Add files using upload-large-folder tool

LTA_openwebtext_dualt/logs/ar_owt_gpt2_len1024_from100k_modelonly_lr1e4_wd0p1_b2p95_cosine_4gpu_smoke.log

@@ -0,0 +1,79 @@
+
+*****************************************
+Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+*****************************************
+resumed_from=runs/ar_owt_gpt2_small_len1024_stream_gbs512_8gpu_1m_nw2/step_0100000.pt start_step=1 model_only=True
+{
+  "task": "ar_lm",
+  "device": "cuda:0",
+  "rank": 0,
+  "world_size": 4,
+  "samples": "wrapped_stream",
+  "vocab_size": 50257,
+  "bos_id": 50256,
+  "eos_id": 50256,
+  "save_dir": "runs/ar_owt_gpt2_len1024_from100k_modelonly_lr1e4_wd0p1_b2p95_cosine_4gpu_smoke",
+  "params": 123597312,
+  "batch_size": 32,
+  "grad_accum": 4,
+  "effective_batch_size": 512,
+  "global_batch_size": 512,
+  "max_len": 1024,
+  "wrap": true,
+  "wrap_mode": "stream",
+  "wrap_record_buffer_size": 200,
+  "text_detokenizer": null,
+  "openwebtext_split": "train_minus_100k",
+  "torch_compile": false
+}
+step=10 micro_steps=40 elapsed=64.8s lr=1.100000e-05 loss=3.0297 ppl=20.7870 acc=0.4218 tokens=32736.0000
+step=20 micro_steps=80 elapsed=40.3s lr=2.100000e-05 loss=3.0076 ppl=20.5603 acc=0.4284 tokens=32736.0000
+step=30 micro_steps=120 elapsed=42.0s lr=3.100000e-05 loss=3.0299 ppl=20.9048 acc=0.4223 tokens=32736.0000
+step=40 micro_steps=160 elapsed=43.8s lr=4.100000e-05 loss=3.0612 ppl=21.4560 acc=0.4161 tokens=32736.0000
+step=50 micro_steps=200 elapsed=39.3s lr=5.100000e-05 loss=3.0069 ppl=20.4106 acc=0.4251 tokens=32736.0000
+step=60 micro_steps=240 elapsed=41.4s lr=6.100000e-05 loss=3.0282 ppl=20.7534 acc=0.4220 tokens=32736.0000
+step=70 micro_steps=280 elapsed=40.5s lr=7.100000e-05 loss=3.0461 ppl=21.1605 acc=0.4203 tokens=32736.0000
+step=80 micro_steps=320 elapsed=39.7s lr=8.100000e-05 loss=3.0507 ppl=21.2206 acc=0.4187 tokens=32736.0000
+step=90 micro_steps=360 elapsed=43.5s lr=9.100000e-05 loss=3.0519 ppl=21.3442 acc=0.4205 tokens=32736.0000
+step=100 micro_steps=400 elapsed=37.2s lr=1.000000e-04 loss=3.0295 ppl=20.8015 acc=0.4238 tokens=32736.0000
+step=110 micro_steps=440 elapsed=54.4s lr=9.996954e-05 loss=3.0593 ppl=21.4342 acc=0.4160 tokens=32736.0000
+step=120 micro_steps=480 elapsed=39.0s lr=9.987820e-05 loss=3.0617 ppl=21.4853 acc=0.4155 tokens=32736.0000
+step=130 micro_steps=520 elapsed=42.9s lr=9.972609e-05 loss=3.0732 ppl=21.7631 acc=0.4173 tokens=32736.0000
+step=140 micro_steps=560 elapsed=39.2s lr=9.951340e-05 loss=3.0504 ppl=21.2292 acc=0.4199 tokens=32736.0000
+step=150 micro_steps=600 elapsed=39.7s lr=9.924039e-05 loss=3.0754 ppl=22.0218 acc=0.4179 tokens=32736.0000
+step=160 micro_steps=640 elapsed=41.2s lr=9.890738e-05 loss=3.0479 ppl=21.2237 acc=0.4195 tokens=32736.0000
+step=170 micro_steps=680 elapsed=38.6s lr=9.851479e-05 loss=3.0272 ppl=20.8744 acc=0.4226 tokens=32736.0000
+W0512 18:20:42.896000 230962 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
+W0512 18:20:42.899000 230962 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 231028 closing signal SIGTERM
+W0512 18:20:42.900000 230962 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 231029 closing signal SIGTERM
+W0512 18:20:42.900000 230962 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 231030 closing signal SIGTERM
+W0512 18:20:42.900000 230962 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 231031 closing signal SIGTERM
+Traceback (most recent call last):
+  File "<frozen runpy>", line 198, in _run_module_as_main
+  File "<frozen runpy>", line 88, in _run_code
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 922, in <module>
+    main()
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
+    return f(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 918, in main
+    run(args)
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 909, in run
+    elastic_launch(
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
+    return launch_agent(self._config, self._entrypoint, list(args))
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
+    result = agent.run()
+             ^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 137, in wrapper
+    result = f(*args, **kwargs)
+             ^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
+    result = self._invoke_run(role)
+             ^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
+    time.sleep(monitor_interval)
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
+    raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
+torch.distributed.elastic.multiprocessing.api.SignalException: Process 230962 got signal: 15

LTA_openwebtext_dualt/logs/elfaligned_t5record_8gpu/lta_owt_t5record_len1024_elfaligned_dditelf_muon_logitnormal_m1p5_s0p8_gbs512_8gpu_5epoch_20260515_223036.log

