediates""][""decoder""][""decoder""]\n\n    for layer_num in range(config.num_decoder_layers):\n      output_metrics[""scalar""][f""activ_fraction_zero/layer_{layer_num:03d}""] = metrics_dict[""activation_fraction_zero""][\n          0\n      ][layer_num]\n      output_metrics[""scalar""][f""activ_mean/layer_{layer_num:03d}""] = metrics_dict[""activation_mean""][0][layer_num]\n      output_metrics[""scalar""][f""activ_stdev/layer_{layer_num:03d}""] = metrics_dict[""activation_stdev""][0][layer_num]\n  else:\n    for layer_num in range(config.num_decoder_layers):\n      layer = intermediate_outputs[""intermediates""][""decoder""][f""layers_{layer_num}""]\n      output_metrics[""scalar""][f""activ_fraction_zero/layer_{layer_num:03d}""] = layer[""activation_fraction_zero""][0]\n      output_metrics[""scalar""][f""activ_mean/layer_{layer_num:03d}""] = layer[""activation_mean""][0]\n      output_metrics[""scalar""][f""activ_stdev/layer_{layer_num:03d}""] = layer[""activation_stdev""][0]\n\n\ndef _split_dpo_state(state):\n  reference_params = state.params[""reference_params""]\n  new_state = state.replace(params={k: v for k, v in state.params.items() if k != ""reference_params""})\n  return new_state, reference_params\n\n\ndef _merge_dpo_state(state, reference_params):\n  return state.replace(params=dict(state.params, reference_params=reference_params))\n\n\ndef dpo_loss_fn(model, config, data, dropout_rng, params, reference_params, is_train=True):\n  """"""loss_fn for both train and eval.\n\n  Args:\n    model: A nn.Module\n    config: Config of parameters\n    data: Batch of data to apply to the model\n    dropout_rng: A key to use to generate rng for dropout\n    params: Model params\n    is_train: True for train_step and False for eval_step\n\n  Returns:\n    loss: average loss\n    aux: a dictionary including intermediate_outputs, total_loss, and total_weights\n  """"""\n  # inputs, targets, segments, positions = apply_args\n  rng1, aqt_rng = jax.random.split(dropout_rng)\n\n  # decimate proportion of data when per_device_batch_size<1\n  if is_train:\n    for k, v in data.items():\n      data[k] = v[: config.micro_batch_size_to_train_on, :]\n\n  # for DPO we don't support packed sequence (they shouldn't be present in the first place)\n  data[""chosen_segmentation""] = (data[""chosen_segmentation""] == 1).astype(jnp.int32)\n  data[""rejected_segmentation""] = (data[""rejected_segmentation""] == 1).astype(jnp.int32)\n  data[""chosen_position""] = data[""chosen_position""] * (data[""chosen_segmentation""] == 1)\n  data[""rejected_position""] = data[""rejected_position""] * (data[""rejected_segmentation""] == 1)\n\n  # concatenated model and reference model forward pass\n  inputs = jnp.concatenate([data[""chosen""], data[""rejected""]], 0)\n  inputs_position = jnp.concatenate([data[""chosen_position""], data[""rejected_position""]], 0)\n  inputs_segmentation = jnp.concatenate([data[""chosen_segmentation""], data[""rejected_segmentation""]], 0)\n\n  logits, intermediate_outputs = model.apply(\n      params,\n      inputs,\n      inputs_position,\n      decoder_segment_ids=inputs_segmentation,\n      enable_dropout=config.enable_dropout if is_train else False,\n      rngs={""dropout"": rng1, ""params"": aqt_rng},\n      mutable=""intermediates"",\n  )\n  ref_logits = model.apply(\n      {""params"": reference_params},\n      inputs,\n      inputs_position,\n      decoder_segment_ids=inputs_segmentation,\n      enable_dropout=False,\n      rngs={""dropout"": rng1, ""params"": aqt_rng},\n  )\n  ref_logits = jax.lax.stop_gradient(ref_logits)\n\n  # extract token ids, segmentation and logits for chosen and rejected sequences\n  chosen_ids = data[""chosen""][..., 1:]\n  rejected_ids = data[""rejected""][..., 1:]\n  chosen_segmentation = data[""chosen_segmentation""][..., 1:]\n  rejected_segmentation = data[""rejected_segmentation""][..., 1:]\n  n_logits = logits.shape[-3] // 2  # [B, S, E] - [batch, sequence, embedding/vocab]\n  chosen_logits, rejected_logits = logits[:n_logits, :, :], logits[n_logits:, :, :]  # [B, S, E], [B, S, E]\n  # ^ [B, S, E], [B, S, E]\n  chosen_ref_logits, rejected_ref_logits = ref_logits[:n_logits, :, :], ref_logits[n_logits:, :, :]\n\n  # common subsequence and padding mask\n  common_prefix_mask = jnp.cumsum(chosen_ids != rejected_ids, axis=-1) == 0  # [B, S]\n  valid_seq_mask = (chosen_segmentation != 0) & (rejected_segmentation != 0) & ~common_prefix_mask  # [B, S]\n\n  # compute logratios from the sequence-reduced observed token log-probability\n  chosen_logps_seq = jnp.take_along_axis(  # [B, S]\n      jax.nn.log_softmax(chosen_logits[..., :-1, :], axis=-1), chosen_ids[..., None], axis=-1\n  )[..., 0]\n  chosen_logps = jnp.sum(chosen_logps_seq * valid_seq_mask, axis=-1)  # [B]\n  chosen_ref_logps_seq = jnp.take_along_axis(  # [B, S]\n      jax.nn.log_softmax(chosen_ref_logits[..., :-1, :], axis=-1), chosen_ids[..., None], axis=-1\n  )[..., 0]\n  chosen_ref_logps = jnp.sum(chosen_ref_logps_seq * valid_seq_mask, axis=-1)  # [B]\n  chosen_logratios = chosen_logps - chosen_ref_logps  # [B]\n\n  rejected_logps_seq = jnp.take_along_axis(  # [B, S]\n      jax.nn.log_softmax(rejected_logits[..., :-1, :], axis=-1), rejected_ids[..., None], axis=-1\n  )[..., 0]\n  rejected_logps = jnp.sum(rejected_logps_seq * valid_seq_mask, axis=-1)  # [B]\n  rejected_ref_logps_seq = jnp.take_along_axis(  # [B, S]\n      jax.nn.log_softmax(rejected_ref_logits[..., :-1, :], axis=-1), rejected_ids[..., None], axis=-1\n  )[..., 0]\n  rejected_ref_logps = jnp.sum(rejected_ref_logps_seq * valid_seq_mask, axis=-1)  # [B]\n  rejected_logratios = rejected_logps - rejected_ref_logps  # [B]\n\n  # DPO loss from chosen and rejected logratios\n  LABEL_SMOOTHING, BETA = config.dpo_label_smoothing, config.dpo_beta\n  logratios_delta = BETA * (chosen_logratios - rejected_logratios)  # [B]\n  losses = (  # [B]\n      -jax.nn.log_sigmoid(BETA * logratios_delta) * (1 - LABEL_SMOOTHING)\n      - jax.nn.log_sigmoid(-BETA * logratios_delta) * LABEL_SMOOTHING\n  )\n  total_loss, total_weights = jnp.mean(losses), losses.shape[0]\n  loss = total_loss\n\n  moe_lb_loss = 0.0\n  if config.num_experts > 1:\n    nested_key = (""intermediates"", ""decoder"", ""layers"", ""moe_lb_loss"")\n    total_moe_lb_loss = maxtext_utils.get_nested_value(intermediate_outputs, nested_key, 0.0)\n    moe_lb_loss = jnp.mean(jnp.array(total_moe_lb_loss))\n    loss += moe_lb_loss\n  reward_accuracy = jnp.mean(chosen_logratios > rejected_logratios)\n  aux = {\n      ""intermediate_outputs"": intermediate_outputs,\n      ""total_loss"": total_loss,\n      ""total_weights"": total_weights,\n      ""moe_lb_loss"": moe_lb_loss,\n      ""reward_accuracy"": reward_accuracy,\n  }\n  return loss, aux\n\n\ndef loss_fn(model, config, data, dropout_rng, params, is_train=True):\n  """"""loss_fn for both train and eval.\n\n  Args:\n    model: A nn.Module\n    config: Config of parameters\n    data: Batch of data to apply to the model\n    dropout_rng: A key to use to generate rng for dropout\n    params: Model params\n    is_train: True for train_step and False for eval_step\n\n  Returns:\n    loss: average loss\n    aux: a dictionary including intermediate_outputs, total_loss, and total_weights\n  """"""\n  # inputs, targets, segments, positions = apply_args\n  rng1, aqt_rng = jax.random.split(dropout_rng)\n\n  # decimate proportion of data when per_device_batch_size<1\n  if is_train:\n    for k, v in data.items():\n      data[k] = v[: config.micro_batch_size_to_train_on, :]\n  else:\n    for k, v in data.items():\n      data[k] = v[: config.micro_batch_size_to_eval_on, :]\n  mutable_collections = [""intermediates""]\n  if config.mtp_num_layers > 0 and is_train:\n    # The single model.apply call now triggers the entire chain if MTP is enabled:\n    # Decoder runs -> returns hidden_state -> MTPBlock uses it -> MTPBlock sows losses -> we reap them here.\n    mutable_collections.append(""mtp_losses"")\n\n  # During evaluation, if the acceptance rate test is enabled, we must\n  # make its specific collection mutable so the MTPBlock can sow into it.\n  if config.mtp_eval_target_module > 0 and not is_train:\n    mutable_collections.append(""mtp_acceptance"")\n\n  logits, intermediate_outputs = model.apply(\n      params,\n      data[""inputs""],\n      data[""inputs_position""],\n      decoder_segment_ids=data[""inputs_segmentation""],\n      encoder_images=data[""images""] if config.use_multimodal else None,\n      enable_dropout=config.enable_dropout if is_train else False,\n      rngs={""dropout"": rng1, ""params"": aqt_rng},\n      mutable=mutable_collections,\n      decoder_target_tokens=data[""targets""],\n      decoder_target_mask=data[""targets_segmentation""],\n  )\n  one_hot_targets = jax.nn.one_hot(data[""targets""], config.vocab_size)\n  xent, _ = max_utils.cross_entropy_with_logits(logits, one_hot_targets, 0.0)\n  xent = nn.with_logical_constraint(xent, (""activation_embed_and_logits_batch"", ""activation_length""))\n  # Mask out paddings at the end of each example.\n  xent = xent * (data[""targets_segmentation""] != 0)\n  total_loss = jnp.sum(xent)\n  total_weights = jnp.sum(data[""targets_segmentation""] != 0)\n\n  # If gradient accumulation is enabled, we don't need to divide total_loss\n  # by total_weights and then multiply the computed gradient by total_weights,\n  # since it's equivalent to computing the gradient from total_loss.\n  # This simplification reduces the number of operations and makes it easier\n  # for XLA to move all-reduce out of the gradient accumulation loop when use\n  # Zero1+GA to reduce communication overhead.\n  # EPS was used to avoid division by zero, but it's not needed when gradient\n  # accumulation is enabled since there's no division.\n  if config.gradient_accumulation_steps > 1:\n    loss = total_loss\n  else:\n    loss = total_loss / (total_weights + EPS)\n\n  # Calculate and Add MTP Loss\n  mtp_loss = 0.0\n  if config.mtp_num_layers > 0 and is_train:\n    mtp_loss = calculate_mtp_loss(intermediate_outputs, config)\n    loss += mtp_loss\n\n  # get moe load balance loss\n  moe_lb_loss = 0.0\n  if config.num_experts > 1:\n    nested_key = (""intermediates"", ""decoder"", ""layers"", ""moe_lb_loss"")\n    total_moe_lb_loss = maxtext_utils.get_nested_value(intermediate_outputs, nested_key, 0.0)\n    moe_lb_loss = jnp.mean(jnp.array(total_moe_lb_loss))\n    loss += moe_lb_loss\n\n  # Add the model's primary output to the intermediates dict so it can be used\n  # by the acceptance rate calculation in eval_step.\n  intermediate_outputs[""logits""] = logits\n\n  aux = {\n      ""intermediate_outputs"": intermediate_outputs,\n      ""total_loss"": total_loss,\n      ""total_weights"": total_weights,\n      ""moe_lb_loss"": moe_lb_loss,\n      ""mtp_loss"": mtp_loss,\n  }\n  return loss, aux\n\n\ndef train_step(model, config, state_mesh_shardings, state, data, dropout_rng):\n  """"""\n\n  Args:\n    model: A nn.Module\n    state: A pytree of the current state of the model\n    data: Batch of data to apply to the model\n    dropout_rng: A key to use to generate rng for dropout\n\n  Returns:\n    new_state: Same format as state.\n    metrics: Dictionary of model metrics such as loss, training rate, etc.\n    rng2: A new rng key that can be used in future calls.\n\n  """"""\n  reference_params, reference_params_sharding, extra_dpo_args, _loss_fn = [], [], [], loss_fn\n  if config.use_dpo:\n    state, reference_params = _split_dpo_state(state)\n    state_mesh_shardings, reference_params_sharding = _split_dpo_state(state_mesh_shardings)\n    extra_dpo_args = [reference_params]\n    _loss_fn = dpo_loss_fn\n\n  if config.gradient_accumulation_steps > 1:\n\n    def accumulate_gradient(acc_grad_and_loss, data):\n      grad_func = jax.value_and_grad(_loss_fn, argnums=4, has_aux=True)\n      (_, aux), cur_batch_gradient = grad_func(\n          model, config, data, dropout_rng, state.params, *extra_dpo_args, is_train=True\n      )\n      acc_grad_and_loss[""loss""] += aux[""total_loss""]\n      acc_grad_and_loss[""moe_lb_loss""] += aux[""moe_lb_loss""]\n      acc_grad_and_loss[""mtp_loss""] += aux[""mtp_loss""]\n      acc_grad_and_loss[""grad""] = jax.tree_util.tree_map(\n          lambda x, y: x + y, cur_batch_gradient, acc_grad_and_loss[""grad""]\n      )\n      acc_grad_and_loss[""total_weights""] += aux[""total_weights""]\n      return acc_grad_and_loss, aux\n\n    def reshape_to_microbatch_accumulations(batch_arr):\n      """"""Reshape global batch to microbatches, assuming batch axis is leading.""""""\n      microbatches = config.gradient_accumulation_steps\n      microbatch_shape = (microbatches, batch_arr.shape[0] // microbatches) + batch_arr.shape[1:]\n      return jnp.reshape(batch_arr, microbatch_shape)\n\n    data = jax.tree_util.tree_map(reshape_to_microbatch_accumulations, data)\n    init_grad = jax.tree_util.tree_map(jnp.zeros_like, state.params)\n    init_grad_and_loss = {""loss"": 0.0, ""grad"": init_grad, ""total_weights"": 0, ""moe_lb_loss"": 0.0, ""mtp_loss"": 0.0}\n\n    grad_and_loss, aux = jax.lax.scan(\n        accumulate_gradient, init_grad_and_loss, data, length=config.gradient_accumulation_steps\n    )\n    loss = (\n        grad_and_loss[""loss""] / grad_and_loss[""total_weights""]\n        + grad_and_loss[""moe_lb_loss""] / config.gradient_accumulation_steps\n        + grad_and_loss[""mtp_loss""] / config.gradient_accumulation_steps\n    )\n    raw_grads = jax.tree_util.tree_map(lambda arr: arr / grad_and_loss[""total_weights""], grad_and_loss[""grad""])\n    aux = jax.tree.map(lambda x: jnp.sum(x, axis=0), aux)  # pytype: disable=module-attr\n  else:\n    if config.optimizer_memory_host_offload:\n      if config.use_dpo:\n        reference_params = jax.device_put(\n            reference_params, max_utils.with_memory_kind(reference_params_sharding, ""device"")\n        )\n        extra_dpo_args = [reference_params]\n    grad_func = jax.value_and_grad(_loss_fn, argnums=4, has_aux=True)\n    (loss, aux), raw_grads = grad_func(model, config, data, dropout_rng, state.params, *extra_dpo_args, is_train=True)\n  intermediate_outputs = aux[""intermediate_outputs""]\n  total_weights = aux[""total_weights""]\n  moe_lb_loss = aux[""moe_lb_loss""]\n  mtp_loss = aux[""mtp_loss""]\n\n  if config.gradient_clipping_threshold > 0:\n    grads = maxtext_utils.apply_gradient_clipping(raw_grads, state, config.gradient_clipping_threshold)\n  else:\n    grads = raw_grads\n  if config.optimizer_memory_host_offload:\n    state = state.replace(\n        opt_state=jax.device_put(\n            state.opt_state,\n            jax.tree_util.tree_map(lambda x: x.with_memory_kind(kind=""device""), state_mesh_shardings.opt_state),\n        )\n    )\n  # Move all parameters to device before optimizer update\n  if config.parameter_memory_host_offload:\n    max_logging.log(""\nMoving all parameters to device before optimizer update"")\n\n    def move(path, value):\n      max_logging.log(f""train.py: Moving f{path} to device"")\n      return value.with_memory_kind(kind=""device"")\n\n    state = state.replace(\n        params=jax.device_put(\n            state.params,\n            jax.tree_util.tree_map_with_path(move, state_mesh_shardings.params),\n        )\n    )\n  new_state = state.apply_gradients(grads=grads)\n\n  scalar_metrics = {\n      ""learning/loss"": loss,\n      ""learning/moe_lb_loss"": moe_lb_loss,\n      ""learning/mtp_loss"": mtp_loss,\n      ""learning/total_weights"": total_weights,\n  }\n  if not config.optimizer_memory_host_offload:\n    scalar_metrics[""learning/grad_norm""] = max_utils.l2norm_pytree(grads)\n    scalar_metrics[""learning/raw_grad_norm""] = max_utils.l2norm_pytree(raw_grads)\n    scalar_metrics[""learning/param_norm""] = max_utils.l2norm_pytree(new_state.params)\n  if config.use_dpo:\n    scalar_metrics[""learning/dpo_reward_accuracy""] = aux[""reward_accuracy""]\n  metrics = {\n      ""scalar"": scalar_metrics,\n      ""scalars"": {},\n  }\n\n  if config.record_internal_nn_metrics:\n    record_activation_metrics(metrics, intermediate_outputs, config)\n\n  if config.use_dpo:\n    new_state = _merge_dpo_state(new_state, reference_params)\n\n  return new_state, metrics\n\n\ndef eval_step(model, config, state, data, dropout_rng):\n  """"""eval_step no backprop and new state compared with train_step.""""""\n\n  reference_params, extra_dpo_args, _loss_fn = [], [], loss_fn\n  if config.use_dpo:\n    state, reference_params = _split_dpo_state(state)\n    extra_dpo_args = [reference_params]\n    _loss_fn = dpo_loss_fn\n\n  eval_loss_fn = functools.partial(_loss_fn, model, config, data, dropout_rng, is_train=False)\n  loss, aux = eval_loss_fn(state.params, *extra_dpo_args)\n\n  mtp_acceptance_rate = 0.0\n  if config.mtp_eval_target_module > 0:\n    mtp_acceptance_rate = calculate_mtp_acceptance_rate(aux[""intermediate_outputs""], config)\n\n  total_loss = aux[""total_loss""]\n  total_weights = aux[""total_weights""]\n  moe_lb_loss = aux[""moe_lb_loss""]\n  mtp_loss = aux[""mtp_loss""]\n  metrics = {\n      ""scalar"": {\n          ""evaluation/loss"": loss,\n          ""evaluation/total_loss"": total_loss,\n          ""evaluation/total_weights"": total_weights,\n          ""evaluation/moe_lb_loss"": moe_lb_loss,\n          ""evaluation/mtp_loss"": mtp_loss,\n          ""evaluation/mtp_acceptance_rate_percent"": mtp_acceptance_rate,\n      },\n  }\n  if config.use_dpo:\n    metrics[""scalar""][""evaluation/dpo_reward_accuracy""] = aux[""reward_accuracy""]\n\n  return metrics\n\n\ndef setup_train_loop(config, recorder, devices=None):\n  """"""Set up prerequisites for the training loop -\n      checkpoint_manager, PRNG keys, Mesh, Model and optimizer.\n      Set up data iterator and tokenizer, initialize the model.\n\n  Args:\n    config\n    recorder\n\n  Returns:\n    init_rng:\n    checkpoint_manager: Orbax checkpointer\n    state_mesh_annotations: the mesh annotations for the train state\n    model:\n    mesh:\n    learning_rate_schedule:\n    data_iterator:\n    state: the initialized train state\n  """"""\n\n  with maybe_record_goodput(recorder, GoodputEvent.TPU_INIT):\n    model = mt.from_pretrained(config, devices)\n    mesh = model.mesh\n    init_rng, checkpoint_manager, learning_rate_schedule, tx = train_utils.create_training_tools(config, model, mesh)\n\n  with maybe_record_goodput(recorder, GoodputEvent.TRAINING_PREPARATION):\n    data_iterator, eval_data_iterator = create_data_iterator(config, mesh)\n    context_parallel_size = config.context_parallel_size\n    # Check if context parallelism is being used with sequence packing\n    if context_parallel_size > 1 and config.packing and config.dataset_type != ""synthetic"":\n      raise ValueError(\n          ""Context parallelism cannot be used with sequence packing except for synthetic data where packing is not applied. ""\n          ""Either disable sequence packing (set packing=False) or disable context parallelism. ""\n          ""Context parallelism with packing support will be added soon.""\n      )\n\n    # Apply reordering wrapper to data iterators if context parallelism is enabled\n    with mesh:\n      if context_parallel_size > 1 and config.context_parallel_load_balance:\n        data_iterator = map(max_utils.get_reorder_callable(context_parallel_size), data_iterator)\n        if eval_data_iterator:\n          eval_data_iterator = map(max_utils.get_reorder_callable(context_parallel_size), eval_data_iterator)\n\n    state, _, state_mesh_shardings, data_iterator = maxtext_utils.setup_training_state(\n        model, data_iterator, tx, config, init_rng, mesh, checkpoint_manager\n    )\n\n    # TODO(aireenmei, hengtaoguo): support sharding in vit for multimodal\n    if not config.using_pipeline_parallelism and not config.use_multimodal:\n      # The vocab tensor(s) of shape [vocab, embed] (and transpose) are not sharded by stage\n      maxtext_utils.assert_params_sufficiently_sharded(state.params, mesh, config.sharding_tolerance)\n\n    if config.use_dpo:\n      abstract_state, _, _ = maxtext_utils.get_abstract_state(model, tx, config, init_rng, mesh, is_training=True)\n      max_logging.log(f""Restoring reference parameters for DPO from '{os.path.join(str(config.checkpoint_dir), str(0))}'"")\n      try:\n        step0_restored, _ = checkpointing.load_state_if_possible(\n            checkpoint_manager,\n            data_iterator,\n            load_parameters_from_path="""",\n            load_full_state_from_path="""",\n            checkpoint_storage_concurrent_gb=config.checkpoint_storage_concurrent_gb,\n            abstract_unboxed_pre_state=abstract_state,\n            enable_single_replica_ckpt_restoring=False,\n            dataset_type=config.dataset_type,\n            step=0,\n            use_ocdbt=config.checkpoint_storage_use_ocdbt,\n            use_zarr3=config.checkpoint_storage_use_zarr3,\n            enable_orbax_v1=config.enable_orbax_v1,\n            checkpoint_conversion_fn=config.checkpoint_conversion_fn,\n            source_checkpoint_layout=config.source_checkpoint_layout,\n        )\n      except FileNotFoundError:\n        step0_restored = None\n      if step0_restored is not None:\n        reference_params = step0_restored[""items""].params[""params""]\n        state = _merge_dpo_state(state, reference_params)\n      else:\n        max_logging.log(\n            f""Could not restore reference parameters for DPO from '{os.path.join(str(config.checkpoint_dir), str(0))}'""\n        )\n\n  return (\n      init_rng,\n      checkpoint_manager,\n      state_mesh_shardings,\n      model,\n      mesh,\n      learning_rate_schedule,\n      data_iterator,\n      eval_data_iterator,\n      state,\n  )\n\n\ndef train_loop(config, recorder, state=None):\n  """"""Main Training loop.""""""\n  (\n      init_rng,\n      checkpoint_manager,\n      state_mesh_shardings,\n      model,\n      mesh,\n      learning_rate_schedule,\n      data_iterator,\n      eval_data_iterator,\n      state,\n  ) = setup_train_loop(config, recorder)\n\n  if config.use_dpo:\n    if ""reference_params"" not in state.params:\n      reference_params = jax.tree.map(jnp.copy, state.params[""params""])\n      state = _merge_dpo_state(state, reference_params)\n    state_mesh_shardings = _merge_dpo_state(state_mesh_shardings, state_mesh_shardings.params[""params""])\n\n  p_train_step, p_eval_step = train_utils.jit_train_and_eval_step(\n      config, model, mesh, state, state_mesh_shardings, train_step, eval_step, eval_data_iterator\n  )\n\n  with mesh, nn_partitioning.axis_rules(config.logical_axis_rules):\n    shaped_batch = maxtext_utils.get_shaped_batch(config)\n    compiled = p_train_step.lower(state, shaped_batch, init_rng).compile()\n    compiled_stats = compiled.memory_analysis()\n    max_utils.print_compiled_memory_stats(compiled_stats)\n\n  start_step = get_first_step(state)  # this is the start_step for training\n  prof = profiler.Profiler(config, offset_step=start_step)\n  data_loader = DataLoader(config, mesh, data_iterator, recorder)\n  metric_logger = MetricLogger(config=config, learning_rate_schedule=learning_rate_schedule)\n\n  # Write train config params, num model params, and XLA flags to tensorboard\n  metric_logger.write_setup_info_to_tensorboard(state.params)\n\n  try:\n    last_step_completion = datetime.datetime.now()\n    for step in np.arange(start_step, config.steps):\n      prof.maybe_activate_profiler(step, state)\n\n      with jax.profiler.StepTraceAnnotation(""train"", step_num=step):\n        example_batch = data_loader.load_next_batch()\n        # pylint: disable=not-callable\n        nextrng = jax.jit(jax.random.fold_in)(init_rng, step)\n        with maybe_record_goodput(recorder, GoodputEvent.STEP, step):\n          with mesh, nn_partitioning.axis_rules(config.logical_axis_rules):\n            state, metrics = p_train_step(state, example_batch, nextrng)\n\n      step_time_delta = datetime.datetime.now() - last_step_completion\n      last_step_completion = datetime.datetime.now()\n\n      state_to_save = state if not config.use_dpo else _split_dpo_state(state)[0]\n      checkpointing.maybe_save_checkpoint(checkpoint_manager, state_to_save, config, data_iterator, step)\n\n      if config.dump_hlo and step == (config.dump_step if config.dump_step >= 0 else start_step):\n        jax.block_until_ready(state)  # Ensure compilation has finished.\n        gcs_utils.upload_dump(\n            config.dump_hlo_local_dir,\n            config.dump_hlo_gcs_dir,\n            module_name=config.dump_hlo_module_name,\n            delete_local_after=config.dump_hlo_delete_local_after,\n            all_host_upload=config.dump_hlo_upload_all,\n        )\n\n      if config.eval_interval > 0 and step > start_step and (step + 1) % config.eval_interval == 0:\n        assert eval_data_iterator\n\n        # Explicitly reset the eval counters before starting the eval loop\n        metric_logger.reset_eval_metrics()\n\n        eval_step_count = 0\n        # pylint: disable=not-callable\n        for eval_batch in eval_data_iterator:\n          if config.eval_steps > 0 and eval_step_count >= config.eval_steps:\n            break\n          with mesh, nn_partitioning.axis_rules(config.logical_axis_rules):\n            eval_metrics = p_eval_step(state, eval_batch, nextrng)\n          metric_logger.record_eval_metrics(step, metrics=eval_metrics)\n          max_logging.log(f""Completed eval step {eval_step_count}"")\n          eval_step_count += 1\n        metric_logger.record_eval_metrics(step, eval_step_count=eval_step_count)\n        if metric_logger.cumulative_eval_metrics[""scalar""][""eval/avg_loss""] <= config.target_eval_loss:\n          prof.deactivate()\n          raise exceptions.StopTraining(f""Target loss {config.target_eval_loss=} is achieved."")\n\n      prof.maybe_deactivate_profiler(step, state)\n\n      if step == start_step:\n        max_utils.print_mem_stats(""After params initialized"")\n\n      metric_logger.buffer_and_write_train_metrics(metrics, step, step_time_delta)\n\n    state_to_save = state if not config.use_dpo else _split_dpo_state(state)[0]\n    checkpointing.maybe_save_checkpoint(checkpoint_manager, state_to_save, config, data_iterator)\n  except exceptions.StopTraining as e:\n    max_logging.log(f""Training stopped: {str(e)}"")\n  finally:\n    metric_logger.flush_metrics_and_cleanup()\n\n  return state\n\n\ndef initialize(argv: Sequence[str]) -> tuple[pyconfig.HyperParameters, Any, Any]:\n  """"""Initialization of hyperparameters and utilities""""""\n  pathwaysutils.initialize()\n  jax.config.update(""jax_default_prng_impl"", ""unsafe_rbg"")\n  # TF allocates extraneous GPU memory when using TFDS data\n  # this leads to CUDA OOMs. WAR for now is to hide GPUs from TF\n  tf.config.set_visible_devices([], ""GPU"")\n  os.environ[""TF_CPP_MIN_LOG_LEVEL""] = ""0""\n  if ""xla_tpu_spmd_rng_bit_generator_unsafe"" not in os.environ.get(""LIBTPU_INIT_ARGS"", """"):\n    os.environ[""LIBTPU_INIT_ARGS""] = (\n        os.environ.get(""LIBTPU_INIT_ARGS"", """") + "" --xla_tpu_spmd_rng_bit_generator_unsafe=true""\n    )\n  # TODO: mazumdera@ : ensure missing mandatory fields in base.yml are filled in in argv,\n  # or fill in here\n  config = pyconfig.initialize(argv)\n  jax.config.update(""jax_use_shardy_partitioner"", config.shardy)\n  max_utils.print_system_information()\n  validate_train_config(config)\n  os.environ[""TFDS_DATA_DIR""] = config.dataset_path or """"\n  vertex_tensorboard_manager = VertexTensorboardManager()\n  if config.use_vertex_tensorboard or os.environ.get(""UPLOAD_DATA_TO_TENSORBOARD""):\n    vertex_tensorboard_manager.configure_vertex_tensorboard(config)\n\n  # Goodput configurations\n  maybe_monitor_goodput(config)\n  recorder = create_goodput_recorder(config)\n\n  # Stack traces configurations\n  debug_config = debug_configuration.DebugConfig(\n      stack_trace_config=stack_trace_configuration.StackTraceConfig(\n          collect_stack_trace=config.collect_stack_trace,\n          stack_trace_to_cloud=config.stack_trace_to_cloud,\n          stack_trace_interval_seconds=config.stack_trace_interval_seconds,\n      )\n  )\n  diagnostic_config = diagnostic_configuration.DiagnosticConfig(debug_config)\n  return config, recorder, diagnostic_config\n\n\ndef run(config, recorder, diagnostic_config):\n  """"""Run the job given hyperparameters and utilities""""""\n  with diagnostic.diagnose(diagnostic_config):\n    with maybe_record_goodput(recorder, GoodputEvent.JOB):\n      train_loop(config, recorder)\n\n\ndef main(argv: Sequence[str]) -> None:\n  config, recorder, diagnostic_config = initialize(argv)\n  run(config, recorder, diagnostic_config)\n\n\nif __name__ == ""__main__"":\n  app.run(main)\n",python,tab
+2,148,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"8:51:59 PM [info] Activating crowd-code\n8:51:59 PM [info] Recording started\n8:51:59 PM [info] Initializing git provider using file system watchers...\n8:51:59 PM [info] Git repository found\n8:51:59 PM [info] Git provider initialized successfully\n8:51:59 PM [info] Initial git state: [object Object]\n",Log,tab
+3,617,"TERMINAL",0,0,"",,terminal_focus
+4,740,"MaxText/train.py",0,0,"",python,tab
+5,1021,"TERMINAL",0,0,"source /home/franz.srambical/maxtext/.venv/bin/activate",,terminal_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-d0283e5c-e49f-44ce-aeca-057010fa8c481755356823321-2025_08_16-17.07.08.205/source.csv

@@ -0,0 +1,4 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"examples/gpt2.py",0,0,"#!/usr/bin/env python3\nimport os, argparse, contextlib\nfrom typing import Optional, Union\nwith contextlib.suppress(ImportError): import tiktoken\nfrom tinygrad import Tensor, TinyJit, Device, GlobalCounters, Variable, dtypes\nfrom tinygrad.uop.ops import UOp\nfrom tinygrad.helpers import Timing, DEBUG, JIT, getenv, fetch, colored, trange\nfrom tinygrad.nn import Embedding, Linear, LayerNorm\nfrom tinygrad.nn.state import gguf_load, torch_load, load_state_dict, get_state_dict\nfrom extra.bench_log import BenchEvent, WallTimeEvent\n\nMAX_CONTEXT = getenv(""MAX_CONTEXT"", 128)\nHALF = getenv(""HALF"")\n\nclass Attention:\n  def __init__(self, dim, n_heads):\n    self.c_attn = Linear(dim, 3*dim, bias=True)\n    self.c_proj = Linear(dim, dim, bias=True)\n    self.n_heads = n_heads\n    self.dim = dim\n    self.head_dim = dim // n_heads\n\n  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]) -> Tensor:\n    if mask is not None or start_pos.val == 0:\n      # no symbolic shape qkv when consuming prompts\n      start_pos = start_pos.val\n\n    if HALF: x = x.half()\n    xqkv = self.c_attn(x)\n    xq, xk, xv = [xqkv.shrink((None, None, (i*self.dim, (i+1)*self.dim))).reshape(None, None, self.n_heads, self.head_dim) for i in range(3)]\n    bsz, seqlen, _, _ = xq.shape\n\n    # create kv cache\n    if not hasattr(self, ""cache_kv""):\n      self.cache_kv = Tensor.zeros(2, bsz, MAX_CONTEXT, self.n_heads, self.head_dim, dtype=x.dtype).contiguous().realize()\n\n    # update the cache\n    self.cache_kv.shrink((None, None,(start_pos,start_pos+seqlen),None,None)).assign(Tensor.stack(xk, xv)).realize()\n\n    if start_pos > 0:\n      keys = self.cache_kv[0].shrink((None, (0, start_pos+seqlen), None, None))\n      values = self.cache_kv[1].shrink((None, (0, start_pos+seqlen), None, None))\n    else:\n      keys = xk\n      values = xv\n\n    xq, keys, values = xq.transpose(1, 2), keys.transpose(1, 2), values.transpose(1, 2)\n    return self.c_proj(xq.scaled_dot_product_attention(keys, values, mask).transpose(1, 2).reshape(bsz, seqlen, self.dim))\n\nclass FeedForward:\n  def __init__(self, dim, hidden_dim):\n    self.c_fc = Linear(dim, hidden_dim, bias=True)\n    self.c_proj = Linear(hidden_dim, dim, bias=True)\n\n  def __call__(self, x:Tensor) -> Tensor:\n    return self.c_proj(self.c_fc(x).gelu())\n\nclass TransformerBlock:\n  def __init__(self, dim, n_heads, norm_eps):\n    self.attn = Attention(dim, n_heads)\n    self.mlp = FeedForward(dim, 4*dim)\n    self.ln_1 = LayerNorm(dim, norm_eps)\n    self.ln_2 = LayerNorm(dim, norm_eps)\n\n  def __call__(self, x:Tensor, start_pos:Variable, mask:Optional[Tensor]):\n    h = x + self.attn(self.ln_1(x), start_pos, mask).float()\n    return (h + self.mlp(self.ln_2(h)))\n\nclass Transformer:\n  def __init__(self, dim, n_heads, n_layers, norm_eps, vocab_size, max_seq_len=1024):\n    self.vocab_size = vocab_size\n    self.wte = Embedding(vocab_size, dim)\n    self.wpe = Embedding(max_seq_len, dim)\n    self.h = [TransformerBlock(dim, n_heads, norm_eps) for _ in range(n_layers)]\n    self.ln_f = LayerNorm(dim, norm_eps)\n    self.lm_head = Linear(dim, vocab_size, bias=False)\n    self.forward_jit = TinyJit(self.forward)\n\n  def forward(self, tokens:Union[Tensor,UOp], start_pos:Variable, temperature:float=0.0):\n    if not hasattr(self, 'allpos'): self.allpos = Tensor.arange(0, MAX_CONTEXT).reshape(1, -1).realize()\n    if isinstance(tokens, UOp):\n      seqlen = 1\n      tok_emb = self.wte.weight.shrink(((tokens, tokens+1), None))\n    else:\n      seqlen = tokens.shape[1]\n      tok_emb = self.wte(tokens)\n\n    # not symbolic when consuming the prompt\n    selected_pos = (0, seqlen) if start_pos.val == 0 else (start_pos, start_pos+1)\n    pos_emb = self.wpe(self.allpos.shrink((None, selected_pos)))\n\n    h = tok_emb + pos_emb\n\n    if HALF: h = h.half()\n\n    mask = Tensor.full((1, 1, seqlen, start_pos.val+seqlen), float(""-inf""), dtype=h.dtype).triu(start_pos.val+1) if seqlen > 1 else None\n\n    for hi in self.h: h = hi(h, start_pos, mask)\n\n    logits = self.lm_head(self.ln_f(h))\n\n    if logits.shape[1] == 0:\n      # special case for empty prompt\n      logits = Tensor.ones((logits.shape[0], self.vocab_size), dtype=logits.dtype, device=logits.device)\n    else:\n      logits = logits[:, -1, :]\n\n    if temperature < 1e-6:\n      ret = logits.argmax(-1)\n    else:\n      ret = (logits / temperature).softmax().multinomial()\n    return ret.flatten().realize()\n\n  def __call__(self, tokens:Union[Tensor,UOp], start_pos:Variable, temperature:float=0.0) -> Tensor:\n    forward = (self.forward_jit if JIT and (isinstance(tokens, UOp) or tokens.shape[1] == 1) else self.forward)\n    return forward(tokens, start_pos, temperature)\n\nVOCAB_SIZE = 50257\nMODEL_PARAMS = {\n  'gpt2':         dict(n_layers=12, n_heads=12, dim=768, norm_eps=1e-5, vocab_size=VOCAB_SIZE),   # 124M params\n  'gpt2-medium':  dict(n_layers=24, n_heads=16, dim=1024, norm_eps=1e-5, vocab_size=VOCAB_SIZE),  # 350M params\n  'gpt2-large':   dict(n_layers=36, n_heads=20, dim=1280, norm_eps=1e-5, vocab_size=VOCAB_SIZE),  # 774M params\n  'gpt2-xl':      dict(n_layers=48, n_heads=25, dim=1600, norm_eps=1e-5, vocab_size=VOCAB_SIZE),  # 1558M params\n}\n\nclass GPT2:\n  @staticmethod\n  def build(model_size=""gpt2""):\n    tokenizer = tiktoken.get_encoding(""gpt2"")\n\n    model = Transformer(**MODEL_PARAMS[model_size])\n    weights = torch_load(fetch(f'https://huggingface.co/{model_size}/resolve/main/pytorch_model.bin'))\n    # special treatment for the Conv1D weights we need to transpose\n    transposed = ('attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight')\n    for k in weights:\n      if k.endswith(transposed):\n        weights[k] = weights[k].T\n    # lm head and wte are tied\n    weights['lm_head.weight'] = weights['wte.weight']\n\n    with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):\n      load_state_dict(model, weights)\n\n      if HALF:\n        for l in get_state_dict(model).values():\n          l.replace(l.half().realize())\n\n    return GPT2(model, tokenizer)\n\n  @staticmethod\n  def build_gguf(model_size: str):\n    q_type = model_size[len(""gpt2_gguf_""):].upper()\n    fn = fetch(f""https://huggingface.co/PrunaAI/gpt2-GGUF-smashed/resolve/main/gpt2.{q_type}.gguf?download=true"")\n    gguf_tensor = Tensor.empty(os.stat(fn).st_size, dtype=dtypes.uint8, device=f""disk:{fn}"").to(Device.DEFAULT)\n    kv_data, state_dict = gguf_load(gguf_tensor)\n\n    gpt2_params = {\n      ""dim"": kv_data[""gpt2.embedding_length""], ""n_heads"": kv_data[""gpt2.attention.head_count""],\n      ""n_layers"": kv_data[""gpt2.block_count""], ""norm_eps"": kv_data[""gpt2.attention.layer_norm_epsilon""],\n      ""vocab_size"": VOCAB_SIZE, ""max_seq_len"": kv_data[""gpt2.context_length""],\n    }\n    def _remap_gguf_key(key: str):\n      replaces = [\n        (""blk."", ""h.""), ("".attn_qkv.bias"", "".attn.c_attn.bias""), ("".attn_qkv.weight"", "".attn.c_attn.weight""),\n        ("".ffn_norm.bias"", "".ln_2.bias""), ("".ffn_norm.weight"", "".ln_2.weight""), ("".attn_norm.bias"", "".ln_1.bias""),\n        ("".attn_norm.weight"", "".ln_1.weight""), ("".attn_output.bias"", "".attn.c_proj.bias""), ("".attn_output.weight"", "".attn.c_proj.weight""),\n        ("".ffn_up.bias"", "".mlp.c_fc.bias""), ("".ffn_up.weight"", "".mlp.c_fc.weight""), ("".ffn_down.bias"", "".mlp.c_proj.bias""),\n        ("".ffn_down.weight"", "".mlp.c_proj.weight""), (""token_embd.weight"", ""wte.weight""), (""output.weight"", ""lm_head.weight""),\n        (""output_norm.bias"", ""ln_f.bias""), (""output_norm.weight"", ""ln_f.weight""), (""position_embd.weight"", ""wpe.weight""),\n      ]\n      for ostr, ns in replaces: key = key.replace(ostr, ns)\n      return key\n    state_dict = { _remap_gguf_key(k): v for k, v in state_dict.items() }\n    model = Transformer(**gpt2_params)\n    with WallTimeEvent(BenchEvent.LOAD_WEIGHTS):\n      load_state_dict(model, state_dict)\n    return GPT2(model, tiktoken.get_encoding(""gpt2""))\n\n  def __init__(self, model, tokenizer):\n    self.model = model\n    self.tokenizer = tokenizer\n\n  def generate(self, prompt:str, max_length:int, temperature:float, timing:bool=False, batch_size:int=1):\n    prompt_tokens = self.tokenizer.encode(prompt, allowed_special={""<|endoftext|>""})\n    toks = [prompt_tokens[:] for _ in range(batch_size)]\n    start_pos = 0\n    for _ in trange(max_length, disable=(timing==True)):\n      GlobalCounters.reset()\n      if timing: print("""")\n      st = GlobalCounters.time_sum_s\n      with Timing(""ran model in "", on_exit=(lambda et: (f"", {(GlobalCounters.time_sum_s-st)*1e3:.2f} ms on GPU"" if DEBUG>=2 else """")+\n                  f"", {GlobalCounters.global_ops*1e-9:.2f} GOPS, {GlobalCounters.global_mem*1e-9:.2f} GB""+\n                  (f"", {GlobalCounters.global_mem*1e-9/(GlobalCounters.time_sum_s-st):.2f} GB/s"" if DEBUG>=2 else """")) if DEBUG else None, enabled=timing):\n        with WallTimeEvent(BenchEvent.STEP):\n          if batch_size == 1 and len(toks[0][start_pos:]) == 1:\n            tokens = Variable(""tokens"", 0, VOCAB_SIZE-1).bind(toks[0][start_pos])\n          else:\n            tokens = Tensor([x[start_pos:] for x in toks])\n          tok = self.model(tokens, Variable(""start_pos"", 1 if start_pos else 0, MAX_CONTEXT-1).bind(start_pos), temperature).tolist()\n      start_pos = len(toks[0])\n      for i,t in enumerate(tok): toks[i].append(t)\n    return [self.tokenizer.decode(x) for x in toks]\n\n# **** main code ****\n\nif __name__ == ""__main__"":\n  print(f""using {Device.DEFAULT} backend"")\n  default_prompt = ""What is the answer to life, the universe, and everything?""\n\n  parser = argparse.ArgumentParser(description='Run GPT2 in tinygrad', formatter_class=argparse.ArgumentDefaultsHelpFormatter)\n  parser.add_argument('--prompt', type=str, default=default_prompt, help=""Phrase to start with"")\n  parser.add_argument('--count', type=int, default=100, help=""Max number of tokens to generate"")\n  parser.add_argument('--temperature', type=float, default=0.8, help=""Temperature in the softmax"")\n  parser.add_argument('--model_size', type=str, default=""gpt2-medium"", help=""Size of model to use [gpt2, gpt2-medium, gpt2-large, gpt2-xl]"")\n  parser.add_argument('--timing', action='store_true', help=""Print timing per token"")\n  parser.add_argument('--seed', type=int, help=""Set the random seed"")\n  parser.add_argument('--batch_size', type=int, default=1, help=""Set the input batch size"")\n  parser.add_argument('--benchmark', type=int, default=-1, help=""Benchmark GPT with the given number of tokens"")\n  parser.add_argument('--noshow', action='store_true', help=""Don't show the output"")\n  args = parser.parse_args()\n\n  if args.seed is not None:\n    Tensor.manual_seed(args.seed)\n\n  print(f""using {args.model_size}"")\n  gpt2 = GPT2.build_gguf(args.model_size) if args.model_size.startswith(""gpt2_gguf_"") else GPT2.build(args.model_size)\n\n  if args.benchmark != -1:\n    gpt2.model(Tensor.rand(args.batch_size, args.benchmark), Variable(""a"", 0, MAX_CONTEXT).bind(0)).realize()\n  else:\n    texts = gpt2.generate(args.prompt, args.count, args.temperature, timing=args.timing, batch_size=args.batch_size)\n    if not args.noshow:\n      print('Generating text...')\n      if len(texts) == 1: print(texts[0])\n      else:\n        for i,text in enumerate(texts): print(colored(f""Response {i}:"", ""green""), text)\n\n    # validate output!\n    if args.temperature == 0 and args.model_size == ""gpt2-medium"" and args.count == 10:\n      expected = {\n        default_prompt: ""What is the answer to life, the universe, and everything?\n\nThe answer is that we are all one"",\n        ""Hello."": ""Hello. I'm a little late to the party, but"",\n      }\n      try:\n        assert texts[0] == expected[args.prompt]\n        print(colored(""output validated"", ""green""))\n      except KeyError:\n        pass\n",python,tab
+2,106,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:07:08 PM [info] Activating crowd-code\n5:07:08 PM [info] Recording started\n5:07:08 PM [info] Initializing git provider using file system watchers...\n5:07:08 PM [info] Git repository found\n5:07:08 PM [info] Git provider initialized successfully\n",Log,tab
+3,160,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"5:07:08 PM [info] Initial git state: [object Object]\n",Log,content

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-dd5e3c9b-4719-489d-821f-6a93177ee2421757927751136-2025_09_15-11.15.59.15/source.csv