@@ -0,0 +1,642 @@
+t-20260516062948-lst5d-worker-0:10241:10241 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10241:10241 [0] NCCL INFO Bootstrap: Using eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10241:10241 [0] NCCL INFO cudaDriverVersion 12080
+t-20260516062948-lst5d-worker-0:10241:10241 [0] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516062948-lst5d-worker-0:10241:10241 [0] NCCL INFO Comm config Blocking set to 1
+t-20260516062948-lst5d-worker-0:10247:10247 [6] NCCL INFO cudaDriverVersion 12080
+t-20260516062948-lst5d-worker-0:10247:10247 [6] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10247:10247 [6] NCCL INFO Bootstrap: Using eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10247:10247 [6] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516062948-lst5d-worker-0:10247:10247 [6] NCCL INFO Comm config Blocking set to 1
+t-20260516062948-lst5d-worker-0:10243:10243 [2] NCCL INFO cudaDriverVersion 12080
+t-20260516062948-lst5d-worker-0:10243:10243 [2] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10243:10243 [2] NCCL INFO Bootstrap: Using eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10243:10243 [2] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516062948-lst5d-worker-0:10243:10243 [2] NCCL INFO Comm config Blocking set to 1
+t-20260516062948-lst5d-worker-0:10242:10242 [1] NCCL INFO cudaDriverVersion 12080
+t-20260516062948-lst5d-worker-0:10242:10242 [1] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10242:10242 [1] NCCL INFO Bootstrap: Using eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10242:10242 [1] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516062948-lst5d-worker-0:10242:10242 [1] NCCL INFO Comm config Blocking set to 1
+t-20260516062948-lst5d-worker-0:10248:10248 [7] NCCL INFO cudaDriverVersion 12080
+t-20260516062948-lst5d-worker-0:10248:10248 [7] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10248:10248 [7] NCCL INFO Bootstrap: Using eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10248:10248 [7] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516062948-lst5d-worker-0:10248:10248 [7] NCCL INFO Comm config Blocking set to 1
+t-20260516062948-lst5d-worker-0:10246:10246 [5] NCCL INFO cudaDriverVersion 12080
+t-20260516062948-lst5d-worker-0:10246:10246 [5] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10246:10246 [5] NCCL INFO Bootstrap: Using eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10246:10246 [5] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516062948-lst5d-worker-0:10245:10245 [4] NCCL INFO cudaDriverVersion 12080
+t-20260516062948-lst5d-worker-0:10245:10245 [4] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10245:10245 [4] NCCL INFO Bootstrap: Using eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10245:10245 [4] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516062948-lst5d-worker-0:10244:10244 [3] NCCL INFO cudaDriverVersion 12080
+t-20260516062948-lst5d-worker-0:10244:10244 [3] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10244:10244 [3] NCCL INFO Bootstrap: Using eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10244:10244 [3] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516062948-lst5d-worker-0:10246:10246 [5] NCCL INFO Comm config Blocking set to 1
+t-20260516062948-lst5d-worker-0:10245:10245 [4] NCCL INFO Comm config Blocking set to 1
+t-20260516062948-lst5d-worker-0:10244:10244 [3] NCCL INFO Comm config Blocking set to 1
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Using network IBext_v9
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO Using network IBext_v9
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO Using network IBext_v9
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO Using network IBext_v9
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO ncclCommInitRankConfig comm 0x9929b80 rank 0 nranks 8 cudaDev 0 nvmlDev 0 busId 65040 commId 0x51bb37a311f021b4 - Init START
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO ncclCommInitRankConfig comm 0xaa53d10 rank 6 nranks 8 cudaDev 6 nvmlDev 6 busId 73020 commId 0x51bb37a311f021b4 - Init START
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO Using network IBext_v9
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO Using network IBext_v9
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO Using network IBext_v9
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.80.4<0>
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO Using network IBext_v9
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO ncclCommInitRankConfig comm 0x9aad980 rank 2 nranks 8 cudaDev 2 nvmlDev 2 busId 69020 commId 0x51bb37a311f021b4 - Init START
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO ncclCommInitRankConfig comm 0xaf65240 rank 1 nranks 8 cudaDev 1 nvmlDev 1 busId 67020 commId 0x51bb37a311f021b4 - Init START
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO ncclCommInitRankConfig comm 0xb1ff580 rank 5 nranks 8 cudaDev 5 nvmlDev 5 busId 71020 commId 0x51bb37a311f021b4 - Init START
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO ncclCommInitRankConfig comm 0x9f87a30 rank 7 nranks 8 cudaDev 7 nvmlDev 7 busId 75020 commId 0x51bb37a311f021b4 - Init START
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO ncclCommInitRankConfig comm 0x9a6d0d0 rank 4 nranks 8 cudaDev 4 nvmlDev 4 busId 6f020 commId 0x51bb37a311f021b4 - Init START
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO ncclCommInitRankConfig comm 0xb291ad0 rank 3 nranks 8 cudaDev 3 nvmlDev 3 busId 6b020 commId 0x51bb37a311f021b4 - Init START
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO Bootstrap timings total 0.013801 (create 0.000021, send 0.000065, recv 0.000101, ring 0.005575, delay 0.000001)
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO Bootstrap timings total 0.006025 (create 0.000021, send 0.000065, recv 0.000083, ring 0.002751, delay 0.000001)
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO Bootstrap timings total 0.003151 (create 0.000014, send 0.000040, recv 0.000083, ring 0.000131, delay 0.000001)
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO Bootstrap timings total 0.498213 (create 0.000021, send 0.000069, recv 0.497686, ring 0.000105, delay 0.000001)
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO Bootstrap timings total 0.006545 (create 0.000022, send 0.000066, recv 0.000083, ring 0.006083, delay 0.000000)
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO Bootstrap timings total 0.275214 (create 0.000023, send 0.000068, recv 0.000123, ring 0.274695, delay 0.000001)
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Bootstrap timings total 0.598817 (create 0.000026, send 0.000074, recv 0.323721, ring 0.005980, delay 0.000001)
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO Bootstrap timings total 0.581473 (create 0.000020, send 0.000060, recv 0.575020, ring 0.005994, delay 0.000001)
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO MNNVL busId 0x71020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO MNNVL busId 0x6f020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO MNNVL busId 0x6b020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO MNNVL busId 0x75020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO MNNVL busId 0x67020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO MNNVL busId 0x65040 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO MNNVL busId 0x69020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO MNNVL busId 0x73020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO Setting affinity for GPU 5 to 0fffff,ffffffff,ffffffff,fc000000,00000000,00000000
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO NVLS multicast support is available on dev 5
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO Setting affinity for GPU 6 to 0fffff,ffffffff,ffffffff,fc000000,00000000,00000000
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO Setting affinity for GPU 2 to 03ffffff,ffffffff,ffffffff
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO NVLS multicast support is available on dev 2
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO Setting affinity for GPU 4 to 0fffff,ffffffff,ffffffff,fc000000,00000000,00000000
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO Setting affinity for GPU 1 to 03ffffff,ffffffff,ffffffff
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO NVLS multicast support is available on dev 6
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO NVLS multicast support is available on dev 1
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Setting affinity for GPU 0 to 03ffffff,ffffffff,ffffffff
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO NVLS multicast support is available on dev 0
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO Setting affinity for GPU 3 to 03ffffff,ffffffff,ffffffff
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO NVLS multicast support is available on dev 4
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO Setting affinity for GPU 7 to 0fffff,ffffffff,ffffffff,fc000000,00000000,00000000
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO NVLS multicast support is available on dev 7
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO NVLS multicast support is available on dev 3
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO comm 0x9aad980 rank 2 nRanks 8 nNodes 1 localRanks 8 localRank 2 MNNVL 0
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO comm 0xaf65240 rank 1 nRanks 8 nNodes 1 localRanks 8 localRank 1 MNNVL 0
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO comm 0x9929b80 rank 0 nRanks 8 nNodes 1 localRanks 8 localRank 0 MNNVL 0
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO comm 0x9f87a30 rank 7 nRanks 8 nNodes 1 localRanks 8 localRank 7 MNNVL 0
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO comm 0xaa53d10 rank 6 nRanks 8 nNodes 1 localRanks 8 localRank 6 MNNVL 0
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO comm 0xb1ff580 rank 5 nRanks 8 nNodes 1 localRanks 8 localRank 5 MNNVL 0
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO comm 0x9a6d0d0 rank 4 nRanks 8 nNodes 1 localRanks 8 localRank 4 MNNVL 0
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO comm 0xb291ad0 rank 3 nRanks 8 nNodes 1 localRanks 8 localRank 3 MNNVL 0
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 00/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 3/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 3/-1/-1->2->1 [4] 3/-1/-1->2->1 [5] 3/-1/-1->2->1 [6] 3/-1/-1->2->1 [7] 3/-1/-1->2->1 [8] 3/-1/-1->2->1 [9] 3/-1/-1->2->1 [10] 3/-1/-1->2->1 [11] 3/-1/-1->2->1 [12] 3/-1/-1->2->1 [13] 3/-1/-1->2->1 [14] 3/-1/-1->2->1 [15] 3/-1/-1->2->1 [16] 3/-1/-1->2->1 [17] 3/-1/-1->2->1 [18] 3/-1/-1->2->1 [19] 3/-1/-1->2->1 [20] 3/-1/-1->2->1 [21] 3/-1/-1->2->1 [22] 3/-1/-1->2->1 [23] 3/-1/-1->2->1
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 01/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO P2P Chunksize set to 524288
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 02/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 03/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 04/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 05/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 06/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 07/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 08/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO Trees [0] 7/-1/-1->6->5 [1] 7/-1/-1->6->5 [2] 7/-1/-1->6->5 [3] 7/-1/-1->6->5 [4] 7/-1/-1->6->5 [5] 7/-1/-1->6->5 [6] 7/-1/-1->6->5 [7] 7/-1/-1->6->5 [8] 7/-1/-1->6->5 [9] 7/-1/-1->6->5 [10] 7/-1/-1->6->5 [11] 7/-1/-1->6->5 [12] 7/-1/-1->6->5 [13] 7/-1/-1->6->5 [14] 7/-1/-1->6->5 [15] 7/-1/-1->6->5 [16] 7/-1/-1->6->5 [17] 7/-1/-1->6->5 [18] 7/-1/-1->6->5 [19] 7/-1/-1->6->5 [20] 7/-1/-1->6->5 [21] 7/-1/-1->6->5 [22] 7/-1/-1->6->5 [23] 7/-1/-1->6->5
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO Trees [0] 6/-1/-1->5->4 [1] 6/-1/-1->5->4 [2] 6/-1/-1->5->4 [3] 6/-1/-1->5->4 [4] 6/-1/-1->5->4 [5] 6/-1/-1->5->4 [6] 6/-1/-1->5->4 [7] 6/-1/-1->5->4 [8] 6/-1/-1->5->4 [9] 6/-1/-1->5->4 [10] 6/-1/-1->5->4 [11] 6/-1/-1->5->4 [12] 6/-1/-1->5->4 [13] 6/-1/-1->5->4 [14] 6/-1/-1->5->4 [15] 6/-1/-1->5->4 [16] 6/-1/-1->5->4 [17] 6/-1/-1->5->4 [18] 6/-1/-1->5->4 [19] 6/-1/-1->5->4 [20] 6/-1/-1->5->4 [21] 6/-1/-1->5->4 [22] 6/-1/-1->5->4 [23] 6/-1/-1->5->4
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 09/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO P2P Chunksize set to 524288
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 10/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO Trees [0] 5/-1/-1->4->3 [1] 5/-1/-1->4->3 [2] 5/-1/-1->4->3 [3] 5/-1/-1->4->3 [4] 5/-1/-1->4->3 [5] 5/-1/-1->4->3 [6] 5/-1/-1->4->3 [7] 5/-1/-1->4->3 [8] 5/-1/-1->4->3 [9] 5/-1/-1->4->3 [10] 5/-1/-1->4->3 [11] 5/-1/-1->4->3 [12] 5/-1/-1->4->3 [13] 5/-1/-1->4->3 [14] 5/-1/-1->4->3 [15] 5/-1/-1->4->3 [16] 5/-1/-1->4->3 [17] 5/-1/-1->4->3 [18] 5/-1/-1->4->3 [19] 5/-1/-1->4->3 [20] 5/-1/-1->4->3 [21] 5/-1/-1->4->3 [22] 5/-1/-1->4->3 [23] 5/-1/-1->4->3
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO Trees [0] -1/-1/-1->7->6 [1] -1/-1/-1->7->6 [2] -1/-1/-1->7->6 [3] -1/-1/-1->7->6 [4] -1/-1/-1->7->6 [5] -1/-1/-1->7->6 [6] -1/-1/-1->7->6 [7] -1/-1/-1->7->6 [8] -1/-1/-1->7->6 [9] -1/-1/-1->7->6 [10] -1/-1/-1->7->6 [11] -1/-1/-1->7->6 [12] -1/-1/-1->7->6 [13] -1/-1/-1->7->6 [14] -1/-1/-1->7->6 [15] -1/-1/-1->7->6 [16] -1/-1/-1->7->6 [17] -1/-1/-1->7->6 [18] -1/-1/-1->7->6 [19] -1/-1/-1->7->6 [20] -1/-1/-1->7->6 [21] -1/-1/-1->7->6 [22] -1/-1/-1->7->6 [23] -1/-1/-1->7->6
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO P2P Chunksize set to 524288
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->0 [2] 2/-1/-1->1->0 [3] 2/-1/-1->1->0 [4] 2/-1/-1->1->0 [5] 2/-1/-1->1->0 [6] 2/-1/-1->1->0 [7] 2/-1/-1->1->0 [8] 2/-1/-1->1->0 [9] 2/-1/-1->1->0 [10] 2/-1/-1->1->0 [11] 2/-1/-1->1->0 [12] 2/-1/-1->1->0 [13] 2/-1/-1->1->0 [14] 2/-1/-1->1->0 [15] 2/-1/-1->1->0 [16] 2/-1/-1->1->0 [17] 2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 11/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO Trees [0] 4/-1/-1->3->2 [1] 4/-1/-1->3->2 [2] 4/-1/-1->3->2 [3] 4/-1/-1->3->2 [4] 4/-1/-1->3->2 [5] 4/-1/-1->3->2 [6] 4/-1/-1->3->2 [7] 4/-1/-1->3->2 [8] 4/-1/-1->3->2 [9] 4/-1/-1->3->2 [10] 4/-1/-1->3->2 [11] 4/-1/-1->3->2 [12] 4/-1/-1->3->2 [13] 4/-1/-1->3->2 [14] 4/-1/-1->3->2 [15] 4/-1/-1->3->2 [16] 4/-1/-1->3->2 [17] 4/-1/-1->3->2 [18] 4/-1/-1->3->2 [19] 4/-1/-1->3->2 [20] 4/-1/-1->3->2 [21] 4/-1/-1->3->2 [22] 4/-1/-1->3->2 [23] 4/-1/-1->3->2
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO P2P Chunksize set to 524288
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO P2P Chunksize set to 524288
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 12/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO P2P Chunksize set to 524288
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO P2P Chunksize set to 524288
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 13/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 14/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 15/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 16/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 17/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 18/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 19/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 20/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 21/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 22/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Channel 23/24 : 0 1 2 3 4 5 6 7
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Trees [0] 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] 1/-1/-1->0->-1 [11] 1/-1/-1->0->-1 [12] 1/-1/-1->0->-1 [13] 1/-1/-1->0->-1 [14] 1/-1/-1->0->-1 [15] 1/-1/-1->0->-1 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO P2P Chunksize set to 524288
+t-20260516062948-lst5d-worker-0:10243:10401 [2] NCCL INFO [Proxy Service] Device 2 CPU core 43
+t-20260516062948-lst5d-worker-0:10246:10402 [5] NCCL INFO [Proxy Service] Device 5 CPU core 100
+t-20260516062948-lst5d-worker-0:10247:10405 [6] NCCL INFO [Proxy Service] Device 6 CPU core 94
+t-20260516062948-lst5d-worker-0:10245:10406 [4] NCCL INFO [Proxy Service] Device 4 CPU core 104
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Check P2P Type intraNodeP2pSupport 1 directMode 0
+t-20260516062948-lst5d-worker-0:10248:10409 [7] NCCL INFO [Proxy Service] Device 7 CPU core 128
+t-20260516062948-lst5d-worker-0:10241:10410 [0] NCCL INFO [Proxy Service] Device 0 CPU core 74
+t-20260516062948-lst5d-worker-0:10245:10408 [4] NCCL INFO [Proxy Service UDS] Device 4 CPU core 106
+t-20260516062948-lst5d-worker-0:10247:10407 [6] NCCL INFO [Proxy Service UDS] Device 6 CPU core 96
+t-20260516062948-lst5d-worker-0:10242:10412 [1] NCCL INFO [Proxy Service] Device 1 CPU core 2
+t-20260516062948-lst5d-worker-0:10243:10403 [2] NCCL INFO [Proxy Service UDS] Device 2 CPU core 46
+t-20260516062948-lst5d-worker-0:10246:10404 [5] NCCL INFO [Proxy Service UDS] Device 5 CPU core 104
+t-20260516062948-lst5d-worker-0:10248:10411 [7] NCCL INFO [Proxy Service UDS] Device 7 CPU core 130
+t-20260516062948-lst5d-worker-0:10241:10414 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 76
+t-20260516062948-lst5d-worker-0:10242:10413 [1] NCCL INFO [Proxy Service UDS] Device 1 CPU core 5
+t-20260516062948-lst5d-worker-0:10244:10415 [3] NCCL INFO [Proxy Service] Device 3 CPU core 2
+t-20260516062948-lst5d-worker-0:10244:10416 [3] NCCL INFO [Proxy Service UDS] Device 3 CPU core 5
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO CC Off, workFifoBytes 1048576
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO ncclCommInitRankConfig comm 0x9aad980 rank 2 nranks 8 cudaDev 2 nvmlDev 2 busId 69020 commId 0x51bb37a311f021b4 - Init COMPLETE
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO ncclCommInitRankConfig comm 0x9f87a30 rank 7 nranks 8 cudaDev 7 nvmlDev 7 busId 75020 commId 0x51bb37a311f021b4 - Init COMPLETE
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO ncclCommInitRankConfig comm 0xb291ad0 rank 3 nranks 8 cudaDev 3 nvmlDev 3 busId 6b020 commId 0x51bb37a311f021b4 - Init COMPLETE
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516062948-lst5d-worker-0:10243:10323 [2] NCCL INFO Init timings - ncclCommInitRankConfig: rank 2 nranks 8 total 2.15 (kernels 0.19, alloc 0.53, bootstrap 0.50, allgathers 0.01, topo 0.53, graphs 0.01, connections 0.36, rest 0.02)
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO ncclCommInitRankConfig comm 0xb1ff580 rank 5 nranks 8 cudaDev 5 nvmlDev 5 busId 71020 commId 0x51bb37a311f021b4 - Init COMPLETE
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO ncclCommInitRankConfig comm 0xaa53d10 rank 6 nranks 8 cudaDev 6 nvmlDev 6 busId 73020 commId 0x51bb37a311f021b4 - Init COMPLETE
+t-20260516062948-lst5d-worker-0:10248:10325 [7] NCCL INFO Init timings - ncclCommInitRankConfig: rank 7 nranks 8 total 2.06 (kernels 0.50, alloc 0.62, bootstrap 0.01, allgathers 0.00, topo 0.53, graphs 0.01, connections 0.36, rest 0.02)
+t-20260516062948-lst5d-worker-0:10244:10328 [3] NCCL INFO Init timings - ncclCommInitRankConfig: rank 3 nranks 8 total 2.04 (kernels 0.49, alloc 0.61, bootstrap 0.00, allgathers 0.00, topo 0.53, graphs 0.01, connections 0.34, rest 0.04)
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO ncclCommInitRankConfig comm 0xaf65240 rank 1 nranks 8 cudaDev 1 nvmlDev 1 busId 67020 commId 0x51bb37a311f021b4 - Init COMPLETE
+t-20260516062948-lst5d-worker-0:10246:10326 [5] NCCL INFO Init timings - ncclCommInitRankConfig: rank 5 nranks 8 total 2.04 (kernels 0.48, alloc 0.61, bootstrap 0.01, allgathers 0.01, topo 0.53, graphs 0.01, connections 0.36, rest 0.02)
+t-20260516062948-lst5d-worker-0:10247:10322 [6] NCCL INFO Init timings - ncclCommInitRankConfig: rank 6 nranks 8 total 2.21 (kernels 0.20, alloc 0.50, bootstrap 0.58, allgathers 0.01, topo 0.53, graphs 0.01, connections 0.36, rest 0.03)
+t-20260516062948-lst5d-worker-0:10242:10324 [1] NCCL INFO Init timings - ncclCommInitRankConfig: rank 1 nranks 8 total 2.11 (kernels 0.30, alloc 0.61, bootstrap 0.28, allgathers 0.01, topo 0.53, graphs 0.01, connections 0.36, rest 0.02)
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO ncclCommInitRankConfig comm 0x9a6d0d0 rank 4 nranks 8 cudaDev 4 nvmlDev 4 busId 6f020 commId 0x51bb37a311f021b4 - Init COMPLETE
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO ncclCommInitRankConfig comm 0x9929b80 rank 0 nranks 8 cudaDev 0 nvmlDev 0 busId 65040 commId 0x51bb37a311f021b4 - Init COMPLETE
+t-20260516062948-lst5d-worker-0:10245:10327 [4] NCCL INFO Init timings - ncclCommInitRankConfig: rank 4 nranks 8 total 2.04 (kernels 0.48, alloc 0.62, bootstrap 0.01, allgathers 0.01, topo 0.53, graphs 0.01, connections 0.36, rest 0.03)
+t-20260516062948-lst5d-worker-0:10241:10321 [0] NCCL INFO Init timings - ncclCommInitRankConfig: rank 0 nranks 8 total 2.21 (kernels 0.20, alloc 0.48, bootstrap 0.60, allgathers 0.01, topo 0.53, graphs 0.01, connections 0.36, rest 0.03)
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 00/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 00/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 01/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 01/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 01/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 02/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 02/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 02/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 03/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 03/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 03/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 04/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 04/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 04/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 05/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 05/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 05/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 06/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 00/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 06/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 06/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 07/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 01/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 07/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 00/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 07/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 08/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 02/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 08/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 01/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 08/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 00/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 09/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 02/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 10/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 00/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 00/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 03/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 09/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 11/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 01/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 01/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 04/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 10/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 12/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 02/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 02/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 05/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 11/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 13/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 03/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 03/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 06/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 12/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 14/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 04/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 04/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 07/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 13/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 15/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 05/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 09/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 05/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 08/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 14/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 16/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 06/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 10/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 06/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 01/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 09/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 15/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 17/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 07/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 03/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 11/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 08/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 04/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 12/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 09/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 05/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 13/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 10/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 06/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 14/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 11/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 07/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 15/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 12/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 08/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 16/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 13/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 09/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 07/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 02/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 17/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 10/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 16/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 14/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 10/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 08/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 03/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 18/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 11/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 17/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 18/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 15/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 11/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 09/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 04/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 19/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 12/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 18/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 19/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 16/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 12/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 10/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 05/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 20/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 13/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 19/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 20/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 17/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 13/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 11/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 06/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 21/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 14/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 20/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 21/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 18/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 14/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 12/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 07/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 22/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 15/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 21/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 22/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 19/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 15/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 13/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 08/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Channel 23/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 16/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 22/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Channel 23/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 20/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 16/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 14/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 09/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 17/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Channel 23/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 21/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 17/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 15/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 10/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 18/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 22/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 18/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 16/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 11/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 19/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Channel 23/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 19/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 17/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 12/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 20/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 20/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 18/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 13/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 21/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 21/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 19/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 14/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 22/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 22/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 20/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 15/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Channel 23/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Channel 23/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 21/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 16/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 22/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 17/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 18/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Channel 23/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 19/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 20/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 21/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 22/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Channel 23/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516062948-lst5d-worker-0:10247:10417 [6] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516062948-lst5d-worker-0:10243:10422 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516062948-lst5d-worker-0:10242:10419 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516062948-lst5d-worker-0:10246:10424 [5] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516062948-lst5d-worker-0:10244:10420 [3] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516062948-lst5d-worker-0:10248:10423 [7] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516062948-lst5d-worker-0:10241:10418 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516062948-lst5d-worker-0:10245:10421 [4] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+{
+  "device": "cuda:0",
+  "rank": 0,
+  "world_size": 8,
+  "samples": "record_pad_truncate:pad=0:add_eos=0:add_special=1:shuffle_buffer=10000",
+  "vocab_size": 32100,
+  "tokenizer_vocab_size": 32100,
+  "save_dir": "runs/lta_owt_t5record_len1024_elfaligned_dditelf_muon_logitnormal_m1p5_s0p8_gbs512_8gpu_5epoch_20260515_223036",
+  "batch_size": 32,
+  "grad_accum": 2,
+  "effective_batch_size": 512,
+  "global_batch_size": 512,
+  "lr_schedule": "constant_warmup",
+  "optimizer": "muon",
+  "epochs": 5.0,
+  "steps_per_epoch": 15457,
+  "total_steps": 77285,
+  "warmup_steps": 7729,
+  "warmup_epochs": 0.5,
+  "min_lr": 0.0,
+  "weight_decay": 0.0,
+  "output_weight_decay": -1.0,
+  "adamw_param_groups": "nanogpt",
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "adam_eps": 1e-08,
+  "muon_impl": "optax",
+  "muon_momentum": 0.95,
+  "muon_ns_steps": 5,
+  "muon_update_scale": 1.0,
+  "muon_nesterov": true,
+  "muon_width_scale": true,
+  "muon_effective_nesterov": true,
+  "muon_effective_width_scale": true,
+  "muon_effective_weight_decay": 0.0,
+  "muon_adam_fallback_nesterov": true,
+  "muon_adam_fallback_weight_decay": 0.0,
+  "ema_decay": 0.9999,
+  "ema_start_step": 0,
+  "model_type": "ddit_elf",
+  "elf_num_time_tokens": 4,
+  "elf_num_model_mode_tokens": 0,
+  "qk_norm": true,
+  "output_bias": false,
+  "norm_type": "rmsnorm",
+  "t_sampling_mode": "logit_normal",
+  "t_sampling_power": 1.0,
+  "t_sampling_eps": 0.0001,
+  "t_sampling_logit_mean": -1.5,
+  "t_sampling_logit_std": 0.8,
+  "dual_t": true,
+  "corrupt_t_mode": "same",
+  "corrupt_min_t": 0.0,
+  "corrupt_max_t": 1.0,
+  "prefix_block_prob": 0.0,
+  "prefix_block_len": 128,
+  "mask_ratio_floor_schedule": "none",
+  "dirichlet_endpoint_mode": "categorical_dual_t",
+  "dirichlet_semantic_t_mode": "same",
+  "dirichlet_semantic_t_value": 0.0,
+  "dirichlet_semantic_t_curve": "linear",
+  "dirichlet_semantic_t_power": 1.0,
+  "endpoint_sequence_random_prob_alpha": 0.0,
+  "categorical_wrong_from_full_vocab": true,
+  "categorical_wrong_from_batch_valid_tokens": false,
+  "categorical_wrong_basin_token_ids": "",
+  "categorical_wrong_basin_prob": 0.0,
+  "categorical_wrong_unigram_prob": 0.0,
+  "categorical_wrong_uniform_prob": 0.0,
+  "categorical_wrong_corpus_unigram_path": "",
+  "categorical_wrong_corpus_unigram_alpha": 1.0,
+  "categorical_wrong_basin_shared_prob": 0.0,
+  "categorical_wrong_unigram_shared_prob": 0.0,
+  "mask_mixture_original_prob": 0.0,
+  "mask_mixture_lowk_prob": 0.0,
+  "mask_mixture_lowcorrupt_prob": 0.0,
+  "mask_mixture_block_prob": 0.0,
+  "mask_mixture_all_prob": 0.0,
+  "mask_mixture_lowk_clean_tokens": "1,2,4,8,16,32,64",
+  "mask_mixture_lowcorrupt_tokens": "1,2,4,8,16,32,64",
+  "mask_mixture_block_tokens": "64,128",
+  "simplex_bridge_sampler": "dirichlet",
+  "logistic_normal_sigma_min": 0.18,
+  "logistic_normal_sigma_max": 2.2,
+  "logistic_normal_tau_min": 0.65,
+  "logistic_normal_tau_max": 1.15,
+  "torch_compile": false,
+  "compile_mode": "max-autotune",
+  "state_format": "prob",
+  "target_loss": "hard_ce",
+  "meanflow_weight": 0.0,
+  "rollout_train_prob": 0.0,
+  "rollout_train_steps": 1,
+  "rollout_train_infer_steps": 64,
+  "rollout_train_temp": 1.45,
+  "rollout_train_max_gamma": 1.0,
+  "rollout_train_corrupt_only": true,
+  "rollout_train_samplewise": false,
+  "rollout_train_compute_always": false,
+  "bridge_noise_init": "logistic_normal",
+  "noise_sigma": -1.0,
+  "allow_tf32": true,
+  "activation_checkpointing": true,
+  "activation_checkpoint_interval": 1,
+  "activation_checkpoint_scope": "mlp",
+  "ddp_static_graph": false,
+  "ddp_gradient_as_bucket_view": true,
+  "blocking_data_transfer": false,
+  "dataloader_prefetch_factor": 4,
+  "full_train_stats": false,
+  "record_pad_truncate": true,
+  "record_add_eos": false,
+  "record_add_special_tokens": true,
+  "record_pad_token": "pad",
+  "record_shuffle_buffer": 10000,
+  "wrap": false,
+  "wrap_mode": "stream",
+  "wrap_record_buffer_size": 200,
+  "owt_cached_chunks": false,
+  "owt_chunk_cache_dir": "",
+  "owt_chunk_cache_rebuild": false,
+  "owt_chunk_cache_write_batch": 4096,
+  "owt_exact_repeat_per_chunk": 0,
+  "online_chunk_shuffle": false,
+  "online_chunk_shuffle_buffer": 10000,
+  "openwebtext_split": "train_minus_100k",
+  "detokenizer": "auto",
+  "resolved_detokenizer": null,
+  "num_workers": 8,
+  "latest_every": 1000,
+  "resume_path": ""
+}
+t-20260516062948-lst5d-worker-0:10241:10814 [0] NCCL INFO NVLS comm 0x9929b80 headRank 0 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516062948-lst5d-worker-0:10242:10870 [1] NCCL INFO NVLS comm 0xaf65240 headRank 1 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516062948-lst5d-worker-0:10244:10960 [3] NCCL INFO NVLS comm 0xb291ad0 headRank 3 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516062948-lst5d-worker-0:10243:10961 [2] NCCL INFO NVLS comm 0x9aad980 headRank 2 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516062948-lst5d-worker-0:10246:10963 [5] NCCL INFO NVLS comm 0xb1ff580 headRank 5 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516062948-lst5d-worker-0:10248:10964 [7] NCCL INFO NVLS comm 0x9f87a30 headRank 7 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516062948-lst5d-worker-0:10247:10965 [6] NCCL INFO NVLS comm 0xaa53d10 headRank 6 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516062948-lst5d-worker-0:10245:10966 [4] NCCL INFO NVLS comm 0x9a6d0d0 headRank 4 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+step=100 micro_steps=200 elapsed=115.3s lr=2.613533e-05 loss=10.3733 loss_recon=10.3733 loss_meanflow=0.0000 mean_model_t=0.2081 mean_corrupt_t=0.2081 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0454 corrupt_frac=0.5549 acc_corrupt=0.0372 loss_corrupt=10.3733 wrong_frac=0.7929 init_acc_corrupt=0.1138 acc_corrupt_t_0p0_0p2=0.0340 corrupt_frac_t_0p0_0p2=0.5679 acc_corrupt_t_0p2_0p4=0.0400 corrupt_frac_t_0p2_0p4=0.3435 out_w_norm=0.0169 out_g_norm=1.0189 acc_corrupt_t_0p4_0p6=0.0467 corrupt_frac_t_0p4_0p6=0.0860 acc_corrupt_t_0p6_0p8=0.0449 corrupt_frac_t_0p6_0p8=0.0342 acc_corrupt_t_0p8_1p0=0.0989 corrupt_frac_t_0p8_1p0=0.0095 loss_all=10.3634 init_gold_top10=0.2154 init_gold_top100=0.4564
+step=200 micro_steps=400 elapsed=86.0s lr=5.201190e-05 loss=10.3446 loss_recon=10.3446 loss_meanflow=0.0000 mean_model_t=0.2102 mean_corrupt_t=0.2102 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0514 corrupt_frac=0.5476 acc_corrupt=0.0442 loss_corrupt=10.3446 wrong_frac=0.7891 init_acc_corrupt=0.1199 acc_corrupt_t_0p0_0p2=0.0422 corrupt_frac_t_0p0_0p2=0.5602 acc_corrupt_t_0p2_0p4=0.0456 corrupt_frac_t_0p2_0p4=0.3460 acc_corrupt_t_0p4_0p6=0.0498 corrupt_frac_t_0p4_0p6=0.0887 out_w_norm=0.0926 out_g_norm=1.6250 acc_corrupt_t_0p6_0p8=0.0561 corrupt_frac_t_0p6_0p8=0.0361 loss_all=10.3182 init_gold_top10=0.1479 init_gold_top100=0.4573
+step=300 micro_steps=600 elapsed=86.1s lr=7.788847e-05 loss=10.2814 loss_recon=10.2814 loss_meanflow=0.0000 mean_model_t=0.2087 mean_corrupt_t=0.2087 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0550 corrupt_frac=0.5486 acc_corrupt=0.0462 loss_corrupt=10.2814 wrong_frac=0.7920 init_acc_corrupt=0.1149 acc_corrupt_t_0p0_0p2=0.0443 corrupt_frac_t_0p0_0p2=0.5626 acc_corrupt_t_0p2_0p4=0.0474 corrupt_frac_t_0p2_0p4=0.3554 acc_corrupt_t_0p4_0p6=0.0526 corrupt_frac_t_0p4_0p6=0.0764 out_w_norm=0.2117 out_g_norm=1.8348 acc_corrupt_t_0p6_0p8=0.0655 corrupt_frac_t_0p6_0p8=0.0338 acc_corrupt_t_0p8_1p0=0.0849 corrupt_frac_t_0p8_1p0=0.0410 loss_all=10.2283 init_gold_top10=0.1761 init_gold_top100=0.4584
+step=400 micro_steps=800 elapsed=86.0s lr=1.037650e-04 loss=10.1819 loss_recon=10.1819 loss_meanflow=0.0000 mean_model_t=0.2101 mean_corrupt_t=0.2101 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0622 corrupt_frac=0.5497 acc_corrupt=0.0480 loss_corrupt=10.1819 wrong_frac=0.7920 init_acc_corrupt=0.1141 acc_corrupt_t_0p0_0p2=0.0451 corrupt_frac_t_0p0_0p2=0.5615 acc_corrupt_t_0p2_0p4=0.0497 corrupt_frac_t_0p2_0p4=0.3516 acc_corrupt_t_0p4_0p6=0.0597 corrupt_frac_t_0p4_0p6=0.0843 out_w_norm=0.3818 out_g_norm=1.8762 acc_corrupt_t_0p6_0p8=0.0649 corrupt_frac_t_0p6_0p8=0.0315 loss_all=10.0834 init_gold_top10=0.2435 init_gold_top100=0.4554
+step=500 micro_steps=1000 elapsed=86.1s lr=1.296416e-04 loss=10.0376 loss_recon=10.0376 loss_meanflow=0.0000 mean_model_t=0.2087 mean_corrupt_t=0.2087 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0844 corrupt_frac=0.5571 acc_corrupt=0.0529 loss_corrupt=10.0376 wrong_frac=0.7934 init_acc_corrupt=0.1133 acc_corrupt_t_0p0_0p2=0.0460 corrupt_frac_t_0p0_0p2=0.5587 acc_corrupt_t_0p2_0p4=0.0575 corrupt_frac_t_0p2_0p4=0.3629 acc_corrupt_t_0p4_0p6=0.0797 corrupt_frac_t_0p4_0p6=0.0744 out_w_norm=0.6182 out_g_norm=1.8807 acc_corrupt_t_0p6_0p8=0.0944 corrupt_frac_t_0p6_0p8=0.0318 acc_corrupt_t_0p8_1p0=0.1441 corrupt_frac_t_0p8_1p0=0.0262 loss_all=9.8675 init_gold_top10=0.1701 init_gold_top100=0.4777
+step=600 micro_steps=1200 elapsed=86.1s lr=1.555182e-04 loss=9.8017 loss_recon=9.8017 loss_meanflow=0.0000 mean_model_t=0.2097 mean_corrupt_t=0.2097 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.1423 corrupt_frac=0.5498 acc_corrupt=0.0682 loss_corrupt=9.8017 wrong_frac=0.7894 init_acc_corrupt=0.1176 acc_corrupt_t_0p0_0p2=0.0488 corrupt_frac_t_0p0_0p2=0.5433 acc_corrupt_t_0p2_0p4=0.0820 corrupt_frac_t_0p2_0p4=0.3728 acc_corrupt_t_0p4_0p6=0.1288 corrupt_frac_t_0p4_0p6=0.0809 out_w_norm=0.9391 out_g_norm=1.8284 acc_corrupt_t_0p6_0p8=0.1608 corrupt_frac_t_0p6_0p8=0.0337 loss_all=9.1658 init_gold_top10=0.1670 init_gold_top100=0.4715
+step=700 micro_steps=1400 elapsed=86.1s lr=1.813947e-04 loss=9.4591 loss_recon=9.4591 loss_meanflow=0.0000 mean_model_t=0.2099 mean_corrupt_t=0.2099 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.1748 corrupt_frac=0.5479 acc_corrupt=0.0775 loss_corrupt=9.4591 wrong_frac=0.7903 init_acc_corrupt=0.1163 acc_corrupt_t_0p0_0p2=0.0524 corrupt_frac_t_0p0_0p2=0.5572 acc_corrupt_t_0p2_0p4=0.0958 corrupt_frac_t_0p2_0p4=0.3513 acc_corrupt_t_0p4_0p6=0.1564 corrupt_frac_t_0p4_0p6=0.0882 out_w_norm=1.3861 out_g_norm=1.7546 acc_corrupt_t_0p6_0p8=0.1990 corrupt_frac_t_0p6_0p8=0.0406 loss_all=8.3915 init_gold_top10=0.2100 init_gold_top100=0.4558
+step=800 micro_steps=1600 elapsed=86.2s lr=2.072713e-04 loss=9.0817 loss_recon=9.0817 loss_meanflow=0.0000 mean_model_t=0.2112 mean_corrupt_t=0.2112 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.1954 corrupt_frac=0.5501 acc_corrupt=0.0859 loss_corrupt=9.0817 wrong_frac=0.7880 init_acc_corrupt=0.1186 acc_corrupt_t_0p0_0p2=0.0551 corrupt_frac_t_0p0_0p2=0.5469 acc_corrupt_t_0p2_0p4=0.1087 corrupt_frac_t_0p2_0p4=0.3664 acc_corrupt_t_0p4_0p6=0.1770 corrupt_frac_t_0p4_0p6=0.0846 acc_corrupt_t_0p6_0p8=0.2349 corrupt_frac_t_0p6_0p8=0.0401 out_w_norm=1.9573 out_g_norm=1.6633 acc_corrupt_t_0p8_1p0=0.3462 corrupt_frac_t_0p8_1p0=0.0142 loss_all=7.7385 init_gold_top10=0.1907 init_gold_top100=0.4702
+step=900 micro_steps=1800 elapsed=86.1s lr=2.331479e-04 loss=8.7010 loss_recon=8.7010 loss_meanflow=0.0000 mean_model_t=0.2089 mean_corrupt_t=0.2089 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.2202 corrupt_frac=0.5522 acc_corrupt=0.1017 loss_corrupt=8.7010 wrong_frac=0.7910 init_acc_corrupt=0.1161 acc_corrupt_t_0p0_0p2=0.0671 corrupt_frac_t_0p0_0p2=0.5570 acc_corrupt_t_0p2_0p4=0.1276 corrupt_frac_t_0p2_0p4=0.3541 acc_corrupt_t_0p4_0p6=0.2083 corrupt_frac_t_0p4_0p6=0.0843 out_w_norm=2.6437 out_g_norm=1.5171 acc_corrupt_t_0p6_0p8=0.2671 corrupt_frac_t_0p6_0p8=0.0450 acc_corrupt_t_0p8_1p0=0.3024 corrupt_frac_t_0p8_1p0=0.0342 loss_all=7.6141 init_gold_top10=0.2222 init_gold_top100=0.4160
+step=1000 micro_steps=2000 elapsed=86.2s lr=2.590245e-04 loss=8.3033 loss_recon=8.3033 loss_meanflow=0.0000 mean_model_t=0.2076 mean_corrupt_t=0.2076 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.2471 corrupt_frac=0.5472 acc_corrupt=0.1205 loss_corrupt=8.3033 wrong_frac=0.7925 init_acc_corrupt=0.1137 acc_corrupt_t_0p0_0p2=0.0828 corrupt_frac_t_0p0_0p2=0.5614 acc_corrupt_t_0p2_0p4=0.1514 corrupt_frac_t_0p2_0p4=0.3565 acc_corrupt_t_0p4_0p6=0.2364 corrupt_frac_t_0p4_0p6=0.0787 acc_corrupt_t_0p6_0p8=0.3057 corrupt_frac_t_0p6_0p8=0.0401 out_w_norm=3.4166 out_g_norm=1.2460 loss_all=7.1860 init_gold_top10=0.2101 init_gold_top100=0.4660
+step=1100 micro_steps=2200 elapsed=87.7s lr=2.849010e-04 loss=7.9125 loss_recon=7.9125 loss_meanflow=0.0000 mean_model_t=0.2096 mean_corrupt_t=0.2096 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.2694 corrupt_frac=0.5487 acc_corrupt=0.1387 loss_corrupt=7.9125 wrong_frac=0.7917 init_acc_corrupt=0.1154 acc_corrupt_t_0p0_0p2=0.0965 corrupt_frac_t_0p0_0p2=0.5529 acc_corrupt_t_0p2_0p4=0.1725 corrupt_frac_t_0p2_0p4=0.3606 acc_corrupt_t_0p4_0p6=0.2619 corrupt_frac_t_0p4_0p6=0.0866 out_w_norm=4.2495 out_g_norm=0.9327 acc_corrupt_t_0p6_0p8=0.3224 corrupt_frac_t_0p6_0p8=0.0308 loss_all=6.9095 init_gold_top10=0.1427 init_gold_top100=0.4703
+step=1200 micro_steps=2400 elapsed=86.4s lr=3.107776e-04 loss=7.5735 loss_recon=7.5735 loss_meanflow=0.0000 mean_model_t=0.2071 mean_corrupt_t=0.2071 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.2900 corrupt_frac=0.5535 acc_corrupt=0.1525 loss_corrupt=7.5735 wrong_frac=0.7916 init_acc_corrupt=0.1151 acc_corrupt_t_0p0_0p2=0.1077 corrupt_frac_t_0p0_0p2=0.5606 acc_corrupt_t_0p2_0p4=0.1906 corrupt_frac_t_0p2_0p4=0.3546 acc_corrupt_t_0p4_0p6=0.2821 corrupt_frac_t_0p4_0p6=0.0834 out_w_norm=5.1992 out_g_norm=0.6850 acc_corrupt_t_0p6_0p8=0.3532 corrupt_frac_t_0p6_0p8=0.0354 acc_corrupt_t_0p8_1p0=0.4220 corrupt_frac_t_0p8_1p0=0.0291 loss_all=6.4896 init_gold_top10=0.1752 init_gold_top100=0.4720

LTA_openwebtext_dualt/logs/elfaligned_t5record_8gpu/lta_owt_t5record_len1024_elfaligned_dditelf_muon_logitnormal_m1p5_s0p8_none_floor0p0_gbs512_8gpu_5epoch_20260516_000824.log