@@ -0,0 +1,390 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"genie.py",0,0,"from typing import Dict\n\nimport einops\nimport jax\nimport jax.numpy as jnp\nimport flax.nnx as nnx\nimport orbax.checkpoint as ocp\n\nfrom models.dynamics import DynamicsMaskGIT, DynamicsCausal\nfrom models.lam import LatentActionModel\nfrom models.tokenizer import TokenizerVQVAE\n\n\nclass Genie(nnx.Module):\n    """"""Genie model""""""\n\n    def __init__(\n        self,\n        in_dim: int,\n        tokenizer_dim: int,\n        tokenizer_ffn_dim: int,\n        latent_patch_dim: int,\n        num_patch_latents: int,\n        patch_size: int,\n        tokenizer_num_blocks: int,\n        tokenizer_num_heads: int,\n        lam_dim: int,\n        lam_ffn_dim: int,\n        latent_action_dim: int,\n        num_latent_actions: int,\n        lam_patch_size: int,\n        lam_num_blocks: int,\n        lam_num_heads: int,\n        lam_co_train: bool,\n        dyna_type: str,\n        dyna_dim: int,\n        dyna_ffn_dim: int,\n        dyna_num_blocks: int,\n        dyna_num_heads: int,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        dropout: float = 0.0,\n        mask_limit: float = 0.0,\n    ):\n        # --- Tokenizer ---\n        self.in_dim = in_dim\n        self.tokenizer_dim = tokenizer_dim\n        self.tokenizer_ffn_dim = tokenizer_ffn_dim\n        self.latent_patch_dim = latent_patch_dim\n        self.num_patch_latents = num_patch_latents\n        self.patch_size = patch_size\n        self.tokenizer_num_blocks = tokenizer_num_blocks\n        self.tokenizer_num_heads = tokenizer_num_heads\n        # --- LAM ---\n        self.lam_dim = lam_dim\n        self.lam_ffn_dim = lam_ffn_dim\n        self.latent_action_dim = latent_action_dim\n        self.num_latent_actions = num_latent_actions\n        self.lam_patch_size = lam_patch_size\n        self.lam_num_blocks = lam_num_blocks\n        self.lam_num_heads = lam_num_heads\n        self.lam_co_train = lam_co_train\n        # --- Dynamics ---\n        self.dyna_type = dyna_type\n        self.dyna_dim = dyna_dim\n        self.dyna_ffn_dim = dyna_ffn_dim\n        self.dyna_num_blocks = dyna_num_blocks\n        self.dyna_num_heads = dyna_num_heads\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.dropout = dropout\n        self.mask_limit = mask_limit\n\n        self.tokenizer = TokenizerVQVAE(\n            in_dim=self.in_dim,\n            model_dim=self.tokenizer_dim,\n            ffn_dim=self.tokenizer_ffn_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_patch_latents,\n            patch_size=self.patch_size,\n            num_blocks=self.tokenizer_num_blocks,\n            num_heads=self.tokenizer_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n        self.lam = LatentActionModel(\n            in_dim=self.in_dim,\n            model_dim=self.lam_dim,\n            ffn_dim=self.lam_ffn_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_latent_actions,\n            patch_size=self.lam_patch_size,\n            num_blocks=self.lam_num_blocks,\n            num_heads=self.lam_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n        if self.dyna_type == ""maskgit"":\n            self.dynamics = DynamicsMaskGIT(\n                model_dim=self.dyna_dim,\n                ffn_dim=self.dyna_ffn_dim,\n                num_latents=self.num_patch_latents,\n                latent_action_dim=self.latent_action_dim,\n                num_blocks=self.dyna_num_blocks,\n                num_heads=self.dyna_num_heads,\n                dropout=self.dropout,\n                mask_limit=self.mask_limit,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n                use_flash_attention=self.use_flash_attention,\n                rngs=rngs,\n            )\n        elif self.dyna_type == ""causal"":\n            self.dynamics = DynamicsCausal(\n                model_dim=self.dyna_dim,\n                ffn_dim=self.dyna_ffn_dim,\n                num_latents=self.num_patch_latents,\n                latent_action_dim=self.latent_action_dim,\n                num_blocks=self.dyna_num_blocks,\n                num_heads=self.dyna_num_heads,\n                dropout=self.dropout,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n                use_flash_attention=self.use_flash_attention,\n                decode=decode,\n                rngs=rngs,\n            )\n        else:\n            raise ValueError(f""Invalid dynamics type: {self.dyna_type}"")\n\n    def __call__(\n        self, batch: Dict[str, jax.Array], training: bool = True\n    ) -> Dict[str, jax.Array]:\n        videos_BTHWC = batch[""videos""]\n        tokenizer_outputs = self.tokenizer.vq_encode(videos_BTHWC, training=False)\n        token_indices_BTN = tokenizer_outputs[""indices""]\n        lam_outputs = self.lam.vq_encode(videos_BTHWC, training=False)\n        z_q_BTm11L = lam_outputs[""z_q""]\n        action_indices_E = lam_outputs[""indices""]\n        latent_actions_BTm11L = jax.lax.cond(\n            self.lam_co_train,\n            lambda: z_q_BTm11L,\n            lambda: jax.lax.stop_gradient(z_q_BTm11L),\n        )\n        outputs = dict(\n            video_tokens=jax.lax.stop_gradient(token_indices_BTN),\n            latent_actions=latent_actions_BTm11L,\n        )\n        outputs[""mask_rng""] = batch[""mask_rng""]\n        dyna_logits_BTNV, dyna_mask = self.dynamics(outputs, training)\n        outputs[""token_logits""] = dyna_logits_BTNV\n        if dyna_mask is not None:\n            outputs[""mask""] = dyna_mask\n        mle_indices_BTN = jnp.argmax(outputs[""token_logits""], axis=-1)\n        H, W = batch[""videos""].shape[2:4]\n        outputs[""recon""] = self.tokenizer.decode(mle_indices_BTN, (H, W))\n        outputs[""lam_indices""] = action_indices_E\n        return outputs\n\n    def sample(\n        self,\n        batch: Dict[str, jax.Array],\n        seq_len: int,\n        steps: int = 25,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ) -> jax.Array:\n        """"""\n        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n\n        - Input frames are tokenized once.\n        - Future frames are generated autoregressively in token space.\n        - All frames are detokenized in a single pass.\n\n        Note:\n        - For interactive or step-wise sampling, detokenization should occur after each action.\n        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n        - Temporal causal structure is preserved by\n            a) reapplying the mask before each decoding step.\n            b) a temporal causal mask is applied within each ST-transformer block.\n\n        Dimension keys:\n            B: batch size\n            T: number of input (conditioning) frames\n            N: number of patches per frame\n            M: model dimension\n            S: sequence length\n            H: height\n            W: width\n            E: B * (S - 1)\n            P: S * N\n        """"""\n        # --- Encode videos and actions ---\n        videos_BTHWC = batch[""videos""]\n        latent_actions_E = batch[""latent_actions""]\n        tokenizer_out = self.tokenizer.vq_encode(videos_BTHWC, training=False)\n        token_idxs_BTN = tokenizer_out[""indices""]\n        B, T, N = token_idxs_BTN.shape\n        pad_shape = (B, seq_len - T, N)\n        pad = jnp.zeros(pad_shape, dtype=token_idxs_BTN.dtype)\n        token_idxs_BSN = jnp.concatenate([token_idxs_BTN, pad], axis=1)\n        action_tokens_EL = self.lam.vq.get_codes(latent_actions_E)\n\n        def maskgit_step_fn(\n            carry: tuple[jax.Array, jax.Array, jax.Array, jax.Array], step: jax.Array\n        ) -> tuple[tuple[jax.Array, jax.Array, jax.Array, jax.Array], None]:\n            rng, token_idxs_BSN, mask_BSN, action_tokens_EL = carry\n            S, N = token_idxs_BSN.shape[1:]\n            L = action_tokens_EL.shape[-1]\n\n            # --- Construct + encode video ---\n            vid_embed_BSNM = self.dynamics.patch_embed(token_idxs_BSN)\n            mask_token_111M = self.dynamics.mask_token.value\n            mask_expanded_BSN1 = mask_BSN[..., None]\n            vid_embed_BSNM = jnp.where(\n                mask_expanded_BSN1, mask_token_111M, vid_embed_BSNM\n            )\n\n            # --- Predict transition ---\n            action_tokens_BSm1L = jnp.reshape(action_tokens_EL, (B, S - 1, L))\n            act_embed_BSm1M = self.dynamics.action_up(action_tokens_BSm1L)\n            act_embed_BSM = jnp.pad(act_embed_BSm1M, ((0, 0), (1, 0), (0, 0)))\n            act_embed_BS1M = jnp.reshape(\n                act_embed_BSM, (B, S, 1, act_embed_BSM.shape[-1])\n            )\n            vid_embed_BSNM += act_embed_BS1M\n            unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (steps * 2))\n            step_temp = temperature * (1.0 - unmasked_ratio)\n            final_logits_BSNV = self.dynamics.transformer(vid_embed_BSNM) / step_temp\n\n            # --- Sample new tokens for final frame ---\n            if sample_argmax:\n                sampled_token_idxs_BSN = jnp.argmax(final_logits_BSNV, axis=-1)\n            else:\n                rng, _rng = jax.random.split(rng)\n                sampled_token_idxs_BSN = jax.random.categorical(_rng, final_logits_BSNV)\n            gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n            final_token_probs_BSN = gather_fn(\n                jax.nn.softmax(final_logits_BSNV), sampled_token_idxs_BSN\n            )\n            final_token_probs_BSN += ~mask_BSN\n            # Update masked tokens only\n            token_idxs_BSN = jnp.where(mask_BSN, sampled_token_idxs_BSN, token_idxs_BSN)\n\n            # --- Update mask ---\n            num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n            final_token_probs_flat_BP = einops.rearrange(\n                final_token_probs_BSN, ""b s n -> b (s n)""\n            )\n            idx_mask_P = (\n                jnp.arange(final_token_probs_flat_BP.shape[-1])\n                <= N - num_unmasked_tokens\n            )\n            sorted_idxs_BP = jnp.argsort(final_token_probs_flat_BP, axis=-1)\n            mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask_P))\n            mask_flat_BP = einops.rearrange(mask_BSN, ""b s n -> b (s n)"")\n            new_mask_flat_BP = mask_update_fn(mask_flat_BP, sorted_idxs_BP)\n            new_mask_BSN = einops.rearrange(new_mask_flat_BP, ""b (s n) -> b s n"", n=N)\n\n            new_carry = (rng, token_idxs_BSN, new_mask_BSN, action_tokens_EL)\n            return new_carry, None\n\n        def generation_step_fn(\n            carry: tuple[jax.Array, jax.Array], step_t: jax.Array\n        ) -> tuple[tuple[jax.Array, jax.Array], None]:\n            rng, current_token_idxs_BSN = carry\n            rng, step_rng = jax.random.split(rng)\n\n            # Mask current frame (i.e., t == step_t)\n            mask_S = jnp.arange(seq_len) == step_t\n            mask_BSN = jnp.broadcast_to(mask_S[None, :, None], (B, seq_len, N)).astype(\n                bool\n            )\n            masked_token_idxs_BSN = current_token_idxs_BSN * ~mask_BSN\n\n            # --- Initialize and run MaskGIT loop ---\n            init_carry_maskgit = (\n                step_rng,\n                masked_token_idxs_BSN,\n                mask_BSN,\n                action_tokens_EL,\n            )\n            final_carry_maskgit, _ = jax.lax.scan(\n                maskgit_step_fn, init_carry_maskgit, jnp.arange(steps)\n            )\n            updated_token_idxs_BSN = final_carry_maskgit[1]\n            new_carry = (rng, updated_token_idxs_BSN)\n            return new_carry, None\n\n        # --- Run the autoregressive generation using jax.lax.scan ---\n        initial_carry = (batch[""rng""], token_idxs_BSN)\n        timesteps_to_scan = jnp.arange(T, seq_len)\n        final_carry, _ = jax.lax.scan(\n            generation_step_fn, initial_carry, timesteps_to_scan\n        )\n        final_token_idxs_BSN = final_carry[1]\n\n        # --- Decode all tokens at once at the end ---\n        H, W = batch[""videos""].shape[2:4]\n        final_frames_BSHWC = self.tokenizer.decode(\n            final_token_idxs_BSN,\n            video_hw=(H, W),\n        )\n        return final_frames_BSHWC\n\n    def sample_causal(\n        self,\n        batch: Dict[str, jax.Array],\n        seq_len: int,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ) -> jax.Array:\n        """"""\n        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n\n        - Input frames are tokenized once.\n        - Future frames are generated autoregressively in token space.\n        - All frames are detokenized in a single pass.\n\n        Note:\n        - For interactive or step-wise sampling, detokenization should occur after each action.\n        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n        - Temporal causal structure is preserved by\n            a) reapplying the mask before each decoding step.\n            b) a temporal causal mask is applied within each ST-transformer block.\n\n        Dimension keys:\n            B: batch size\n            T: number of input (conditioning) frames\n            N: number of patches per frame\n            M: model dimension\n            S: sequence length\n            H: height\n            W: width\n            E: B * (S - 1)\n        """"""\n        assert isinstance(self.dynamics, DynamicsCausal)\n        # --- Encode videos and actions ---\n        videos_BTHWC = batch[""videos""]\n        latent_actions_E = batch[""latent_actions""]\n        tokenizer_out = self.tokenizer.vq_encode(videos_BTHWC, training=False)\n        token_idxs_BTN = tokenizer_out[""indices""]\n        B, T, N = token_idxs_BTN.shape\n        pad_shape = (B, seq_len - T, N)\n        pad = jnp.zeros(pad_shape, dtype=token_idxs_BTN.dtype)\n        token_idxs_BSN = jnp.concatenate([token_idxs_BTN, pad], axis=1)\n        action_tokens_EL = self.lam.vq.get_codes(latent_actions_E)\n        dynamics_causal: DynamicsCausal = self.dynamics\n\n        def causal_step_fn(\n            carry: tuple[jax.Array, jax.Array, jax.Array, jax.Array], step_n: jax.Array\n        ) -> tuple[tuple[jax.Array, jax.Array, jax.Array, jax.Array], None]:\n            rng, token_idxs_BSN, action_tokens_EL, step_t = carry\n            S, N = token_idxs_BSN.shape[1:]\n            L = action_tokens_EL.shape[-1]\n\n            # --- Construct + encode video ---\n            vid_embed_BSNM = dynamics_causal.patch_embed(token_idxs_BSN)\n\n            # --- Predict transition ---\n            action_tokens_BSm1L = jnp.reshape(action_tokens_EL, (B, S - 1, L))\n            act_embed_BSm1M = dynamics_causal.action_up(action_tokens_BSm1L)\n            act_embed_BSM = jnp.pad(act_embed_BSm1M, ((0, 0), (1, 0), (0, 0)))\n            act_embed_BS1M = jnp.reshape(\n                act_embed_BSM, (B, S, 1, act_embed_BSM.shape[-1])\n            )\n            vid_embed_BSNp1M = jnp.concatenate([act_embed_BS1M, vid_embed_BSNM], axis=2)\n            final_logits_BTNp1V = (\n                dynamics_causal.transformer(vid_embed_BSNp1M, (step_t, step_n))\n                / temperature\n            )\n            final_logits_BV = final_logits_BTNp1V[:, step_t, step_n, :]\n\n            # --- Sample new tokens for final frame ---\n            if sample_argmax:\n                sampled_token_idxs_B = jnp.argmax(final_logits_BV, axis=-1)\n            else:\n                rng, _rng = jax.random.split(rng)\n                sampled_token_idxs_B = jax.random.categorical(_rng, final_logits_BV)\n            # Update next tokens only\n            token_idxs_BSN = token_idxs_BSN.at[:, step_t, step_n].set(\n                sampled_token_idxs_B\n            )\n\n            new_carry = (rng, token_idxs_BSN, action_tokens_EL, step_t)\n            return new_carry, None\n\n        def generation_step_fn(\n            carry: tuple[jax.Array, jax.Array], step_t: jax.Array\n        ) -> tuple[tuple[jax.Array, jax.Array], None]:\n            rng, current_token_idxs_BSN = carry\n            rng, step_rng = jax.random.split(rng)\n\n            # --- Initialize and run causal loop ---\n            init_carry_causal = (\n                step_rng,\n                current_token_idxs_BSN,\n                action_tokens_EL,\n                step_t,\n            )\n            final_carry_causal, _ = jax.lax.scan(\n                causal_step_fn, init_carry_causal, jnp.arange(N)\n            )\n            updated_token_idxs_BSN = final_carry_causal[1]\n            new_carry = (rng, updated_token_idxs_BSN)\n            return new_carry, None\n\n        # --- Run the autoregressive generation using jax.lax.scan ---\n        initial_carry = (batch[""rng""], token_idxs_BSN)\n        timesteps_to_scan = jnp.arange(T, seq_len)\n        final_carry, _ = jax.lax.scan(\n            generation_step_fn, initial_carry, timesteps_to_scan\n        )\n        final_token_idxs_BSN = final_carry[1]\n\n        # --- Decode all tokens at once at the end ---\n        H, W = batch[""videos""].shape[2:4]\n        final_frames_BSHWC = self.tokenizer.decode(\n            final_token_idxs_BSN,\n            video_hw=(H, W),\n        )\n        return final_frames_BSHWC\n\n    def vq_encode(self, batch: Dict[str, jax.Array], training: bool) -> jax.Array:\n        # --- Preprocess videos ---\n        video_BTHWC = batch[""videos""]\n        lam_output = self.lam.vq_encode(video_BTHWC, training=training)\n        lam_indices_E = lam_output[""indices""]\n        return lam_indices_E\n\n\n# FIXME (f.srambical): add conversion script for old checkpoints\ndef restore_genie_components(\n    optimizer: nnx.Optimizer,\n    sharding: jax.sharding.NamedSharding,\n    rng: jax.Array,\n    args,\n) -> nnx.Optimizer:\n    """"""Restore pre-trained Genie components""""""\n    rng_tokenizer, rng_lam = jax.random.split(rng)\n    rngs_tokenizer = nnx.Rngs(rng_tokenizer)\n    rngs_lam = nnx.Rngs(rng_lam)\n\n    tx = optimizer.tx\n    model = optimizer.model\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        step_format_fixed_length=6,\n    )\n    tokenizer_checkpoint_manager = ocp.CheckpointManager(\n        directory=args.tokenizer_checkpoint,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    dummy_tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.tokenizer_dim,\n        ffn_dim=args.tokenizer_ffn_dim,\n        latent_dim=args.latent_patch_dim,\n        num_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.tokenizer_num_blocks,\n        num_heads=args.tokenizer_num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs_tokenizer,\n    )\n    dummy_tokenizer_optimizer = nnx.Optimizer(dummy_tokenizer, tx)\n    dummy_tokenizer_optimizer_state = nnx.state(dummy_tokenizer_optimizer)\n    abstract_sharded_tokenizer_optimizer_state = _create_abstract_sharded_pytree(\n        dummy_tokenizer_optimizer_state, sharding\n    )\n    restored_tokenizer = tokenizer_checkpoint_manager.restore(\n        step=tokenizer_checkpoint_manager.latest_step(),\n        args=ocp.args.Composite(\n            model_state=ocp.args.PyTreeRestore(  # type: ignore\n                abstract_sharded_tokenizer_optimizer_state  # type: ignore\n            ),\n        ),\n    )[""model_state""]\n    nnx.update(dummy_tokenizer_optimizer.model, restored_tokenizer.model)\n    model.tokenizer = dummy_tokenizer_optimizer.model\n    tokenizer_checkpoint_manager.close()\n\n    if args.lam_checkpoint:\n        lam_checkpoint_manager = ocp.CheckpointManager(\n            directory=args.lam_checkpoint,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        dummy_lam = LatentActionModel(\n            in_dim=args.image_channels,\n            model_dim=args.lam_dim,\n            ffn_dim=args.lam_ffn_dim,\n            latent_dim=args.latent_patch_dim,\n            num_latents=args.num_latent_actions,\n            patch_size=args.lam_patch_size,\n            num_blocks=args.lam_num_blocks,\n            num_heads=args.lam_num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs_lam,\n        )\n        dummy_lam_optimizer = nnx.Optimizer(dummy_lam, tx)\n        dummy_lam_optimizer_state = nnx.state(dummy_lam_optimizer)\n        abstract_sharded_lam_optimizer_state = _create_abstract_sharded_pytree(\n            dummy_lam_optimizer_state, sharding\n        )\n        restored_lam_optimizer = lam_checkpoint_manager.restore(\n            step=lam_checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(  # type: ignore\n                    abstract_sharded_lam_optimizer_state  # type: ignore\n                ),\n            ),\n        )[""model_state""]\n        nnx.update(dummy_lam_optimizer.model, restored_lam_optimizer.model)\n        model.lam = dummy_lam_optimizer.model\n        # Remove the LAM decoder to save memory and avoid unnecessary computation.\n        del model.lam.decoder\n        lam_checkpoint_manager.close()\n\n    # Reinitialize the optimizer states\n    optimizer = nnx.Optimizer(model, tx)\n    return optimizer\n\n\ndef _create_abstract_sharded_pytree(\n    pytree_template: nnx.GraphState, sharding_spec: jax.sharding.NamedSharding\n) -> jax.Array:\n    """"""Replaces arrays in a pytree with ShapeDtypeStructs having the given sharding.""""""\n\n    def map_fn(leaf_template):\n        if hasattr(leaf_template, ""shape"") and hasattr(leaf_template, ""dtype""):\n            return jax.ShapeDtypeStruct(\n                leaf_template.shape, leaf_template.dtype, sharding=sharding_spec\n            )\n        return leaf_template\n\n    return jax.tree_util.tree_map(map_fn, pytree_template)\n",python,tab
+2,210,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"11:15:59 AM [info] Activating crowd-code\n11:15:59 AM [info] Recording started\n11:15:59 AM [info] Initializing git provider using file system watchers...\n11:15:59 AM [info] Git repository found\n11:15:59 AM [info] Git provider initialized successfully\n",Log,tab
+3,257,"extension-output-pdoom-org.crowd-code-#1-crowd-code",250,0,"11:15:59 AM [info] Initial git state: [object Object]\n",Log,content
+4,3811,"TERMINAL",0,0,"",,terminal_command
+5,10608,"TERMINAL",0,0,"",,terminal_command
+6,33220,"genie.py",0,0,"",python,tab
+7,118615,"TERMINAL",0,0,"",,terminal_command
+8,391009,"genie.py",14389,0,"",python,selection_command
+9,391255,"genie.py",14322,0,"",python,selection_command
+10,391283,"genie.py",14250,0,"",python,selection_command
+11,391319,"genie.py",14187,0,"",python,selection_command
+12,391351,"genie.py",14147,0,"",python,selection_command
+13,391384,"genie.py",14108,0,"",python,selection_command
+14,391417,"genie.py",14058,0,"",python,selection_command
+15,391451,"genie.py",13979,0,"",python,selection_command
+16,391484,"genie.py",13928,0,"",python,selection_command
+17,391520,"genie.py",13889,0,"",python,selection_command
+18,391553,"genie.py",13845,0,"",python,selection_command
+19,391584,"genie.py",13788,0,"",python,selection_command
+20,391616,"genie.py",13776,0,"",python,selection_command
+21,391650,"genie.py",13749,0,"",python,selection_command
+22,391684,"genie.py",13728,0,"",python,selection_command
+23,391717,"genie.py",13706,0,"",python,selection_command
+24,391751,"genie.py",13675,0,"",python,selection_command
+25,392003,"genie.py",13706,0,"",python,selection_command
+26,392167,"genie.py",13728,0,"",python,selection_command
+27,392324,"genie.py",13749,0,"",python,selection_command
+28,392467,"genie.py",13776,0,"",python,selection_command
+29,393820,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+30,396588,"genie.py",0,0,"",python,tab
+31,396591,"TERMINAL",0,0,"",,terminal_focus
+32,397130,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+33,397138,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+34,946805,"genie.py",13788,0,"",python,selection_command
+35,946951,"genie.py",13776,0,"",python,selection_command
+36,949220,"genie.py",13749,0,"",python,selection_command
+37,949475,"genie.py",13776,0,"",python,selection_command
+38,949647,"genie.py",13788,0,"",python,selection_command
+39,949806,"genie.py",13845,0,"",python,selection_command
+40,949972,"genie.py",13788,0,"",python,selection_command
+41,950133,"genie.py",13776,0,"",python,selection_command
+42,1000017,"input_pipeline/download/download_array_records.sh",0,0,"#!/bin/bash\n\n# Download and extract array records from Hugging Face\n# \n# This script performs a two-step process:\n# 1. Downloads compressed array records from a Hugging Face dataset repository\n# 2. Extracts the compressed tar files in parallel for better performance\n#\n# Usage:\n#   ./download_array_records.sh [hf_download_dir] [final_dataset_dir]\n#\n# Arguments:\n#   hf_download_dir    - Directory to store compressed downloads (default: data/minecraft_arrayrecords_compressed)\n#   final_dataset_dir  - Directory for extracted array records (default: data/minecraft_arrayrecords)\n\n# Set default directories if not provided as arguments\nhf_download_dir=""${1:-data/minecraft_arrayrecords_compressed}"" \nfinal_dataset_dir=""${2:-data/minecraft_arrayrecords}""          \n\nmkdir -p $hf_download_dir\nmkdir -p $final_dataset_dir\n\n# Step 1: Download compressed dataset from Hugging Face\necho ""Starting download from Hugging Face...""\nrepo_id=p-doom/open_ai_minecraft_arrayrecords_chunked\nstart_time_hf_download=$(date +%s)\n\nHF_HUB_ENABLE_HF_TRANSFER=1 HF_HUB_DISABLE_SYMLINKS=1 \\nhuggingface-cli download --repo-type dataset $repo_id --local-dir $hf_download_dir\n\nend_time_hf_download=$(date +%s)\necho ""Download completed. Time taken: $((end_time_hf_download - start_time_hf_download)) seconds""\n\n# Step 2: Extract compressed array records in parallel\necho ""Starting parallel extraction of tar files...""\nnum_workers=64  # Number of parallel extraction processes\nstart_time_uncompress=$(date +%s)\n\n# Find all shard tar files and extract them in parallel:\nxargs -0 -P $num_workers -I {} bash -c 'echo ""Extracting {}""; tar -xf ""{}"" -C ""'$final_dataset_dir'""'\n\nend_time_uncompress=$(date +%s)\n\n# Display timing summary\necho ""================================""\necho ""Extraction completed successfully!""\necho ""Uncompress time: $((end_time_uncompress - start_time_uncompress)) seconds""\necho ""Download time: $((end_time_hf_download - start_time_hf_download)) seconds""\necho ""Total time: $((end_time_uncompress - start_time_hf_download)) seconds""\necho ""Final dataset location: $final_dataset_dir""\n",shellscript,tab
+43,1002647,"input_pipeline/download/download_array_records.sh",527,0,"",shellscript,selection_mouse
+44,1003335,"input_pipeline/download/download_array_records.sh",580,0,"",shellscript,selection_mouse
+45,1005434,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+46,1006016,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+47,1006205,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+48,1006819,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+49,1006979,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+50,1007434,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+51,1007955,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+52,1008469,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+53,1008820,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+54,1009401,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+55,1009598,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+56,1010173,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+57,1010358,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+58,1010854,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+59,1010994,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+60,1012464,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+61,1012621,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+62,1013085,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+63,1013271,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+64,1013620,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+65,1013789,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+66,1014203,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+67,1014405,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+68,1014817,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+69,1014964,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+70,1015363,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+71,1015494,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+72,1015723,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+73,1015873,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+74,1016295,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+75,1016443,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+76,1016728,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+77,1016894,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+78,1017350,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+79,1017503,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+80,1032720,"input_pipeline/download/download_array_records.sh",1484,0,"",shellscript,selection_keyboard
+81,1032891,"input_pipeline/download/download_array_records.sh",2073,0,"",shellscript,selection_keyboard
+82,1035307,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_keyboard
+83,1035477,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_keyboard
+84,1036405,"input_pipeline/download/download_array_records.sh",12,0,"",shellscript,selection_command
+85,1036675,"input_pipeline/download/download_array_records.sh",13,0,"",shellscript,selection_command
+86,1036701,"input_pipeline/download/download_array_records.sh",68,0,"",shellscript,selection_command
+87,1036730,"input_pipeline/download/download_array_records.sh",71,0,"",shellscript,selection_command
+88,1036752,"input_pipeline/download/download_array_records.sh",114,0,"",shellscript,selection_command
+89,1036784,"input_pipeline/download/download_array_records.sh",193,0,"",shellscript,selection_command
+90,1036823,"input_pipeline/download/download_array_records.sh",267,0,"",shellscript,selection_command
+91,1036856,"input_pipeline/download/download_array_records.sh",269,0,"",shellscript,selection_command
+92,1036884,"input_pipeline/download/download_array_records.sh",278,0,"",shellscript,selection_command
+93,1036918,"input_pipeline/download/download_array_records.sh",348,0,"",shellscript,selection_command
+94,1036959,"input_pipeline/download/download_array_records.sh",350,0,"",shellscript,selection_command
+95,1036993,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_command
+96,1037150,"input_pipeline/download/download_array_records.sh",478,0,"",shellscript,selection_command
+97,1037316,"input_pipeline/download/download_array_records.sh",580,0,"",shellscript,selection_command
+98,1037452,"input_pipeline/download/download_array_records.sh",581,0,"",shellscript,selection_command
+99,1037630,"input_pipeline/download/download_array_records.sh",636,0,"",shellscript,selection_command
+100,1038251,"input_pipeline/download/download_array_records.sh",581,0,"",shellscript,selection_command
+101,1038366,"input_pipeline/download/download_array_records.sh",580,0,"",shellscript,selection_command
+102,1038504,"input_pipeline/download/download_array_records.sh",478,0,"",shellscript,selection_command
+103,1038761,"input_pipeline/download/download_array_records.sh",363,0,"",shellscript,selection_command
+104,1038787,"input_pipeline/download/download_array_records.sh",350,0,"",shellscript,selection_command
+105,1038821,"input_pipeline/download/download_array_records.sh",348,0,"",shellscript,selection_command
+106,1038847,"input_pipeline/download/download_array_records.sh",278,0,"",shellscript,selection_command
+107,1038883,"input_pipeline/download/download_array_records.sh",269,0,"",shellscript,selection_command
+108,1038917,"input_pipeline/download/download_array_records.sh",267,0,"",shellscript,selection_command
+109,1038950,"input_pipeline/download/download_array_records.sh",193,0,"",shellscript,selection_command
+110,1038984,"input_pipeline/download/download_array_records.sh",114,0,"",shellscript,selection_command
+111,1039019,"input_pipeline/download/download_array_records.sh",71,0,"",shellscript,selection_command
+112,1039048,"input_pipeline/download/download_array_records.sh",68,0,"",shellscript,selection_command
+113,1039083,"input_pipeline/download/download_array_records.sh",13,0,"",shellscript,selection_command
+114,1039116,"input_pipeline/download/download_array_records.sh",12,0,"",shellscript,selection_command
+115,1039147,"input_pipeline/download/download_array_records.sh",0,0,"",shellscript,selection_command
+116,1039352,"input_pipeline/download/download_array_records.sh",12,0,"",shellscript,selection_command
+117,1543400,"input_pipeline/download/openai/download_actions_files.py",0,0,"import subprocess\nimport json\nimport tyro\nfrom dataclasses import dataclass\nimport os\nfrom multiprocessing import Pool, cpu_count\nfrom tqdm import tqdm\n\n\n@dataclass\nclass Args:\n    index_file: str = ""data/open_ai_index_files/all_6xx_Jun_29.json""\n    output_dir: str = ""data/open_ai_minecraft_actions_files""\n    num_workers: int = -1  # -1 means use all available cores\n\n\ndef flatten_path(relpath):\n    """"""Convert nested path to flattened filename with subdirectory as prefix\n    e.g. data/6.10/filename.mp4 -> 6.10_filename.mp4\n    """"""\n\n    parts = relpath.split(""/"")\n\n    if len(parts) >= 3:\n        subdir = parts[1]\n        filename = parts[2]\n        return f""{subdir}_{filename}""\n    else:\n        return relpath.replace(""/"", ""_"")\n\n\ndef download_file(args):\n    try:\n        url, base_dir, output_dir = args\n        jsonl_url = url.rsplit(""."", 1)[0] + "".jsonl""\n        filename = flatten_path(jsonl_url)\n        output_file = os.path.join(output_dir, filename)\n        subprocess.run(\n            [""wget"", ""-q"", base_dir + jsonl_url, ""-O"", output_file], check=True\n        )\n        return {""file"": jsonl_url, ""success"": True}\n    except subprocess.CalledProcessError as e:\n        # delete file if it exists\n        if os.path.exists(output_file):\n            os.remove(output_file)\n        return {""file"": jsonl_url, ""success"": False, ""error"": str(e)}\n\n\ndef download_actions_files(index_file: str, output_dir: str, num_workers: int):\n    # load json file\n    with open(index_file, ""r"") as f:\n        data = json.load(f)\n\n    base_dir = data[""basedir""]\n    urls = data[""relpaths""]\n\n    # Prepare arguments for each process\n    args_list = [(url, base_dir, output_dir) for url in urls]\n\n    results = []\n    with tqdm(total=len(args_list), desc=""Downloading actions files"") as pbar:\n        with Pool(processes=num_workers) as pool:\n            for result in pool.imap_unordered(download_file, args_list):\n                results.append(result)\n                pbar.update(1)\n\n    # save results to json\n    meta_data_file_name = index_file.split(""/"")[-1].split(""."")[0] + ""_metadata.json""\n    with open(os.path.join(output_dir, meta_data_file_name), ""w"") as f:\n        json.dump(results, f)\n\n    # print number of failed downloads\n    failed_downloads = [result for result in results if not result[""success""]]\n    print(f""Number of failed downloads: {len(failed_downloads)}"")\n\n    # print number of successful downloads\n    successful_downloads = [result for result in results if result[""success""]]\n    print(f""Number of successful downloads: {len(successful_downloads)}"")\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    if not os.path.exists(args.output_dir):\n        os.makedirs(args.output_dir)\n\n    if args.num_workers == -1:\n        args.num_workers = cpu_count()\n\n    print(f""Index file: {args.index_file}"")\n    print(f""Output directory: {args.output_dir}"")\n    print(f""Number of workers: {args.num_workers}"")\n\n    download_actions_files(args.index_file, args.output_dir, args.num_workers)\n",python,tab
+118,1555091,"input_pipeline/download/openai/download_index_files.sh",0,0,"#!/bin/bash\n# Download index files from OpenAI Video-Pre-Training dataset\n# https://github.com/openai/Video-Pre-Training\n#\n# This script downloads specific JSON index files containing metadata\n# for the OpenAI Video-Pre-Training dataset from Microsoft Azure blob storage.\n#\n# Usage:\n#   ./download_index_files.sh [output_dir]\n#\n# Arguments:\n#   output_dir: Directory to save downloaded files (default: data/open_ai_index_files)\n#\n# Example:\n#   ./download_index_files.sh /path/to/custom/directory\n\n# Set output directory, use default if not provided\noutput_dir=""${1:-data/open_ai_index_files}"" \nmkdir -p $output_dir\n\n# List of index files to download\n# These files contain metadata for different video ranges (6xx, 7xx, etc.)\nindex_files=(\n    ""all_6xx_Jun_29.json""    \n    ""all_7xx_Apr_6.json""     \n    ""all_8xx_Jun_29.json""    \n    ""all_9xx_Jun_29.json""    \n    ""all_10xx_Jun_29.json""   \n)\n\n# Download each index file from Azure blob storage\nfor index_file in ""${index_files[@]}""; do\n    echo ""Downloading $index_file...""\n    wget https://openaipublic.blob.core.windows.net/minecraft-rl/snapshots/$index_file -O $output_dir/$index_file\ndone\n\necho ""Download complete. Files saved to: $output_dir""\n",shellscript,tab
+119,1558712,"input_pipeline/download/openai/download_videos.py",0,0,"import json\nimport requests\nimport os\nimport tyro\nimport logging\nfrom urllib.parse import urljoin\nfrom dataclasses import dataclass\nfrom pathlib import Path\nfrom tqdm import tqdm\nfrom multiprocessing import Pool, cpu_count\nimport time\n\n\n@dataclass\nclass DownloadVideos:\n    index_file_path: str = ""data/open_ai_index_files/all_6xx_Jun_29.json""\n    num_workers: int = -1  # -1 means use all available cores\n    output_dir: str = ""data/minecraft_videos/""\n\n\ndef download_single_file(args):\n    """"""Download a single file - designed to be used with multiprocessing""""""\n    relpath, url, output_path = args\n\n    if os.path.exists(output_path):\n        return f""Skipped {relpath} (already exists)""\n\n    # No need to create parent directories since we're flattening the structure\n    try:\n        response = requests.get(url, stream=True, timeout=30)\n        if response.status_code == 200:\n            file_size = 0\n            with open(output_path, ""wb"") as f:\n                for chunk in response.iter_content(chunk_size=8192):\n                    if chunk:\n                        f.write(chunk)\n                        file_size += len(chunk)\n\n            # Convert to MB for logging\n            file_size_mb = file_size / (1024 * 1024)\n            return f""Downloaded {relpath} ({file_size_mb:.2f} MB)""\n        else:\n            return f""Failed to download {relpath}: HTTP {response.status_code}""\n    except requests.exceptions.RequestException as e:\n        return f""Request failed for {relpath}: {e}""\n    except Exception as e:\n        return f""Unexpected error downloading {relpath}: {e}""\n\n\ndef flatten_path(relpath):\n    """"""Convert nested path to flattened filename with subdirectory as prefix\n    e.g. data/6.10/filename.mp4 -> 6.10_filename.mp4\n    """"""\n\n    parts = relpath.split(""/"")\n\n    if len(parts) >= 3:\n        subdir = parts[1]\n        filename = parts[2]\n        return f""{subdir}_{filename}""\n    else:\n        return relpath.replace(""/"", ""_"")\n\n\ndef download_dataset(index_file_path, output_dir, num_workers=64):\n    # Load the index file\n    with open(index_file_path, ""r"") as f:\n        index_data = json.load(f)\n\n    basedir = index_data[""basedir""]\n    relpaths = index_data[""relpaths""]\n\n    # Filter for mp4 files only and flatten the path structure\n    mp4_files = []\n    for relpath in relpaths:\n        if relpath.endswith("".mp4""):\n            url = urljoin(basedir, relpath)\n            flattened_filename = flatten_path(relpath)\n            output_path = os.path.join(output_dir, flattened_filename)\n            mp4_files.append((relpath, url, output_path))\n\n    print(f""Found {len(mp4_files)} MP4 files to download"")\n    print(f""Using {num_workers} workers for parallel downloads"")\n\n    start_time = time.time()\n\n    if num_workers > len(mp4_files):\n        num_workers = len(mp4_files)\n\n    with tqdm(\n        total=len(mp4_files), desc=""Overall Download Progress"", unit=""files""\n    ) as pbar:\n        with Pool(processes=num_workers) as pool:\n            results = []\n            for result in pool.imap_unordered(\n                download_single_file,\n                [\n                    (relpath, url, output_path)\n                    for relpath, url, output_path in mp4_files\n                ],\n            ):\n                results.append(result)\n                pbar.update(1)\n    # Print final results summary\n    successful_downloads = sum(1 for r in results if ""Downloaded"" in r)\n    skipped_files = sum(1 for r in results if ""Skipped"" in r)\n    failed_downloads = len(results) - successful_downloads - skipped_files\n\n    print(f""\nDownload Summary:"")\n    print(f""  Successful downloads: {successful_downloads}"")\n    print(f""  Skipped files: {skipped_files}"")\n    print(f""  Failed downloads: {failed_downloads}"")\n\n    end_time = time.time()\n    total_time = end_time - start_time\n    print(f""Download completed in {total_time:.2f} seconds"")\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(DownloadVideos)\n    os.makedirs(args.output_dir, exist_ok=True)\n\n    if args.num_workers == -1:\n        args.num_workers = cpu_count()\n\n    print(f""Index file path: {args.index_file_path}"")\n    print(f""Output directory: {args.output_dir}"")\n    print(f""Number of workers: {args.num_workers}"")\n\n    download_dataset(args.index_file_path, args.output_dir, args.num_workers)\n",python,tab
+120,1642636,"input_pipeline/download/huggingface/download_openai_array_records.sh",0,0,"#!/bin/bash\n\n# Download and extract array records from Hugging Face\n# \n# This script performs a two-step process:\n# 1. Downloads compressed array records from a Hugging Face dataset repository\n# 2. Extracts the compressed tar files in parallel for better performance\n#\n# Usage:\n#   ./download_array_records.sh [hf_download_dir] [final_dataset_dir]\n#\n# Arguments:\n#   hf_download_dir    - Directory to store compressed downloads (default: data/minecraft_arrayrecords_compressed)\n#   final_dataset_dir  - Directory for extracted array records (default: data/minecraft_arrayrecords)\n\n# Set default directories if not provided as arguments\nhf_download_dir=""${1:-data/minecraft_arrayrecords_compressed}"" \nfinal_dataset_dir=""${2:-data/minecraft_arrayrecords}""          \n\nmkdir -p $hf_download_dir\nmkdir -p $final_dataset_dir\n\n# Step 1: Download compressed dataset from Hugging Face\necho ""Starting download from Hugging Face...""\nrepo_id=p-doom/open_ai_minecraft_arrayrecords_chunked\nstart_time_hf_download=$(date +%s)\n\nHF_HUB_ENABLE_HF_TRANSFER=1 HF_HUB_DISABLE_SYMLINKS=1 \\nhuggingface-cli download --repo-type dataset $repo_id --local-dir $hf_download_dir\n\nend_time_hf_download=$(date +%s)\necho ""Download completed. Time taken: $((end_time_hf_download - start_time_hf_download)) seconds""\n\n# Step 2: Extract compressed array records in parallel\necho ""Starting parallel extraction of tar files...""\nnum_workers=64  # Number of parallel extraction processes\nstart_time_uncompress=$(date +%s)\n\n# Find all shard tar files and extract them in parallel:\nxargs -0 -P $num_workers -I {} bash -c 'echo ""Extracting {}""; tar -xf ""{}"" -C ""'$final_dataset_dir'""'\n\nend_time_uncompress=$(date +%s)\n\n# Display timing summary\necho ""================================""\necho ""Extraction completed successfully!""\necho ""Uncompress time: $((end_time_uncompress - start_time_uncompress)) seconds""\necho ""Download time: $((end_time_hf_download - start_time_hf_download)) seconds""\necho ""Total time: $((end_time_uncompress - start_time_hf_download)) seconds""\necho ""Final dataset location: $final_dataset_dir""\n",shellscript,tab
+121,1643410,"input_pipeline/download/huggingface/download_openai_array_records.sh",393,0,"",shellscript,selection_mouse