@@ -0,0 +1,636 @@
+t-20260516080640-ks2k6-worker-0:10250:10250 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10250:10250 [0] NCCL INFO Bootstrap: Using eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10250:10250 [0] NCCL INFO cudaDriverVersion 12080
+t-20260516080640-ks2k6-worker-0:10250:10250 [0] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516080640-ks2k6-worker-0:10250:10250 [0] NCCL INFO Comm config Blocking set to 1
+t-20260516080640-ks2k6-worker-0:10256:10256 [6] NCCL INFO cudaDriverVersion 12080
+t-20260516080640-ks2k6-worker-0:10256:10256 [6] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10251:10251 [1] NCCL INFO cudaDriverVersion 12080
+t-20260516080640-ks2k6-worker-0:10251:10251 [1] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10254:10254 [4] NCCL INFO cudaDriverVersion 12080
+t-20260516080640-ks2k6-worker-0:10254:10254 [4] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10256:10256 [6] NCCL INFO Bootstrap: Using eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10256:10256 [6] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516080640-ks2k6-worker-0:10251:10251 [1] NCCL INFO Bootstrap: Using eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10254:10254 [4] NCCL INFO Bootstrap: Using eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10251:10251 [1] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516080640-ks2k6-worker-0:10254:10254 [4] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516080640-ks2k6-worker-0:10256:10256 [6] NCCL INFO Comm config Blocking set to 1
+t-20260516080640-ks2k6-worker-0:10254:10254 [4] NCCL INFO Comm config Blocking set to 1
+t-20260516080640-ks2k6-worker-0:10251:10251 [1] NCCL INFO Comm config Blocking set to 1
+t-20260516080640-ks2k6-worker-0:10252:10252 [2] NCCL INFO cudaDriverVersion 12080
+t-20260516080640-ks2k6-worker-0:10252:10252 [2] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10252:10252 [2] NCCL INFO Bootstrap: Using eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10252:10252 [2] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516080640-ks2k6-worker-0:10252:10252 [2] NCCL INFO Comm config Blocking set to 1
+t-20260516080640-ks2k6-worker-0:10253:10253 [3] NCCL INFO cudaDriverVersion 12080
+t-20260516080640-ks2k6-worker-0:10253:10253 [3] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10253:10253 [3] NCCL INFO Bootstrap: Using eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10253:10253 [3] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516080640-ks2k6-worker-0:10253:10253 [3] NCCL INFO Comm config Blocking set to 1
+t-20260516080640-ks2k6-worker-0:10255:10255 [5] NCCL INFO cudaDriverVersion 12080
+t-20260516080640-ks2k6-worker-0:10255:10255 [5] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10255:10255 [5] NCCL INFO Bootstrap: Using eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10255:10255 [5] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516080640-ks2k6-worker-0:10255:10255 [5] NCCL INFO Comm config Blocking set to 1
+t-20260516080640-ks2k6-worker-0:10257:10257 [7] NCCL INFO cudaDriverVersion 12080
+t-20260516080640-ks2k6-worker-0:10257:10257 [7] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10257:10257 [7] NCCL INFO Bootstrap: Using eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10257:10257 [7] NCCL INFO NCCL version 2.25.1+cuda12.8
+t-20260516080640-ks2k6-worker-0:10257:10257 [7] NCCL INFO Comm config Blocking set to 1
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Using network IBext_v9
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO Using network IBext_v9
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO Using network IBext_v9
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO Using network IBext_v9
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO Using network IBext_v9
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO Using network IBext_v9
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO Using network IBext_v9
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO NET/Plugin: Loaded net plugin NCCL RDMA Plugin v9 (v9)
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO NET/Plugin: Loaded collnet plugin SHARP (v9)
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO Plugin Path : /opt/hpcx/nccl_rdma_sharp_plugin/lib/libnccl-net.so
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO P2P plugin v9 IBext_v9
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth1
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO ncclCommInitRankConfig comm 0x9f91be0 rank 0 nranks 8 cudaDev 0 nvmlDev 0 busId 65040 commId 0x888384a9cc51f1d4 - Init START
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO NCCL_IB_PCI_RELAXED_ORDERING set by environment to 1.
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [1]mlx5_4:1/RoCE [2]mlx5_5:1/RoCE [3]mlx5_6:1/RoCE [4]mlx5_7:1/RoCE [5]mlx5_8:1/RoCE [6]mlx5_9:1/RoCE [7]mlx5_10:1/RoCE [RO]; OOB eth1:10.82.32.15<0>
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so.
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO Using network IBext_v9
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO ncclCommInitRankConfig comm 0xb506970 rank 4 nranks 8 cudaDev 4 nvmlDev 4 busId 6f020 commId 0x888384a9cc51f1d4 - Init START
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO ncclCommInitRankConfig comm 0x9a45db0 rank 1 nranks 8 cudaDev 1 nvmlDev 1 busId 67020 commId 0x888384a9cc51f1d4 - Init START
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO ncclCommInitRankConfig comm 0xb812680 rank 6 nranks 8 cudaDev 6 nvmlDev 6 busId 73020 commId 0x888384a9cc51f1d4 - Init START
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO ncclCommInitRankConfig comm 0xb2c0700 rank 2 nranks 8 cudaDev 2 nvmlDev 2 busId 69020 commId 0x888384a9cc51f1d4 - Init START
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO ncclCommInitRankConfig comm 0xa06a9c0 rank 3 nranks 8 cudaDev 3 nvmlDev 3 busId 6b020 commId 0x888384a9cc51f1d4 - Init START
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO ncclCommInitRankConfig comm 0xa632cd0 rank 5 nranks 8 cudaDev 5 nvmlDev 5 busId 71020 commId 0x888384a9cc51f1d4 - Init START
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO ncclCommInitRankConfig comm 0xa6619b0 rank 7 nranks 8 cudaDev 7 nvmlDev 7 busId 75020 commId 0x888384a9cc51f1d4 - Init START
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO RAS client listening socket at ::1<28028>
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO Bootstrap timings total 0.035510 (create 0.000022, send 0.000069, recv 0.000083, ring 0.035078, delay 0.000001)
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Bootstrap timings total 0.429759 (create 0.000025, send 0.000071, recv 0.033914, ring 0.000166, delay 0.000001)
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO Bootstrap timings total 0.042301 (create 0.000021, send 0.000068, recv 0.006856, ring 0.035065, delay 0.000001)
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO Bootstrap timings total 0.000623 (create 0.000021, send 0.000070, recv 0.000073, ring 0.000163, delay 0.000001)
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO Bootstrap timings total 0.024796 (create 0.000020, send 0.000065, recv 0.000077, ring 0.024312, delay 0.000001)
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO Bootstrap timings total 0.387672 (create 0.000020, send 0.000068, recv 0.387128, ring 0.000126, delay 0.000001)
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO Bootstrap timings total 0.398382 (create 0.000022, send 0.000066, recv 0.373675, ring 0.024179, delay 0.000001)
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO Bootstrap timings total 0.395926 (create 0.000021, send 0.000082, recv 0.353710, ring 0.041723, delay 0.000001)
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO MNNVL busId 0x69020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO MNNVL busId 0x71020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO MNNVL busId 0x65040 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO MNNVL busId 0x6b020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO MNNVL busId 0x75020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO MNNVL busId 0x73020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO MNNVL busId 0x67020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO MNNVL busId 0x6f020 fabric UUID 0.0 cliqueId 0x0 state 3 healthMask 0x0
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO NCCL_TOPO_FILE set by environment to /var/run/nvidia-topologyd/virtualTopology.xml
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO Setting affinity for GPU 3 to 03ffffff,ffffffff,ffffffff
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO NVLS multicast support is available on dev 3
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO Setting affinity for GPU 1 to 03ffffff,ffffffff,ffffffff
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO Setting affinity for GPU 6 to 0fffff,ffffffff,ffffffff,fc000000,00000000,00000000
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO NVLS multicast support is available on dev 6
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO NVLS multicast support is available on dev 1
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Setting affinity for GPU 0 to 03ffffff,ffffffff,ffffffff
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO Setting affinity for GPU 7 to 0fffff,ffffffff,ffffffff,fc000000,00000000,00000000
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO NVLS multicast support is available on dev 7
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO Setting affinity for GPU 2 to 03ffffff,ffffffff,ffffffff
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO Setting affinity for GPU 4 to 0fffff,ffffffff,ffffffff,fc000000,00000000,00000000
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO Setting affinity for GPU 5 to 0fffff,ffffffff,ffffffff,fc000000,00000000,00000000
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO NVLS multicast support is available on dev 5
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO NVLS multicast support is available on dev 4
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO NVLS multicast support is available on dev 0
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO NVLS multicast support is available on dev 2
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO comm 0xb2c0700 rank 2 nRanks 8 nNodes 1 localRanks 8 localRank 2 MNNVL 0
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO comm 0x9a45db0 rank 1 nRanks 8 nNodes 1 localRanks 8 localRank 1 MNNVL 0
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO comm 0xb812680 rank 6 nRanks 8 nNodes 1 localRanks 8 localRank 6 MNNVL 0
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO comm 0x9f91be0 rank 0 nRanks 8 nNodes 1 localRanks 8 localRank 0 MNNVL 0
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO comm 0xa6619b0 rank 7 nRanks 8 nNodes 1 localRanks 8 localRank 7 MNNVL 0
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO comm 0xa632cd0 rank 5 nRanks 8 nNodes 1 localRanks 8 localRank 5 MNNVL 0
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO comm 0xb506970 rank 4 nRanks 8 nNodes 1 localRanks 8 localRank 4 MNNVL 0
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO comm 0xa06a9c0 rank 3 nRanks 8 nNodes 1 localRanks 8 localRank 3 MNNVL 0
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 00/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 01/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 02/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 03/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 04/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 05/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO Trees [0] 7/-1/-1->6->5 [1] 7/-1/-1->6->5 [2] 7/-1/-1->6->5 [3] 7/-1/-1->6->5 [4] 7/-1/-1->6->5 [5] 7/-1/-1->6->5 [6] 7/-1/-1->6->5 [7] 7/-1/-1->6->5 [8] 7/-1/-1->6->5 [9] 7/-1/-1->6->5 [10] 7/-1/-1->6->5 [11] 7/-1/-1->6->5 [12] 7/-1/-1->6->5 [13] 7/-1/-1->6->5 [14] 7/-1/-1->6->5 [15] 7/-1/-1->6->5 [16] 7/-1/-1->6->5 [17] 7/-1/-1->6->5 [18] 7/-1/-1->6->5 [19] 7/-1/-1->6->5 [20] 7/-1/-1->6->5 [21] 7/-1/-1->6->5 [22] 7/-1/-1->6->5 [23] 7/-1/-1->6->5
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO Trees [0] -1/-1/-1->7->6 [1] -1/-1/-1->7->6 [2] -1/-1/-1->7->6 [3] -1/-1/-1->7->6 [4] -1/-1/-1->7->6 [5] -1/-1/-1->7->6 [6] -1/-1/-1->7->6 [7] -1/-1/-1->7->6 [8] -1/-1/-1->7->6 [9] -1/-1/-1->7->6 [10] -1/-1/-1->7->6 [11] -1/-1/-1->7->6 [12] -1/-1/-1->7->6 [13] -1/-1/-1->7->6 [14] -1/-1/-1->7->6 [15] -1/-1/-1->7->6 [16] -1/-1/-1->7->6 [17] -1/-1/-1->7->6 [18] -1/-1/-1->7->6 [19] -1/-1/-1->7->6 [20] -1/-1/-1->7->6 [21] -1/-1/-1->7->6 [22] -1/-1/-1->7->6 [23] -1/-1/-1->7->6
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO Trees [0] 5/-1/-1->4->3 [1] 5/-1/-1->4->3 [2] 5/-1/-1->4->3 [3] 5/-1/-1->4->3 [4] 5/-1/-1->4->3 [5] 5/-1/-1->4->3 [6] 5/-1/-1->4->3 [7] 5/-1/-1->4->3 [8] 5/-1/-1->4->3 [9] 5/-1/-1->4->3 [10] 5/-1/-1->4->3 [11] 5/-1/-1->4->3 [12] 5/-1/-1->4->3 [13] 5/-1/-1->4->3 [14] 5/-1/-1->4->3 [15] 5/-1/-1->4->3 [16] 5/-1/-1->4->3 [17] 5/-1/-1->4->3 [18] 5/-1/-1->4->3 [19] 5/-1/-1->4->3 [20] 5/-1/-1->4->3 [21] 5/-1/-1->4->3 [22] 5/-1/-1->4->3 [23] 5/-1/-1->4->3
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 06/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO P2P Chunksize set to 524288
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO P2P Chunksize set to 524288
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->0 [2] 2/-1/-1->1->0 [3] 2/-1/-1->1->0 [4] 2/-1/-1->1->0 [5] 2/-1/-1->1->0 [6] 2/-1/-1->1->0 [7] 2/-1/-1->1->0 [8] 2/-1/-1->1->0 [9] 2/-1/-1->1->0 [10] 2/-1/-1->1->0 [11] 2/-1/-1->1->0 [12] 2/-1/-1->1->0 [13] 2/-1/-1->1->0 [14] 2/-1/-1->1->0 [15] 2/-1/-1->1->0 [16] 2/-1/-1->1->0 [17] 2/-1/-1->1->0 [18] 2/-1/-1->1->0 [19] 2/-1/-1->1->0 [20] 2/-1/-1->1->0 [21] 2/-1/-1->1->0 [22] 2/-1/-1->1->0 [23] 2/-1/-1->1->0
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO Trees [0] 6/-1/-1->5->4 [1] 6/-1/-1->5->4 [2] 6/-1/-1->5->4 [3] 6/-1/-1->5->4 [4] 6/-1/-1->5->4 [5] 6/-1/-1->5->4 [6] 6/-1/-1->5->4 [7] 6/-1/-1->5->4 [8] 6/-1/-1->5->4 [9] 6/-1/-1->5->4 [10] 6/-1/-1->5->4 [11] 6/-1/-1->5->4 [12] 6/-1/-1->5->4 [13] 6/-1/-1->5->4 [14] 6/-1/-1->5->4 [15] 6/-1/-1->5->4 [16] 6/-1/-1->5->4 [17] 6/-1/-1->5->4 [18] 6/-1/-1->5->4 [19] 6/-1/-1->5->4 [20] 6/-1/-1->5->4 [21] 6/-1/-1->5->4 [22] 6/-1/-1->5->4 [23] 6/-1/-1->5->4
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO P2P Chunksize set to 524288
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 07/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO P2P Chunksize set to 524288
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 3/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 3/-1/-1->2->1 [4] 3/-1/-1->2->1 [5] 3/-1/-1->2->1 [6] 3/-1/-1->2->1 [7] 3/-1/-1->2->1 [8] 3/-1/-1->2->1 [9] 3/-1/-1->2->1 [10] 3/-1/-1->2->1 [11] 3/-1/-1->2->1 [12] 3/-1/-1->2->1 [13] 3/-1/-1->2->1 [14] 3/-1/-1->2->1 [15] 3/-1/-1->2->1 [16] 3/-1/-1->2->1 [17] 3/-1/-1->2->1 [18] 3/-1/-1->2->1 [19] 3/-1/-1->2->1 [20] 3/-1/-1->2->1 [21] 3/-1/-1->2->1 [22] 3/-1/-1->2->1 [23] 3/-1/-1->2->1
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO P2P Chunksize set to 524288
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 08/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO Trees [0] 4/-1/-1->3->2 [1] 4/-1/-1->3->2 [2] 4/-1/-1->3->2 [3] 4/-1/-1->3->2 [4] 4/-1/-1->3->2 [5] 4/-1/-1->3->2 [6] 4/-1/-1->3->2 [7] 4/-1/-1->3->2 [8] 4/-1/-1->3->2 [9] 4/-1/-1->3->2 [10] 4/-1/-1->3->2 [11] 4/-1/-1->3->2 [12] 4/-1/-1->3->2 [13] 4/-1/-1->3->2 [14] 4/-1/-1->3->2 [15] 4/-1/-1->3->2 [16] 4/-1/-1->3->2 [17] 4/-1/-1->3->2 [18] 4/-1/-1->3->2 [19] 4/-1/-1->3->2 [20] 4/-1/-1->3->2 [21] 4/-1/-1->3->2 [22] 4/-1/-1->3->2 [23] 4/-1/-1->3->2
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 09/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO P2P Chunksize set to 524288
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 10/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO P2P Chunksize set to 524288
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 11/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 12/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 13/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 14/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 15/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 16/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 17/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 18/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 19/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 20/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 21/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 22/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Channel 23/24 : 0 1 2 3 4 5 6 7
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Trees [0] 1/-1/-1->0->-1 [1] 1/-1/-1->0->-1 [2] 1/-1/-1->0->-1 [3] 1/-1/-1->0->-1 [4] 1/-1/-1->0->-1 [5] 1/-1/-1->0->-1 [6] 1/-1/-1->0->-1 [7] 1/-1/-1->0->-1 [8] 1/-1/-1->0->-1 [9] 1/-1/-1->0->-1 [10] 1/-1/-1->0->-1 [11] 1/-1/-1->0->-1 [12] 1/-1/-1->0->-1 [13] 1/-1/-1->0->-1 [14] 1/-1/-1->0->-1 [15] 1/-1/-1->0->-1 [16] 1/-1/-1->0->-1 [17] 1/-1/-1->0->-1 [18] 1/-1/-1->0->-1 [19] 1/-1/-1->0->-1 [20] 1/-1/-1->0->-1 [21] 1/-1/-1->0->-1 [22] 1/-1/-1->0->-1 [23] 1/-1/-1->0->-1
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO P2P Chunksize set to 524288
+t-20260516080640-ks2k6-worker-0:10257:10411 [7] NCCL INFO [Proxy Service] Device 7 CPU core 123
+t-20260516080640-ks2k6-worker-0:10254:10410 [4] NCCL INFO [Proxy Service] Device 4 CPU core 160
+t-20260516080640-ks2k6-worker-0:10256:10414 [6] NCCL INFO [Proxy Service] Device 6 CPU core 92
+t-20260516080640-ks2k6-worker-0:10255:10415 [5] NCCL INFO [Proxy Service] Device 5 CPU core 114
+t-20260516080640-ks2k6-worker-0:10251:10418 [1] NCCL INFO [Proxy Service] Device 1 CPU core 68
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Check P2P Type intraNodeP2pSupport 1 directMode 0
+t-20260516080640-ks2k6-worker-0:10253:10421 [3] NCCL INFO [Proxy Service] Device 3 CPU core 2
+t-20260516080640-ks2k6-worker-0:10252:10422 [2] NCCL INFO [Proxy Service UDS] Device 2 CPU core 8
+t-20260516080640-ks2k6-worker-0:10250:10423 [0] NCCL INFO [Proxy Service] Device 0 CPU core 84
+t-20260516080640-ks2k6-worker-0:10251:10419 [1] NCCL INFO [Proxy Service UDS] Device 1 CPU core 72
+t-20260516080640-ks2k6-worker-0:10255:10417 [5] NCCL INFO [Proxy Service UDS] Device 5 CPU core 116
+t-20260516080640-ks2k6-worker-0:10257:10413 [7] NCCL INFO [Proxy Service UDS] Device 7 CPU core 124
+t-20260516080640-ks2k6-worker-0:10253:10424 [3] NCCL INFO [Proxy Service UDS] Device 3 CPU core 19
+t-20260516080640-ks2k6-worker-0:10252:10420 [2] NCCL INFO [Proxy Service] Device 2 CPU core 2
+t-20260516080640-ks2k6-worker-0:10254:10412 [4] NCCL INFO [Proxy Service UDS] Device 4 CPU core 164
+t-20260516080640-ks2k6-worker-0:10256:10416 [6] NCCL INFO [Proxy Service UDS] Device 6 CPU core 94
+t-20260516080640-ks2k6-worker-0:10250:10425 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 87
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO threadThresholds 8/8/64 | 64/8/64 | 512 | 512
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO 24 coll channels, 24 collnet channels, 16 nvls channels, 32 p2p channels, 32 p2p channels per peer
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO CC Off, workFifoBytes 1048576
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO ncclCommInitRankConfig comm 0xb2c0700 rank 2 nranks 8 cudaDev 2 nvmlDev 2 busId 69020 commId 0x888384a9cc51f1d4 - Init COMPLETE
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO ncclCommInitRankConfig comm 0xa06a9c0 rank 3 nranks 8 cudaDev 3 nvmlDev 3 busId 6b020 commId 0x888384a9cc51f1d4 - Init COMPLETE
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO ncclCommInitRankConfig comm 0xa632cd0 rank 5 nranks 8 cudaDev 5 nvmlDev 5 busId 71020 commId 0x888384a9cc51f1d4 - Init COMPLETE
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v4 symbol.
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO ncclCommInitRankConfig comm 0xb506970 rank 4 nranks 8 cudaDev 4 nvmlDev 4 busId 6f020 commId 0x888384a9cc51f1d4 - Init COMPLETE
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO ncclCommInitRankConfig comm 0x9a45db0 rank 1 nranks 8 cudaDev 1 nvmlDev 1 busId 67020 commId 0x888384a9cc51f1d4 - Init COMPLETE
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516080640-ks2k6-worker-0:10253:10335 [3] NCCL INFO Init timings - ncclCommInitRankConfig: rank 3 nranks 8 total 2.14 (kernels 0.46, alloc 0.65, bootstrap 0.04, allgathers 0.01, topo 0.55, graphs 0.01, connections 0.40, rest 0.02)
+t-20260516080640-ks2k6-worker-0:10252:10334 [2] NCCL INFO Init timings - ncclCommInitRankConfig: rank 2 nranks 8 total 2.15 (kernels 0.45, alloc 0.66, bootstrap 0.04, allgathers 0.00, topo 0.55, graphs 0.01, connections 0.40, rest 0.03)
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO ncclCommInitRankConfig comm 0xa6619b0 rank 7 nranks 8 cudaDev 7 nvmlDev 7 busId 75020 commId 0x888384a9cc51f1d4 - Init COMPLETE
+t-20260516080640-ks2k6-worker-0:10255:10336 [5] NCCL INFO Init timings - ncclCommInitRankConfig: rank 5 nranks 8 total 2.14 (kernels 0.50, alloc 0.63, bootstrap 0.02, allgathers 0.00, topo 0.55, graphs 0.01, connections 0.40, rest 0.02)
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v3 symbol.
+t-20260516080640-ks2k6-worker-0:10251:10333 [1] NCCL INFO Init timings - ncclCommInitRankConfig: rank 1 nranks 8 total 2.26 (kernels 0.19, alloc 0.68, bootstrap 0.40, allgathers 0.01, topo 0.55, graphs 0.01, connections 0.40, rest 0.02)
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516080640-ks2k6-worker-0:10254:10332 [4] NCCL INFO Init timings - ncclCommInitRankConfig: rank 4 nranks 8 total 2.26 (kernels 0.19, alloc 0.68, bootstrap 0.40, allgathers 0.01, topo 0.55, graphs 0.01, connections 0.40, rest 0.02)
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO TUNER/Plugin: Failed to find ncclTunerPlugin_v2 symbol, using internal tuner instead.
+t-20260516080640-ks2k6-worker-0:10257:10337 [7] NCCL INFO Init timings - ncclCommInitRankConfig: rank 7 nranks 8 total 2.13 (kernels 0.66, alloc 0.48, bootstrap 0.00, allgathers 0.01, topo 0.55, graphs 0.01, connections 0.40, rest 0.02)
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO ncclCommInitRankConfig comm 0xb812680 rank 6 nranks 8 cudaDev 6 nvmlDev 6 busId 73020 commId 0x888384a9cc51f1d4 - Init COMPLETE
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO ncclCommInitRankConfig comm 0x9f91be0 rank 0 nranks 8 cudaDev 0 nvmlDev 0 busId 65040 commId 0x888384a9cc51f1d4 - Init COMPLETE
+t-20260516080640-ks2k6-worker-0:10256:10331 [6] NCCL INFO Init timings - ncclCommInitRankConfig: rank 6 nranks 8 total 2.26 (kernels 0.20, alloc 0.69, bootstrap 0.39, allgathers 0.01, topo 0.55, graphs 0.01, connections 0.40, rest 0.02)
+t-20260516080640-ks2k6-worker-0:10250:10330 [0] NCCL INFO Init timings - ncclCommInitRankConfig: rank 0 nranks 8 total 2.27 (kernels 0.19, alloc 0.65, bootstrap 0.43, allgathers 0.01, topo 0.55, graphs 0.01, connections 0.40, rest 0.02)
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 00/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 01/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 02/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 03/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 04/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 05/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 06/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 07/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 08/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 09/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 10/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 11/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 12/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 13/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 14/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 15/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 00/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 16/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 01/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 17/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 00/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 00/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 18/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 01/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 01/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 19/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 02/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 02/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 20/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 02/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 21/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 00/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 03/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 03/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 03/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 04/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 04/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 04/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 05/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 05/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 06/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 05/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 06/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 07/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 06/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 07/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 08/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 07/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 08/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 09/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 08/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 09/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 00/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 10/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 09/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 10/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 01/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 11/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 00/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 10/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 11/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 02/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 12/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 22/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 01/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 11/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 12/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 01/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 03/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 13/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Channel 23/0 : 2[2] -> 3[3] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 02/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 12/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 01/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 13/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 02/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 04/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 14/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 03/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 13/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 05/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 04/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 14/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 06/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 05/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 15/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 07/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 06/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 16/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 08/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 07/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 17/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 09/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 08/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 18/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 10/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 09/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 19/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 11/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 10/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 20/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 12/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 11/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 21/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 13/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 12/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 22/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 14/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 13/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Channel 23/0 : 4[4] -> 5[5] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 15/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 03/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 02/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 14/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 14/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 15/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 16/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 04/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 03/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 15/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 15/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 16/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 17/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 05/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 04/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 16/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 16/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 17/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 18/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 06/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 05/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 17/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 17/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 18/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 19/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 07/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 06/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 18/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 18/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 19/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 20/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 08/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 07/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 19/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 19/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 20/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 21/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 09/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 08/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 20/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 20/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 21/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 22/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 10/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 09/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 21/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 21/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 22/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Channel 23/0 : 5[5] -> 6[6] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 11/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 10/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 22/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 22/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Channel 23/0 : 6[6] -> 7[7] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 12/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 11/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Channel 23/0 : 7[7] -> 0[0] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Channel 23/0 : 3[3] -> 4[4] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 13/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 12/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 14/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 13/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 15/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 14/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 16/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 17/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 15/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 16/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 18/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 17/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 19/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 18/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 20/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 19/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 21/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 20/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 22/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 21/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Channel 23/0 : 1[1] -> 2[2] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 22/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Channel 23/0 : 0[0] -> 1[1] via P2P/CUMEM
+t-20260516080640-ks2k6-worker-0:10256:10427 [6] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516080640-ks2k6-worker-0:10257:10430 [7] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516080640-ks2k6-worker-0:10252:10426 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516080640-ks2k6-worker-0:10251:10431 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516080640-ks2k6-worker-0:10255:10433 [5] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516080640-ks2k6-worker-0:10254:10432 [4] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516080640-ks2k6-worker-0:10250:10428 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+t-20260516080640-ks2k6-worker-0:10253:10429 [3] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+{
+  "device": "cuda:0",
+  "rank": 0,
+  "world_size": 8,
+  "samples": "record_pad_truncate:pad=0:add_eos=0:add_special=0:shuffle_buffer=10000",
+  "vocab_size": 32100,
+  "tokenizer_vocab_size": 32100,
+  "save_dir": "runs/lta_owt_t5record_len1024_elfaligned_dditelf_muon_logitnormal_m1p5_s0p8_none_floor0p0_gbs512_8gpu_5epoch_20260516_000824",
+  "batch_size": 32,
+  "grad_accum": 2,
+  "effective_batch_size": 512,
+  "global_batch_size": 512,
+  "lr_schedule": "constant_warmup",
+  "optimizer": "muon",
+  "epochs": 5.0,
+  "steps_per_epoch": 15457,
+  "total_steps": 77285,
+  "warmup_steps": 7729,
+  "warmup_epochs": 0.5,
+  "min_lr": 0.0,
+  "weight_decay": 0.0,
+  "output_weight_decay": -1.0,
+  "adamw_param_groups": "nanogpt",
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "adam_eps": 1e-08,
+  "muon_impl": "optax",
+  "muon_momentum": 0.95,
+  "muon_ns_steps": 5,
+  "muon_update_scale": 1.0,
+  "muon_nesterov": true,
+  "muon_width_scale": true,
+  "muon_effective_nesterov": true,
+  "muon_effective_width_scale": true,
+  "muon_effective_weight_decay": 0.0,
+  "muon_adam_fallback_nesterov": true,
+  "muon_adam_fallback_weight_decay": 0.0,
+  "ema_decay": 0.9999,
+  "ema_start_step": 0,
+  "model_type": "ddit_elf",
+  "elf_num_time_tokens": 4,
+  "elf_num_model_mode_tokens": 0,
+  "qk_norm": true,
+  "output_bias": false,
+  "output_init_std": 0.02,
+  "norm_type": "rmsnorm",
+  "t_sampling_mode": "logit_normal",
+  "t_sampling_power": 1.0,
+  "t_sampling_eps": 0.0001,
+  "t_sampling_logit_mean": -1.5,
+  "t_sampling_logit_std": 0.8,
+  "dual_t": true,
+  "corrupt_t_mode": "same",
+  "corrupt_min_t": 0.0,
+  "corrupt_max_t": 1.0,
+  "prefix_block_prob": 0.0,
+  "prefix_block_len": 128,
+  "mask_ratio_floor_schedule": "none",
+  "dirichlet_endpoint_mode": "categorical_dual_t",
+  "dirichlet_semantic_t_mode": "same",
+  "dirichlet_semantic_t_value": 0.0,
+  "dirichlet_semantic_t_curve": "linear",
+  "dirichlet_semantic_t_power": 1.0,
+  "endpoint_sequence_random_prob_alpha": 0.0,
+  "categorical_wrong_from_full_vocab": true,
+  "categorical_wrong_from_batch_valid_tokens": false,
+  "categorical_wrong_basin_token_ids": "",
+  "categorical_wrong_basin_prob": 0.0,
+  "categorical_wrong_unigram_prob": 0.0,
+  "categorical_wrong_uniform_prob": 0.0,
+  "categorical_wrong_corpus_unigram_path": "",
+  "categorical_wrong_corpus_unigram_alpha": 1.0,
+  "categorical_wrong_basin_shared_prob": 0.0,
+  "categorical_wrong_unigram_shared_prob": 0.0,
+  "mask_mixture_original_prob": 0.0,
+  "mask_mixture_lowk_prob": 0.0,
+  "mask_mixture_lowcorrupt_prob": 0.0,
+  "mask_mixture_block_prob": 0.0,
+  "mask_mixture_all_prob": 0.0,
+  "mask_mixture_lowk_clean_tokens": "1,2,4,8,16,32,64",
+  "mask_mixture_lowcorrupt_tokens": "1,2,4,8,16,32,64",
+  "mask_mixture_block_tokens": "64,128",
+  "simplex_bridge_sampler": "dirichlet",
+  "logistic_normal_sigma_min": 0.18,
+  "logistic_normal_sigma_max": 2.2,
+  "logistic_normal_tau_min": 0.65,
+  "logistic_normal_tau_max": 1.15,
+  "torch_compile": false,
+  "compile_mode": "max-autotune",
+  "state_format": "prob",
+  "target_loss": "hard_ce",
+  "meanflow_weight": 0.0,
+  "rollout_train_prob": 0.0,
+  "rollout_train_steps": 1,
+  "rollout_train_infer_steps": 64,
+  "rollout_train_temp": 1.45,
+  "rollout_train_max_gamma": 1.0,
+  "rollout_train_corrupt_only": true,
+  "rollout_train_samplewise": false,
+  "rollout_train_compute_always": false,
+  "bridge_noise_init": "logistic_normal",
+  "noise_sigma": -1.0,
+  "allow_tf32": false,
+  "activation_checkpointing": true,
+  "activation_checkpoint_interval": 1,
+  "activation_checkpoint_scope": "mlp",
+  "ddp_static_graph": false,
+  "ddp_gradient_as_bucket_view": true,
+  "blocking_data_transfer": false,
+  "dataloader_prefetch_factor": 4,
+  "full_train_stats": false,
+  "record_pad_truncate": true,
+  "record_add_eos": false,
+  "record_add_special_tokens": false,
+  "record_pad_token": "pad",
+  "record_shuffle_buffer": 10000,
+  "wrap": false,
+  "wrap_mode": "stream",
+  "wrap_record_buffer_size": 200,
+  "owt_cached_chunks": false,
+  "owt_chunk_cache_dir": "",
+  "owt_chunk_cache_rebuild": false,
+  "owt_chunk_cache_write_batch": 4096,
+  "owt_exact_repeat_per_chunk": 0,
+  "online_chunk_shuffle": false,
+  "online_chunk_shuffle_buffer": 10000,
+  "openwebtext_split": "train_minus_100k",
+  "detokenizer": "auto",
+  "resolved_detokenizer": null,
+  "num_workers": 8,
+  "latest_every": 1000,
+  "resume_path": ""
+}
+t-20260516080640-ks2k6-worker-0:10250:10958 [0] NCCL INFO NVLS comm 0x9f91be0 headRank 0 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516080640-ks2k6-worker-0:10251:10960 [1] NCCL INFO NVLS comm 0x9a45db0 headRank 1 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516080640-ks2k6-worker-0:10253:10963 [3] NCCL INFO NVLS comm 0xa06a9c0 headRank 3 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516080640-ks2k6-worker-0:10252:10964 [2] NCCL INFO NVLS comm 0xb2c0700 headRank 2 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516080640-ks2k6-worker-0:10256:10965 [6] NCCL INFO NVLS comm 0xb812680 headRank 6 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516080640-ks2k6-worker-0:10255:10966 [5] NCCL INFO NVLS comm 0xa632cd0 headRank 5 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516080640-ks2k6-worker-0:10254:10967 [4] NCCL INFO NVLS comm 0xb506970 headRank 4 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+t-20260516080640-ks2k6-worker-0:10257:10998 [7] NCCL INFO NVLS comm 0xa6619b0 headRank 7 nHeads 8 buffSize 1048576 nvlsPerRankSize 33554432 nvlsTotalSize 268435456
+step=100 epoch=1/5 epoch_step=100/15457 micro_steps=200 elapsed=209.7s lr=2.613533e-05 loss=10.4988 loss_recon=10.4988 loss_meanflow=0.0000 mean_model_t=0.2081 mean_corrupt_t=0.2081 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0001 corrupt_frac=0.5549 acc_corrupt=0.0001 loss_corrupt=10.4988 wrong_frac=0.7929 init_acc_corrupt=0.1138 acc_corrupt_t_0p0_0p2=0.0001 corrupt_frac_t_0p0_0p2=0.5679 acc_corrupt_t_0p2_0p4=0.0001 corrupt_frac_t_0p2_0p4=0.3435 out_w_norm=99.2788 out_g_norm=0.8302 acc_corrupt_t_0p4_0p6=0.0000 corrupt_frac_t_0p4_0p6=0.0860 acc_corrupt_t_0p6_0p8=0.0000 corrupt_frac_t_0p6_0p8=0.0342 acc_corrupt_t_0p8_1p0=0.0000 corrupt_frac_t_0p8_1p0=0.0095 loss_all=10.3770 init_gold_top10=0.2156 init_gold_top100=0.4562
+step=200 epoch=1/5 epoch_step=200/15457 micro_steps=400 elapsed=179.4s lr=5.201190e-05 loss=10.1387 loss_recon=10.1387 loss_meanflow=0.0000 mean_model_t=0.2102 mean_corrupt_t=0.2102 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0249 corrupt_frac=0.5476 acc_corrupt=0.0248 loss_corrupt=10.1387 wrong_frac=0.7891 init_acc_corrupt=0.1199 acc_corrupt_t_0p0_0p2=0.0244 corrupt_frac_t_0p0_0p2=0.5602 acc_corrupt_t_0p2_0p4=0.0252 corrupt_frac_t_0p2_0p4=0.3460 acc_corrupt_t_0p4_0p6=0.0264 corrupt_frac_t_0p4_0p6=0.0887 out_w_norm=99.2659 out_g_norm=1.1446 acc_corrupt_t_0p6_0p8=0.0224 corrupt_frac_t_0p6_0p8=0.0361 loss_all=9.9286 init_gold_top10=0.1478 init_gold_top100=0.4570
+step=300 epoch=1/5 epoch_step=300/15457 micro_steps=600 elapsed=179.5s lr=7.788847e-05 loss=9.6553 loss_recon=9.6553 loss_meanflow=0.0000 mean_model_t=0.2087 mean_corrupt_t=0.2087 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0436 corrupt_frac=0.5486 acc_corrupt=0.0428 loss_corrupt=9.6553 wrong_frac=0.7920 init_acc_corrupt=0.1149 acc_corrupt_t_0p0_0p2=0.0427 corrupt_frac_t_0p0_0p2=0.5626 acc_corrupt_t_0p2_0p4=0.0429 corrupt_frac_t_0p2_0p4=0.3554 acc_corrupt_t_0p4_0p6=0.0419 corrupt_frac_t_0p4_0p6=0.0764 out_w_norm=99.2536 out_g_norm=1.5325 acc_corrupt_t_0p6_0p8=0.0511 corrupt_frac_t_0p6_0p8=0.0338 acc_corrupt_t_0p8_1p0=0.0702 corrupt_frac_t_0p8_1p0=0.0410 loss_all=9.4007 init_gold_top10=0.1763 init_gold_top100=0.4579
+step=400 epoch=1/5 epoch_step=400/15457 micro_steps=800 elapsed=179.5s lr=1.037650e-04 loss=9.2582 loss_recon=9.2582 loss_meanflow=0.0000 mean_model_t=0.2101 mean_corrupt_t=0.2101 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0497 corrupt_frac=0.5498 acc_corrupt=0.0446 loss_corrupt=9.2582 wrong_frac=0.7920 init_acc_corrupt=0.1141 acc_corrupt_t_0p0_0p2=0.0438 corrupt_frac_t_0p0_0p2=0.5615 acc_corrupt_t_0p2_0p4=0.0449 corrupt_frac_t_0p2_0p4=0.3516 acc_corrupt_t_0p4_0p6=0.0475 corrupt_frac_t_0p4_0p6=0.0843 out_w_norm=99.2475 out_g_norm=1.5639 acc_corrupt_t_0p6_0p8=0.0543 corrupt_frac_t_0p6_0p8=0.0315 loss_all=9.0300 init_gold_top10=0.2436 init_gold_top100=0.4551
+step=500 epoch=1/5 epoch_step=500/15457 micro_steps=1000 elapsed=179.3s lr=1.296416e-04 loss=8.9755 loss_recon=8.9755 loss_meanflow=0.0000 mean_model_t=0.2087 mean_corrupt_t=0.2087 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 rollout_train_applied=0.0000 grad_enabled_before_rollout=1.0000 grad_enabled_after_rollout=1.0000 logits_requires_grad=1.0000 raw_loss_requires_grad=1.0000 acc_all=0.0741 corrupt_frac=0.5571 acc_corrupt=0.0502 loss_corrupt=8.9755 wrong_frac=0.7934 init_acc_corrupt=0.1133 acc_corrupt_t_0p0_0p2=0.0456 corrupt_frac_t_0p0_0p2=0.5588 acc_corrupt_t_0p2_0p4=0.0524 corrupt_frac_t_0p2_0p4=0.3629 acc_corrupt_t_0p4_0p6=0.0708 corrupt_frac_t_0p4_0p6=0.0744 out_w_norm=99.2468 out_g_norm=1.4180 acc_corrupt_t_0p6_0p8=0.0897 corrupt_frac_t_0p6_0p8=0.0318 acc_corrupt_t_0p8_1p0=0.1306 corrupt_frac_t_0p8_1p0=0.0262 loss_all=8.7811 init_gold_top10=0.1701 init_gold_top100=0.4774