+122,1654595,"README.md",0,0,"<h1 align=""center"">🧞‍♀️ Jasmine: A simple, performant and scalable JAX-based world modeling codebase 🧞‍♀️</h1>\n\n<p align=""center"">\n    <a href= ""https://github.com/FLAIROx/jafar/blob/main/LICENSE"">\n        <img src=""https://img.shields.io/badge/license-Apache2.0-blue.svg"" /></a>\n    <a href= ""https://github.com/psf/black"">\n        <img src=""https://img.shields.io/badge/code%20style-black-000000.svg"" /></a>\n</p>\n\nJasmine is a production-ready JAX-based world modeling codebase. It currently implements the high-level architecture of [Genie: Generative Interactive Environments](https://arxiv.org/abs/2402.15391) (Bruce et al., 2024) with [MaskGIT](https://arxiv.org/abs/2202.04200) (Chang et al., 2022), as well as an autoregressive (causal) baseline. A diffusion baseline is coming soon.\n\nJasmine scales from single hosts to hundreds of xPUs thanks to XLA and strives to be an easily hackable, batteries-included foundation for world modeling research.\n\n<h2 name=""overview"" id=""overview"">Overview</h2>\n\n- Asynchronous & distributed checkpointing thanks to [orbax.checkpoint](https://github.com/google/orbax)\n    - Jasmine also supports mixing and matching hardware topologies (e.g. train on four nodes, load the checkpoint on a single node)\n- Optimized dataloading thanks to [Grain](https://github.com/google/grain)\n    - Dataloading scales with the number of processes (i.e. nodes/xPUs)\n- Checkpointing of model weights, optimizer and dataloader states\n- Full reproducibility with **identical** training curves (thanks to seeded dataloading and training, and [JAX' approach to pseudo random numbers](https://docs.jax.dev/en/latest/random-numbers.html))\n- Automatic checkpoint deletion/retention according to specified retention policy thanks to `orbax.checkpoint.CheckpointManager`\n- Mixed precision training using `bfloat16`\n    - `int8` training is on the roadmap via [aqt](https://github.com/google/aqt)\n- FlashAttention thanks to [cuDNN SDPA](https://github.com/jax-ml/jax/blob/a155c5a9997924170e0067d552351a9833c12c11/jax/_src/cudnn/fused_attention_stablehlo.py#L842)\n- Frame-level KV cache resets for accelerated spatiotemporal attention in causal baseline (still in PR)\n- Activation checkpointing (even onto host memory if desired)\n- DDP (changing to FSDP requires changing **a single line of code**)\n- WSD learning rate schedule\n    -  No need to retrain from scratch if you want to train for longer\n- Index-shuffling during dataloading\n- Google-native stack\n    - https://github.com/google/orbax for checkpointing\n    - https://github.com/google/grain for dataloading\n    - https://github.com/google-deepmind/dm_pix for image manipulation\n    - https://github.com/google/array_record as the data format\n- Easy model inspection thanks to [treescope](https://github.com/google-deepmind/treescope)\n- Modularized training script for easy inspection using notebooks ([demo notebook](https://colab.research.google.com/drive/1zHkciFIZxXloJgue9F5LtFlA0m00rJIf?usp=sharing))\n- Easy model surgery thanks to the new [flax.nnx](https://flax.readthedocs.io/en/latest/migrating/linen_to_nnx.html) API\n- [Shape suffixes](https://medium.com/@NoamShazeer/shape-suffixes-good-coding-style-f836e72e24fd) throughout the repository\n\n<h2 name=""start"" id=""start"">Setup 🧗</h2>\n\nJasmine requires `python 3.10`, `jax 0.6.2`, and `flax 0.10.7`. To install the requirements, run:\n\n```bash\npip install -r requirements.txt\npre-commit install\n```\n\n---\n\n<h2 name=""dataset"" id=""dataset"">Dataset 📂</h2>\n\nYou can either download our preprocessed dataset from [Hugging Face](https://huggingface.co/datasets/p-doom/open_ai_minecraft_arrayrecords_chunked) or preprocess [OpenAI's VPT dataset](https://github.com/openai/Video-Pre-Training) manually.\n\n### Option 1: Use Preprocessed Dataset (Recommended)\n\nThe easiest way to get started is to download our preprocessed dataset from Hugging Face. This script will handle downloading and extracting it:\n\n```bash\nbash input_pipeline/download/download_array_records.sh\n```\n\n---\n\n### Option 2: Manual Download & Preprocessing of OpenAI's VPT Dataset\n\nIf you prefer to use the raw VPT dataset from OpenAI and preprocess it yourself, follow these steps:\n\n1. **Download index files:**\n   This will download the initial index file:\n\n   ```bash\n   bash input_pipeline/download/openai/download_index_files.sh\n   ```\n\n2. **Download from all index files:**\n   This may take a long time depending on your bandwidth:\n\n   ```bash\n   python input_pipeline/download/openai/download_videos.py --index_file_path data/open_ai_index_files/all_7xx_Apr_6.json\n   python input_pipeline/download/openai/download_videos.py --index_file_path data/open_ai_index_files/all_8xx_Jun_29.json\n   python input_pipeline/download/openai/download_videos.py --index_file_path data/open_ai_index_files/all_9xx_Jun_29.json\n   python input_pipeline/download/openai/download_videos.py --index_file_path data/open_ai_index_files/all_10xx_Jun_29.json\n   ```\n\n3. **Preprocess videos into ArrayRecords:**\n   For efficient distributed training, convert the raw videos into the arrayrecord format (make sure to have [ffmpeg](https://github.com/FFmpeg/FFmpeg) installed on your machine):\n\n   ```bash\n   python input_pipeline/preprocess/video_to_array_records.py\n   ```\n\n> **Note:** This is a large dataset and may take considerable time and storage to download and process.\n\n\n<h2 name=""train"" id=""train"">Quick Start 🚀 </h2>\n\nGenie has three components: a [video tokenizer](models/tokenizer.py), a [latent action model](models/lam.py), and a [dynamics model](models/dynamics.py). Each of these components are trained separately, however, the dynamics model requires a pre-trained video tokenizer (and latent action model).\n\nTo train the video tokenizer, run:\n\n```bash\npython train_tokenizer.py --ckpt_dir <path>\n```\n\nTo train the latent action model, run:\n\n```bash\npython train_lam.py --ckpt_dir <path>\n```\n\nOnce the tokenizer and LAM are trained, the dynamics model can be trained with:\n\n```bash\npython train_dynamics.py --tokenizer_checkpoint <path> --lam_checkpoint <path>\n```\n\nLogging with `wandb` is supported. To enable logging, set the `WANDB_API_KEY` environment variable or run:\n\n```bash\nwandb login\n```\n\nTraining can then be logged by setting the `--log` flag:\n\n```bash\npython train_tokenizer.py --log --entity <wandb-entity> --project <wandb-project>\n```\n\n<h2 name=""cite"" id=""cite"">Citing 📜 </h2>\n\nJasmine was built by [Mihir Mahajan](https://maharajamihir.github.io/), [Alfred Nguyen](https://avocadoali.github.io/) and [Franz Srambical](https://srambical.fr/), but started as a fork of [Jafar](https://github.com/flairox/jafar), built by [Matthew Jackson](https://matthewtjackson.com) and [Timon Willi](https://www.timonwilli.com).\n\nIf you use Jasmine in your work, please cite us, Jafar, and the original Genie paper as follows:\n\n```\n@article{\n    mahajan2025jasmine,\n    title={Jasmine: A simple, performant and scalable JAX-based world modeling codebase},\n    author={Mihir Mahajan and Alfred Nguyen and Franz Srambical and Stefan Bauer},\n    journal = {p(doom) blog},\n    year={2025},\n    url={https://pdoom.org/jasmine.html},\n    note = {https://pdoom.org/blog.html}\n}\n```\n```\n@inproceedings{\n    willi2024jafar,\n    title={Jafar: An Open-Source Genie Reimplemention in Jax},\n    author={Timon Willi and Matthew Thomas Jackson and Jakob Nicolaus Foerster},\n    booktitle={First Workshop on Controllable Video Generation @ ICML 2024},\n    year={2024},\n    url={https://openreview.net/forum?id=ZZGaQHs9Jb}\n}\n```\n```\n@inproceedings{\n    bruce2024genie,\n    title={Genie: Generative Interactive Environments},\n    author={Jake Bruce and Michael D Dennis and Ashley Edwards and Jack Parker-Holder and Yuge Shi and Edward Hughes and Matthew Lai and Aditi Mavalankar and Richie Steigerwald and Chris Apps and Yusuf Aytar and Sarah Maria Elisabeth Bechtle and Feryal Behbahani and Stephanie C.Y. Chan and Nicolas Heess and Lucy Gonzalez and Simon Osindero and Sherjil Ozair and Scott Reed and Jingwei Zhang and Konrad Zolna and Jeff Clune and Nando de Freitas and Satinder Singh and Tim Rockt{\""a}schel},\n    booktitle={Forty-first International Conference on Machine Learning},\n    year={2024},\n    url={https://openreview.net/forum?id=bJbSbJskOS}\n}\n```\n",markdown,tab
+123,1654596,"README.md",3967,20,"download_array_recor",markdown,selection_command
+124,1656443,"README.md",3969,0,"",markdown,selection_mouse
+125,1657458,"README.md",3968,0,"",markdown,selection_mouse
+126,1658262,"README.md",3967,0,"",markdown,selection_command
+127,1659981,"README.md",3967,0,"h",markdown,content
+128,1659982,"README.md",3968,0,"",markdown,selection_keyboard
+129,1660114,"README.md",3968,0,"u",markdown,content
+130,1660114,"README.md",3969,0,"",markdown,selection_keyboard
+131,1660220,"README.md",3969,0,"g",markdown,content
+132,1660220,"README.md",3970,0,"",markdown,selection_keyboard
+133,1660343,"README.md",3970,0,"g",markdown,content
+134,1660343,"README.md",3971,0,"",markdown,selection_keyboard
+135,1660439,"README.md",3971,0,"i",markdown,content
+136,1660439,"README.md",3972,0,"",markdown,selection_keyboard
+137,1660617,"README.md",3972,0,"g",markdown,content
+138,1660618,"README.md",3973,0,"",markdown,selection_keyboard
+139,1660622,"README.md",3973,0,"n",markdown,content
+140,1660622,"README.md",3974,0,"",markdown,selection_keyboard
+141,1660992,"README.md",3974,0,"f",markdown,content
+142,1660992,"README.md",3975,0,"",markdown,selection_keyboard
+143,1660995,"README.md",3975,0,"a",markdown,content
+144,1660995,"README.md",3976,0,"",markdown,selection_keyboard
+145,1661355,"README.md",3975,1,"",markdown,content
+146,1661538,"README.md",3974,1,"",markdown,content
+147,1661662,"README.md",3973,1,"",markdown,content
+148,1662166,"README.md",3972,1,"",markdown,content
+149,1662269,"README.md",3972,0,"n",markdown,content
+150,1662269,"README.md",3973,0,"",markdown,selection_keyboard
+151,1662311,"README.md",3973,0,"g",markdown,content
+152,1662311,"README.md",3974,0,"",markdown,selection_keyboard
+153,1662825,"README.md",3974,0,"f",markdown,content
+154,1662825,"README.md",3975,0,"",markdown,selection_keyboard
+155,1662831,"README.md",3975,0,"a",markdown,content
+156,1662831,"README.md",3976,0,"",markdown,selection_keyboard
+157,1662961,"README.md",3976,0,"c",markdown,content
+158,1662961,"README.md",3977,0,"",markdown,selection_keyboard
+159,1663145,"README.md",3977,0,"e",markdown,content
+160,1663145,"README.md",3978,0,"",markdown,selection_keyboard
+161,1663197,"README.md",3978,0,"/",markdown,content
+162,1663197,"README.md",3979,0,"",markdown,selection_keyboard
+163,1663431,"README.md",3978,0,"",markdown,selection_command
+164,1667673,"README.md",3979,0,"",markdown,selection_command
+165,1667768,"README.md",3980,0,"",markdown,selection_command
+166,1667918,"README.md",3981,0,"",markdown,selection_command
+167,1668062,"README.md",3982,0,"",markdown,selection_command
+168,1668197,"README.md",3983,0,"",markdown,selection_command
+169,1668334,"README.md",3984,0,"",markdown,selection_command
+170,1668510,"README.md",3985,0,"",markdown,selection_command
+171,1668854,"README.md",3986,0,"",markdown,selection_command
+172,1669284,"README.md",3987,0,"",markdown,selection_command
+173,1669608,"README.md",3987,0,"_",markdown,content
+174,1669608,"README.md",3988,0,"",markdown,selection_keyboard
+175,1669902,"README.md",3988,0,"o",markdown,content
+176,1669902,"README.md",3989,0,"",markdown,selection_keyboard
+177,1669961,"README.md",3989,0,"p",markdown,content
+178,1669961,"README.md",3990,0,"",markdown,selection_keyboard
+179,1670127,"README.md",3990,0,"e",markdown,content
+180,1670127,"README.md",3991,0,"",markdown,selection_keyboard
+181,1670241,"README.md",3991,0,"n",markdown,content
+182,1670241,"README.md",3992,0,"",markdown,selection_keyboard
+183,1670341,"README.md",3992,0,"a",markdown,content
+184,1670341,"README.md",3993,0,"",markdown,selection_keyboard
+185,1670441,"README.md",3993,0,"i",markdown,content
+186,1670441,"README.md",3994,0,"",markdown,selection_keyboard
+187,1670873,"README.md",3993,0,"",markdown,selection_command
+188,1678944,"input_pipeline/download/huggingface/download_openai_array_records.sh",0,0,"",shellscript,tab
+189,1678944,"input_pipeline/download/huggingface/download_openai_array_records.sh",284,20,"download_array_recor",shellscript,selection_command
+190,1681059,"input_pipeline/download/huggingface/download_openai_array_records.sh",308,0,"",shellscript,selection_mouse
+191,1681865,"input_pipeline/download/huggingface/download_openai_array_records.sh",307,0,"",shellscript,selection_command
+192,1682094,"input_pipeline/download/huggingface/download_openai_array_records.sh",306,0,"",shellscript,selection_command
+193,1682125,"input_pipeline/download/huggingface/download_openai_array_records.sh",305,0,"",shellscript,selection_command
+194,1682159,"input_pipeline/download/huggingface/download_openai_array_records.sh",304,0,"",shellscript,selection_command
+195,1682194,"input_pipeline/download/huggingface/download_openai_array_records.sh",303,0,"",shellscript,selection_command
+196,1682228,"input_pipeline/download/huggingface/download_openai_array_records.sh",302,0,"",shellscript,selection_command
+197,1682262,"input_pipeline/download/huggingface/download_openai_array_records.sh",301,0,"",shellscript,selection_command
+198,1682304,"input_pipeline/download/huggingface/download_openai_array_records.sh",300,0,"",shellscript,selection_command
+199,1682327,"input_pipeline/download/huggingface/download_openai_array_records.sh",299,0,"",shellscript,selection_command
+200,1682361,"input_pipeline/download/huggingface/download_openai_array_records.sh",298,0,"",shellscript,selection_command
+201,1682447,"input_pipeline/download/huggingface/download_openai_array_records.sh",297,0,"",shellscript,selection_command
+202,1682621,"input_pipeline/download/huggingface/download_openai_array_records.sh",296,0,"",shellscript,selection_command
+203,1682776,"input_pipeline/download/huggingface/download_openai_array_records.sh",295,0,"",shellscript,selection_command
+204,1682908,"input_pipeline/download/huggingface/download_openai_array_records.sh",294,0,"",shellscript,selection_command
+205,1683066,"input_pipeline/download/huggingface/download_openai_array_records.sh",293,0,"",shellscript,selection_command
+206,1683245,"input_pipeline/download/huggingface/download_openai_array_records.sh",292,0,"",shellscript,selection_command
+207,1684057,"input_pipeline/download/huggingface/download_openai_array_records.sh",292,0,"_",shellscript,content
+208,1684058,"input_pipeline/download/huggingface/download_openai_array_records.sh",293,0,"",shellscript,selection_keyboard
+209,1684318,"input_pipeline/download/huggingface/download_openai_array_records.sh",293,0,"o",shellscript,content
+210,1684319,"input_pipeline/download/huggingface/download_openai_array_records.sh",294,0,"",shellscript,selection_keyboard
+211,1684379,"input_pipeline/download/huggingface/download_openai_array_records.sh",294,0,"p",shellscript,content
+212,1684380,"input_pipeline/download/huggingface/download_openai_array_records.sh",295,0,"",shellscript,selection_keyboard
+213,1684494,"input_pipeline/download/huggingface/download_openai_array_records.sh",295,0,"e",shellscript,content
+214,1684494,"input_pipeline/download/huggingface/download_openai_array_records.sh",296,0,"",shellscript,selection_keyboard
+215,1684587,"input_pipeline/download/huggingface/download_openai_array_records.sh",296,0,"n",shellscript,content
+216,1684588,"input_pipeline/download/huggingface/download_openai_array_records.sh",297,0,"",shellscript,selection_keyboard
+217,1684713,"input_pipeline/download/huggingface/download_openai_array_records.sh",297,0,"a",shellscript,content
+218,1684714,"input_pipeline/download/huggingface/download_openai_array_records.sh",298,0,"",shellscript,selection_keyboard
+219,1684778,"input_pipeline/download/huggingface/download_openai_array_records.sh",298,0,"i",shellscript,content
+220,1684778,"input_pipeline/download/huggingface/download_openai_array_records.sh",299,0,"",shellscript,selection_keyboard
+221,1685038,"input_pipeline/download/huggingface/download_openai_array_records.sh",298,0,"",shellscript,selection_command
+222,1707371,"input_pipeline/download/huggingface/download_openai_array_records.sh",355,0,"",shellscript,selection_command
+223,1707604,"input_pipeline/download/huggingface/download_openai_array_records.sh",298,0,"",shellscript,selection_command
+224,1708263,"input_pipeline/download/huggingface/download_openai_array_records.sh",1892,0,"",shellscript,selection_keyboard
+225,1708455,"input_pipeline/download/huggingface/download_openai_array_records.sh",2080,0,"",shellscript,selection_keyboard
+226,1710118,"input_pipeline/download/huggingface/download_openai_array_records.sh",390,0,"",shellscript,selection_keyboard
+227,1712367,"input_pipeline/download/huggingface/download_openai_array_records.sh",505,0,"",shellscript,selection_command
+228,1712603,"input_pipeline/download/huggingface/download_openai_array_records.sh",587,0,"",shellscript,selection_command
+229,1712638,"input_pipeline/download/huggingface/download_openai_array_records.sh",608,0,"",shellscript,selection_command
+230,1712744,"input_pipeline/download/huggingface/download_openai_array_records.sh",587,0,"",shellscript,selection_command
+231,1712968,"input_pipeline/download/huggingface/download_openai_array_records.sh",608,0,"",shellscript,selection_command
+232,1713150,"input_pipeline/download/huggingface/download_openai_array_records.sh",663,0,"",shellscript,selection_command
+233,1717329,"input_pipeline/download/huggingface/download_openai_array_records.sh",727,0,"",shellscript,selection_command
+234,1717441,"input_pipeline/download/huggingface/download_openai_array_records.sh",771,0,"",shellscript,selection_command
+235,1719057,"input_pipeline/download/huggingface/download_openai_array_records.sh",2080,0,"",shellscript,selection_keyboard
+236,1719529,"input_pipeline/download/huggingface/download_openai_array_records.sh",370,0,"",shellscript,selection_keyboard
+237,1719698,"input_pipeline/download/huggingface/download_openai_array_records.sh",0,0,"",shellscript,selection_keyboard
+238,1720202,"input_pipeline/download/huggingface/download_openai_array_records.sh",12,0,"",shellscript,selection_command
+239,1720350,"input_pipeline/download/huggingface/download_openai_array_records.sh",13,0,"",shellscript,selection_command
+240,1720603,"input_pipeline/download/huggingface/download_openai_array_records.sh",68,0,"",shellscript,selection_command
+241,1720634,"input_pipeline/download/huggingface/download_openai_array_records.sh",71,0,"",shellscript,selection_command
+242,1720657,"input_pipeline/download/huggingface/download_openai_array_records.sh",114,0,"",shellscript,selection_command
+243,1720693,"input_pipeline/download/huggingface/download_openai_array_records.sh",193,0,"",shellscript,selection_command
+244,1720730,"input_pipeline/download/huggingface/download_openai_array_records.sh",267,0,"",shellscript,selection_command
+245,1720760,"input_pipeline/download/huggingface/download_openai_array_records.sh",269,0,"",shellscript,selection_command
+246,1720794,"input_pipeline/download/huggingface/download_openai_array_records.sh",278,0,"",shellscript,selection_command
+247,1720822,"input_pipeline/download/huggingface/download_openai_array_records.sh",355,0,"",shellscript,selection_command
+248,1720855,"input_pipeline/download/huggingface/download_openai_array_records.sh",357,0,"",shellscript,selection_command
+249,1720891,"input_pipeline/download/huggingface/download_openai_array_records.sh",370,0,"",shellscript,selection_command
+250,1720922,"input_pipeline/download/huggingface/download_openai_array_records.sh",485,0,"",shellscript,selection_command
+251,1720958,"input_pipeline/download/huggingface/download_openai_array_records.sh",587,0,"",shellscript,selection_command
+252,1720994,"input_pipeline/download/huggingface/download_openai_array_records.sh",588,0,"",shellscript,selection_command
+253,1721112,"input_pipeline/download/huggingface/download_openai_array_records.sh",643,0,"",shellscript,selection_command
+254,1730694,"input_pipeline/download/huggingface/download_openai_array_records.sh",707,0,"",shellscript,selection_command
+255,1730947,"input_pipeline/download/huggingface/download_openai_array_records.sh",771,0,"",shellscript,selection_command
+256,1730969,"input_pipeline/download/huggingface/download_openai_array_records.sh",772,0,"",shellscript,selection_command
+257,1731194,"input_pipeline/download/huggingface/download_openai_array_records.sh",798,0,"",shellscript,selection_command
+258,1731435,"input_pipeline/download/huggingface/download_openai_array_records.sh",826,0,"",shellscript,selection_command
+259,1731553,"input_pipeline/download/huggingface/download_openai_array_records.sh",827,0,"",shellscript,selection_command
+260,1731726,"input_pipeline/download/huggingface/download_openai_array_records.sh",883,0,"",shellscript,selection_command
+261,1731914,"input_pipeline/download/huggingface/download_openai_array_records.sh",929,0,"",shellscript,selection_command
+262,1732351,"input_pipeline/download/huggingface/download_openai_array_records.sh",983,0,"",shellscript,selection_command
+263,1732489,"input_pipeline/download/huggingface/download_openai_array_records.sh",1018,0,"",shellscript,selection_command
+264,1732636,"input_pipeline/download/huggingface/download_openai_array_records.sh",1019,0,"",shellscript,selection_command
+265,1732796,"input_pipeline/download/huggingface/download_openai_array_records.sh",1075,0,"",shellscript,selection_command
+266,1746594,"input_pipeline/download/huggingface/download_openai_array_records.sh",1158,0,"",shellscript,selection_command
+267,1746840,"input_pipeline/download/huggingface/download_openai_array_records.sh",1159,0,"",shellscript,selection_command
+268,1746864,"input_pipeline/download/huggingface/download_openai_array_records.sh",1192,0,"",shellscript,selection_command
+269,1747025,"input_pipeline/download/huggingface/download_openai_array_records.sh",1290,0,"",shellscript,selection_command
+270,1747194,"input_pipeline/download/huggingface/download_openai_array_records.sh",1291,0,"",shellscript,selection_command
+271,1747612,"input_pipeline/download/huggingface/download_openai_array_records.sh",1346,0,"",shellscript,selection_command
+272,1747793,"input_pipeline/download/huggingface/download_openai_array_records.sh",1398,0,"",shellscript,selection_command
+273,1748189,"input_pipeline/download/huggingface/download_openai_array_records.sh",1456,0,"",shellscript,selection_command
+274,1748368,"input_pipeline/download/huggingface/download_openai_array_records.sh",1490,0,"",shellscript,selection_command
+275,1748542,"input_pipeline/download/huggingface/download_openai_array_records.sh",1491,0,"",shellscript,selection_command
+276,1748829,"input_pipeline/download/huggingface/download_openai_array_records.sh",1548,0,"",shellscript,selection_command
+277,1749895,"input_pipeline/download/huggingface/download_openai_array_records.sh",1650,0,"",shellscript,selection_command
+278,1750041,"input_pipeline/download/huggingface/download_openai_array_records.sh",1651,0,"",shellscript,selection_command
+279,1750233,"input_pipeline/download/huggingface/download_openai_array_records.sh",1683,0,"",shellscript,selection_command
+280,1750438,"input_pipeline/download/huggingface/download_openai_array_records.sh",1684,0,"",shellscript,selection_command
+281,1750867,"input_pipeline/download/huggingface/download_openai_array_records.sh",1709,0,"",shellscript,selection_command
+282,1751121,"input_pipeline/download/huggingface/download_openai_array_records.sh",1749,0,"",shellscript,selection_command
+283,1751151,"input_pipeline/download/huggingface/download_openai_array_records.sh",1791,0,"",shellscript,selection_command
+284,1751179,"input_pipeline/download/huggingface/download_openai_array_records.sh",1872,0,"",shellscript,selection_command
+285,1751213,"input_pipeline/download/huggingface/download_openai_array_records.sh",1953,0,"",shellscript,selection_command
+286,1751245,"input_pipeline/download/huggingface/download_openai_array_records.sh",2030,0,"",shellscript,selection_command
+287,1751282,"input_pipeline/download/huggingface/download_openai_array_records.sh",2080,0,"",shellscript,selection_command
+288,1751471,"input_pipeline/download/huggingface/download_openai_array_records.sh",2030,0,"",shellscript,selection_command
+289,1751721,"input_pipeline/download/huggingface/download_openai_array_records.sh",1953,0,"",shellscript,selection_command
+290,1751752,"input_pipeline/download/huggingface/download_openai_array_records.sh",1872,0,"",shellscript,selection_command
+291,1751775,"input_pipeline/download/huggingface/download_openai_array_records.sh",1791,0,"",shellscript,selection_command
+292,1751811,"input_pipeline/download/huggingface/download_openai_array_records.sh",1749,0,"",shellscript,selection_command
+293,1751846,"input_pipeline/download/huggingface/download_openai_array_records.sh",1709,0,"",shellscript,selection_command
+294,1751879,"input_pipeline/download/huggingface/download_openai_array_records.sh",1684,0,"",shellscript,selection_command
+295,1752010,"input_pipeline/download/huggingface/download_openai_array_records.sh",1683,0,"",shellscript,selection_command
+296,1752200,"input_pipeline/download/huggingface/download_openai_array_records.sh",1651,0,"",shellscript,selection_command
+297,1752345,"input_pipeline/download/huggingface/download_openai_array_records.sh",1650,0,"",shellscript,selection_command
+298,1752590,"input_pipeline/download/huggingface/download_openai_array_records.sh",1548,0,"",shellscript,selection_command
+299,1752618,"input_pipeline/download/huggingface/download_openai_array_records.sh",1491,0,"",shellscript,selection_command
+300,1752653,"input_pipeline/download/huggingface/download_openai_array_records.sh",1490,0,"",shellscript,selection_command
+301,1752684,"input_pipeline/download/huggingface/download_openai_array_records.sh",1456,0,"",shellscript,selection_command
+302,1752717,"input_pipeline/download/huggingface/download_openai_array_records.sh",1398,0,"",shellscript,selection_command
+303,1752750,"input_pipeline/download/huggingface/download_openai_array_records.sh",1346,0,"",shellscript,selection_command
+304,1752785,"input_pipeline/download/huggingface/download_openai_array_records.sh",1291,0,"",shellscript,selection_command
+305,1752818,"input_pipeline/download/huggingface/download_openai_array_records.sh",1290,0,"",shellscript,selection_command
+306,1752856,"input_pipeline/download/huggingface/download_openai_array_records.sh",1192,0,"",shellscript,selection_command
+307,1753307,"input_pipeline/download/huggingface/download_openai_array_records.sh",1290,0,"",shellscript,selection_command
+308,1787200,"input_pipeline/download/huggingface/download_openai_array_records.sh",0,0,"",shellscript,selection_keyboard
+309,1788022,"input_pipeline/download/huggingface/download_openai_array_records.sh",12,0,"",shellscript,selection_command
+310,1788270,"input_pipeline/download/huggingface/download_openai_array_records.sh",13,0,"",shellscript,selection_command
+311,1788300,"input_pipeline/download/huggingface/download_openai_array_records.sh",68,0,"",shellscript,selection_command
+312,1788336,"input_pipeline/download/huggingface/download_openai_array_records.sh",71,0,"",shellscript,selection_command
+313,1788358,"input_pipeline/download/huggingface/download_openai_array_records.sh",114,0,"",shellscript,selection_command
+314,1788395,"input_pipeline/download/huggingface/download_openai_array_records.sh",193,0,"",shellscript,selection_command
+315,1788427,"input_pipeline/download/huggingface/download_openai_array_records.sh",267,0,"",shellscript,selection_command
+316,1788462,"input_pipeline/download/huggingface/download_openai_array_records.sh",269,0,"",shellscript,selection_command
+317,1788495,"input_pipeline/download/huggingface/download_openai_array_records.sh",278,0,"",shellscript,selection_command
+318,1788528,"input_pipeline/download/huggingface/download_openai_array_records.sh",355,0,"",shellscript,selection_command
+319,1788561,"input_pipeline/download/huggingface/download_openai_array_records.sh",357,0,"",shellscript,selection_command
+320,1788598,"input_pipeline/download/huggingface/download_openai_array_records.sh",370,0,"",shellscript,selection_command
+321,1788627,"input_pipeline/download/huggingface/download_openai_array_records.sh",485,0,"",shellscript,selection_command
+322,1788664,"input_pipeline/download/huggingface/download_openai_array_records.sh",587,0,"",shellscript,selection_command
+323,1788699,"input_pipeline/download/huggingface/download_openai_array_records.sh",588,0,"",shellscript,selection_command
+324,1788727,"input_pipeline/download/huggingface/download_openai_array_records.sh",643,0,"",shellscript,selection_command
+325,1788932,"input_pipeline/download/huggingface/download_openai_array_records.sh",588,0,"",shellscript,selection_command
+326,1789148,"input_pipeline/download/huggingface/download_openai_array_records.sh",587,0,"",shellscript,selection_command
+327,1789502,"input_pipeline/download/huggingface/download_openai_array_records.sh",588,0,"",shellscript,selection_command
+328,1789686,"input_pipeline/download/huggingface/download_openai_array_records.sh",643,0,"",shellscript,selection_command
+329,1789858,"input_pipeline/download/huggingface/download_openai_array_records.sh",707,0,"",shellscript,selection_command
+330,1790012,"input_pipeline/download/huggingface/download_openai_array_records.sh",771,0,"",shellscript,selection_command
+331,1790420,"input_pipeline/download/huggingface/download_openai_array_records.sh",707,0,"",shellscript,selection_command
+332,1790790,"input_pipeline/download/huggingface/download_openai_array_records.sh",771,0,"",shellscript,selection_command
+333,1791834,"input_pipeline/download/huggingface/download_openai_array_records.sh",2080,0,"",shellscript,selection_keyboard
+334,1792651,"input_pipeline/download/huggingface/download_openai_array_records.sh",2030,0,"",shellscript,selection_command
+335,1792902,"input_pipeline/download/huggingface/download_openai_array_records.sh",1953,0,"",shellscript,selection_command
+336,1792927,"input_pipeline/download/huggingface/download_openai_array_records.sh",1872,0,"",shellscript,selection_command
+337,1792958,"input_pipeline/download/huggingface/download_openai_array_records.sh",1791,0,"",shellscript,selection_command
+338,1792985,"input_pipeline/download/huggingface/download_openai_array_records.sh",1749,0,"",shellscript,selection_command
+339,1793021,"input_pipeline/download/huggingface/download_openai_array_records.sh",1709,0,"",shellscript,selection_command
+340,1793052,"input_pipeline/download/huggingface/download_openai_array_records.sh",1684,0,"",shellscript,selection_command
+341,1793088,"input_pipeline/download/huggingface/download_openai_array_records.sh",1683,0,"",shellscript,selection_command
+342,1793120,"input_pipeline/download/huggingface/download_openai_array_records.sh",1651,0,"",shellscript,selection_command
+343,1793158,"input_pipeline/download/huggingface/download_openai_array_records.sh",1650,0,"",shellscript,selection_command
+344,1793187,"input_pipeline/download/huggingface/download_openai_array_records.sh",1548,0,"",shellscript,selection_command
+345,1793223,"input_pipeline/download/huggingface/download_openai_array_records.sh",1491,0,"",shellscript,selection_command
+346,1793254,"input_pipeline/download/huggingface/download_openai_array_records.sh",1490,0,"",shellscript,selection_command
+347,1793290,"input_pipeline/download/huggingface/download_openai_array_records.sh",1456,0,"",shellscript,selection_command
+348,1793321,"input_pipeline/download/huggingface/download_openai_array_records.sh",1398,0,"",shellscript,selection_command
+349,1793357,"input_pipeline/download/huggingface/download_openai_array_records.sh",1346,0,"",shellscript,selection_command
+350,1793388,"input_pipeline/download/huggingface/download_openai_array_records.sh",1291,0,"",shellscript,selection_command
+351,1793424,"input_pipeline/download/huggingface/download_openai_array_records.sh",1290,0,"",shellscript,selection_command
+352,1793467,"input_pipeline/download/huggingface/download_openai_array_records.sh",1192,0,"",shellscript,selection_command
+353,1793492,"input_pipeline/download/huggingface/download_openai_array_records.sh",1159,0,"",shellscript,selection_command
+354,1793538,"input_pipeline/download/huggingface/download_openai_array_records.sh",1158,0,"",shellscript,selection_command
+355,1793561,"input_pipeline/download/huggingface/download_openai_array_records.sh",1075,0,"",shellscript,selection_command
+356,1793589,"input_pipeline/download/huggingface/download_openai_array_records.sh",1019,0,"",shellscript,selection_command
+357,1823637,"input_pipeline/download/huggingface/download_openai_array_records.sh",0,0,"",shellscript,selection_keyboard
+358,1824075,"input_pipeline/download/huggingface/download_openai_array_records.sh",1491,0,"",shellscript,selection_keyboard
+359,1824222,"input_pipeline/download/huggingface/download_openai_array_records.sh",2080,0,"",shellscript,selection_keyboard
+360,1824862,"input_pipeline/download/huggingface/download_openai_array_records.sh",2030,0,"",shellscript,selection_command
+361,1825112,"input_pipeline/download/huggingface/download_openai_array_records.sh",1953,0,"",shellscript,selection_command
+362,1825135,"input_pipeline/download/huggingface/download_openai_array_records.sh",1872,0,"",shellscript,selection_command
+363,1825169,"input_pipeline/download/huggingface/download_openai_array_records.sh",1791,0,"",shellscript,selection_command
+364,1825203,"input_pipeline/download/huggingface/download_openai_array_records.sh",1749,0,"",shellscript,selection_command
+365,1825237,"input_pipeline/download/huggingface/download_openai_array_records.sh",1709,0,"",shellscript,selection_command
+366,1825271,"input_pipeline/download/huggingface/download_openai_array_records.sh",1684,0,"",shellscript,selection_command
+367,1825304,"input_pipeline/download/huggingface/download_openai_array_records.sh",1683,0,"",shellscript,selection_command
+368,1825339,"input_pipeline/download/huggingface/download_openai_array_records.sh",1651,0,"",shellscript,selection_command
+369,1825368,"input_pipeline/download/huggingface/download_openai_array_records.sh",1650,0,"",shellscript,selection_command
+370,1825403,"input_pipeline/download/huggingface/download_openai_array_records.sh",1548,0,"",shellscript,selection_command
+371,1825436,"input_pipeline/download/huggingface/download_openai_array_records.sh",1491,0,"",shellscript,selection_command
+372,1825470,"input_pipeline/download/huggingface/download_openai_array_records.sh",1490,0,"",shellscript,selection_command
+373,1825503,"input_pipeline/download/huggingface/download_openai_array_records.sh",1456,0,"",shellscript,selection_command
+374,1825538,"input_pipeline/download/huggingface/download_openai_array_records.sh",1398,0,"",shellscript,selection_command
+375,1825572,"input_pipeline/download/huggingface/download_openai_array_records.sh",1346,0,"",shellscript,selection_command
+376,1825601,"input_pipeline/download/huggingface/download_openai_array_records.sh",1291,0,"",shellscript,selection_command
+377,1825639,"input_pipeline/download/huggingface/download_openai_array_records.sh",1290,0,"",shellscript,selection_command
+378,1825673,"input_pipeline/download/huggingface/download_openai_array_records.sh",1192,0,"",shellscript,selection_command
+379,1825704,"input_pipeline/download/huggingface/download_openai_array_records.sh",1159,0,"",shellscript,selection_command
+380,1825737,"input_pipeline/download/huggingface/download_openai_array_records.sh",1158,0,"",shellscript,selection_command
+381,1825771,"input_pipeline/download/huggingface/download_openai_array_records.sh",1075,0,"",shellscript,selection_command
+382,1825805,"input_pipeline/download/huggingface/download_openai_array_records.sh",1019,0,"",shellscript,selection_command
+383,1825849,"input_pipeline/download/huggingface/download_openai_array_records.sh",1018,0,"",shellscript,selection_command
+384,1825868,"input_pipeline/download/huggingface/download_openai_array_records.sh",983,0,"",shellscript,selection_command
+385,1825905,"input_pipeline/download/huggingface/download_openai_array_records.sh",929,0,"",shellscript,selection_command
+386,1825939,"input_pipeline/download/huggingface/download_openai_array_records.sh",883,0,"",shellscript,selection_command
+387,1826229,"input_pipeline/download/huggingface/download_openai_array_records.sh",929,0,"",shellscript,selection_command
+388,1826392,"input_pipeline/download/huggingface/download_openai_array_records.sh",983,0,"",shellscript,selection_command
+389,1826595,"input_pipeline/download/huggingface/download_openai_array_records.sh",929,0,"",shellscript,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-e590a6b9-4abf-46d8-be83-6ab91f3e20861755118129671-2025_08_13-22.49.18.19/source.csv

@@ -0,0 +1,14 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"test/test.py",0,0,"# import jax\n# with jax.profiler.trace(""tensorboard""):\n  # key = jax.random.key(0)\n  # x = jax.random.normal(key, (1024, 1024))\n  # y = x @ x\n  # y.block_until_ready()\n\nimport datetime\nimport jax.numpy as jnp\nimport jax\n\nMATRIX_DIM = 32768\nSTEPS = 10\n\nA = jnp.ones((MATRIX_DIM, MATRIX_DIM))\nB = jnp.ones((MATRIX_DIM, MATRIX_DIM))\n\nnum_bytes = A.size * 4\ntotal_num_bytes_crossing_to_hbm = num_bytes * 3\n\ntotal_num_flops = 2 * MATRIX_DIM * MATRIX_DIM**2\n\ndef matmul(A, B):\n  return A @ B\n\nmatmul(A, B) # warmup\n\nstart_time = datetime.datetime.now()\nfor i in range(STEPS):\n  C = A @ B\n  C.block_until_ready()\nend_time = datetime.datetime.now()\n\naverage_time_per_step = (end_time - start_time).total_seconds() / STEPS\n\nprint(f""{average_time_per_step}, teraflops per second: {total_num_flops / average_time_per_step / 1e12}, gigabytes per second: {total_num_bytes_crossing_to_hbm / average_time_per_step / 1e9}"")",python,tab
+2,1286,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:49:18 PM [info] Activating crowd-code\n10:49:18 PM [info] Recording started\n10:49:18 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,2287,"extension-output-pdoom-org.crowd-code-#1-crowd-code",153,0,"10:49:19 PM [info] Git repository found\n10:49:19 PM [info] Git provider initialized successfully\n10:49:19 PM [info] Initial git state: [object Object]\n",Log,content
+4,2330,"test/test.py",0,0,"",python,tab
+5,2333,"TERMINAL",0,0,"",,terminal_focus
+6,3521,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+7,3529,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+8,4418,"TERMINAL",0,0,"squee",,terminal_command
+9,4510,"TERMINAL",0,0,"]633;Cbash: squee: command not found\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+10,5865,"TERMINAL",0,0,"squeue",,terminal_command
+11,5866,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  18743 franz.sram interacti     1     1  R 2025-08-13T21:45:17 2025-08-13T21:45:17  1:04:06 1-00:00:00 hai002\r\n  18721   xiao.liu interacti     1    32  R 2025-08-13T16:44:08 2025-08-13T16:44:08  6:05:15   23:59:00 hai001\r\n  18363   xiao.liu interacti     1    32  R 2025-08-13T13:42:17 2025-08-13T13:42:17  9:07:06   23:59:00 hai006\r\n  18742 alfred.ngu  standard     1    20  R 2025-08-13T19:59:10 2025-08-13T19:59:10  2:50:13 1-00:00:00 hai002\r\n  18725 alfred.ngu  standard     1    32  R 2025-08-13T18:15:19 2025-08-13T18:15:19  4:34:04 1-00:00:00 hai003\r\n  18726 alfred.ngu  standard     1    32  R 2025-08-13T18:15:19 2025-08-13T18:15:19  4:34:04 1-00:00:00 hai008\r\n  18620 nishant.ku  standard     3   192  R 2025-08-13T16:09:20 2025-08-13T16:09:20  6:40:03 1-00:00:00 hai[004-005,007]\r\n]0;franz.srambical@hai-login1:~/jafar",,terminal_output
+12,12250,"TERMINAL",0,0,"scancel --me",,terminal_command
+13,12264,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login1:~/jafar",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-ee737e55-5872-414d-b78b-de2c071ad9761758201107741-2025_09_18-15.11.58.718/source.csv

@@ -0,0 +1,244 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"train_dynamics.py",0,0,"import os\n\n\nos.environ.setdefault(""XLA_PYTHON_CLIENT_MEM_FRACTION"", ""0.98"")\n\nfrom dataclasses import dataclass, field\nimport itertools\nfrom typing import cast, Optional\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_type: str = ""maskgit""  # supported options: maskgit, causal\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    val_data_dir: str = """"\n    val_interval: int = 20_000\n    val_steps: int = 50\n    eval_full_frame: bool = False\n    val_maskgit_steps: int = 25\n    val_temperature: float = 1\n    val_sample_argmax: bool = False\n    wandb_id: str = """"\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[Genie, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        # Dynamics\n        dyna_type=args.dyna_type,\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        decode=False,\n        rngs=rngs,\n    )\n    del genie.lam.decoder\n    return genie, rng\n\n\ndef build_optimizer(genie: Genie, args: Args) -> tuple[nnx.Optimizer, optax.Schedule]:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.Optimizer(genie, tx)\n    return optimizer, lr_schedule\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.Optimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args, data_dir: str) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(data_dir, x)\n        for x in os.listdir(data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> ocp.CheckpointManager:\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""train_dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""train_dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    if args.val_data_dir:\n        handler_registry.add(\n            ""val_dataloader_state"",\n            grain.checkpoint.CheckpointSave,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        handler_registry.add(\n            ""val_dataloader_state"",\n            grain.checkpoint.CheckpointRestore,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    return checkpoint_manager\n\n\ndef restore_or_initialize_components(\n    args: Args,\n    checkpoint_manager: ocp.CheckpointManager,\n    optimizer: nnx.Optimizer,\n    train_iterator: grain.DataLoaderIterator,\n    rng: jax.Array,\n    replicated_sharding: NamedSharding,\n    val_iterator: Optional[grain.DataLoaderIterator],\n    restore_step: Optional[int] = None,\n) -> tuple[\n    int, nnx.Optimizer, grain.DataLoaderIterator, grain.DataLoaderIterator, jax.Array\n]:\n    step = 0\n    if restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        if val_iterator:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n                val_dataloader_state=grain.checkpoint.CheckpointRestore(val_iterator),  # type: ignore\n            )\n        else:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n            )\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(), args=restore_args\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        train_iterator = restored[""train_dataloader_state""]\n        if val_iterator:\n            val_iterator = restored[""val_dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        rng, _rng = jax.random.split(rng)\n        optimizer = restore_genie_components(optimizer, replicated_sharding, _rng, args)\n        # NOTE: We have to remove the (unused) tokenizer vq dropout due flax.nnx lazily initializing modules.