LTA_openwebtext_dualt/logs/lta_lm1b_classic_dirichlet_len256_gbs512_4gpu_10k_save1k_20260523.watch.pid

@@ -0,0 +1 @@
+994420

LTA_openwebtext_dualt/logs/train_lta_owt_cached_fast10k_4gpu_500step.log

@@ -0,0 +1,178 @@
+[launch] method=owt_categorical_fullvocab_c1024_fullycoupled host=di-20260411014000-djqhq time=2026-05-12T16:34:51+00:00
+[launch] cwd=/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt
+[launch] run_name=lta_owt_c1024_gpt2_cached_chunks_len1024_fast10k_4gpu_500step
+[launch] save_dir=runs/lta_owt_c1024_gpt2_cached_chunks_len1024_fast10k_4gpu_500step
+[launch] log_file=logs/lta_owt_c1024_gpt2_cached_chunks_len1024_fast10k_4gpu_500step.log
+[launch] data_path=/e2e-data/evad-tech-vla/wanghan58/data/small_benchmarks/langflow_2604_11748/openwebtext
+[launch] tokenizer=/e2e-data/evad-tech-vla/wanghan58/models/flowtext_scorers/gpt2-standard/tokenizer.json
+[launch] split=train_minus_100k text_column=text
+[launch] owt_cached_chunks=1 cache_dir=/e2e-data/evad-tech-vla/wanghan58/data/small_benchmarks/langflow_2604_11748/openwebtext_lta_cached_chunks/gpt2_len1024_train_minus_100k_fast10k
+[launch] nproc_per_node=4 global_batch_size=512 per_gpu_batch_size=32
+
+*****************************************
+Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+*****************************************
+[rank0]:[W512 16:35:00.260053068 ProcessGroupNCCL.cpp:4571] [PG ID 0 PG GUID 0 Rank 0]  using GPU 0 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+NCCL version 2.25.1+cuda12.8
+[rank1]:[W512 16:35:00.299852507 ProcessGroupNCCL.cpp:4571] [PG ID 0 PG GUID 0 Rank 1]  using GPU 1 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+[rank2]:[W512 16:35:00.301136581 ProcessGroupNCCL.cpp:4571] [PG ID 0 PG GUID 0 Rank 2]  using GPU 2 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+[rank3]:[W512 16:35:00.304399895 ProcessGroupNCCL.cpp:4571] [PG ID 0 PG GUID 0 Rank 3]  using GPU 3 to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. Specify device_ids in barrier() to force use of a particular device, or call init_process_group() with a device_id.
+{
+  "device": "cuda:0",
+  "rank": 0,
+  "world_size": 4,
+  "samples": "owt_cached_chunks:10904",
+  "vocab_size": 50257,
+  "save_dir": "runs/lta_owt_c1024_gpt2_cached_chunks_len1024_fast10k_4gpu_500step",
+  "batch_size": 32,
+  "grad_accum": 4,
+  "effective_batch_size": 512,
+  "global_batch_size": 512,
+  "lr_schedule": "constant_warmup",
+  "warmup_steps": 50,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "adam_eps": 1e-08,
+  "model_type": "ddit",
+  "dual_t": true,
+  "corrupt_t_mode": "same",
+  "corrupt_min_t": 0.0,
+  "corrupt_max_t": 1.0,
+  "dirichlet_endpoint_mode": "categorical_dual_t",
+  "dirichlet_semantic_t_mode": "same",
+  "dirichlet_semantic_t_value": 0.0,
+  "categorical_wrong_from_full_vocab": true,
+  "simplex_bridge_sampler": "dirichlet",
+  "logistic_normal_sigma_min": 0.18,
+  "logistic_normal_sigma_max": 2.2,
+  "logistic_normal_tau_min": 0.65,
+  "logistic_normal_tau_max": 1.15,
+  "torch_compile": false,
+  "compile_mode": "max-autotune",
+  "state_format": "prob",
+  "target_loss": "hard_ce",
+  "meanflow_weight": 0.0,
+  "bridge_noise_init": "logistic_normal",
+  "noise_sigma": -1.0,
+  "wrap": true,
+  "wrap_mode": "stream",
+  "wrap_record_buffer_size": 200,
+  "owt_cached_chunks": true,
+  "owt_chunk_cache_dir": "/e2e-data/evad-tech-vla/wanghan58/data/small_benchmarks/langflow_2604_11748/openwebtext_lta_cached_chunks/gpt2_len1024_train_minus_100k_fast10k",
+  "owt_chunk_cache_rebuild": false,
+  "owt_chunk_cache_write_batch": 4096,
+  "online_chunk_shuffle": false,
+  "online_chunk_shuffle_buffer": 10000,
+  "openwebtext_split": "train_minus_100k",
+  "detokenizer": "auto",
+  "resolved_detokenizer": null,
+  "num_workers": 2,
+  "latest_every": 50,
+  "resume_path": ""
+}
+step=10 micro_steps=40 elapsed=67.7s lr=6.600000e-05 loss_all=10.7775 acc_all=0.6098 loss_corrupt=10.7889 acc_corrupt=0.4141 corrupt_frac=0.5619 loss=10.7889 loss_recon=10.7889 loss_meanflow=0.0000 mean_model_t=0.4986 mean_corrupt_t=0.4986 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 wrong_frac=0.4990 init_acc_corrupt=0.4668 init_gold_top10=0.4955 init_gold_top100=0.5245
+step=20 micro_steps=80 elapsed=57.9s lr=1.260000e-04 loss_all=10.3772 acc_all=0.3528 loss_corrupt=10.4082 acc_corrupt=0.2188 corrupt_frac=0.5554 loss=10.4082 loss_recon=10.4082 loss_meanflow=0.0000 mean_model_t=0.4975 mean_corrupt_t=0.4975 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 wrong_frac=0.4996 init_acc_corrupt=0.4680 init_gold_top10=0.4957 init_gold_top100=0.5226
+step=30 micro_steps=120 elapsed=59.3s lr=1.860000e-04 loss_all=9.3684 acc_all=0.2037 loss_corrupt=9.4072 acc_corrupt=0.1250 corrupt_frac=0.5514 loss=9.4072 loss_recon=9.4072 loss_meanflow=0.0000 mean_model_t=0.4946 mean_corrupt_t=0.4946 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 wrong_frac=0.5065 init_acc_corrupt=0.4578 init_gold_top10=0.4874 init_gold_top100=0.5184
+step=40 micro_steps=160 elapsed=62.9s lr=2.460000e-04 loss_all=8.1540 acc_all=0.2264 loss_corrupt=8.2181 acc_corrupt=0.1478 corrupt_frac=0.5382 loss=8.2181 loss_recon=8.2181 loss_meanflow=0.0000 mean_model_t=0.4880 mean_corrupt_t=0.4880 mean_loss_t_weight=1.0000 prior_center_loss_beta=0.0000 wrong_frac=0.5150 init_acc_corrupt=0.4486 init_gold_top10=0.4790 init_gold_top100=0.5111
+[rank0]: Traceback (most recent call last):
+[rank0]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 797, in <module>
+[rank0]:     main()
+[rank0]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 693, in main
+[rank0]:     bridge = make_bridge(
+[rank0]:              ^^^^^^^^^^^^
+[rank0]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 313, in make_bridge
+[rank0]:     return make_dirichlet_bridge_batch(
+[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+[rank0]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/flowtext_lab/bridges.py", line 529, in make_dirichlet_bridge_batch
+[rank0]:     state_probs = sample_dirichlet_bridge_simplex(
+[rank0]:                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+[rank0]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/flowtext_lab/bridges.py", line 263, in sample_dirichlet_bridge_simplex
+[rank0]:     sample = sample.clamp_min(eps)
+[rank0]:              ^^^^^^^^^^^^^^^^^^^^^
+[rank0]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 6.14 GiB. GPU 0 has a total capacity of 95.22 GiB of which 573.56 MiB is free. Process 971246 has 94.66 GiB memory in use. Of the allocated memory 61.51 GiB is allocated by PyTorch, and 31.86 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
+[rank1]: Traceback (most recent call last):
+[rank1]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 797, in <module>
+[rank1]:     main()
+[rank1]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 693, in main
+[rank1]:     bridge = make_bridge(
+[rank1]:              ^^^^^^^^^^^^
+[rank1]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 313, in make_bridge
+[rank1]:     return make_dirichlet_bridge_batch(
+[rank1]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+[rank1]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/flowtext_lab/bridges.py", line 529, in make_dirichlet_bridge_batch
+[rank1]:     state_probs = sample_dirichlet_bridge_simplex(
+[rank1]:                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+[rank1]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/flowtext_lab/bridges.py", line 263, in sample_dirichlet_bridge_simplex
+[rank1]:     sample = sample.clamp_min(eps)
+[rank1]:              ^^^^^^^^^^^^^^^^^^^^^
+[rank1]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 6.14 GiB. GPU 1 has a total capacity of 95.22 GiB of which 573.56 MiB is free. Process 971247 has 94.66 GiB memory in use. Of the allocated memory 61.51 GiB is allocated by PyTorch, and 31.86 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
+[rank3]: Traceback (most recent call last):
+[rank3]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 797, in <module>
+[rank3]:     main()
+[rank3]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 693, in main
+[rank3]:     bridge = make_bridge(
+[rank3]:              ^^^^^^^^^^^^
+[rank3]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 313, in make_bridge
+[rank3]:     return make_dirichlet_bridge_batch(
+[rank3]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+[rank3]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/flowtext_lab/bridges.py", line 529, in make_dirichlet_bridge_batch
+[rank3]:     state_probs = sample_dirichlet_bridge_simplex(
+[rank3]:                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+[rank3]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/flowtext_lab/bridges.py", line 263, in sample_dirichlet_bridge_simplex
+[rank3]:     sample = sample.clamp_min(eps)
+[rank3]:              ^^^^^^^^^^^^^^^^^^^^^
+[rank3]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 6.14 GiB. GPU 3 has a total capacity of 95.22 GiB of which 573.56 MiB is free. Process 971249 has 94.66 GiB memory in use. Of the allocated memory 61.51 GiB is allocated by PyTorch, and 31.86 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
+[rank2]: Traceback (most recent call last):
+[rank2]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 797, in <module>
+[rank2]:     main()
+[rank2]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 693, in main
+[rank2]:     bridge = make_bridge(
+[rank2]:              ^^^^^^^^^^^^
+[rank2]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/train.py", line 313, in make_bridge
+[rank2]:     return make_dirichlet_bridge_batch(
+[rank2]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+[rank2]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/flowtext_lab/bridges.py", line 529, in make_dirichlet_bridge_batch
+[rank2]:     state_probs = sample_dirichlet_bridge_simplex(
+[rank2]:                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+[rank2]:   File "/e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt/flowtext_lab/bridges.py", line 263, in sample_dirichlet_bridge_simplex
+[rank2]:     sample = sample.clamp_min(eps)
+[rank2]:              ^^^^^^^^^^^^^^^^^^^^^
+[rank2]: torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 6.14 GiB. GPU 2 has a total capacity of 95.22 GiB of which 573.56 MiB is free. Process 971248 has 94.66 GiB memory in use. Of the allocated memory 61.51 GiB is allocated by PyTorch, and 31.86 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
+[rank0]:[W512 16:40:20.480386792 ProcessGroupNCCL.cpp:1487] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
+W0512 16:40:21.089000 208015 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 208083 closing signal SIGTERM
+W0512 16:40:21.090000 208015 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 208084 closing signal SIGTERM
+W0512 16:40:21.090000 208015 torch/distributed/elastic/multiprocessing/api.py:898] Sending process 208085 closing signal SIGTERM
+E0512 16:40:21.469000 208015 torch/distributed/elastic/multiprocessing/api.py:870] failed (exitcode: 1) local_rank: 0 (pid: 208082) of binary: /usr/bin/python
+Traceback (most recent call last):
+  File "<frozen runpy>", line 198, in _run_module_as_main
+  File "<frozen runpy>", line 88, in _run_code
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 922, in <module>
+    main()
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
+    return f(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 918, in main
+    run(args)
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 909, in run
+    elastic_launch(
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
+    return launch_agent(self._config, self._entrypoint, list(args))
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 270, in launch_agent
+    raise ChildFailedError(
+torch.distributed.elastic.multiprocessing.errors.ChildFailedError: 
+============================================================
+train.py FAILED
+------------------------------------------------------------
+Failures:
+  <NO_OTHER_FAILURES>
+------------------------------------------------------------
+Root Cause (first observed failure):
+[0]:
+  time      : 2026-05-12_16:40:21
+  host      : localhost
+  rank      : 0 (local_rank: 0)
+  exitcode  : 1 (pid: 208082)
+  error_file: <N/A>
+  traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
+============================================================

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/fsspec/implementations/http.py