\n        # Specifically, the first dynamics model checkpoint will contain the vq dropout module,\n        # but the first full restore will fail due to nnx not initializing the module when\n        # dropout is set to 0.0.\n        del optimizer.model.tokenizer.vq.drop\n    return step, optimizer, train_iterator, val_iterator, rng\n\n\ndef _calculate_step_metrics(\n    outputs: dict[str, jax.Array],\n    gt: jax.Array,\n    num_latent_actions: int,\n    num_patch_latents: int,\n) -> tuple[jax.Array, dict]:\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    gt_val = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt_val, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt_val, recon)).mean()\n    _, index_counts_lam = jnp.unique_counts(\n        jnp.ravel(outputs[""lam_indices""]),\n        size=num_latent_actions,\n        fill_value=0,\n    )\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]),\n        size=num_patch_latents,\n        fill_value=0,\n    )\n    codebook_usage_lam = (index_counts_lam != 0).mean()\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_lam=codebook_usage_lam,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    return ce_loss, metrics\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    genie, rng = build_model(args, rng)\n    _, params, _ = nnx.split(genie, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer, lr_schedule = build_optimizer(genie, args)\n    del genie\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    _, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    train_iterator = build_dataloader(args, args.data_dir)\n    val_iterator = None\n    if args.val_data_dir:\n        val_iterator = build_dataloader(args, args.val_data_dir)\n\n    # --- Restore checkpoint ---\n    step, optimizer, train_iterator, val_iterator, rng = (\n        restore_or_initialize_components(\n            args,\n            checkpoint_manager,\n            optimizer,\n            train_iterator,\n            rng,\n            replicated_sharding,\n            val_iterator,\n        )\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def dynamics_loss_fn(\n        model: Genie,\n        inputs: dict,\n        training: bool = False,\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        outputs = model(inputs, training=training)\n        ce_loss, metrics = _calculate_step_metrics(\n            outputs, gt, args.num_latent_actions, args.num_patch_latents\n        )\n        return ce_loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.Optimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: Genie) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            model.train()\n            return dynamics_loss_fn(model, inputs, training=True)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""dynamics""]\n            )\n        return loss, recon, metrics\n\n    @nnx.jit()\n    def val_step(genie: Genie, inputs: dict) -> dict:\n        """"""Evaluate model and compute metrics""""""\n        genie.eval()\n        (loss, (recon, metrics)) = dynamics_loss_fn(genie, inputs, training=False)\n        val_output = {""loss"": loss, ""recon"": recon, ""metrics"": metrics}\n\n        # --- Evaluate full frame prediction (sampling) ---\n        if args.eval_full_frame:\n            lam_indices = genie.vq_encode(inputs, training=False)\n            tokenizer_outputs = genie.tokenizer.vq_encode(\n                inputs[""videos""], training=False\n            )\n            tokens_full_frame = tokenizer_outputs[""indices""]\n            inputs[""latent_actions""] = lam_indices\n            gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n            inputs[""videos""] = gt[:, :-1].astype(\n                args.dtype\n            )  # remove last frame for generation\n            recon_full_frame, logits_full_frame = genie.sample(\n                inputs,\n                args.seq_len,\n                args.val_temperature,\n                args.val_sample_argmax,\n                args.val_maskgit_steps,\n            )\n            step_outputs = {\n                ""recon"": recon_full_frame,\n                ""token_logits"": logits_full_frame,\n                ""video_tokens"": tokens_full_frame,\n                ""mask"": jnp.zeros_like(tokens_full_frame).at[:, -1].set(True),\n                ""lam_indices"": lam_indices,\n            }\n            loss_full_frame, metrics_full_frame = _calculate_step_metrics(\n                step_outputs, gt, args.num_latent_actions, args.num_patch_latents\n            )\n            val_output.update(\n                {\n                    ""loss_full_frame"": loss_full_frame,\n                    ""recon_full_frame"": recon_full_frame,\n                    ""metrics_full_frame"": metrics_full_frame,\n                }\n            )\n        return val_output\n\n    def calculate_validation_metrics(val_dataloader, genie, rng):\n        step = 0\n        loss_per_step = []\n        metrics_per_step = []\n        loss_full_frame_per_step = []\n        metrics_full_frame_per_step = []\n        inputs = None\n        recon = None\n        recon_full_frame = None\n        for videos in val_dataloader:\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(videos=videos, rng=_rng_mask)\n            val_outputs = val_step(genie, inputs)\n            loss_per_step.append(val_outputs[""loss""])\n            metrics_per_step.append(val_outputs[""metrics""])\n            if args.eval_full_frame:\n                loss_full_frame_per_step.append(val_outputs[""loss_full_frame""])\n                metrics_full_frame_per_step.append(val_outputs[""metrics_full_frame""])\n            step += 1\n            if step > args.val_steps:\n                break\n\n        if step < args.val_steps:\n            print(\n                f""Warning: Your validation dataset is too small to make val_steps many steps. Made {step} steps, expected {args.val_steps}""\n            )\n\n        val_metrics = {\n            f""val_{key}"": np.mean([float(m[key]) for m in metrics_per_step])\n            for key in metrics_per_step[0].keys()\n        }\n        val_metrics[""val_loss""] = np.mean(loss_per_step)\n        if args.eval_full_frame:\n            val_metrics_full_frame = {\n                f""val_full_frame_{key}"": np.mean(\n                    [float(m[key]) for m in metrics_full_frame_per_step]\n                )\n                for key in metrics_full_frame_per_step[0].keys()\n            }\n            val_metrics.update(val_metrics_full_frame)\n            val_metrics[""val_loss_full_frame""] = np.mean(loss_full_frame_per_step)\n        return val_metrics, inputs, recon, recon_full_frame\n\n    # --- TRAIN LOOP ---\n    dataloader_train = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in train_iterator\n    )\n    dataloader_val = None\n    if val_iterator:\n        dataloader_val = (\n            jax.make_array_from_process_local_data(videos_sharding, elem)\n            for elem in val_iterator\n        )\n    if jax.process_index() == 0:\n        first_videos = next(dataloader_train)\n        sample_inputs = dict(videos=first_videos, rng=rng)\n        compiled = train_step.lower(optimizer, sample_inputs).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader_train = itertools.chain([first_videos], dataloader_train)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for videos in dataloader_train:\n            # --- Train step ---\n            rng, _rng_mask = jax.random.split(rng, 2)\n            inputs = dict(videos=videos, rng=_rng_mask)\n            loss, recon, metrics = train_step(optimizer, inputs)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Validation loss ---\n            val_results = {}\n            if dataloader_val and step % args.val_interval == 0:\n                rng, _rng_mask_val = jax.random.split(rng, 2)\n                print(""Calculating validation metrics..."")\n                val_metrics, val_gt_batch, val_recon, val_recon_full_frame = (\n                    calculate_validation_metrics(\n                        dataloader_val, optimizer.model, _rng_mask_val\n                    )\n                )\n                print(f""Step {step}, validation loss: {val_metrics['val_loss']}"")\n                val_results = {\n                    ""metrics"": val_metrics,\n                    ""gt_batch"": val_gt_batch,\n                    ""recon"": val_recon,\n                    ""full_frame"": val_recon_full_frame,\n                }\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    log_dict = {""loss"": loss, ""step"": step, **metrics}\n                    if val_results:\n                        log_dict.update(val_results[""metrics""])\n                    wandb.log(log_dict)\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if val_results:\n                        val_results[""gt_seq_val""] = (\n                            val_results[""gt_batch""][""videos""][0].astype(jnp.float32)\n                            / 255.0\n                        )\n                        val_results[""recon_seq_val""] = val_results[""recon""].clip(0, 1)\n                        val_comparison_seq = jnp.concatenate(\n                            (val_results[""gt_seq_val""], val_results[""recon_seq_val""]),\n                            axis=1,\n                        )\n                        val_results[""val_comparison_seq""] = einops.rearrange(\n                            val_comparison_seq * 255, ""t h w c -> h (t w) c""\n                        )\n                        if args.eval_full_frame:\n                            val_results[""full_frame_seq_val""] = val_results[\n                                ""full_frame""\n                            ][0].clip(0, 1)\n                            val_results[""val_full_frame_comparison_seq""] = (\n                                jnp.concatenate(\n                                    (\n                                        val_results[""gt_seq_val""],\n                                        val_results[""full_frame_seq_val""],\n                                    ),\n                                    axis=1,\n                                )\n                            )\n                            val_results[""val_full_frame_comparison_seq""] = (\n                                einops.rearrange(\n                                    val_results[""val_full_frame_comparison_seq""] * 255,\n                                    ""t h w c -> h (t w) c"",\n                                )\n                            )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                            recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        if val_results:\n                            log_images.update(\n                                dict(\n                                    val_image=wandb.Image(\n                                        np.asarray(\n                                            val_results[""gt_seq_val""][args.seq_len - 1]\n                                        )\n                                    ),\n                                    val_recon=wandb.Image(\n                                        np.asarray(\n                                            val_results[""recon_seq_val""][\n                                                args.seq_len - 1\n                                            ]\n                                        )\n                                    ),\n                                    val_true_vs_recon=wandb.Image(\n                                        np.asarray(\n                                            val_results[""val_comparison_seq""].astype(\n                                                np.uint8\n                                            )\n                                        )\n                                    ),\n                                )\n                            )\n                            if args.eval_full_frame:\n                                log_images.update(\n                                    dict(\n                                        val_full_frame=wandb.Image(\n                                            np.asarray(\n                                                val_results[""full_frame_seq_val""][\n                                                    args.seq_len - 1\n                                                ]\n                                            )\n                                        ),\n                                        val_true_vs_full_frame=wandb.Image(\n                                            np.asarray(\n                                                val_results[\n                                                    ""val_full_frame_comparison_seq""\n                                                ].astype(np.uint8)\n                                            )\n                                        ),\n                                    )\n                                )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                if val_iterator:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                        val_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            val_iterator  # type: ignore\n                        ),\n                    )\n                else:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                    )\n                checkpoint_manager.save(step, args=ckpt_manager_args)\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+2,281,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"3:11:58 PM [info] Activating crowd-code\n3:11:58 PM [info] Recording started\n3:11:58 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,465,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"3:11:58 PM [info] Git repository found\n3:11:58 PM [info] Git provider initialized successfully\n3:11:58 PM [info] Initial git state: [object Object]\n",Log,content
+4,1574,"train_dynamics.py",0,0,"",python,tab
+5,2175,"TERMINAL",0,0,"",,terminal_command
+6,5067,"utils/dataloader.py",0,0,"import jax\nimport numpy as np\nimport grain\nfrom typing import Any\nimport pickle\n\n\nclass EpisodeLengthFilter(grain.transforms.Filter):\n    """"""\n    A Grain Filter that keeps only episodes with sufficient length.\n    """"""\n\n    def __init__(self, seq_len: int, image_h: int, image_w: int, image_c: int):\n        """"""Initializes the filter with sequence length requirements.""""""\n        self.seq_len = seq_len\n        self.image_h = image_h\n        self.image_w = image_w\n        self.image_c = image_c\n\n    def filter(self, element: Any) -> bool:\n        """"""\n        Filters episodes based on length.\n\n        Args:\n            element: A dictionary representing one record from the DataSource.\n                     Expected to contain 'raw_video' (bytes) and 'sequence_length' (int)\n\n        Returns:\n            True if the episode has sufficient length, False otherwise.\n        """"""\n        assert isinstance(element, bytes)\n        element = pickle.loads(element)\n\n        current_episode_len = element[""sequence_length""]\n        if current_episode_len < self.seq_len:\n            print(\n                f""Filtering out episode with length {current_episode_len}, which is ""\n                f""shorter than the requested sequence length {self.seq_len}.""\n            )\n            return False\n\n        return True\n\n\nclass ProcessEpisodeAndSlice(grain.transforms.RandomMap):\n    """"""\n    A Grain Transformation that combines parsing, slicing, and normalizing.\n    """"""\n\n    def __init__(self, seq_len: int, image_h: int, image_w: int, image_c: int):\n        """"""Initializes the transformation with processing parameters.""""""\n        self.seq_len = seq_len\n        self.image_h = image_h\n        self.image_w = image_w\n        self.image_c = image_c\n\n    def random_map(self, element: dict, rng: np.random.Generator) -> Any:\n        """"""\n        Processes a single raw episode from the data source.\n\n        Args:\n            element: A dictionary representing one record from the DataSource.\n                     Expected to contain 'raw_video' (bytes) and 'sequence_length' (int)\n            rng: A per-record random number generator provided by the Grain sampler.\n\n        Returns:\n            A processed video sequence as a NumPy array with shape\n            (seq_len, height, width, channels) and dtype float32.\n        """"""\n        assert isinstance(element, bytes)\n        element = pickle.loads(element)\n\n        video_shape = (\n            element[""sequence_length""],\n            self.image_h,\n            self.image_w,\n            self.image_c,\n        )\n        episode_tensor = np.frombuffer(element[""raw_video""], dtype=np.uint8)\n        episode_tensor = episode_tensor.reshape(video_shape)\n\n        current_episode_len = episode_tensor.shape[0]\n        if current_episode_len < self.seq_len:\n            raise ValueError(\n                f""Episode length {current_episode_len} is shorter than ""\n                f""requested sequence length {self.seq_len}. This should ""\n                f""have been filtered out.""\n            )\n\n        max_start_idx = current_episode_len - self.seq_len\n\n        start_idx = rng.integers(0, max_start_idx + 1)\n\n        seq = episode_tensor[start_idx : start_idx + self.seq_len]\n\n        return seq\n\n\ndef get_dataloader(\n    array_record_paths: list[str],\n    seq_len: int,\n    global_batch_size: int,\n    image_h: int,\n    image_w: int,\n    image_c: int,\n    num_workers: int = 1,\n    prefetch_buffer_size: int = 1,\n    seed: int = 42,\n):\n    """"""\n    Creates a data loading pipeline using Grain.\n    """"""\n    if not array_record_paths:\n        raise ValueError(""array_record_paths list cannot be empty."")\n\n    num_processes = jax.process_count()\n\n    if global_batch_size % num_processes != 0:\n        raise ValueError(\n            f""Global batch size {global_batch_size} must be divisible by ""\n            f""the number of JAX processes {num_processes} for proper sharding.""\n        )\n    per_process_batch_size = global_batch_size // num_processes\n\n    source = grain.sources.ArrayRecordDataSource(array_record_paths)\n\n    sampler = grain.samplers.IndexSampler(\n        num_records=len(source),\n        shard_options=grain.sharding.ShardByJaxProcess(drop_remainder=True),\n        shuffle=True,\n        num_epochs=None,\n        seed=seed,\n    )\n\n    operations = [\n        EpisodeLengthFilter(\n            seq_len=seq_len, image_h=image_h, image_w=image_w, image_c=image_c\n        ),\n        ProcessEpisodeAndSlice(\n            seq_len=seq_len, image_h=image_h, image_w=image_w, image_c=image_c\n        ),\n        grain.transforms.Batch(batch_size=per_process_batch_size, drop_remainder=True),\n    ]\n\n    read_options = grain.ReadOptions(\n        prefetch_buffer_size=prefetch_buffer_size,\n        num_threads=1,\n    )\n    dataloader = grain.DataLoader(\n        data_source=source,\n        sampler=sampler,\n        operations=operations,\n        worker_count=num_workers,\n        worker_buffer_size=1,\n        read_options=read_options,\n    )\n\n    return dataloader\n",python,tab
+7,9501,"TERMINAL",0,0,"",,terminal_command
+8,11665,"utils/dataloader.py",3212,0,"",python,selection_keyboard
+9,12159,"utils/dataloader.py",1814,0,"",python,selection_keyboard
+10,13609,"utils/dataloader.py",496,0,"",python,selection_keyboard
+11,15927,"utils/dataloader.py",1814,0,"",python,selection_keyboard
+12,16439,"utils/dataloader.py",3212,0,"",python,selection_keyboard
+13,17206,"utils/dataloader.py",1814,0,"",python,selection_keyboard
+14,19465,"utils/dataloader.py",1740,0,"",python,selection_command
+15,19551,"utils/dataloader.py",1739,0,"",python,selection_command
+16,19663,"utils/dataloader.py",1708,0,"",python,selection_command
+17,19928,"utils/dataloader.py",1677,0,"",python,selection_command
+18,20097,"utils/dataloader.py",1646,0,"",python,selection_command
+19,20097,"utils/dataloader.py",1615,0,"",python,selection_command
+20,20098,"utils/dataloader.py",1542,0,"",python,selection_command
+21,20099,"utils/dataloader.py",1462,0,"",python,selection_command
+22,20099,"utils/dataloader.py",1461,0,"",python,selection_command
+23,20102,"utils/dataloader.py",1453,0,"",python,selection_command
+24,20166,"utils/dataloader.py",1377,0,"",python,selection_command
+25,20505,"utils/dataloader.py",1453,0,"",python,selection_command
+26,20573,"utils/dataloader.py",1377,0,"",python,selection_command
+27,20742,"utils/dataloader.py",1369,0,"",python,selection_command
+28,20982,"utils/dataloader.py",1311,0,"",python,selection_command
+29,21086,"utils/dataloader.py",1310,0,"",python,selection_command
+30,22825,"utils/dataloader.py",80,0,"",python,selection_keyboard
+31,23030,"utils/dataloader.py",0,0,"",python,selection_keyboard
+32,24103,"utils/dataloader.py",11,0,"",python,selection_command
+33,24358,"utils/dataloader.py",30,0,"",python,selection_command
+34,24360,"utils/dataloader.py",43,0,"",python,selection_command
+35,24559,"utils/dataloader.py",1289,0,"",python,selection_keyboard
+36,24727,"utils/dataloader.py",2630,0,"",python,selection_keyboard
+37,24863,"utils/dataloader.py",3679,0,"",python,selection_keyboard
+38,25049,"utils/dataloader.py",4969,0,"",python,selection_keyboard
+39,25129,"utils/dataloader.py",4998,0,"",python,selection_keyboard
+40,26615,"utils/dataloader.py",3827,0,"",python,selection_keyboard
+41,27165,"utils/dataloader.py",4998,0,"",python,selection_keyboard
+42,27882,"utils/dataloader.py",4976,0,"",python,selection_command