@@ -0,0 +1,902 @@
+import asyncio
+import io
+import logging
+import re
+import weakref
+from copy import copy
+from urllib.parse import urlparse
+
+import aiohttp
+import yarl
+
+from fsspec.asyn import AbstractAsyncStreamedFile, AsyncFileSystem, sync, sync_wrapper
+from fsspec.callbacks import DEFAULT_CALLBACK
+from fsspec.exceptions import FSTimeoutError
+from fsspec.spec import AbstractBufferedFile
+from fsspec.utils import (
+    DEFAULT_BLOCK_SIZE,
+    glob_translate,
+    isfilelike,
+    nullcontext,
+    tokenize,
+)
+
+from ..caching import AllBytes
+
+# https://stackoverflow.com/a/15926317/3821154
+ex = re.compile(r"""<(a|A)\s+(?:[^>]*?\s+)?(href|HREF)=["'](?P<url>[^"']+)""")
+ex2 = re.compile(r"""(?P<url>http[s]?://[-a-zA-Z0-9@:%_+.~#?&/=]+)""")
+logger = logging.getLogger("fsspec.http")
+
+
+async def get_client(**kwargs):
+    return aiohttp.ClientSession(**kwargs)
+
+
+class HTTPFileSystem(AsyncFileSystem):
+    """
+    Simple File-System for fetching data via HTTP(S)
+
+    ``ls()`` is implemented by loading the parent page and doing a regex
+    match on the result. If simple_link=True, anything of the form
+    "http(s)://server.com/stuff?thing=other"; otherwise only links within
+    HTML href tags will be used.
+
+    URLs are passed unfiltered to aiohttp, so all addresses are accessible. Where URLs are
+    supplied by a user, the calling application may wish to filter to prevent scanning.
+    """
+
+    protocol = ("http", "https")
+    sep = "/"
+
+    def __init__(
+        self,
+        simple_links=True,
+        block_size=None,
+        same_scheme=True,
+        size_policy=None,
+        cache_type="bytes",
+        cache_options=None,
+        asynchronous=False,
+        loop=None,
+        client_kwargs=None,
+        get_client=get_client,
+        encoded=False,
+        **storage_options,
+    ):
+        """
+        NB: if this is called async, you must await set_client
+
+        Parameters
+        ----------
+        block_size: int
+            Blocks to read bytes; if 0, will default to raw requests file-like
+            objects instead of HTTPFile instances
+        simple_links: bool
+            If True, will consider both HTML <a> tags and anything that looks
+            like a URL; if False, will consider only the former.
+        same_scheme: True
+            When doing ls/glob, if this is True, only consider paths that have
+            http/https matching the input URLs.
+        size_policy: this argument is deprecated
+        client_kwargs: dict
+            Passed to aiohttp.ClientSession, see
+            https://docs.aiohttp.org/en/stable/client_reference.html
+            For example, ``{'auth': aiohttp.BasicAuth('user', 'pass')}``
+        get_client: Callable[..., aiohttp.ClientSession]
+            A callable, which takes keyword arguments and constructs
+            an aiohttp.ClientSession. Its state will be managed by
+            the HTTPFileSystem class.
+        storage_options: key-value
+            Any other parameters passed on to requests
+        cache_type, cache_options: defaults used in open()
+        """
+        super().__init__(self, asynchronous=asynchronous, loop=loop, **storage_options)
+        self.block_size = block_size if block_size is not None else DEFAULT_BLOCK_SIZE
+        self.simple_links = simple_links
+        self.same_schema = same_scheme
+        self.cache_type = cache_type
+        self.cache_options = cache_options
+        self.client_kwargs = client_kwargs or {}
+        self.get_client = get_client
+        self.encoded = encoded
+        self.kwargs = storage_options
+        self._session = None
+
+        # Clean caching-related parameters from `storage_options`
+        # before propagating them as `request_options` through `self.kwargs`.
+        # TODO: Maybe rename `self.kwargs` to `self.request_options` to make
+        #       it clearer.
+        request_options = copy(storage_options)
+        self.use_listings_cache = request_options.pop("use_listings_cache", False)
+        request_options.pop("listings_expiry_time", None)
+        request_options.pop("max_paths", None)
+        request_options.pop("skip_instance_cache", None)
+        self.kwargs = request_options
+
+    @property
+    def fsid(self):
+        return "http"
+
+    def encode_url(self, url):
+        return yarl.URL(url, encoded=self.encoded)
+
+    @staticmethod
+    def close_session(loop, session):
+        if loop is not None and loop.is_running():
+            try:
+                sync(loop, session.close, timeout=0.1)
+                return
+            except (TimeoutError, FSTimeoutError, NotImplementedError):
+                pass
+        connector = getattr(session, "_connector", None)
+        if connector is not None:
+            # close after loop is dead
+            connector._close()
+
+    async def set_session(self):
+        if self._session is None:
+            self._session = await self.get_client(loop=self.loop, **self.client_kwargs)
+            if not self.asynchronous:
+                weakref.finalize(self, self.close_session, self.loop, self._session)
+        return self._session
+
+    @classmethod
+    def _strip_protocol(cls, path):
+        """For HTTP, we always want to keep the full URL"""
+        return path
+
+    @classmethod
+    def _parent(cls, path):
+        # override, since _strip_protocol is different for URLs
+        par = super()._parent(path)
+        if len(par) > 7:  # "http://..."
+            return par
+        return ""
+
+    async def _ls_real(self, url, detail=True, **kwargs):
+        # ignoring URL-encoded arguments
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        logger.debug(url)
+        session = await self.set_session()
+        async with session.get(self.encode_url(url), **self.kwargs) as r:
+            self._raise_not_found_for_status(r, url)
+
+            if "Content-Type" in r.headers:
+                mimetype = r.headers["Content-Type"].partition(";")[0]
+            else:
+                mimetype = None
+
+            if mimetype in ("text/html", None):
+                try:
+                    text = await r.text(errors="ignore")
+                    if self.simple_links:
+                        links = ex2.findall(text) + [u[2] for u in ex.findall(text)]
+                    else:
+                        links = [u[2] for u in ex.findall(text)]
+                except UnicodeDecodeError:
+                    links = []  # binary, not HTML
+            else:
+                links = []
+
+        out = set()
+        parts = urlparse(url)
+        for l in links:
+            if isinstance(l, tuple):
+                l = l[1]
+            if l.startswith("/") and len(l) > 1:
+                # absolute URL on this server
+                l = f"{parts.scheme}://{parts.netloc}{l}"
+            if l.startswith("http"):
+                if self.same_schema and l.startswith(url.rstrip("/") + "/"):
+                    out.add(l)
+                elif l.replace("https", "http").startswith(
+                    url.replace("https", "http").rstrip("/") + "/"
+                ):
+                    # allowed to cross http <-> https
+                    out.add(l)
+            else:
+                if l not in ["..", "../"]:
+                    # Ignore FTP-like "parent"
+                    out.add("/".join([url.rstrip("/"), l.lstrip("/")]))
+        if not out and url.endswith("/"):
+            out = await self._ls_real(url.rstrip("/"), detail=False)
+        if detail:
+            return [
+                {
+                    "name": u,
+                    "size": None,
+                    "type": "directory" if u.endswith("/") else "file",
+                }
+                for u in out
+            ]
+        else:
+            return sorted(out)
+
+    async def _ls(self, url, detail=True, **kwargs):
+        if self.use_listings_cache and url in self.dircache:
+            out = self.dircache[url]
+        else:
+            out = await self._ls_real(url, detail=detail, **kwargs)
+            self.dircache[url] = out
+        return out
+
+    ls = sync_wrapper(_ls)
+
+    def _raise_not_found_for_status(self, response, url):
+        """
+        Raises FileNotFoundError for 404s, otherwise uses raise_for_status.
+        """
+        if response.status == 404:
+            raise FileNotFoundError(url)
+        response.raise_for_status()
+
+    async def _cat_file(self, url, start=None, end=None, **kwargs):
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        logger.debug(url)
+
+        if start is not None or end is not None:
+            if start == end:
+                return b""
+            headers = kw.pop("headers", {}).copy()
+
+            headers["Range"] = await self._process_limits(url, start, end)
+            kw["headers"] = headers
+        session = await self.set_session()
+        async with session.get(self.encode_url(url), **kw) as r:
+            out = await r.read()
+            self._raise_not_found_for_status(r, url)
+        return out
+
+    async def _get_file(
+        self, rpath, lpath, chunk_size=5 * 2**20, callback=DEFAULT_CALLBACK, **kwargs
+    ):
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        logger.debug(rpath)
+        session = await self.set_session()
+        async with session.get(self.encode_url(rpath), **kw) as r:
+            try:
+                size = int(r.headers["content-length"])
+            except (ValueError, KeyError):
+                size = None
+
+            callback.set_size(size)
+            self._raise_not_found_for_status(r, rpath)
+            if isfilelike(lpath):
+                outfile = lpath
+            else:
+                outfile = open(lpath, "wb")  # noqa: ASYNC230
+
+            try:
+                chunk = True
+                while chunk:
+                    chunk = await r.content.read(chunk_size)
+                    outfile.write(chunk)
+                    callback.relative_update(len(chunk))
+            finally:
+                if not isfilelike(lpath):
+                    outfile.close()
+
+    async def _put_file(
+        self,
+        lpath,
+        rpath,
+        chunk_size=5 * 2**20,
+        callback=DEFAULT_CALLBACK,
+        method="post",
+        mode="overwrite",
+        **kwargs,
+    ):
+        if mode != "overwrite":
+            raise NotImplementedError("Exclusive write")
+
+        async def gen_chunks():
+            # Support passing arbitrary file-like objects
+            # and use them instead of streams.
+            if isinstance(lpath, io.IOBase):
+                context = nullcontext(lpath)
+                use_seek = False  # might not support seeking
+            else:
+                context = open(lpath, "rb")  # noqa: ASYNC230
+                use_seek = True
+
+            with context as f:
+                if use_seek:
+                    callback.set_size(f.seek(0, 2))
+                    f.seek(0)
+                else:
+                    callback.set_size(getattr(f, "size", None))
+
+                chunk = f.read(chunk_size)
+                while chunk:
+                    yield chunk
+                    callback.relative_update(len(chunk))
+                    chunk = f.read(chunk_size)
+
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        session = await self.set_session()
+
+        method = method.lower()
+        if method not in ("post", "put"):
+            raise ValueError(
+                f"method has to be either 'post' or 'put', not: {method!r}"
+            )
+
+        meth = getattr(session, method)
+        async with meth(self.encode_url(rpath), data=gen_chunks(), **kw) as resp:
+            self._raise_not_found_for_status(resp, rpath)
+
+    async def _exists(self, path, strict=False, **kwargs):
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        try:
+            logger.debug(path)
+            session = await self.set_session()
+            r = await session.get(self.encode_url(path), **kw)
+            async with r:
+                if strict:
+                    self._raise_not_found_for_status(r, path)
+                return r.status < 400
+        except FileNotFoundError:
+            return False
+        except aiohttp.ClientError:
+            if strict:
+                raise
+            return False
+
+    async def _isfile(self, path, **kwargs):
+        return await self._exists(path, **kwargs)
+
+    def _open(
+        self,
+        path,
+        mode="rb",
+        block_size=None,
+        autocommit=None,  # XXX: This differs from the base class.
+        cache_type=None,
+        cache_options=None,
+        size=None,
+        **kwargs,
+    ):
+        """Make a file-like object
+
+        Parameters
+        ----------
+        path: str
+            Full URL with protocol
+        mode: string
+            must be "rb"
+        block_size: int or None
+            Bytes to download in one request; use instance value if None. If
+            zero, will return a streaming Requests file-like instance.
+        kwargs: key-value
+            Any other parameters, passed to requests calls
+        """
+        if mode != "rb":
+            raise NotImplementedError
+        block_size = block_size if block_size is not None else self.block_size
+        kw = self.kwargs.copy()
+        kw["asynchronous"] = self.asynchronous
+        kw.update(kwargs)
+        info = {}
+        size = size or info.update(self.info(path, **kwargs)) or info["size"]
+        session = sync(self.loop, self.set_session)
+        if block_size and size and info.get("partial", True):
+            return HTTPFile(
+                self,
+                path,
+                session=session,
+                block_size=block_size,
+                mode=mode,
+                size=size,
+                cache_type=cache_type or self.cache_type,
+                cache_options=cache_options or self.cache_options,
+                loop=self.loop,
+                **kw,
+            )
+        else:
+            return HTTPStreamFile(
+                self,
+                path,
+                mode=mode,
+                loop=self.loop,
+                session=session,
+                **kw,
+            )
+
+    async def open_async(self, path, mode="rb", size=None, **kwargs):
+        session = await self.set_session()
+        if size is None:
+            try:
+                size = (await self._info(path, **kwargs))["size"]
+            except FileNotFoundError:
+                pass
+        return AsyncStreamFile(
+            self,
+            path,
+            loop=self.loop,
+            session=session,
+            size=size,
+            **kwargs,
+        )
+
+    def ukey(self, url):
+        """Unique identifier; assume HTTP files are static, unchanging"""
+        return tokenize(url, self.kwargs, self.protocol)
+
+    async def _info(self, url, **kwargs):
+        """Get info of URL
+
+        Tries to access location via HEAD, and then GET methods, but does
+        not fetch the data.
+
+        It is possible that the server does not supply any size information, in
+        which case size will be given as None (and certain operations on the
+        corresponding file will not work).
+        """
+        info = {}
+        session = await self.set_session()
+
+        for policy in ["head", "get"]:
+            try:
+                info.update(
+                    await _file_info(
+                        self.encode_url(url),
+                        size_policy=policy,
+                        session=session,
+                        **self.kwargs,
+                        **kwargs,
+                    )
+                )
+                if info.get("size") is not None:
+                    break
+            except Exception as exc:
+                if policy == "get":
+                    # If get failed, then raise a FileNotFoundError
+                    raise FileNotFoundError(url) from exc
+                logger.debug("", exc_info=exc)
+
+        return {"name": url, "size": None, **info, "type": "file"}
+
+    async def _glob(self, path, maxdepth=None, **kwargs):
+        """
+        Find files by glob-matching.
+
+        This implementation is idntical to the one in AbstractFileSystem,
+        but "?" is not considered as a character for globbing, because it is
+        so common in URLs, often identifying the "query" part.
+        """
+        if maxdepth is not None and maxdepth < 1:
+            raise ValueError("maxdepth must be at least 1")
+        import re
+
+        ends_with_slash = path.endswith("/")  # _strip_protocol strips trailing slash
+        path = self._strip_protocol(path)
+        append_slash_to_dirname = ends_with_slash or path.endswith(("/**", "/*"))
+        idx_star = path.find("*") if path.find("*") >= 0 else len(path)
+        idx_brace = path.find("[") if path.find("[") >= 0 else len(path)
+
+        min_idx = min(idx_star, idx_brace)
+
+        detail = kwargs.pop("detail", False)
+
+        if not has_magic(path):
+            if await self._exists(path, **kwargs):
+                if not detail:
+                    return [path]
+                else:
+                    return {path: await self._info(path, **kwargs)}
+            else:
+                if not detail:
+                    return []  # glob of non-existent returns empty
+                else:
+                    return {}
+        elif "/" in path[:min_idx]:
+            min_idx = path[:min_idx].rindex("/")
+            root = path[: min_idx + 1]
+            depth = path[min_idx + 1 :].count("/") + 1
+        else:
+            root = ""
+            depth = path[min_idx + 1 :].count("/") + 1
+
+        if "**" in path:
+            if maxdepth is not None:
+                idx_double_stars = path.find("**")
+                depth_double_stars = path[idx_double_stars:].count("/") + 1
+                depth = depth - depth_double_stars + maxdepth
+            else:
+                depth = None
+
+        allpaths = await self._find(
+            root, maxdepth=depth, withdirs=True, detail=True, **kwargs
+        )
+
+        pattern = glob_translate(path + ("/" if ends_with_slash else ""))
+        pattern = re.compile(pattern)
+
+        out = {
+            (
+                p.rstrip("/")
+                if not append_slash_to_dirname
+                and info["type"] == "directory"
+                and p.endswith("/")
+                else p
+            ): info
+            for p, info in sorted(allpaths.items())
+            if pattern.match(p.rstrip("/"))
+        }
+
+        if detail:
+            return out
+        else:
+            return list(out)
+
+    async def _isdir(self, path):
+        # override, since all URLs are (also) files
+        try:
+            return bool(await self._ls(path))
+        except (FileNotFoundError, ValueError):
+            return False
+
+    async def _pipe_file(self, path, value, mode="overwrite", **kwargs):
+        """
+        Write bytes to a remote file over HTTP.
+
+        Parameters
+        ----------
+        path : str
+            Target URL where the data should be written
+        value : bytes
+            Data to be written
+        mode : str
+            How to write to the file - 'overwrite' or 'append'
+        **kwargs : dict
+            Additional parameters to pass to the HTTP request
+        """
+        url = self._strip_protocol(path)
+        headers = kwargs.pop("headers", {})
+        headers["Content-Length"] = str(len(value))
+
+        session = await self.set_session()
+
+        async with session.put(
+            self.encode_url(url), data=value, headers=headers, **kwargs
+        ) as r:
+            r.raise_for_status()
+
+
+class HTTPFile(AbstractBufferedFile):
+    """
+    A file-like object pointing to a remote HTTP(S) resource
+
+    Supports only reading, with read-ahead of a predetermined block-size.
+
+    In the case that the server does not supply the filesize, only reading of
+    the complete file in one go is supported.
+
+    Parameters
+    ----------
+    url: str
+        Full URL of the remote resource, including the protocol
+    session: aiohttp.ClientSession or None
+        All calls will be made within this session, to avoid restarting
+        connections where the server allows this
+    block_size: int or None
+        The amount of read-ahead to do, in bytes. Default is 5MB, or the value
+        configured for the FileSystem creating this file
+    size: None or int
+        If given, this is the size of the file in bytes, and we don't attempt
+        to call the server to find the value.
+    kwargs: all other key-values are passed to requests calls.
+    """
+
+    def __init__(
+        self,
+        fs,
+        url,
+        session=None,
+        block_size=None,
+        mode="rb",
+        cache_type="bytes",
+        cache_options=None,
+        size=None,
+        loop=None,
+        asynchronous=False,
+        **kwargs,
+    ):
+        if mode != "rb":
+            raise NotImplementedError("File mode not supported")
+        self.asynchronous = asynchronous
+        self.loop = loop
+        self.url = url
+        self.session = session
+        self.details = {"name": url, "size": size, "type": "file"}
+        super().__init__(
+            fs=fs,
+            path=url,
+            mode=mode,
+            block_size=block_size,
+            cache_type=cache_type,
+            cache_options=cache_options,
+            **kwargs,
+        )
+
+    def read(self, length=-1):
+        """Read bytes from file
+
+        Parameters
+        ----------
+        length: int
+            Read up to this many bytes. If negative, read all content to end of
+            file. If the server has not supplied the filesize, attempting to
+            read only part of the data will raise a ValueError.
+        """
+        if (
+            (length < 0 and self.loc == 0)  # explicit read all
+            # but not when the size is known and fits into a block anyways
+            and not (self.size is not None and self.size <= self.blocksize)
+        ):
+            self._fetch_all()
+        if self.size is None:
+            if length < 0:
+                self._fetch_all()
+        else:
+            length = min(self.size - self.loc, length)
+        return super().read(length)
+
+    async def async_fetch_all(self):
+        """Read whole file in one shot, without caching
+
+        This is only called when position is still at zero,
+        and read() is called without a byte-count.
+        """
+        logger.debug(f"Fetch all for {self}")
+        if not isinstance(self.cache, AllBytes):
+            r = await self.session.get(self.fs.encode_url(self.url), **self.kwargs)
+            async with r:
+                r.raise_for_status()
+                out = await r.read()
+                self.cache = AllBytes(
+                    size=len(out), fetcher=None, blocksize=None, data=out
+                )
+                self.size = len(out)
+
+    _fetch_all = sync_wrapper(async_fetch_all)
+
+    def _parse_content_range(self, headers):
+        """Parse the Content-Range header"""
+        s = headers.get("Content-Range", "")
+        m = re.match(r"bytes (\d+-\d+|\*)/(\d+|\*)", s)
+        if not m:
+            return None, None, None
+
+        if m[1] == "*":
+            start = end = None
+        else:
+            start, end = [int(x) for x in m[1].split("-")]
+        total = None if m[2] == "*" else int(m[2])
+        return start, end, total
+
+    async def async_fetch_range(self, start, end):
+        """Download a block of data
+
+        The expectation is that the server returns only the requested bytes,
+        with HTTP code 206. If this is not the case, we first check the headers,
+        and then stream the output - if the data size is bigger than we
+        requested, an exception is raised.
+        """
+        logger.debug(f"Fetch range for {self}: {start}-{end}")
+        kwargs = self.kwargs.copy()
+        headers = kwargs.pop("headers", {}).copy()
+        headers["Range"] = f"bytes={start}-{end - 1}"
+        logger.debug(f"{self.url} : {headers['Range']}")
+        r = await self.session.get(
+            self.fs.encode_url(self.url), headers=headers, **kwargs
+        )
+        async with r:
+            if r.status == 416:
+                # range request outside file
+                return b""
+            r.raise_for_status()
+
+            # If the server has handled the range request, it should reply
+            # with status 206 (partial content). But we'll guess that a suitable
+            # Content-Range header or a Content-Length no more than the
+            # requested range also mean we have got the desired range.
+            response_is_range = (
+                r.status == 206
+                or self._parse_content_range(r.headers)[0] == start
+                or int(r.headers.get("Content-Length", end + 1)) <= end - start
+            )
+
+            if response_is_range:
+                # partial content, as expected
+                out = await r.read()
+            elif start > 0:
+                raise ValueError(
+                    "The HTTP server doesn't appear to support range requests. "
+                    "Only reading this file from the beginning is supported. "
+                    "Open with block_size=0 for a streaming file interface."
+                )
+            else:
+                # Response is not a range, but we want the start of the file,
+                # so we can read the required amount anyway.
+                cl = 0
+                out = []
+                while True:
+                    chunk = await r.content.read(2**20)
+                    # data size unknown, let's read until we have enough
+                    if chunk:
+                        out.append(chunk)
+                        cl += len(chunk)
+                        if cl > end - start:
+                            break
+                    else:
+                        break
+                out = b"".join(out)[: end - start]
+            return out
+
+    _fetch_range = sync_wrapper(async_fetch_range)
+
+
+magic_check = re.compile("([*[])")
+
+
+def has_magic(s):
+    match = magic_check.search(s)
+    return match is not None
+
+
+class HTTPStreamFile(AbstractBufferedFile):
+    def __init__(self, fs, url, mode="rb", loop=None, session=None, **kwargs):
+        self.asynchronous = kwargs.pop("asynchronous", False)
+        self.url = url
+        self.loop = loop
+        self.session = session
+        if mode != "rb":
+            raise ValueError
+        self.details = {"name": url, "size": None}
+        super().__init__(fs=fs, path=url, mode=mode, cache_type="none", **kwargs)
+
+        async def cor():
+            r = await self.session.get(self.fs.encode_url(url), **kwargs).__aenter__()
+            self.fs._raise_not_found_for_status(r, url)
+            return r
+
+        self.r = sync(self.loop, cor)
+        self.loop = fs.loop
+
+    def seek(self, loc, whence=0):
+        if loc == 0 and whence == 1:
+            return
+        if loc == self.loc and whence == 0:
+            return
+        raise ValueError("Cannot seek streaming HTTP file")
+
+    async def _read(self, num=-1):
+        out = await self.r.content.read(num)
+        self.loc += len(out)
+        return out
+
+    read = sync_wrapper(_read)
+
+    async def _close(self):
+        self.r.close()
+
+    def close(self):
+        asyncio.run_coroutine_threadsafe(self._close(), self.loop)
+        super().close()
+
+
+class AsyncStreamFile(AbstractAsyncStreamedFile):
+    def __init__(
+        self, fs, url, mode="rb", loop=None, session=None, size=None, **kwargs
+    ):
+        self.url = url
+        self.session = session
+        self.r = None
+        if mode != "rb":
+            raise ValueError
+        self.details = {"name": url, "size": None}
+        self.kwargs = kwargs
+        super().__init__(fs=fs, path=url, mode=mode, cache_type="none")
+        self.size = size
+
+    async def read(self, num=-1):
+        if self.r is None:
+            r = await self.session.get(
+                self.fs.encode_url(self.url), **self.kwargs
+            ).__aenter__()
+            self.fs._raise_not_found_for_status(r, self.url)
+            self.r = r
+        out = await self.r.content.read(num)
+        self.loc += len(out)
+        return out
+
+    async def close(self):
+        if self.r is not None:
+            self.r.close()
+            self.r = None
+        await super().close()
+
+
+async def get_range(session, url, start, end, file=None, **kwargs):
+    # explicit get a range when we know it must be safe
+    kwargs = kwargs.copy()
+    headers = kwargs.pop("headers", {}).copy()
+    headers["Range"] = f"bytes={start}-{end - 1}"
+    r = await session.get(url, headers=headers, **kwargs)
+    r.raise_for_status()
+    async with r:
+        out = await r.read()
+    if file:
+        with open(file, "r+b") as f:  # noqa: ASYNC230
+            f.seek(start)
+            f.write(out)
+    else:
+        return out
+
+
+async def _file_info(url, session, size_policy="head", **kwargs):
+    """Call HEAD on the server to get details about the file (size/checksum etc.)
+
+    Default operation is to explicitly allow redirects and use encoding
+    'identity' (no compression) to get the true size of the target.
+    """
+    logger.debug("Retrieve file size for %s", url)
+    kwargs = kwargs.copy()
+    ar = kwargs.pop("allow_redirects", True)
+    head = kwargs.get("headers", {}).copy()
+    head["Accept-Encoding"] = "identity"
+    kwargs["headers"] = head
+
+    info = {}
+    if size_policy == "head":
+        r = await session.head(url, allow_redirects=ar, **kwargs)
+    elif size_policy == "get":
+        r = await session.get(url, allow_redirects=ar, **kwargs)
+    else:
+        raise TypeError(f'size_policy must be "head" or "get", got {size_policy}')
+    async with r:
+        r.raise_for_status()
+
+        if "Content-Length" in r.headers:
+            # Some servers may choose to ignore Accept-Encoding and return
+            # compressed content, in which case the returned size is unreliable.
+            if "Content-Encoding" not in r.headers or r.headers["Content-Encoding"] in [
+                "identity",
+                "",
+            ]:
+                info["size"] = int(r.headers["Content-Length"])
+        elif "Content-Range" in r.headers:
+            info["size"] = int(r.headers["Content-Range"].split("/")[1])
+
+        if "Content-Type" in r.headers:
+            info["mimetype"] = r.headers["Content-Type"].partition(";")[0]
+
+        if r.headers.get("Accept-Ranges") == "none":
+            # Some servers may explicitly discourage partial content requests, but
+            # the lack of "Accept-Ranges" does not always indicate they would fail
+            info["partial"] = False
+
+        info["url"] = str(r.url)
+
+        for checksum_field in ["ETag", "Content-MD5", "Digest", "Last-Modified"]:
+            if r.headers.get(checksum_field):
+                info[checksum_field] = r.headers[checksum_field]
+
+    return info
+
+
+async def _file_size(url, session=None, *args, **kwargs):
+    if session is None:
+        session = await get_client()
+    info = await _file_info(url, session=session, *args, **kwargs)
+    return info.get("size")
+
+
+file_size = sync_wrapper(_file_size)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/fsspec/implementations/http_sync.py

@@ -0,0 +1,937 @@
+"""This file is largely copied from http.py"""
+
+import io
+import logging
+import re
+import urllib.error
+import urllib.parse
+from copy import copy
+from json import dumps, loads
+from urllib.parse import urlparse
+
+try:
+    import yarl
+except (ImportError, ModuleNotFoundError, OSError):
+    yarl = False
+
+from fsspec.callbacks import _DEFAULT_CALLBACK
+from fsspec.registry import register_implementation
+from fsspec.spec import AbstractBufferedFile, AbstractFileSystem
+from fsspec.utils import DEFAULT_BLOCK_SIZE, isfilelike, nullcontext, tokenize
+
+from ..caching import AllBytes
+
+# https://stackoverflow.com/a/15926317/3821154
+ex = re.compile(r"""<(a|A)\s+(?:[^>]*?\s+)?(href|HREF)=["'](?P<url>[^"']+)""")
+ex2 = re.compile(r"""(?P<url>http[s]?://[-a-zA-Z0-9@:%_+.~#?&/=]+)""")
+logger = logging.getLogger("fsspec.http")
+
+
+class JsHttpException(urllib.error.HTTPError): ...
+
+
+class StreamIO(io.BytesIO):
+    # fake class, so you can set attributes on it
+    # will eventually actually stream
+    ...
+
+
+class ResponseProxy:
+    """Looks like a requests response"""
+
+    def __init__(self, req, stream=False):
+        self.request = req
+        self.stream = stream
+        self._data = None
+        self._headers = None
+
+    @property
+    def raw(self):
+        if self._data is None:
+            b = self.request.response.to_bytes()
+            if self.stream:
+                self._data = StreamIO(b)
+            else:
+                self._data = b
+        return self._data
+
+    def close(self):
+        if hasattr(self, "_data"):
+            del self._data
+
+    @property
+    def headers(self):
+        if self._headers is None:
+            self._headers = dict(
+                [
+                    _.split(": ")
+                    for _ in self.request.getAllResponseHeaders().strip().split("\r\n")
+                ]
+            )
+        return self._headers
+
+    @property
+    def status_code(self):
+        return int(self.request.status)
+
+    def raise_for_status(self):
+        if not self.ok:
+            raise JsHttpException(
+                self.url, self.status_code, self.reason, self.headers, None
+            )
+
+    def iter_content(self, chunksize, *_, **__):
+        while True:
+            out = self.raw.read(chunksize)
+            if out:
+                yield out
+            else:
+                break
+
+    @property
+    def reason(self):
+        return self.request.statusText
+
+    @property
+    def ok(self):
+        return self.status_code < 400
+
+    @property
+    def url(self):
+        return self.request.response.responseURL
+
+    @property
+    def text(self):
+        # TODO: encoding from headers
+        return self.content.decode()
+
+    @property
+    def content(self):
+        self.stream = False
+        return self.raw
+
+    def json(self):
+        return loads(self.text)
+
+
+class RequestsSessionShim:
+    def __init__(self):
+        self.headers = {}
+
+    def request(
+        self,
+        method,
+        url,
+        params=None,
+        data=None,
+        headers=None,
+        cookies=None,
+        files=None,
+        auth=None,
+        timeout=None,
+        allow_redirects=None,
+        proxies=None,
+        hooks=None,
+        stream=None,
+        verify=None,
+        cert=None,
+        json=None,
+    ):
+        from js import Blob, XMLHttpRequest
+
+        logger.debug("JS request: %s %s", method, url)
+
+        if cert or verify or proxies or files or cookies or hooks:
+            raise NotImplementedError
+        if data and json:
+            raise ValueError("Use json= or data=, not both")
+        req = XMLHttpRequest.new()
+        extra = auth if auth else ()
+        if params:
+            url = f"{url}?{urllib.parse.urlencode(params)}"
+        req.open(method, url, False, *extra)
+        if timeout:
+            req.timeout = timeout
+        if headers:
+            for k, v in headers.items():
+                req.setRequestHeader(k, v)
+
+        req.setRequestHeader("Accept", "application/octet-stream")
+        req.responseType = "arraybuffer"
+        if json:
+            blob = Blob.new([dumps(data)], {type: "application/json"})
+            req.send(blob)
+        elif data:
+            if isinstance(data, io.IOBase):
+                data = data.read()
+            blob = Blob.new([data], {type: "application/octet-stream"})
+            req.send(blob)
+        else:
+            req.send(None)
+        return ResponseProxy(req, stream=stream)
+
+    def get(self, url, **kwargs):
+        return self.request("GET", url, **kwargs)
+
+    def head(self, url, **kwargs):
+        return self.request("HEAD", url, **kwargs)
+
+    def post(self, url, **kwargs):
+        return self.request("POST}", url, **kwargs)
+
+    def put(self, url, **kwargs):
+        return self.request("PUT", url, **kwargs)
+
+    def patch(self, url, **kwargs):
+        return self.request("PATCH", url, **kwargs)
+
+    def delete(self, url, **kwargs):
+        return self.request("DELETE", url, **kwargs)
+
+
+class HTTPFileSystem(AbstractFileSystem):
+    """
+    Simple File-System for fetching data via HTTP(S)
+
+    This is the BLOCKING version of the normal HTTPFileSystem. It uses
+    requests in normal python and the JS runtime in pyodide.
+
+    ***This implementation is extremely experimental, do not use unless
+    you are testing pyodide/pyscript integration***
+    """
+
+    protocol = ("http", "https", "sync-http", "sync-https")
+    sep = "/"
+
+    def __init__(
+        self,
+        simple_links=True,
+        block_size=None,
+        same_scheme=True,
+        cache_type="readahead",
+        cache_options=None,
+        client_kwargs=None,
+        encoded=False,
+        **storage_options,
+    ):
+        """
+
+        Parameters
+        ----------
+        block_size: int
+            Blocks to read bytes; if 0, will default to raw requests file-like
+            objects instead of HTTPFile instances
+        simple_links: bool
+            If True, will consider both HTML <a> tags and anything that looks
+            like a URL; if False, will consider only the former.
+        same_scheme: True
+            When doing ls/glob, if this is True, only consider paths that have
+            http/https matching the input URLs.
+        size_policy: this argument is deprecated
+        client_kwargs: dict
+            Passed to aiohttp.ClientSession, see
+            https://docs.aiohttp.org/en/stable/client_reference.html
+            For example, ``{'auth': aiohttp.BasicAuth('user', 'pass')}``
+        storage_options: key-value
+            Any other parameters passed on to requests
+        cache_type, cache_options: defaults used in open
+        """
+        super().__init__(self, **storage_options)
+        self.block_size = block_size if block_size is not None else DEFAULT_BLOCK_SIZE
+        self.simple_links = simple_links
+        self.same_schema = same_scheme
+        self.cache_type = cache_type
+        self.cache_options = cache_options
+        self.client_kwargs = client_kwargs or {}
+        self.encoded = encoded
+        self.kwargs = storage_options
+
+        try:
+            import js  # noqa: F401
+
+            logger.debug("Starting JS session")
+            self.session = RequestsSessionShim()
+            self.js = True
+        except Exception as e:
+            import requests
+
+            logger.debug("Starting cpython session because of: %s", e)
+            self.session = requests.Session(**(client_kwargs or {}))
+            self.js = False
+
+        request_options = copy(storage_options)
+        self.use_listings_cache = request_options.pop("use_listings_cache", False)
+        request_options.pop("listings_expiry_time", None)
+        request_options.pop("max_paths", None)
+        request_options.pop("skip_instance_cache", None)
+        self.kwargs = request_options
+
+    @property
+    def fsid(self):
+        return "sync-http"
+
+    def encode_url(self, url):
+        if yarl:
+            return yarl.URL(url, encoded=self.encoded)
+        return url
+
+    @classmethod
+    def _strip_protocol(cls, path: str) -> str:
+        """For HTTP, we always want to keep the full URL"""
+        path = path.replace("sync-http://", "http://").replace(
+            "sync-https://", "https://"
+        )
+        return path
+
+    @classmethod
+    def _parent(cls, path):
+        # override, since _strip_protocol is different for URLs
+        par = super()._parent(path)
+        if len(par) > 7:  # "http://..."
+            return par
+        return ""
+
+    def _ls_real(self, url, detail=True, **kwargs):
+        # ignoring URL-encoded arguments
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        logger.debug(url)
+        r = self.session.get(self.encode_url(url), **self.kwargs)
+        self._raise_not_found_for_status(r, url)
+        text = r.text
+        if self.simple_links:
+            links = ex2.findall(text) + [u[2] for u in ex.findall(text)]
+        else:
+            links = [u[2] for u in ex.findall(text)]
+        out = set()
+        parts = urlparse(url)
+        for l in links:
+            if isinstance(l, tuple):
+                l = l[1]
+            if l.startswith("/") and len(l) > 1:
+                # absolute URL on this server
+                l = parts.scheme + "://" + parts.netloc + l
+            if l.startswith("http"):
+                if self.same_schema and l.startswith(url.rstrip("/") + "/"):
+                    out.add(l)
+                elif l.replace("https", "http").startswith(
+                    url.replace("https", "http").rstrip("/") + "/"
+                ):
+                    # allowed to cross http <-> https
+                    out.add(l)
+            else:
+                if l not in ["..", "../"]:
+                    # Ignore FTP-like "parent"
+                    out.add("/".join([url.rstrip("/"), l.lstrip("/")]))
+        if not out and url.endswith("/"):
+            out = self._ls_real(url.rstrip("/"), detail=False)
+        if detail:
+            return [
+                {
+                    "name": u,
+                    "size": None,
+                    "type": "directory" if u.endswith("/") else "file",
+                }
+                for u in out
+            ]
+        else:
+            return sorted(out)
+
+    def ls(self, url, detail=True, **kwargs):
+        if self.use_listings_cache and url in self.dircache:
+            out = self.dircache[url]
+        else:
+            out = self._ls_real(url, detail=detail, **kwargs)
+            self.dircache[url] = out
+        return out
+
+    def _raise_not_found_for_status(self, response, url):
+        """
+        Raises FileNotFoundError for 404s, otherwise uses raise_for_status.
+        """
+        if response.status_code == 404:
+            raise FileNotFoundError(url)
+        response.raise_for_status()
+
+    def cat_file(self, url, start=None, end=None, **kwargs):
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        logger.debug(url)
+
+        if start is not None or end is not None:
+            if start == end:
+                return b""
+            headers = kw.pop("headers", {}).copy()
+
+            headers["Range"] = self._process_limits(url, start, end)
+            kw["headers"] = headers
+        r = self.session.get(self.encode_url(url), **kw)
+        self._raise_not_found_for_status(r, url)
+        return r.content
+
+    def get_file(
+        self, rpath, lpath, chunk_size=5 * 2**20, callback=_DEFAULT_CALLBACK, **kwargs
+    ):
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        logger.debug(rpath)
+        r = self.session.get(self.encode_url(rpath), **kw)
+        try:
+            size = int(
+                r.headers.get("content-length", None)
+                or r.headers.get("Content-Length", None)
+            )
+        except (ValueError, KeyError, TypeError):
+            size = None
+
+        callback.set_size(size)
+        self._raise_not_found_for_status(r, rpath)
+        if not isfilelike(lpath):
+            lpath = open(lpath, "wb")
+        for chunk in r.iter_content(chunk_size, decode_unicode=False):
+            lpath.write(chunk)
+            callback.relative_update(len(chunk))
+
+    def put_file(
+        self,
+        lpath,
+        rpath,
+        chunk_size=5 * 2**20,
+        callback=_DEFAULT_CALLBACK,
+        method="post",
+        **kwargs,
+    ):
+        def gen_chunks():
+            # Support passing arbitrary file-like objects
+            # and use them instead of streams.
+            if isinstance(lpath, io.IOBase):
+                context = nullcontext(lpath)
+                use_seek = False  # might not support seeking
+            else:
+                context = open(lpath, "rb")
+                use_seek = True
+
+            with context as f:
+                if use_seek:
+                    callback.set_size(f.seek(0, 2))
+                    f.seek(0)
+                else:
+                    callback.set_size(getattr(f, "size", None))
+
+                chunk = f.read(chunk_size)
+                while chunk:
+                    yield chunk
+                    callback.relative_update(len(chunk))
+                    chunk = f.read(chunk_size)
+
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+
+        method = method.lower()
+        if method not in ("post", "put"):
+            raise ValueError(
+                f"method has to be either 'post' or 'put', not: {method!r}"
+            )
+
+        meth = getattr(self.session, method)
+        resp = meth(rpath, data=gen_chunks(), **kw)
+        self._raise_not_found_for_status(resp, rpath)
+
+    def _process_limits(self, url, start, end):
+        """Helper for "Range"-based _cat_file"""
+        size = None
+        suff = False
+        if start is not None and start < 0:
+            # if start is negative and end None, end is the "suffix length"
+            if end is None:
+                end = -start
+                start = ""
+                suff = True
+            else:
+                size = size or self.info(url)["size"]
+                start = size + start
+        elif start is None:
+            start = 0
+        if not suff:
+            if end is not None and end < 0:
+                if start is not None:
+                    size = size or self.info(url)["size"]
+                    end = size + end
+            elif end is None:
+                end = ""
+            if isinstance(end, int):
+                end -= 1  # bytes range is inclusive
+        return f"bytes={start}-{end}"
+
+    def exists(self, path, strict=False, **kwargs):
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        try:
+            logger.debug(path)
+            r = self.session.get(self.encode_url(path), **kw)
+            if strict:
+                self._raise_not_found_for_status(r, path)
+            return r.status_code < 400
+        except FileNotFoundError:
+            return False
+        except Exception:
+            if strict:
+                raise
+            return False
+
+    def isfile(self, path, **kwargs):
+        return self.exists(path, **kwargs)
+
+    def _open(
+        self,
+        path,
+        mode="rb",
+        block_size=None,
+        autocommit=None,  # XXX: This differs from the base class.
+        cache_type=None,
+        cache_options=None,
+        size=None,
+        **kwargs,
+    ):
+        """Make a file-like object
+
+        Parameters
+        ----------
+        path: str
+            Full URL with protocol
+        mode: string
+            must be "rb"
+        block_size: int or None
+            Bytes to download in one request; use instance value if None. If
+            zero, will return a streaming Requests file-like instance.
+        kwargs: key-value
+            Any other parameters, passed to requests calls
+        """
+        if mode != "rb":
+            raise NotImplementedError
+        block_size = block_size if block_size is not None else self.block_size
+        kw = self.kwargs.copy()
+        kw.update(kwargs)
+        size = size or self.info(path, **kwargs)["size"]
+        if block_size and size:
+            return HTTPFile(
+                self,
+                path,
+                session=self.session,
+                block_size=block_size,
+                mode=mode,
+                size=size,
+                cache_type=cache_type or self.cache_type,
+                cache_options=cache_options or self.cache_options,
+                **kw,
+            )
+        else:
+            return HTTPStreamFile(
+                self,
+                path,
+                mode=mode,
+                session=self.session,
+                **kw,
+            )
+
+    def ukey(self, url):
+        """Unique identifier; assume HTTP files are static, unchanging"""
+        return tokenize(url, self.kwargs, self.protocol)
+
+    def info(self, url, **kwargs):
+        """Get info of URL
+
+        Tries to access location via HEAD, and then GET methods, but does
+        not fetch the data.
+
+        It is possible that the server does not supply any size information, in
+        which case size will be given as None (and certain operations on the
+        corresponding file will not work).
+        """
+        info = {}
+        for policy in ["head", "get"]:
+            try:
+                info.update(
+                    _file_info(
+                        self.encode_url(url),
+                        size_policy=policy,
+                        session=self.session,
+                        **self.kwargs,
+                        **kwargs,
+                    )
+                )
+                if info.get("size") is not None:
+                    break
+            except Exception as exc:
+                if policy == "get":
+                    # If get failed, then raise a FileNotFoundError
+                    raise FileNotFoundError(url) from exc
+                logger.debug(str(exc))
+
+        return {"name": url, "size": None, **info, "type": "file"}
+
+    def glob(self, path, maxdepth=None, **kwargs):
+        """
+        Find files by glob-matching.
+
+        This implementation is idntical to the one in AbstractFileSystem,
+        but "?" is not considered as a character for globbing, because it is
+        so common in URLs, often identifying the "query" part.
+        """
+        import re
+
+        ends = path.endswith("/")
+        path = self._strip_protocol(path)
+        indstar = path.find("*") if path.find("*") >= 0 else len(path)
+        indbrace = path.find("[") if path.find("[") >= 0 else len(path)
+
+        ind = min(indstar, indbrace)
+
+        detail = kwargs.pop("detail", False)
+
+        if not has_magic(path):
+            root = path
+            depth = 1
+            if ends:
+                path += "/*"
+            elif self.exists(path):
+                if not detail:
+                    return [path]
+                else:
+                    return {path: self.info(path)}
+            else:
+                if not detail:
+                    return []  # glob of non-existent returns empty
+                else:
+                    return {}
+        elif "/" in path[:ind]:
+            ind2 = path[:ind].rindex("/")
+            root = path[: ind2 + 1]
+            depth = None if "**" in path else path[ind2 + 1 :].count("/") + 1
+        else:
+            root = ""
+            depth = None if "**" in path else path[ind + 1 :].count("/") + 1
+
+        allpaths = self.find(
+            root, maxdepth=maxdepth or depth, withdirs=True, detail=True, **kwargs
+        )
+        # Escape characters special to python regex, leaving our supported
+        # special characters in place.
+        # See https://www.gnu.org/software/bash/manual/html_node/Pattern-Matching.html
+        # for shell globbing details.
+        pattern = (
+            "^"
+            + (
+                path.replace("\\", r"\\")
+                .replace(".", r"\.")
+                .replace("+", r"\+")
+                .replace("//", "/")
+                .replace("(", r"\(")
+                .replace(")", r"\)")
+                .replace("|", r"\|")
+                .replace("^", r"\^")
+                .replace("$", r"\$")
+                .replace("{", r"\{")
+                .replace("}", r"\}")
+                .rstrip("/")
+            )
+            + "$"
+        )
+        pattern = re.sub("[*]{2}", "=PLACEHOLDER=", pattern)
+        pattern = re.sub("[*]", "[^/]*", pattern)
+        pattern = re.compile(pattern.replace("=PLACEHOLDER=", ".*"))
+        out = {
+            p: allpaths[p]
+            for p in sorted(allpaths)
+            if pattern.match(p.replace("//", "/").rstrip("/"))
+        }
+        if detail:
+            return out
+        else:
+            return list(out)
+
+    def isdir(self, path):
+        # override, since all URLs are (also) files
+        try:
+            return bool(self.ls(path))
+        except (FileNotFoundError, ValueError):
+            return False
+
+
+class HTTPFile(AbstractBufferedFile):
+    """
+    A file-like object pointing to a remove HTTP(S) resource
+
+    Supports only reading, with read-ahead of a predermined block-size.
+
+    In the case that the server does not supply the filesize, only reading of
+    the complete file in one go is supported.
+
+    Parameters
+    ----------
+    url: str
+        Full URL of the remote resource, including the protocol
+    session: requests.Session or None
+        All calls will be made within this session, to avoid restarting
+        connections where the server allows this
+    block_size: int or None
+        The amount of read-ahead to do, in bytes. Default is 5MB, or the value
+        configured for the FileSystem creating this file
+    size: None or int
+        If given, this is the size of the file in bytes, and we don't attempt
+        to call the server to find the value.
+    kwargs: all other key-values are passed to requests calls.
+    """
+
+    def __init__(
+        self,
+        fs,
+        url,
+        session=None,
+        block_size=None,
+        mode="rb",
+        cache_type="bytes",
+        cache_options=None,
+        size=None,
+        **kwargs,
+    ):
+        if mode != "rb":
+            raise NotImplementedError("File mode not supported")
+        self.url = url
+        self.session = session
+        self.details = {"name": url, "size": size, "type": "file"}
+        super().__init__(
+            fs=fs,
+            path=url,
+            mode=mode,
+            block_size=block_size,
+            cache_type=cache_type,
+            cache_options=cache_options,
+            **kwargs,
+        )
+
+    def read(self, length=-1):
+        """Read bytes from file
+
+        Parameters
+        ----------
+        length: int
+            Read up to this many bytes. If negative, read all content to end of
+            file. If the server has not supplied the filesize, attempting to
+            read only part of the data will raise a ValueError.
+        """
+        if (
+            (length < 0 and self.loc == 0)  # explicit read all
+            # but not when the size is known and fits into a block anyways
+            and not (self.size is not None and self.size <= self.blocksize)
+        ):
+            self._fetch_all()
+        if self.size is None:
+            if length < 0:
+                self._fetch_all()
+        else:
+            length = min(self.size - self.loc, length)
+        return super().read(length)
+
+    def _fetch_all(self):
+        """Read whole file in one shot, without caching
+
+        This is only called when position is still at zero,
+        and read() is called without a byte-count.
+        """
+        logger.debug(f"Fetch all for {self}")
+        if not isinstance(self.cache, AllBytes):
+            r = self.session.get(self.fs.encode_url(self.url), **self.kwargs)
+            r.raise_for_status()
+            out = r.content
+            self.cache = AllBytes(size=len(out), fetcher=None, blocksize=None, data=out)
+            self.size = len(out)
+
+    def _parse_content_range(self, headers):
+        """Parse the Content-Range header"""
+        s = headers.get("Content-Range", "")
+        m = re.match(r"bytes (\d+-\d+|\*)/(\d+|\*)", s)
+        if not m:
+            return None, None, None
+
+        if m[1] == "*":
+            start = end = None
+        else:
+            start, end = [int(x) for x in m[1].split("-")]
+        total = None if m[2] == "*" else int(m[2])
+        return start, end, total
+
+    def _fetch_range(self, start, end):
+        """Download a block of data
+
+        The expectation is that the server returns only the requested bytes,
+        with HTTP code 206. If this is not the case, we first check the headers,
+        and then stream the output - if the data size is bigger than we
+        requested, an exception is raised.
+        """
+        logger.debug(f"Fetch range for {self}: {start}-{end}")
+        kwargs = self.kwargs.copy()
+        headers = kwargs.pop("headers", {}).copy()
+        headers["Range"] = f"bytes={start}-{end - 1}"
+        logger.debug("%s : %s", self.url, headers["Range"])
+        r = self.session.get(self.fs.encode_url(self.url), headers=headers, **kwargs)
+        if r.status_code == 416:
+            # range request outside file
+            return b""
+        r.raise_for_status()
+
+        # If the server has handled the range request, it should reply
+        # with status 206 (partial content). But we'll guess that a suitable
+        # Content-Range header or a Content-Length no more than the
+        # requested range also mean we have got the desired range.
+        cl = r.headers.get("Content-Length", r.headers.get("content-length", end + 1))
+        response_is_range = (
+            r.status_code == 206
+            or self._parse_content_range(r.headers)[0] == start
+            or int(cl) <= end - start
+        )
+
+        if response_is_range:
+            # partial content, as expected
+            out = r.content
+        elif start > 0:
+            raise ValueError(
+                "The HTTP server doesn't appear to support range requests. "
+                "Only reading this file from the beginning is supported. "
+                "Open with block_size=0 for a streaming file interface."
+            )
+        else:
+            # Response is not a range, but we want the start of the file,
+            # so we can read the required amount anyway.
+            cl = 0
+            out = []
+            for chunk in r.iter_content(2**20, False):
+                out.append(chunk)
+                cl += len(chunk)
+            out = b"".join(out)[: end - start]
+        return out
+
+
+magic_check = re.compile("([*[])")
+
+
+def has_magic(s):
+    match = magic_check.search(s)
+    return match is not None
+
+
+class HTTPStreamFile(AbstractBufferedFile):
+    def __init__(self, fs, url, mode="rb", session=None, **kwargs):
+        self.url = url
+        self.session = session
+        if mode != "rb":
+            raise ValueError
+        self.details = {"name": url, "size": None}
+        super().__init__(fs=fs, path=url, mode=mode, cache_type="readahead", **kwargs)
+
+        r = self.session.get(self.fs.encode_url(url), stream=True, **kwargs)
+        self.fs._raise_not_found_for_status(r, url)
+        self.it = r.iter_content(1024, False)
+        self.leftover = b""
+
+        self.r = r
+
+    def seek(self, *args, **kwargs):
+        raise ValueError("Cannot seek streaming HTTP file")
+
+    def read(self, num=-1):
+        bufs = [self.leftover]
+        leng = len(self.leftover)
+        while leng < num or num < 0:
+            try:
+                out = self.it.__next__()
+            except StopIteration:
+                break
+            if out:
+                bufs.append(out)
+            else:
+                break
+            leng += len(out)
+        out = b"".join(bufs)
+        if num >= 0:
+            self.leftover = out[num:]
+            out = out[:num]
+        else:
+            self.leftover = b""
+        self.loc += len(out)
+        return out
+
+    def close(self):
+        self.r.close()
+        self.closed = True
+
+
+def get_range(session, url, start, end, **kwargs):
+    # explicit get a range when we know it must be safe
+    kwargs = kwargs.copy()
+    headers = kwargs.pop("headers", {}).copy()
+    headers["Range"] = f"bytes={start}-{end - 1}"
+    r = session.get(url, headers=headers, **kwargs)
+    r.raise_for_status()
+    return r.content
+
+
+def _file_info(url, session, size_policy="head", **kwargs):
+    """Call HEAD on the server to get details about the file (size/checksum etc.)
+
+    Default operation is to explicitly allow redirects and use encoding
+    'identity' (no compression) to get the true size of the target.
+    """
+    logger.debug("Retrieve file size for %s", url)
+    kwargs = kwargs.copy()
+    ar = kwargs.pop("allow_redirects", True)
+    head = kwargs.get("headers", {}).copy()
+    # TODO: not allowed in JS
+    # head["Accept-Encoding"] = "identity"
+    kwargs["headers"] = head
+
+    info = {}
+    if size_policy == "head":
+        r = session.head(url, allow_redirects=ar, **kwargs)
+    elif size_policy == "get":
+        r = session.get(url, allow_redirects=ar, **kwargs)
+    else:
+        raise TypeError(f'size_policy must be "head" or "get", got {size_policy}')
+    r.raise_for_status()
+
+    # TODO:
+    #  recognise lack of 'Accept-Ranges',
+    #                 or 'Accept-Ranges': 'none' (not 'bytes')
+    #  to mean streaming only, no random access => return None
+    if "Content-Length" in r.headers:
+        info["size"] = int(r.headers["Content-Length"])
+    elif "Content-Range" in r.headers:
+        info["size"] = int(r.headers["Content-Range"].split("/")[1])
+    elif "content-length" in r.headers:
+        info["size"] = int(r.headers["content-length"])
+    elif "content-range" in r.headers:
+        info["size"] = int(r.headers["content-range"].split("/")[1])
+
+    for checksum_field in ["ETag", "Content-MD5", "Digest"]:
+        if r.headers.get(checksum_field):
+            info[checksum_field] = r.headers[checksum_field]
+
+    return info
+
+
+# importing this is enough to register it
+def register():
+    register_implementation("http", HTTPFileSystem, clobber=True)
+    register_implementation("https", HTTPFileSystem, clobber=True)
+    register_implementation("sync-http", HTTPFileSystem, clobber=True)
+    register_implementation("sync-https", HTTPFileSystem, clobber=True)
+
+
+register()
+
+
+def unregister():
+    from fsspec.implementations.http import HTTPFileSystem
+
+    register_implementation("http", HTTPFileSystem, clobber=True)
+    register_implementation("https", HTTPFileSystem, clobber=True)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/fsspec/implementations/memory.py