+43,28153,"utils/dataloader.py",4975,0,"",python,selection_command
+44,28166,"utils/dataloader.py",4969,0,"",python,selection_command
+45,28167,"utils/dataloader.py",4934,0,"",python,selection_command
+46,28204,"utils/dataloader.py",4904,0,"",python,selection_command
+47,28271,"utils/dataloader.py",4870,0,"",python,selection_command
+48,28284,"utils/dataloader.py",4839,0,"",python,selection_command
+49,28301,"utils/dataloader.py",4814,0,"",python,selection_command
+50,28344,"utils/dataloader.py",4786,0,"",python,selection_command
+51,28390,"utils/dataloader.py",4751,0,"",python,selection_command
+52,28446,"utils/dataloader.py",4745,0,"",python,selection_command
+53,28594,"utils/dataloader.py",4722,0,"",python,selection_command
+54,28595,"utils/dataloader.py",4671,0,"",python,selection_command
+55,28595,"utils/dataloader.py",4633,0,"",python,selection_command
+56,28681,"utils/dataloader.py",4671,0,"",python,selection_command
+57,28984,"utils/dataloader.py",4722,0,"",python,selection_command
+58,29143,"utils/dataloader.py",4745,0,"",python,selection_command
+59,29270,"utils/dataloader.py",4751,0,"",python,selection_command
+60,29463,"utils/dataloader.py",4745,0,"",python,selection_command
+61,29721,"utils/dataloader.py",4722,0,"",python,selection_command
+62,29725,"utils/dataloader.py",4671,0,"",python,selection_command
+63,29792,"utils/dataloader.py",4633,0,"",python,selection_command
+64,29912,"utils/dataloader.py",4632,0,"",python,selection_command
+65,30104,"utils/dataloader.py",4626,0,"",python,selection_command
+66,30361,"utils/dataloader.py",4538,0,"",python,selection_command
+67,30362,"utils/dataloader.py",4527,0,"",python,selection_command
+68,30381,"utils/dataloader.py",4448,0,"",python,selection_command
+69,30447,"utils/dataloader.py",4416,0,"",python,selection_command
+70,30582,"utils/dataloader.py",4405,0,"",python,selection_command
+71,31033,"utils/dataloader.py",4326,0,"",python,selection_command
+72,31035,"utils/dataloader.py",4297,0,"",python,selection_command
+73,31287,"utils/dataloader.py",4326,0,"",python,selection_command
+74,31431,"utils/dataloader.py",4405,0,"",python,selection_command
+75,31543,"utils/dataloader.py",4416,0,"",python,selection_command
+76,31832,"utils/dataloader.py",4424,0,"",python,selection_command
+77,31993,"utils/dataloader.py",4446,0,"",python,selection_command
+78,32306,"utils/dataloader.py",4424,0,"",python,selection_command
+79,32713,"utils/dataloader.py",1317,0,"",python,selection_command
+80,35543,"utils/dataloader.py",1375,0,"",python,selection_command
+81,35790,"utils/dataloader.py",1383,0,"",python,selection_command
+82,35815,"utils/dataloader.py",1459,0,"",python,selection_command
+83,35990,"utils/dataloader.py",1461,0,"",python,selection_command
+84,36247,"utils/dataloader.py",1468,0,"",python,selection_command
+85,36263,"utils/dataloader.py",1548,0,"",python,selection_command
+86,36310,"utils/dataloader.py",1621,0,"",python,selection_command
+87,36344,"utils/dataloader.py",1652,0,"",python,selection_command
+88,36345,"utils/dataloader.py",1683,0,"",python,selection_command
+89,36390,"utils/dataloader.py",1714,0,"",python,selection_command
+90,36519,"utils/dataloader.py",1739,0,"",python,selection_command
+91,36706,"utils/dataloader.py",1746,0,"",python,selection_command
+92,37128,"utils/dataloader.py",1739,0,"",python,selection_command
+93,46592,"utils/dataloader.py",1746,0,"",python,selection_command
+94,46600,"utils/dataloader.py",1820,0,"",python,selection_command
+95,47112,"utils/dataloader.py",1832,0,"",python,selection_command
+96,47480,"utils/dataloader.py",1887,0,"",python,selection_command
+97,47481,"utils/dataloader.py",1894,0,"",python,selection_command
+98,47672,"utils/dataloader.py",1908,0,"",python,selection_command
+99,47673,"utils/dataloader.py",1987,0,"",python,selection_command
+100,47673,"utils/dataloader.py",2076,0,"",python,selection_command
+101,47674,"utils/dataloader.py",2155,0,"",python,selection_command
+102,47674,"utils/dataloader.py",2162,0,"",python,selection_command
+103,47675,"utils/dataloader.py",2179,0,"",python,selection_command
+104,47675,"utils/dataloader.py",2246,0,"",python,selection_command
+105,47676,"utils/dataloader.py",2312,0,"",python,selection_command
+106,47869,"utils/dataloader.py",2324,0,"",python,selection_command
+107,48004,"utils/dataloader.py",2366,0,"",python,selection_command
+108,48164,"utils/dataloader.py",2400,0,"",python,selection_command
+109,65559,"utils/dataloader.py",2407,0,"",python,selection_command
+110,65788,"utils/dataloader.py",2431,0,"",python,selection_command
+111,65818,"utils/dataloader.py",2471,0,"",python,selection_command
+112,65880,"utils/dataloader.py",2497,0,"",python,selection_command
+113,66076,"utils/dataloader.py",2471,0,"",python,selection_command
+114,66364,"utils/dataloader.py",2431,0,"",python,selection_command
+115,66365,"utils/dataloader.py",2407,0,"",python,selection_command
+116,66404,"utils/dataloader.py",2400,0,"",python,selection_command
+117,66618,"utils/dataloader.py",2366,0,"",python,selection_command
+118,66846,"utils/dataloader.py",2400,0,"",python,selection_command
+119,67110,"utils/dataloader.py",2407,0,"",python,selection_command
+120,67129,"utils/dataloader.py",2431,0,"",python,selection_command
+121,67135,"utils/dataloader.py",2471,0,"",python,selection_command
+122,67303,"utils/dataloader.py",2497,0,"",python,selection_command
+123,67463,"utils/dataloader.py",2523,0,"",python,selection_command
+124,67605,"utils/dataloader.py",2549,0,"",python,selection_command
+125,67800,"utils/dataloader.py",2559,0,"",python,selection_command
+126,67912,"utils/dataloader.py",2636,0,"",python,selection_command
+127,68727,"utils/dataloader.py",2691,0,"",python,selection_command
+128,68920,"utils/dataloader.py",2698,0,"",python,selection_command
+129,69032,"utils/dataloader.py",2752,0,"",python,selection_command
+130,69212,"utils/dataloader.py",2799,0,"",python,selection_command
+131,69444,"utils/dataloader.py",2829,0,"",python,selection_command
+132,69527,"utils/dataloader.py",2902,0,"",python,selection_command
+133,69527,"utils/dataloader.py",2976,0,"",python,selection_command
+134,69542,"utils/dataloader.py",3019,0,"",python,selection_command
+135,69817,"utils/dataloader.py",3027,0,"",python,selection_command
+136,69854,"utils/dataloader.py",3034,0,"",python,selection_command
+137,70696,"utils/dataloader.py",3036,0,"",python,selection_command
+138,71104,"utils/dataloader.py",3050,0,"",python,selection_command
+139,71824,"utils/dataloader.py",3036,0,"",python,selection_command
+140,72248,"utils/dataloader.py",3087,0,"",python,selection_command
+141,72374,"utils/dataloader.py",3096,0,"",python,selection_command
+142,72545,"utils/dataloader.py",3106,0,"",python,selection_command
+143,72743,"utils/dataloader.py",3108,0,"",python,selection_command
+144,72861,"utils/dataloader.py",3111,0,"",python,selection_command
+145,73047,"utils/dataloader.py",3112,0,"",python,selection_command
+146,88223,"utils/dataloader.py",3111,0,"",python,selection_command
+147,116063,"TERMINAL",0,0,"",,terminal_command
+148,986707,"utils/dataloader.py",3212,19,"        data_dict = {""videos"": seq}\n        if ""actions"" in element.keys():\n            actions_tensor = np.array(element[""actions""])\n            actions = actions_tensor[start_idx : start_idx + self.seq_len]\n            data_dict[""actions""] = actions\n\n        return data_dict\n",python,content
+149,990332,"utils/dataloader.py",0,0,"Switched from branch 'main' to 'gt-actions'",python,git_branch_checkout
+150,999370,"input_pipeline/generate_coinrun_dataset.py",0,0,"""""""\nGenerates a dataset of random-action CoinRun episodes.\nEpisodes are saved individually as memory-mapped files for efficient loading.\n""""""\n\nfrom dataclasses import dataclass\n\nfrom gym3 import types_np\nimport numpy as np\nfrom procgen import ProcgenGym3Env\nimport tyro\nimport json\nimport os\nfrom utils import save_chunks\n\n\n@dataclass\nclass Args:\n    num_episodes_train: int = 10000\n    num_episodes_val: int = 500\n    num_episodes_test: int = 500\n    output_dir: str = ""data/coinrun_episodes""\n    min_episode_length: int = 1000\n    max_episode_length: int = 1000\n    chunk_size: int = 100\n    chunks_per_file: int = 100\n    seed: int = 0\n\n\nargs = tyro.cli(Args)\nassert (\n    args.max_episode_length >= args.min_episode_length\n), ""Maximum episode length must be greater than or equal to minimum episode length.""\n\nif args.min_episode_length < args.chunk_size:\n    print(\n        ""Warning: Minimum episode length is smaller than chunk size. Note that episodes shorter than the chunk size will be discarded.""\n    )\n\n\n# --- Generate episodes ---\ndef generate_episodes(num_episodes, split):\n    episode_idx = 0\n    episode_metadata = []\n    obs_chunks = []\n    act_chunks = []\n    file_idx = 0\n    output_dir_split = os.path.join(args.output_dir, split)\n    while episode_idx < num_episodes:\n        seed = np.random.randint(0, 10000)\n        env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=seed)\n\n        observations_seq = []\n        actions_seq = []\n        episode_obs_chunks = []\n        episode_act_chunks = []\n\n        # --- Run episode ---\n        step_t = 0\n        for step_t in range(args.max_episode_length):\n            action = types_np.sample(env.ac_space, bshape=(env.num,))\n            env.act(action)\n            _, obs, first = env.observe()\n            observations_seq.append(obs[""rgb""])\n            actions_seq.append(action)\n            if len(observations_seq) == args.chunk_size:\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n                observations_seq = []\n                actions_seq = []\n            if first:\n                break\n\n        # --- Save episode ---\n        if step_t + 1 >= args.min_episode_length:\n            if observations_seq:\n                if len(observations_seq) < args.chunk_size:\n                    print(\n                        f""Warning: Inconsistent chunk_sizes. Episode has {len(observations_seq)} frames, ""\n                        f""which is smaller than the requested chunk_size: {args.chunk_size}. ""\n                        ""This might lead to performance degradation during training.""\n                    )\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n\n            obs_chunks_data = [\n                np.concatenate(seq, axis=0).astype(np.uint8)\n                for seq in episode_obs_chunks\n            ]\n            act_chunks_data = [\n                np.concatenate(act, axis=0) for act in episode_act_chunks\n            ]\n            obs_chunks.extend(obs_chunks_data)\n            act_chunks.extend(act_chunks_data)\n\n            ep_metadata, obs_chunks, file_idx, act_chunks = save_chunks(\n                obs_chunks, file_idx, args.chunks_per_file, output_dir_split, act_chunks\n            )\n            episode_metadata.extend(ep_metadata)\n\n            print(f""Episode {episode_idx} completed, length: {step_t + 1}."")\n            episode_idx += 1\n        else:\n            print(f""Episode too short ({step_t + 1}), resampling..."")\n\n    if len(obs_chunks) > 0:\n        print(\n            f""Warning: Dropping {len(obs_chunks)} chunks for consistent number of chunks per file."",\n            ""Consider changing the chunk_size and chunks_per_file parameters to prevent data-loss."",\n        )\n\n    print(f""Done generating {split} split"")\n    return episode_metadata\n\n\ndef get_action_space():\n    env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=0)\n    return env.ac_space.eltype.n\n\n\ndef main():\n    # Set random seed and create dataset directories\n    np.random.seed(args.seed)\n    # --- Generate episodes ---\n    train_episode_metadata = generate_episodes(args.num_episodes_train, ""train"")\n    val_episode_metadata = generate_episodes(args.num_episodes_val, ""val"")\n    test_episode_metadata = generate_episodes(args.num_episodes_test, ""test"")\n\n    # --- Save metadata ---\n    metadata = {\n        ""env"": ""coinrun"",\n        ""num_actions"": get_action_space(),\n        ""num_episodes_train"": args.num_episodes_train,\n        ""num_episodes_val"": args.num_episodes_val,\n        ""num_episodes_test"": args.num_episodes_test,\n        ""avg_episode_len_train"": np.mean(\n            [ep[""avg_seq_len""] for ep in train_episode_metadata]\n        ),\n        ""avg_episode_len_val"": np.mean(\n            [ep[""avg_seq_len""] for ep in val_episode_metadata]\n        ),\n        ""avg_episode_len_test"": np.mean(\n            [ep[""avg_seq_len""] for ep in test_episode_metadata]\n        ),\n        ""episode_metadata_train"": train_episode_metadata,\n        ""episode_metadata_val"": val_episode_metadata,\n        ""episode_metadata_test"": test_episode_metadata,\n    }\n    with open(os.path.join(args.output_dir, ""metadata.json""), ""w"") as f:\n        json.dump(metadata, f)\n\n    print(f""Done generating dataset."")\n\n\nif __name__ == ""__main__"":\n    main()\n",python,tab
+151,1002105,"input_pipeline/generate_coinrun_dataset.py",323,0,"",python,selection_keyboard
+152,1003826,"input_pipeline/generate_coinrun_dataset.py",1252,0,"",python,selection_command
+153,1005699,"input_pipeline/generate_coinrun_dataset.py",1827,0,"",python,selection_command
+154,1006307,"input_pipeline/generate_coinrun_dataset.py",2569,0,"",python,selection_command
+155,1006769,"input_pipeline/generate_coinrun_dataset.py",1827,0,"",python,selection_command
+156,1006933,"input_pipeline/generate_coinrun_dataset.py",1252,0,"",python,selection_command
+157,1009410,"input_pipeline/generate_coinrun_dataset.py",4014,0,"",python,selection_command
+158,1012656,"input_pipeline/generate_coinrun_dataset.py",4026,0,"",python,selection_command
+159,1012887,"input_pipeline/generate_coinrun_dataset.py",4079,0,"",python,selection_command
+160,1012943,"input_pipeline/generate_coinrun_dataset.py",4109,0,"",python,selection_command
+161,1013073,"input_pipeline/generate_coinrun_dataset.py",4141,0,"",python,selection_command
+162,1014866,"input_pipeline/generate_coinrun_dataset.py",4164,0,"",python,selection_command
+163,1015048,"input_pipeline/generate_coinrun_dataset.py",4166,0,"",python,selection_command
+164,1015514,"input_pipeline/generate_coinrun_dataset.py",1045,0,"",python,selection_command
+165,1017729,"input_pipeline/generate_coinrun_dataset.py",1532,0,"",python,selection_command
+166,1018560,"input_pipeline/generate_coinrun_dataset.py",1523,0,"",python,selection_command
+167,1018755,"input_pipeline/generate_coinrun_dataset.py",1499,0,"",python,selection_command
+168,1018760,"input_pipeline/generate_coinrun_dataset.py",1467,0,"",python,selection_command
+169,1018952,"input_pipeline/generate_coinrun_dataset.py",1442,0,"",python,selection_command
+170,1018955,"input_pipeline/generate_coinrun_dataset.py",1412,0,"",python,selection_command
+171,1018955,"input_pipeline/generate_coinrun_dataset.py",1403,0,"",python,selection_command
+172,1018956,"input_pipeline/generate_coinrun_dataset.py",1337,0,"",python,selection_command
+173,1019030,"input_pipeline/generate_coinrun_dataset.py",1294,0,"",python,selection_command
+174,1019366,"input_pipeline/generate_coinrun_dataset.py",1337,0,"",python,selection_command
+175,1019674,"input_pipeline/generate_coinrun_dataset.py",1341,0,"",python,selection_command
+176,1021167,"input_pipeline/generate_coinrun_dataset.py",1343,0,"",python,selection_command
+177,1024417,"input_pipeline/generate_coinrun_dataset.py",1341,0,"",python,selection_command
+178,1024665,"input_pipeline/generate_coinrun_dataset.py",1337,0,"",python,selection_command
+179,1028081,"input_pipeline/generate_coinrun_dataset.py",3888,0,"",python,selection_command
+180,1030249,"input_pipeline/generate_coinrun_dataset.py",4466,0,"",python,selection_command
+181,1037266,"input_pipeline/generate_coinrun_dataset.py",4440,0,"",python,selection_command
+182,1037388,"input_pipeline/generate_coinrun_dataset.py",4466,0,"",python,selection_command
+183,1037983,"input_pipeline/generate_coinrun_dataset.py",4440,0,"",python,selection_command
+184,1038160,"input_pipeline/generate_coinrun_dataset.py",4466,0,"",python,selection_command
+185,1043747,"input_pipeline/generate_coinrun_dataset.py",4440,0,"",python,selection_command
+186,1044017,"input_pipeline/generate_coinrun_dataset.py",4415,0,"",python,selection_command
+187,1044017,"input_pipeline/generate_coinrun_dataset.py",4395,0,"",python,selection_command
+188,1044113,"input_pipeline/generate_coinrun_dataset.py",4371,0,"",python,selection_command
+189,1044113,"input_pipeline/generate_coinrun_dataset.py",4316,0,"",python,selection_command
+190,1044132,"input_pipeline/generate_coinrun_dataset.py",4241,0,"",python,selection_command
+191,1044324,"input_pipeline/generate_coinrun_dataset.py",4160,0,"",python,selection_command
+192,1044704,"input_pipeline/generate_coinrun_dataset.py",4164,0,"",python,selection_command
+193,1045098,"input_pipeline/generate_coinrun_dataset.py",4166,0,"",python,selection_command
+194,1045968,"input_pipeline/generate_coinrun_dataset.py",4134,0,"",python,selection_command
+195,1046075,"input_pipeline/generate_coinrun_dataset.py",4103,0,"",python,selection_command
+196,1046099,"input_pipeline/generate_coinrun_dataset.py",4051,0,"",python,selection_command
+197,1046152,"input_pipeline/generate_coinrun_dataset.py",4020,0,"",python,selection_command
+198,1046170,"input_pipeline/generate_coinrun_dataset.py",4009,0,"",python,selection_command
+199,1046234,"input_pipeline/generate_coinrun_dataset.py",4008,0,"",python,selection_command
+200,1046234,"input_pipeline/generate_coinrun_dataset.py",4004,0,"",python,selection_command
+201,1046387,"input_pipeline/generate_coinrun_dataset.py",4008,0,"",python,selection_command
+202,1046633,"input_pipeline/generate_coinrun_dataset.py",4009,0,"",python,selection_command
+203,1046658,"input_pipeline/generate_coinrun_dataset.py",4020,0,"",python,selection_command
+204,1046688,"input_pipeline/generate_coinrun_dataset.py",4051,0,"",python,selection_command
+205,1046768,"input_pipeline/generate_coinrun_dataset.py",4103,0,"",python,selection_command
+206,1046907,"input_pipeline/generate_coinrun_dataset.py",4134,0,"",python,selection_command
+207,1047104,"input_pipeline/generate_coinrun_dataset.py",4166,0,"",python,selection_command
+208,1049570,"input_pipeline/generate_coinrun_dataset.py",1045,0,"",python,selection_command
+209,1052806,"input_pipeline/generate_coinrun_dataset.py",1089,0,"",python,selection_command
+210,1053120,"input_pipeline/generate_coinrun_dataset.py",1109,0,"",python,selection_command
+211,1053121,"input_pipeline/generate_coinrun_dataset.py",1135,0,"",python,selection_command
+212,1053122,"input_pipeline/generate_coinrun_dataset.py",1155,0,"",python,selection_command
+213,1053179,"input_pipeline/generate_coinrun_dataset.py",1175,0,"",python,selection_command
+214,1053218,"input_pipeline/generate_coinrun_dataset.py",1192,0,"",python,selection_command
+215,1053219,"input_pipeline/generate_coinrun_dataset.py",1252,0,"",python,selection_command
+216,1053450,"input_pipeline/generate_coinrun_dataset.py",1290,0,"",python,selection_command
+217,1053585,"input_pipeline/generate_coinrun_dataset.py",1333,0,"",python,selection_command
+218,1054064,"input_pipeline/generate_coinrun_dataset.py",1337,0,"",python,selection_command
+219,1054265,"input_pipeline/generate_coinrun_dataset.py",1341,0,"",python,selection_command
+220,1054584,"input_pipeline/generate_coinrun_dataset.py",1343,0,"",python,selection_command
+221,1062563,"input_pipeline/generate_coinrun_dataset.py",4166,0,"",python,selection_command
+222,1067237,"input_pipeline/generate_coinrun_dataset.py",1045,0,"",python,selection_command
+223,1105393,"input_pipeline/generate_coinrun_dataset.py",1089,0,"",python,selection_command
+224,1105628,"input_pipeline/generate_coinrun_dataset.py",1109,0,"",python,selection_command
+225,1105657,"input_pipeline/generate_coinrun_dataset.py",1135,0,"",python,selection_command
+226,1105688,"input_pipeline/generate_coinrun_dataset.py",1155,0,"",python,selection_command
+227,1105753,"input_pipeline/generate_coinrun_dataset.py",1175,0,"",python,selection_command
+228,1105778,"input_pipeline/generate_coinrun_dataset.py",1192,0,"",python,selection_command
+229,1105795,"input_pipeline/generate_coinrun_dataset.py",1252,0,"",python,selection_command
+230,1105841,"input_pipeline/generate_coinrun_dataset.py",1290,0,"",python,selection_command
+231,1105898,"input_pipeline/generate_coinrun_dataset.py",1333,0,"",python,selection_command
+232,1105904,"input_pipeline/generate_coinrun_dataset.py",1403,0,"",python,selection_command
+233,1106016,"input_pipeline/generate_coinrun_dataset.py",1408,0,"",python,selection_command
+234,1106241,"input_pipeline/generate_coinrun_dataset.py",1403,0,"",python,selection_command
+235,1106410,"input_pipeline/generate_coinrun_dataset.py",1412,0,"",python,selection_command
+236,1106608,"input_pipeline/generate_coinrun_dataset.py",1403,0,"",python,selection_command
+237,1106785,"input_pipeline/generate_coinrun_dataset.py",1337,0,"",python,selection_command
+238,1106986,"input_pipeline/generate_coinrun_dataset.py",1294,0,"",python,selection_command
+239,1107195,"input_pipeline/generate_coinrun_dataset.py",1337,0,"",python,selection_command
+240,1107474,"input_pipeline/generate_coinrun_dataset.py",1341,0,"",python,selection_command
+241,1107704,"input_pipeline/generate_coinrun_dataset.py",1343,0,"",python,selection_command
+242,1107921,"input_pipeline/generate_coinrun_dataset.py",1357,0,"",python,selection_command
+243,1108321,"input_pipeline/generate_coinrun_dataset.py",1343,0,"",python,selection_command