@@ -0,0 +1,311 @@
+from __future__ import annotations
+
+import logging
+from datetime import datetime, timezone
+from errno import ENOTEMPTY
+from io import BytesIO
+from pathlib import PurePath, PureWindowsPath
+from typing import Any, ClassVar
+
+from fsspec import AbstractFileSystem
+from fsspec.implementations.local import LocalFileSystem
+from fsspec.utils import stringify_path
+
+logger = logging.getLogger("fsspec.memoryfs")
+
+
+class MemoryFileSystem(AbstractFileSystem):
+    """A filesystem based on a dict of BytesIO objects
+
+    This is a global filesystem so instances of this class all point to the same
+    in memory filesystem.
+    """
+
+    store: ClassVar[dict[str, Any]] = {}  # global, do not overwrite!
+    pseudo_dirs = [""]  # global, do not overwrite!
+    protocol = "memory"
+    root_marker = "/"
+
+    @classmethod
+    def _strip_protocol(cls, path):
+        if isinstance(path, PurePath):
+            if isinstance(path, PureWindowsPath):
+                return LocalFileSystem._strip_protocol(path)
+            else:
+                path = stringify_path(path)
+
+        path = path.removeprefix("memory://")
+        if "::" in path or "://" in path:
+            return path.rstrip("/")
+        path = path.lstrip("/").rstrip("/")
+        return "/" + path if path else ""
+
+    def ls(self, path, detail=True, **kwargs):
+        path = self._strip_protocol(path)
+        if path in self.store:
+            # there is a key with this exact name
+            if not detail:
+                return [path]
+            return [
+                {
+                    "name": path,
+                    "size": self.store[path].size,
+                    "type": "file",
+                    "created": self.store[path].created.timestamp(),
+                }
+            ]
+        paths = set()
+        starter = path + "/"
+        out = []
+        for p2 in tuple(self.store):
+            if p2.startswith(starter):
+                if "/" not in p2[len(starter) :]:
+                    # exact child
+                    out.append(
+                        {
+                            "name": p2,
+                            "size": self.store[p2].size,
+                            "type": "file",
+                            "created": self.store[p2].created.timestamp(),
+                        }
+                    )
+                elif len(p2) > len(starter):
+                    # implied child directory
+                    ppath = starter + p2[len(starter) :].split("/", 1)[0]
+                    if ppath not in paths:
+                        out = out or []
+                        out.append(
+                            {
+                                "name": ppath,
+                                "size": 0,
+                                "type": "directory",
+                            }
+                        )
+                        paths.add(ppath)
+        for p2 in self.pseudo_dirs:
+            if p2.startswith(starter):
+                if "/" not in p2[len(starter) :]:
+                    # exact child pdir
+                    if p2 not in paths:
+                        out.append({"name": p2, "size": 0, "type": "directory"})
+                        paths.add(p2)
+                else:
+                    # directory implied by deeper pdir
+                    ppath = starter + p2[len(starter) :].split("/", 1)[0]
+                    if ppath not in paths:
+                        out.append({"name": ppath, "size": 0, "type": "directory"})
+                        paths.add(ppath)
+        if not out:
+            if path in self.pseudo_dirs:
+                # empty dir
+                return []
+            raise FileNotFoundError(path)
+        if detail:
+            return out
+        return sorted([f["name"] for f in out])
+
+    def mkdir(self, path, create_parents=True, **kwargs):
+        path = self._strip_protocol(path)
+        if path in self.store or path in self.pseudo_dirs:
+            raise FileExistsError(path)
+        if self._parent(path).strip("/") and self.isfile(self._parent(path)):
+            raise NotADirectoryError(self._parent(path))
+        if create_parents and self._parent(path).strip("/"):
+            try:
+                self.mkdir(self._parent(path), create_parents, **kwargs)
+            except FileExistsError:
+                pass
+        if path and path not in self.pseudo_dirs:
+            self.pseudo_dirs.append(path)
+
+    def makedirs(self, path, exist_ok=False):
+        try:
+            self.mkdir(path, create_parents=True)
+        except FileExistsError:
+            if not exist_ok:
+                raise
+
+    def pipe_file(self, path, value, mode="overwrite", **kwargs):
+        """Set the bytes of given file
+
+        Avoids copies of the data if possible
+        """
+        mode = "xb" if mode == "create" else "wb"
+        self.open(path, mode=mode, data=value)
+
+    def rmdir(self, path):
+        path = self._strip_protocol(path)
+        if path == "":
+            # silently avoid deleting FS root
+            return
+        if path in self.pseudo_dirs:
+            if not self.ls(path):
+                self.pseudo_dirs.remove(path)
+            else:
+                raise OSError(ENOTEMPTY, "Directory not empty", path)
+        else:
+            raise FileNotFoundError(path)
+
+    def info(self, path, **kwargs):
+        logger.debug("info: %s", path)
+        path = self._strip_protocol(path)
+        if path in self.pseudo_dirs or any(
+            p.startswith(path + "/") for p in list(self.store) + self.pseudo_dirs
+        ):
+            return {
+                "name": path,
+                "size": 0,
+                "type": "directory",
+            }
+        elif path in self.store:
+            filelike = self.store[path]
+            return {
+                "name": path,
+                "size": filelike.size,
+                "type": "file",
+                "created": getattr(filelike, "created", None),
+            }
+        else:
+            raise FileNotFoundError(path)
+
+    def _open(
+        self,
+        path,
+        mode="rb",
+        block_size=None,
+        autocommit=True,
+        cache_options=None,
+        **kwargs,
+    ):
+        path = self._strip_protocol(path)
+        if "x" in mode and self.exists(path):
+            raise FileExistsError
+        if path in self.pseudo_dirs:
+            raise IsADirectoryError(path)
+        parent = path
+        while len(parent) > 1:
+            parent = self._parent(parent)
+            if self.isfile(parent):
+                raise FileExistsError(parent)
+        if mode in ["rb", "ab", "r+b", "a+b"]:
+            if path in self.store:
+                f = self.store[path]
+                if "a" in mode:
+                    # position at the end of file
+                    f.seek(0, 2)
+                else:
+                    # position at the beginning of file
+                    f.seek(0)
+                return f
+            else:
+                raise FileNotFoundError(path)
+        elif mode in {"wb", "w+b", "xb", "x+b"}:
+            if "x" in mode and self.exists(path):
+                raise FileExistsError
+            m = MemoryFile(self, path, kwargs.get("data"))
+            if not self._intrans:
+                m.commit()
+            return m
+        else:
+            name = self.__class__.__name__
+            raise ValueError(f"unsupported file mode for {name}: {mode!r}")
+
+    def cp_file(self, path1, path2, **kwargs):
+        path1 = self._strip_protocol(path1)
+        path2 = self._strip_protocol(path2)
+        if self.isfile(path1):
+            self.store[path2] = MemoryFile(
+                self, path2, self.store[path1].getvalue()
+            )  # implicit copy
+        elif self.isdir(path1):
+            if path2 not in self.pseudo_dirs:
+                self.pseudo_dirs.append(path2)
+        else:
+            raise FileNotFoundError(path1)
+
+    def cat_file(self, path, start=None, end=None, **kwargs):
+        logger.debug("cat: %s", path)
+        path = self._strip_protocol(path)
+        try:
+            return bytes(self.store[path].getbuffer()[start:end])
+        except KeyError as e:
+            raise FileNotFoundError(path) from e
+
+    def _rm(self, path):
+        path = self._strip_protocol(path)
+        try:
+            del self.store[path]
+        except KeyError as e:
+            raise FileNotFoundError(path) from e
+
+    def modified(self, path):
+        path = self._strip_protocol(path)
+        try:
+            return self.store[path].modified
+        except KeyError as e:
+            raise FileNotFoundError(path) from e
+
+    def created(self, path):
+        path = self._strip_protocol(path)
+        try:
+            return self.store[path].created
+        except KeyError as e:
+            raise FileNotFoundError(path) from e
+
+    def isfile(self, path):
+        path = self._strip_protocol(path)
+        return path in self.store
+
+    def rm(self, path, recursive=False, maxdepth=None):
+        if isinstance(path, str):
+            path = self._strip_protocol(path)
+        else:
+            path = [self._strip_protocol(p) for p in path]
+        paths = self.expand_path(path, recursive=recursive, maxdepth=maxdepth)
+        for p in reversed(paths):
+            if self.isfile(p):
+                self.rm_file(p)
+            # If the expanded path doesn't exist, it is only because the expanded
+            # path was a directory that does not exist in self.pseudo_dirs. This
+            # is possible if you directly create files without making the
+            # directories first.
+            elif not self.exists(p):
+                continue
+            else:
+                self.rmdir(p)
+
+
+class MemoryFile(BytesIO):
+    """A BytesIO which can't close and works as a context manager
+
+    Can initialise with data. Each path should only be active once at any moment.
+
+    No need to provide fs, path if auto-committing (default)
+    """
+
+    def __init__(self, fs=None, path=None, data=None):
+        logger.debug("open file %s", path)
+        self.fs = fs
+        self.path = path
+        self.created = datetime.now(tz=timezone.utc)
+        self.modified = datetime.now(tz=timezone.utc)
+        if data:
+            super().__init__(data)
+            self.seek(0)
+
+    @property
+    def size(self):
+        return self.getbuffer().nbytes
+
+    def __enter__(self):
+        return self
+
+    def close(self):
+        pass
+
+    def discard(self):
+        pass
+
+    def commit(self):
+        self.fs.store[self.path] = self
+        self.modified = datetime.now(tz=timezone.utc)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/fsspec/implementations/zip.py

@@ -0,0 +1,183 @@
+import os
+import zipfile
+
+import fsspec
+from fsspec.archive import AbstractArchiveFileSystem
+
+
+class ZipFileSystem(AbstractArchiveFileSystem):
+    """Read/Write contents of ZIP archive as a file-system
+
+    Keeps file object open while instance lives.
+
+    This class is pickleable, but not necessarily thread-safe
+    """
+
+    root_marker = ""
+    protocol = "zip"
+    cachable = False
+
+    def __init__(
+        self,
+        fo="",
+        mode="r",
+        target_protocol=None,
+        target_options=None,
+        compression=zipfile.ZIP_STORED,
+        allowZip64=True,
+        compresslevel=None,
+        **kwargs,
+    ):
+        """
+        Parameters
+        ----------
+        fo: str or file-like
+            Contains ZIP, and must exist. If a str, will fetch file using
+            :meth:`~fsspec.open_files`, which must return one file exactly.
+        mode: str
+            Accept: "r", "w", "a"
+        target_protocol: str (optional)
+            If ``fo`` is a string, this value can be used to override the
+            FS protocol inferred from a URL
+        target_options: dict (optional)
+            Kwargs passed when instantiating the target FS, if ``fo`` is
+            a string.
+        compression, allowZip64, compresslevel: passed to ZipFile
+            Only relevant when creating a ZIP
+        """
+        super().__init__(self, **kwargs)
+        if mode not in set("rwa"):
+            raise ValueError(f"mode '{mode}' no understood")
+        self.mode = mode
+        if isinstance(fo, (str, os.PathLike)):
+            if mode == "a":
+                m = "r+b"
+            else:
+                m = mode + "b"
+            fo = fsspec.open(
+                fo, mode=m, protocol=target_protocol, **(target_options or {})
+            )
+        self.force_zip_64 = allowZip64
+        self.of = fo
+        self.fo = fo.__enter__()  # the whole instance is a context
+        self.zip = zipfile.ZipFile(
+            self.fo,
+            mode=mode,
+            compression=compression,
+            allowZip64=allowZip64,
+            compresslevel=compresslevel,
+        )
+        self.dir_cache = None
+
+    @classmethod
+    def _strip_protocol(cls, path):
+        # zip file paths are always relative to the archive root
+        return super()._strip_protocol(path).lstrip("/")
+
+    def __del__(self):
+        if hasattr(self, "zip"):
+            self.close()
+            del self.zip
+        if hasattr(self, "of") and hasattr(self.of, "__exit__"):
+            self.of.__exit__(None, None, None)
+
+    def close(self):
+        """Commits any write changes to the file. Done on ``del`` too."""
+        self.zip.close()
+
+    def _get_dirs(self):
+        if self.dir_cache is None or self.mode in set("wa"):
+            # when writing, dir_cache is always in the ZipFile's attributes,
+            # not read from the file.
+            files = self.zip.infolist()
+            self.dir_cache = {
+                dirname.rstrip("/"): {
+                    "name": dirname.rstrip("/"),
+                    "size": 0,
+                    "type": "directory",
+                }
+                for dirname in self._all_dirnames(self.zip.namelist())
+            }
+            for z in files:
+                f = {s: getattr(z, s, None) for s in zipfile.ZipInfo.__slots__}
+                f.update(
+                    {
+                        "name": z.filename.rstrip("/"),
+                        "size": z.file_size,
+                        "type": ("directory" if z.is_dir() else "file"),
+                    }
+                )
+                self.dir_cache[f["name"]] = f
+
+    def pipe_file(self, path, value, **kwargs):
+        # override upstream, because we know the exact file size in this case
+        self.zip.writestr(path, value, **kwargs)
+
+    def _open(
+        self,
+        path,
+        mode="rb",
+        block_size=None,
+        autocommit=True,
+        cache_options=None,
+        **kwargs,
+    ):
+        path = self._strip_protocol(path)
+        if "r" in mode and self.mode in set("wa"):
+            if self.exists(path):
+                raise OSError("ZipFS can only be open for reading or writing, not both")
+            raise FileNotFoundError(path)
+        if "r" in self.mode and "w" in mode:
+            raise OSError("ZipFS can only be open for reading or writing, not both")
+        out = self.zip.open(path, mode.strip("b"), force_zip64=self.force_zip_64)
+        if "r" in mode:
+            info = self.info(path)
+            out.size = info["size"]
+            out.name = info["name"]
+        return out
+
+    def find(self, path, maxdepth=None, withdirs=False, detail=False, **kwargs):
+        if maxdepth is not None and maxdepth < 1:
+            raise ValueError("maxdepth must be at least 1")
+
+        def to_parts(_path: str):
+            return list(filter(None, _path.replace("\\", "/").split("/")))
+
+        if not isinstance(path, str):
+            path = str(path)
+
+        # Remove the leading slash, as the zip file paths are always
+        # given without a leading slash
+        path = path.lstrip("/")
+        path_parts = to_parts(path)
+        path_depth = len(path_parts)
+
+        self._get_dirs()
+
+        result = {}
+        # To match posix find, if an exact file name is given, we should
+        # return only that file
+        if path in self.dir_cache and self.dir_cache[path]["type"] == "file":
+            result[path] = self.dir_cache[path]
+            return result if detail else [path]
+
+        for file_path, file_info in self.dir_cache.items():
+            if len(file_parts := to_parts(file_path)) < path_depth or any(
+                a != b for a, b in zip(path_parts, file_parts)
+            ):
+                # skip parent folders and mismatching paths
+                continue
+
+            if file_info["type"] == "directory":
+                if withdirs and file_path not in result:
+                    result[file_path.strip("/")] = file_info
+                continue
+
+            if file_path not in result:
+                result[file_path] = file_info if detail else None
+
+        if maxdepth:
+            result = {
+                k: v for k, v in result.items() if k.count("/") < maxdepth + path_depth
+            }
+        return result if detail else sorted(result)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/_version.py

@@ -0,0 +1,155 @@
+"""Utility to compare (NumPy) version strings.
+
+The NumpyVersion class allows properly comparing numpy version strings.
+The LooseVersion and StrictVersion classes that distutils provides don't
+work; they don't recognize anything like alpha/beta/rc/dev versions.
+
+"""
+import re
+
+
+__all__ = ['NumpyVersion']
+
+
+class NumpyVersion():
+    """Parse and compare numpy version strings.
+
+    NumPy has the following versioning scheme (numbers given are examples; they
+    can be > 9 in principle):
+
+    - Released version: '1.8.0', '1.8.1', etc.
+    - Alpha: '1.8.0a1', '1.8.0a2', etc.
+    - Beta: '1.8.0b1', '1.8.0b2', etc.
+    - Release candidates: '1.8.0rc1', '1.8.0rc2', etc.
+    - Development versions: '1.8.0.dev-f1234afa' (git commit hash appended)
+    - Development versions after a1: '1.8.0a1.dev-f1234afa',
+                                     '1.8.0b2.dev-f1234afa',
+                                     '1.8.1rc1.dev-f1234afa', etc.
+    - Development versions (no git hash available): '1.8.0.dev-Unknown'
+
+    Comparing needs to be done against a valid version string or other
+    `NumpyVersion` instance. Note that all development versions of the same
+    (pre-)release compare equal.
+
+    .. versionadded:: 1.9.0
+
+    Parameters
+    ----------
+    vstring : str
+        NumPy version string (``np.__version__``).
+
+    Examples
+    --------
+    >>> from numpy.lib import NumpyVersion
+    >>> if NumpyVersion(np.__version__) < '1.7.0':
+    ...     print('skip')
+    >>> # skip
+
+    >>> NumpyVersion('1.7')  # raises ValueError, add ".0"
+    Traceback (most recent call last):
+        ...
+    ValueError: Not a valid numpy version string
+
+    """
+
+    def __init__(self, vstring):
+        self.vstring = vstring
+        ver_main = re.match(r'\d+\.\d+\.\d+', vstring)
+        if not ver_main:
+            raise ValueError("Not a valid numpy version string")
+
+        self.version = ver_main.group()
+        self.major, self.minor, self.bugfix = [int(x) for x in
+            self.version.split('.')]
+        if len(vstring) == ver_main.end():
+            self.pre_release = 'final'
+        else:
+            alpha = re.match(r'a\d', vstring[ver_main.end():])
+            beta = re.match(r'b\d', vstring[ver_main.end():])
+            rc = re.match(r'rc\d', vstring[ver_main.end():])
+            pre_rel = [m for m in [alpha, beta, rc] if m is not None]
+            if pre_rel:
+                self.pre_release = pre_rel[0].group()
+            else:
+                self.pre_release = ''
+
+        self.is_devversion = bool(re.search(r'.dev', vstring))
+
+    def _compare_version(self, other):
+        """Compare major.minor.bugfix"""
+        if self.major == other.major:
+            if self.minor == other.minor:
+                if self.bugfix == other.bugfix:
+                    vercmp = 0
+                elif self.bugfix > other.bugfix:
+                    vercmp = 1
+                else:
+                    vercmp = -1
+            elif self.minor > other.minor:
+                vercmp = 1
+            else:
+                vercmp = -1
+        elif self.major > other.major:
+            vercmp = 1
+        else:
+            vercmp = -1
+
+        return vercmp
+
+    def _compare_pre_release(self, other):
+        """Compare alpha/beta/rc/final."""
+        if self.pre_release == other.pre_release:
+            vercmp = 0
+        elif self.pre_release == 'final':
+            vercmp = 1
+        elif other.pre_release == 'final':
+            vercmp = -1
+        elif self.pre_release > other.pre_release:
+            vercmp = 1
+        else:
+            vercmp = -1
+
+        return vercmp
+
+    def _compare(self, other):
+        if not isinstance(other, (str, NumpyVersion)):
+            raise ValueError("Invalid object to compare with NumpyVersion.")
+
+        if isinstance(other, str):
+            other = NumpyVersion(other)
+
+        vercmp = self._compare_version(other)
+        if vercmp == 0:
+            # Same x.y.z version, check for alpha/beta/rc
+            vercmp = self._compare_pre_release(other)
+            if vercmp == 0:
+                # Same version and same pre-release, check if dev version
+                if self.is_devversion is other.is_devversion:
+                    vercmp = 0
+                elif self.is_devversion:
+                    vercmp = -1
+                else:
+                    vercmp = 1
+
+        return vercmp
+
+    def __lt__(self, other):
+        return self._compare(other) < 0
+
+    def __le__(self, other):
+        return self._compare(other) <= 0
+
+    def __eq__(self, other):
+        return self._compare(other) == 0
+
+    def __ne__(self, other):
+        return self._compare(other) != 0
+
+    def __gt__(self, other):
+        return self._compare(other) > 0
+
+    def __ge__(self, other):
+        return self._compare(other) >= 0
+
+    def __repr__(self):
+        return "NumpyVersion(%s)" % self.vstring

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/histograms.py

@@ -0,0 +1,1072 @@
+"""
+Histogram-related functions
+"""
+import contextlib
+import functools
+import operator
+import warnings
+
+import numpy as np
+from numpy.core import overrides
+
+__all__ = ['histogram', 'histogramdd', 'histogram_bin_edges']
+
+array_function_dispatch = functools.partial(
+    overrides.array_function_dispatch, module='numpy')
+
+# range is a keyword argument to many functions, so save the builtin so they can
+# use it.
+_range = range
+
+
+def _ptp(x):
+    """Peak-to-peak value of x.
+
+    This implementation avoids the problem of signed integer arrays having a
+    peak-to-peak value that cannot be represented with the array's data type.
+    This function returns an unsigned value for signed integer arrays.
+    """
+    return _unsigned_subtract(x.max(), x.min())
+
+
+def _hist_bin_sqrt(x, range):
+    """
+    Square root histogram bin estimator.
+
+    Bin width is inversely proportional to the data size. Used by many
+    programs for its simplicity.
+
+    Parameters
+    ----------
+    x : array_like
+        Input data that is to be histogrammed, trimmed to range. May not
+        be empty.
+
+    Returns
+    -------
+    h : An estimate of the optimal bin width for the given data.
+    """
+    del range  # unused
+    return _ptp(x) / np.sqrt(x.size)
+
+
+def _hist_bin_sturges(x, range):
+    """
+    Sturges histogram bin estimator.
+
+    A very simplistic estimator based on the assumption of normality of
+    the data. This estimator has poor performance for non-normal data,
+    which becomes especially obvious for large data sets. The estimate
+    depends only on size of the data.
+
+    Parameters
+    ----------
+    x : array_like
+        Input data that is to be histogrammed, trimmed to range. May not
+        be empty.
+
+    Returns
+    -------
+    h : An estimate of the optimal bin width for the given data.
+    """
+    del range  # unused
+    return _ptp(x) / (np.log2(x.size) + 1.0)
+
+
+def _hist_bin_rice(x, range):
+    """
+    Rice histogram bin estimator.
+
+    Another simple estimator with no normality assumption. It has better
+    performance for large data than Sturges, but tends to overestimate
+    the number of bins. The number of bins is proportional to the cube
+    root of data size (asymptotically optimal). The estimate depends
+    only on size of the data.
+
+    Parameters
+    ----------
+    x : array_like
+        Input data that is to be histogrammed, trimmed to range. May not
+        be empty.
+
+    Returns
+    -------
+    h : An estimate of the optimal bin width for the given data.
+    """
+    del range  # unused
+    return _ptp(x) / (2.0 * x.size ** (1.0 / 3))
+
+
+def _hist_bin_scott(x, range):
+    """
+    Scott histogram bin estimator.
+
+    The binwidth is proportional to the standard deviation of the data
+    and inversely proportional to the cube root of data size
+    (asymptotically optimal).
+
+    Parameters
+    ----------
+    x : array_like
+        Input data that is to be histogrammed, trimmed to range. May not
+        be empty.
+
+    Returns
+    -------
+    h : An estimate of the optimal bin width for the given data.
+    """
+    del range  # unused
+    return (24.0 * np.pi**0.5 / x.size)**(1.0 / 3.0) * np.std(x)
+
+
+def _hist_bin_stone(x, range):
+    """
+    Histogram bin estimator based on minimizing the estimated integrated squared error (ISE).
+
+    The number of bins is chosen by minimizing the estimated ISE against the unknown true distribution.
+    The ISE is estimated using cross-validation and can be regarded as a generalization of Scott's rule.
+    https://en.wikipedia.org/wiki/Histogram#Scott.27s_normal_reference_rule
+
+    This paper by Stone appears to be the origination of this rule.
+    http://digitalassets.lib.berkeley.edu/sdtr/ucb/text/34.pdf
+
+    Parameters
+    ----------
+    x : array_like
+        Input data that is to be histogrammed, trimmed to range. May not
+        be empty.
+    range : (float, float)
+        The lower and upper range of the bins.
+
+    Returns
+    -------
+    h : An estimate of the optimal bin width for the given data.
+    """
+
+    n = x.size
+    ptp_x = _ptp(x)
+    if n <= 1 or ptp_x == 0:
+        return 0
+
+    def jhat(nbins):
+        hh = ptp_x / nbins
+        p_k = np.histogram(x, bins=nbins, range=range)[0] / n
+        return (2 - (n + 1) * p_k.dot(p_k)) / hh
+
+    nbins_upper_bound = max(100, int(np.sqrt(n)))
+    nbins = min(_range(1, nbins_upper_bound + 1), key=jhat)
+    if nbins == nbins_upper_bound:
+        warnings.warn("The number of bins estimated may be suboptimal.",
+                      RuntimeWarning, stacklevel=3)
+    return ptp_x / nbins
+
+
+def _hist_bin_doane(x, range):
+    """
+    Doane's histogram bin estimator.
+
+    Improved version of Sturges' formula which works better for
+    non-normal data. See
+    stats.stackexchange.com/questions/55134/doanes-formula-for-histogram-binning
+
+    Parameters
+    ----------
+    x : array_like
+        Input data that is to be histogrammed, trimmed to range. May not
+        be empty.
+
+    Returns
+    -------
+    h : An estimate of the optimal bin width for the given data.
+    """
+    del range  # unused
+    if x.size > 2:
+        sg1 = np.sqrt(6.0 * (x.size - 2) / ((x.size + 1.0) * (x.size + 3)))
+        sigma = np.std(x)
+        if sigma > 0.0:
+            # These three operations add up to
+            # g1 = np.mean(((x - np.mean(x)) / sigma)**3)
+            # but use only one temp array instead of three
+            temp = x - np.mean(x)
+            np.true_divide(temp, sigma, temp)
+            np.power(temp, 3, temp)
+            g1 = np.mean(temp)
+            return _ptp(x) / (1.0 + np.log2(x.size) +
+                                    np.log2(1.0 + np.absolute(g1) / sg1))
+    return 0.0
+
+
+def _hist_bin_fd(x, range):
+    """
+    The Freedman-Diaconis histogram bin estimator.
+
+    The Freedman-Diaconis rule uses interquartile range (IQR) to
+    estimate binwidth. It is considered a variation of the Scott rule
+    with more robustness as the IQR is less affected by outliers than
+    the standard deviation. However, the IQR depends on fewer points
+    than the standard deviation, so it is less accurate, especially for
+    long tailed distributions.
+
+    If the IQR is 0, this function returns 0 for the bin width.
+    Binwidth is inversely proportional to the cube root of data size
+    (asymptotically optimal).
+
+    Parameters
+    ----------
+    x : array_like
+        Input data that is to be histogrammed, trimmed to range. May not
+        be empty.
+
+    Returns
+    -------
+    h : An estimate of the optimal bin width for the given data.
+    """
+    del range  # unused
+    iqr = np.subtract(*np.percentile(x, [75, 25]))
+    return 2.0 * iqr * x.size ** (-1.0 / 3.0)
+
+
+def _hist_bin_auto(x, range):
+    """
+    Histogram bin estimator that uses the minimum width of the
+    Freedman-Diaconis and Sturges estimators if the FD bin width is non-zero.
+    If the bin width from the FD estimator is 0, the Sturges estimator is used.
+
+    The FD estimator is usually the most robust method, but its width
+    estimate tends to be too large for small `x` and bad for data with limited
+    variance. The Sturges estimator is quite good for small (<1000) datasets
+    and is the default in the R language. This method gives good off-the-shelf
+    behaviour.
+
+    .. versionchanged:: 1.15.0
+    If there is limited variance the IQR can be 0, which results in the
+    FD bin width being 0 too. This is not a valid bin width, so
+    ``np.histogram_bin_edges`` chooses 1 bin instead, which may not be optimal.
+    If the IQR is 0, it's unlikely any variance-based estimators will be of
+    use, so we revert to the Sturges estimator, which only uses the size of the
+    dataset in its calculation.
+
+    Parameters
+    ----------
+    x : array_like
+        Input data that is to be histogrammed, trimmed to range. May not
+        be empty.
+
+    Returns
+    -------
+    h : An estimate of the optimal bin width for the given data.
+
+    See Also
+    --------
+    _hist_bin_fd, _hist_bin_sturges
+    """
+    fd_bw = _hist_bin_fd(x, range)
+    sturges_bw = _hist_bin_sturges(x, range)
+    del range  # unused
+    if fd_bw:
+        return min(fd_bw, sturges_bw)
+    else:
+        # limited variance, so we return a len dependent bw estimator
+        return sturges_bw
+
+# Private dict initialized at module load time
+_hist_bin_selectors = {'stone': _hist_bin_stone,
+                       'auto': _hist_bin_auto,
+                       'doane': _hist_bin_doane,
+                       'fd': _hist_bin_fd,
+                       'rice': _hist_bin_rice,
+                       'scott': _hist_bin_scott,
+                       'sqrt': _hist_bin_sqrt,
+                       'sturges': _hist_bin_sturges}
+
+
+def _ravel_and_check_weights(a, weights):
+    """ Check a and weights have matching shapes, and ravel both """
+    a = np.asarray(a)
+
+    # Ensure that the array is a "subtractable" dtype
+    if a.dtype == np.bool_:
+        warnings.warn("Converting input from {} to {} for compatibility."
+                      .format(a.dtype, np.uint8),
+                      RuntimeWarning, stacklevel=3)
+        a = a.astype(np.uint8)
+
+    if weights is not None:
+        weights = np.asarray(weights)
+        if weights.shape != a.shape:
+            raise ValueError(
+                'weights should have the same shape as a.')
+        weights = weights.ravel()
+    a = a.ravel()
+    return a, weights
+
+
+def _get_outer_edges(a, range):
+    """
+    Determine the outer bin edges to use, from either the data or the range
+    argument
+    """
+    if range is not None:
+        first_edge, last_edge = range
+        if first_edge > last_edge:
+            raise ValueError(
+                'max must be larger than min in range parameter.')
+        if not (np.isfinite(first_edge) and np.isfinite(last_edge)):
+            raise ValueError(
+                "supplied range of [{}, {}] is not finite".format(first_edge, last_edge))
+    elif a.size == 0:
+        # handle empty arrays. Can't determine range, so use 0-1.
+        first_edge, last_edge = 0, 1
+    else:
+        first_edge, last_edge = a.min(), a.max()
+        if not (np.isfinite(first_edge) and np.isfinite(last_edge)):
+            raise ValueError(
+                "autodetected range of [{}, {}] is not finite".format(first_edge, last_edge))
+
+    # expand empty range to avoid divide by zero
+    if first_edge == last_edge:
+        first_edge = first_edge - 0.5
+        last_edge = last_edge + 0.5
+
+    return first_edge, last_edge
+
+
+def _unsigned_subtract(a, b):
+    """
+    Subtract two values where a >= b, and produce an unsigned result
+
+    This is needed when finding the difference between the upper and lower
+    bound of an int16 histogram
+    """
+    # coerce to a single type
+    signed_to_unsigned = {
+        np.byte: np.ubyte,
+        np.short: np.ushort,
+        np.intc: np.uintc,
+        np.int_: np.uint,
+        np.longlong: np.ulonglong
+    }
+    dt = np.result_type(a, b)
+    try:
+        dt = signed_to_unsigned[dt.type]
+    except KeyError:
+        return np.subtract(a, b, dtype=dt)
+    else:
+        # we know the inputs are integers, and we are deliberately casting
+        # signed to unsigned
+        return np.subtract(a, b, casting='unsafe', dtype=dt)
+
+
+def _get_bin_edges(a, bins, range, weights):
+    """
+    Computes the bins used internally by `histogram`.
+
+    Parameters
+    ==========
+    a : ndarray
+        Ravelled data array
+    bins, range
+        Forwarded arguments from `histogram`.
+    weights : ndarray, optional
+        Ravelled weights array, or None
+
+    Returns
+    =======
+    bin_edges : ndarray
+        Array of bin edges
+    uniform_bins : (Number, Number, int):
+        The upper bound, lowerbound, and number of bins, used in the optimized
+        implementation of `histogram` that works on uniform bins.
+    """
+    # parse the overloaded bins argument
+    n_equal_bins = None
+    bin_edges = None
+
+    if isinstance(bins, str):
+        bin_name = bins
+        # if `bins` is a string for an automatic method,
+        # this will replace it with the number of bins calculated
+        if bin_name not in _hist_bin_selectors:
+            raise ValueError(
+                "{!r} is not a valid estimator for `bins`".format(bin_name))
+        if weights is not None:
+            raise TypeError("Automated estimation of the number of "
+                            "bins is not supported for weighted data")
+
+        first_edge, last_edge = _get_outer_edges(a, range)
+
+        # truncate the range if needed
+        if range is not None:
+            keep = (a >= first_edge)
+            keep &= (a <= last_edge)
+            if not np.logical_and.reduce(keep):
+                a = a[keep]
+
+        if a.size == 0:
+            n_equal_bins = 1
+        else:
+            # Do not call selectors on empty arrays
+            width = _hist_bin_selectors[bin_name](a, (first_edge, last_edge))
+            if width:
+                n_equal_bins = int(np.ceil(_unsigned_subtract(last_edge, first_edge) / width))
+            else:
+                # Width can be zero for some estimators, e.g. FD when
+                # the IQR of the data is zero.
+                n_equal_bins = 1
+
+    elif np.ndim(bins) == 0:
+        try:
+            n_equal_bins = operator.index(bins)
+        except TypeError as e:
+            raise TypeError(
+                '`bins` must be an integer, a string, or an array') from e
+        if n_equal_bins < 1:
+            raise ValueError('`bins` must be positive, when an integer')
+
+        first_edge, last_edge = _get_outer_edges(a, range)
+
+    elif np.ndim(bins) == 1:
+        bin_edges = np.asarray(bins)
+        if np.any(bin_edges[:-1] > bin_edges[1:]):
+            raise ValueError(
+                '`bins` must increase monotonically, when an array')
+
+    else:
+        raise ValueError('`bins` must be 1d, when an array')
+
+    if n_equal_bins is not None:
+        # gh-10322 means that type resolution rules are dependent on array
+        # shapes. To avoid this causing problems, we pick a type now and stick
+        # with it throughout.
+        bin_type = np.result_type(first_edge, last_edge, a)
+        if np.issubdtype(bin_type, np.integer):
+            bin_type = np.result_type(bin_type, float)
+
+        # bin edges must be computed
+        bin_edges = np.linspace(
+            first_edge, last_edge, n_equal_bins + 1,
+            endpoint=True, dtype=bin_type)
+        return bin_edges, (first_edge, last_edge, n_equal_bins)
+    else:
+        return bin_edges, None
+
+
+def _search_sorted_inclusive(a, v):
+    """
+    Like `searchsorted`, but where the last item in `v` is placed on the right.
+
+    In the context of a histogram, this makes the last bin edge inclusive
+    """
+    return np.concatenate((
+        a.searchsorted(v[:-1], 'left'),
+        a.searchsorted(v[-1:], 'right')
+    ))
+
+
+def _histogram_bin_edges_dispatcher(a, bins=None, range=None, weights=None):
+    return (a, bins, weights)
+
+
+@array_function_dispatch(_histogram_bin_edges_dispatcher)
+def histogram_bin_edges(a, bins=10, range=None, weights=None):
+    r"""
+    Function to calculate only the edges of the bins used by the `histogram`
+    function.
+
+    Parameters
+    ----------
+    a : array_like
+        Input data. The histogram is computed over the flattened array.
+    bins : int or sequence of scalars or str, optional
+        If `bins` is an int, it defines the number of equal-width
+        bins in the given range (10, by default). If `bins` is a
+        sequence, it defines the bin edges, including the rightmost
+        edge, allowing for non-uniform bin widths.
+
+        If `bins` is a string from the list below, `histogram_bin_edges` will use
+        the method chosen to calculate the optimal bin width and
+        consequently the number of bins (see `Notes` for more detail on
+        the estimators) from the data that falls within the requested
+        range. While the bin width will be optimal for the actual data
+        in the range, the number of bins will be computed to fill the
+        entire range, including the empty portions. For visualisation,
+        using the 'auto' option is suggested. Weighted data is not
+        supported for automated bin size selection.
+
+        'auto'
+            Maximum of the 'sturges' and 'fd' estimators. Provides good
+            all around performance.
+
+        'fd' (Freedman Diaconis Estimator)
+            Robust (resilient to outliers) estimator that takes into
+            account data variability and data size.
+
+        'doane'
+            An improved version of Sturges' estimator that works better
+            with non-normal datasets.
+
+        'scott'
+            Less robust estimator that takes into account data variability
+            and data size.
+
+        'stone'
+            Estimator based on leave-one-out cross-validation estimate of
+            the integrated squared error. Can be regarded as a generalization
+            of Scott's rule.
+
+        'rice'
+            Estimator does not take variability into account, only data
+            size. Commonly overestimates number of bins required.
+
+        'sturges'
+            R's default method, only accounts for data size. Only
+            optimal for gaussian data and underestimates number of bins
+            for large non-gaussian datasets.
+
+        'sqrt'
+            Square root (of data size) estimator, used by Excel and
+            other programs for its speed and simplicity.
+
+    range : (float, float), optional
+        The lower and upper range of the bins.  If not provided, range
+        is simply ``(a.min(), a.max())``.  Values outside the range are
+        ignored. The first element of the range must be less than or
+        equal to the second. `range` affects the automatic bin
+        computation as well. While bin width is computed to be optimal
+        based on the actual data within `range`, the bin count will fill
+        the entire range including portions containing no data.
+
+    weights : array_like, optional
+        An array of weights, of the same shape as `a`.  Each value in
+        `a` only contributes its associated weight towards the bin count
+        (instead of 1). This is currently not used by any of the bin estimators,
+        but may be in the future.
+
+    Returns
+    -------
+    bin_edges : array of dtype float
+        The edges to pass into `histogram`
+
+    See Also
+    --------
+    histogram
+
+    Notes
+    -----
+    The methods to estimate the optimal number of bins are well founded
+    in literature, and are inspired by the choices R provides for
+    histogram visualisation. Note that having the number of bins
+    proportional to :math:`n^{1/3}` is asymptotically optimal, which is
+    why it appears in most estimators. These are simply plug-in methods
+    that give good starting points for number of bins. In the equations
+    below, :math:`h` is the binwidth and :math:`n_h` is the number of
+    bins. All estimators that compute bin counts are recast to bin width
+    using the `ptp` of the data. The final bin count is obtained from
+    ``np.round(np.ceil(range / h))``. The final bin width is often less
+    than what is returned by the estimators below.
+
+    'auto' (maximum of the 'sturges' and 'fd' estimators)
+        A compromise to get a good value. For small datasets the Sturges
+        value will usually be chosen, while larger datasets will usually
+        default to FD.  Avoids the overly conservative behaviour of FD
+        and Sturges for small and large datasets respectively.
+        Switchover point is usually :math:`a.size \approx 1000`.
+
+    'fd' (Freedman Diaconis Estimator)
+        .. math:: h = 2 \frac{IQR}{n^{1/3}}
+
+        The binwidth is proportional to the interquartile range (IQR)
+        and inversely proportional to cube root of a.size. Can be too
+        conservative for small datasets, but is quite good for large
+        datasets. The IQR is very robust to outliers.
+
+    'scott'
+        .. math:: h = \sigma \sqrt[3]{\frac{24 \sqrt{\pi}}{n}}
+
+        The binwidth is proportional to the standard deviation of the
+        data and inversely proportional to cube root of ``x.size``. Can
+        be too conservative for small datasets, but is quite good for
+        large datasets. The standard deviation is not very robust to
+        outliers. Values are very similar to the Freedman-Diaconis
+        estimator in the absence of outliers.
+
+    'rice'
+        .. math:: n_h = 2n^{1/3}
+
+        The number of bins is only proportional to cube root of
+        ``a.size``. It tends to overestimate the number of bins and it
+        does not take into account data variability.
+
+    'sturges'
+        .. math:: n_h = \log _{2}(n) + 1
+
+        The number of bins is the base 2 log of ``a.size``.  This
+        estimator assumes normality of data and is too conservative for
+        larger, non-normal datasets. This is the default method in R's
+        ``hist`` method.
+
+    'doane'
+        .. math:: n_h = 1 + \log_{2}(n) +
+                        \log_{2}\left(1 + \frac{|g_1|}{\sigma_{g_1}}\right)
+
+            g_1 = mean\left[\left(\frac{x - \mu}{\sigma}\right)^3\right]
+
+            \sigma_{g_1} = \sqrt{\frac{6(n - 2)}{(n + 1)(n + 3)}}
+
+        An improved version of Sturges' formula that produces better
+        estimates for non-normal datasets. This estimator attempts to
+        account for the skew of the data.
+
+    'sqrt'
+        .. math:: n_h = \sqrt n
+
+        The simplest and fastest estimator. Only takes into account the
+        data size.
+
+    Examples
+    --------
+    >>> arr = np.array([0, 0, 0, 1, 2, 3, 3, 4, 5])
+    >>> np.histogram_bin_edges(arr, bins='auto', range=(0, 1))
+    array([0.  , 0.25, 0.5 , 0.75, 1.  ])
+    >>> np.histogram_bin_edges(arr, bins=2)
+    array([0. , 2.5, 5. ])
+
+    For consistency with histogram, an array of pre-computed bins is
+    passed through unmodified:
+
+    >>> np.histogram_bin_edges(arr, [1, 2])
+    array([1, 2])
+
+    This function allows one set of bins to be computed, and reused across
+    multiple histograms:
+
+    >>> shared_bins = np.histogram_bin_edges(arr, bins='auto')
+    >>> shared_bins
+    array([0., 1., 2., 3., 4., 5.])
+
+    >>> group_id = np.array([0, 1, 1, 0, 1, 1, 0, 1, 1])
+    >>> hist_0, _ = np.histogram(arr[group_id == 0], bins=shared_bins)
+    >>> hist_1, _ = np.histogram(arr[group_id == 1], bins=shared_bins)
+
+    >>> hist_0; hist_1
+    array([1, 1, 0, 1, 0])
+    array([2, 0, 1, 1, 2])
+
+    Which gives more easily comparable results than using separate bins for
+    each histogram:
+
+    >>> hist_0, bins_0 = np.histogram(arr[group_id == 0], bins='auto')
+    >>> hist_1, bins_1 = np.histogram(arr[group_id == 1], bins='auto')
+    >>> hist_0; hist_1
+    array([1, 1, 1])
+    array([2, 1, 1, 2])
+    >>> bins_0; bins_1
+    array([0., 1., 2., 3.])
+    array([0.  , 1.25, 2.5 , 3.75, 5.  ])
+
+    """
+    a, weights = _ravel_and_check_weights(a, weights)
+    bin_edges, _ = _get_bin_edges(a, bins, range, weights)
+    return bin_edges
+
+
+def _histogram_dispatcher(
+        a, bins=None, range=None, density=None, weights=None):
+    return (a, bins, weights)
+
+
+@array_function_dispatch(_histogram_dispatcher)
+def histogram(a, bins=10, range=None, density=None, weights=None):
+    r"""
+    Compute the histogram of a dataset.
+
+    Parameters
+    ----------
+    a : array_like
+        Input data. The histogram is computed over the flattened array.
+    bins : int or sequence of scalars or str, optional
+        If `bins` is an int, it defines the number of equal-width
+        bins in the given range (10, by default). If `bins` is a
+        sequence, it defines a monotonically increasing array of bin edges,
+        including the rightmost edge, allowing for non-uniform bin widths.
+
+        .. versionadded:: 1.11.0
+
+        If `bins` is a string, it defines the method used to calculate the
+        optimal bin width, as defined by `histogram_bin_edges`.
+
+    range : (float, float), optional
+        The lower and upper range of the bins.  If not provided, range
+        is simply ``(a.min(), a.max())``.  Values outside the range are
+        ignored. The first element of the range must be less than or
+        equal to the second. `range` affects the automatic bin
+        computation as well. While bin width is computed to be optimal
+        based on the actual data within `range`, the bin count will fill
+        the entire range including portions containing no data.
+    weights : array_like, optional
+        An array of weights, of the same shape as `a`.  Each value in
+        `a` only contributes its associated weight towards the bin count
+        (instead of 1). If `density` is True, the weights are
+        normalized, so that the integral of the density over the range
+        remains 1.
+    density : bool, optional
+        If ``False``, the result will contain the number of samples in
+        each bin. If ``True``, the result is the value of the
+        probability *density* function at the bin, normalized such that
+        the *integral* over the range is 1. Note that the sum of the
+        histogram values will not be equal to 1 unless bins of unity
+        width are chosen; it is not a probability *mass* function.
+
+    Returns
+    -------
+    hist : array
+        The values of the histogram. See `density` and `weights` for a
+        description of the possible semantics.
+    bin_edges : array of dtype float
+        Return the bin edges ``(length(hist)+1)``.
+
+
+    See Also
+    --------
+    histogramdd, bincount, searchsorted, digitize, histogram_bin_edges
+
+    Notes
+    -----
+    All but the last (righthand-most) bin is half-open.  In other words,
+    if `bins` is::
+
+      [1, 2, 3, 4]
+
+    then the first bin is ``[1, 2)`` (including 1, but excluding 2) and
+    the second ``[2, 3)``.  The last bin, however, is ``[3, 4]``, which
+    *includes* 4.
+
+
+    Examples
+    --------
+    >>> np.histogram([1, 2, 1], bins=[0, 1, 2, 3])
+    (array([0, 2, 1]), array([0, 1, 2, 3]))
+    >>> np.histogram(np.arange(4), bins=np.arange(5), density=True)
+    (array([0.25, 0.25, 0.25, 0.25]), array([0, 1, 2, 3, 4]))
+    >>> np.histogram([[1, 2, 1], [1, 0, 1]], bins=[0,1,2,3])
+    (array([1, 4, 1]), array([0, 1, 2, 3]))
+
+    >>> a = np.arange(5)
+    >>> hist, bin_edges = np.histogram(a, density=True)
+    >>> hist
+    array([0.5, 0. , 0.5, 0. , 0. , 0.5, 0. , 0.5, 0. , 0.5])
+    >>> hist.sum()
+    2.4999999999999996
+    >>> np.sum(hist * np.diff(bin_edges))
+    1.0
+
+    .. versionadded:: 1.11.0
+
+    Automated Bin Selection Methods example, using 2 peak random data
+    with 2000 points:
+
+    >>> import matplotlib.pyplot as plt
+    >>> rng = np.random.RandomState(10)  # deterministic random data
+    >>> a = np.hstack((rng.normal(size=1000),
+    ...                rng.normal(loc=5, scale=2, size=1000)))
+    >>> _ = plt.hist(a, bins='auto')  # arguments are passed to np.histogram
+    >>> plt.title("Histogram with 'auto' bins")
+    Text(0.5, 1.0, "Histogram with 'auto' bins")
+    >>> plt.show()
+
+    """
+    a, weights = _ravel_and_check_weights(a, weights)
+
+    bin_edges, uniform_bins = _get_bin_edges(a, bins, range, weights)
+
+    # Histogram is an integer or a float array depending on the weights.
+    if weights is None:
+        ntype = np.dtype(np.intp)
+    else:
+        ntype = weights.dtype
+
+    # We set a block size, as this allows us to iterate over chunks when
+    # computing histograms, to minimize memory usage.
+    BLOCK = 65536
+
+    # The fast path uses bincount, but that only works for certain types
+    # of weight
+    simple_weights = (
+        weights is None or
+        np.can_cast(weights.dtype, np.double) or
+        np.can_cast(weights.dtype, complex)
+    )
+
+    if uniform_bins is not None and simple_weights:
+        # Fast algorithm for equal bins
+        # We now convert values of a to bin indices, under the assumption of
+        # equal bin widths (which is valid here).
+        first_edge, last_edge, n_equal_bins = uniform_bins
+
+        # Initialize empty histogram
+        n = np.zeros(n_equal_bins, ntype)
+
+        # Pre-compute histogram scaling factor
+        norm_numerator = n_equal_bins
+        norm_denom = _unsigned_subtract(last_edge, first_edge)
+
+        # We iterate over blocks here for two reasons: the first is that for
+        # large arrays, it is actually faster (for example for a 10^8 array it
+        # is 2x as fast) and it results in a memory footprint 3x lower in the
+        # limit of large arrays.
+        for i in _range(0, len(a), BLOCK):
+            tmp_a = a[i:i+BLOCK]
+            if weights is None:
+                tmp_w = None
+            else:
+                tmp_w = weights[i:i + BLOCK]
+
+            # Only include values in the right range
+            keep = (tmp_a >= first_edge)
+            keep &= (tmp_a <= last_edge)
+            if not np.logical_and.reduce(keep):
+                tmp_a = tmp_a[keep]
+                if tmp_w is not None:
+                    tmp_w = tmp_w[keep]
+
+            # This cast ensures no type promotions occur below, which gh-10322
+            # make unpredictable. Getting it wrong leads to precision errors
+            # like gh-8123.
+            tmp_a = tmp_a.astype(bin_edges.dtype, copy=False)
+
+            # Compute the bin indices, and for values that lie exactly on
+            # last_edge we need to subtract one
+            f_indices = ((_unsigned_subtract(tmp_a, first_edge) / norm_denom)
+                         * norm_numerator)
+            indices = f_indices.astype(np.intp)
+            indices[indices == n_equal_bins] -= 1
+
+            # The index computation is not guaranteed to give exactly
+            # consistent results within ~1 ULP of the bin edges.
+            decrement = tmp_a < bin_edges[indices]
+            indices[decrement] -= 1
+            # The last bin includes the right edge. The other bins do not.
+            increment = ((tmp_a >= bin_edges[indices + 1])
+                         & (indices != n_equal_bins - 1))
+            indices[increment] += 1
+
+            # We now compute the histogram using bincount
+            if ntype.kind == 'c':
+                n.real += np.bincount(indices, weights=tmp_w.real,
+                                      minlength=n_equal_bins)
+                n.imag += np.bincount(indices, weights=tmp_w.imag,
+                                      minlength=n_equal_bins)
+            else:
+                n += np.bincount(indices, weights=tmp_w,
+                                 minlength=n_equal_bins).astype(ntype)
+    else:
+        # Compute via cumulative histogram
+        cum_n = np.zeros(bin_edges.shape, ntype)
+        if weights is None:
+            for i in _range(0, len(a), BLOCK):
+                sa = np.sort(a[i:i+BLOCK])
+                cum_n += _search_sorted_inclusive(sa, bin_edges)
+        else:
+            zero = np.zeros(1, dtype=ntype)
+            for i in _range(0, len(a), BLOCK):
+                tmp_a = a[i:i+BLOCK]
+                tmp_w = weights[i:i+BLOCK]
+                sorting_index = np.argsort(tmp_a)
+                sa = tmp_a[sorting_index]
+                sw = tmp_w[sorting_index]
+                cw = np.concatenate((zero, sw.cumsum()))
+                bin_index = _search_sorted_inclusive(sa, bin_edges)
+                cum_n += cw[bin_index]
+
+        n = np.diff(cum_n)
+
+    if density:
+        db = np.array(np.diff(bin_edges), float)
+        return n/db/n.sum(), bin_edges
+
+    return n, bin_edges
+
+
+def _histogramdd_dispatcher(sample, bins=None, range=None, density=None,
+                            weights=None):
+    if hasattr(sample, 'shape'):  # same condition as used in histogramdd
+        yield sample
+    else:
+        yield from sample
+    with contextlib.suppress(TypeError):
+        yield from bins
+    yield weights
+
+
+@array_function_dispatch(_histogramdd_dispatcher)
+def histogramdd(sample, bins=10, range=None, density=None, weights=None):
+    """
+    Compute the multidimensional histogram of some data.
+
+    Parameters
+    ----------
+    sample : (N, D) array, or (N, D) array_like
+        The data to be histogrammed.
+
+        Note the unusual interpretation of sample when an array_like:
+
+        * When an array, each row is a coordinate in a D-dimensional space -
+          such as ``histogramdd(np.array([p1, p2, p3]))``.
+        * When an array_like, each element is the list of values for single
+          coordinate - such as ``histogramdd((X, Y, Z))``.
+
+        The first form should be preferred.
+
+    bins : sequence or int, optional
+        The bin specification:
+
+        * A sequence of arrays describing the monotonically increasing bin
+          edges along each dimension.
+        * The number of bins for each dimension (nx, ny, ... =bins)
+        * The number of bins for all dimensions (nx=ny=...=bins).
+
+    range : sequence, optional
+        A sequence of length D, each an optional (lower, upper) tuple giving
+        the outer bin edges to be used if the edges are not given explicitly in
+        `bins`.
+        An entry of None in the sequence results in the minimum and maximum
+        values being used for the corresponding dimension.
+        The default, None, is equivalent to passing a tuple of D None values.
+    density : bool, optional
+        If False, the default, returns the number of samples in each bin.
+        If True, returns the probability *density* function at the bin,
+        ``bin_count / sample_count / bin_volume``.
+    weights : (N,) array_like, optional
+        An array of values `w_i` weighing each sample `(x_i, y_i, z_i, ...)`.
+        Weights are normalized to 1 if density is True. If density is False,
+        the values of the returned histogram are equal to the sum of the
+        weights belonging to the samples falling into each bin.
+
+    Returns
+    -------
+    H : ndarray
+        The multidimensional histogram of sample x. See density and weights
+        for the different possible semantics.
+    edges : list
+        A list of D arrays describing the bin edges for each dimension.
+
+    See Also
+    --------
+    histogram: 1-D histogram
+    histogram2d: 2-D histogram
+
+    Examples
+    --------
+    >>> r = np.random.randn(100,3)
+    >>> H, edges = np.histogramdd(r, bins = (5, 8, 4))
+    >>> H.shape, edges[0].size, edges[1].size, edges[2].size
+    ((5, 8, 4), 6, 9, 5)
+
+    """
+
+    try:
+        # Sample is an ND-array.
+        N, D = sample.shape
+    except (AttributeError, ValueError):
+        # Sample is a sequence of 1D arrays.
+        sample = np.atleast_2d(sample).T
+        N, D = sample.shape
+
+    nbin = np.empty(D, np.intp)
+    edges = D*[None]
+    dedges = D*[None]
+    if weights is not None:
+        weights = np.asarray(weights)
+
+    try:
+        M = len(bins)
+        if M != D:
+            raise ValueError(
+                'The dimension of bins must be equal to the dimension of the '
+                'sample x.')
+    except TypeError:
+        # bins is an integer
+        bins = D*[bins]
+
+    # normalize the range argument
+    if range is None:
+        range = (None,) * D
+    elif len(range) != D:
+        raise ValueError('range argument must have one entry per dimension')
+
+    # Create edge arrays
+    for i in _range(D):
+        if np.ndim(bins[i]) == 0:
+            if bins[i] < 1:
+                raise ValueError(
+                    '`bins[{}]` must be positive, when an integer'.format(i))
+            smin, smax = _get_outer_edges(sample[:,i], range[i])
+            try:
+                n = operator.index(bins[i])
+
+            except TypeError as e:
+                raise TypeError(
+                	"`bins[{}]` must be an integer, when a scalar".format(i)
+                ) from e
+
+            edges[i] = np.linspace(smin, smax, n + 1)
+        elif np.ndim(bins[i]) == 1:
+            edges[i] = np.asarray(bins[i])
+            if np.any(edges[i][:-1] > edges[i][1:]):
+                raise ValueError(
+                    '`bins[{}]` must be monotonically increasing, when an array'
+                    .format(i))
+        else:
+            raise ValueError(
+                '`bins[{}]` must be a scalar or 1d array'.format(i))
+
+        nbin[i] = len(edges[i]) + 1  # includes an outlier on each end
+        dedges[i] = np.diff(edges[i])
+
+    # Compute the bin number each sample falls into.
+    Ncount = tuple(
+        # avoid np.digitize to work around gh-11022
+        np.searchsorted(edges[i], sample[:, i], side='right')
+        for i in _range(D)
+    )
+
+    # Using digitize, values that fall on an edge are put in the right bin.
+    # For the rightmost bin, we want values equal to the right edge to be
+    # counted in the last bin, and not as an outlier.
+    for i in _range(D):
+        # Find which points are on the rightmost edge.
+        on_edge = (sample[:, i] == edges[i][-1])
+        # Shift these points one bin to the left.
+        Ncount[i][on_edge] -= 1
+
+    # Compute the sample indices in the flattened histogram matrix.
+    # This raises an error if the array is too large.
+    xy = np.ravel_multi_index(Ncount, nbin)
+
+    # Compute the number of repetitions in xy and assign it to the
+    # flattened histmat.
+    hist = np.bincount(xy, weights, minlength=nbin.prod())
+
+    # Shape into a proper matrix
+    hist = hist.reshape(nbin)
+
+    # This preserves the (bad) behavior observed in gh-7845, for now.
+    hist = hist.astype(float, casting='safe')
+
+    # Remove outliers (indices 0 and -1 for each dimension).
+    core = D*(slice(1, -1),)
+    hist = hist[core]
+
+    if density:
+        # calculate the probability density function
+        s = hist.sum()
+        for i in _range(D):
+            shape = np.ones(D, int)
+            shape[i] = nbin[i] - 2
+            hist = hist / dedges[i].reshape(shape)
+        hist /= s
+
+    if (hist.shape != nbin - 2).any():
+        raise RuntimeError(
+            "Internal Shape Error")
+    return hist, edges

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/index_tricks.py

@@ -0,0 +1,1046 @@
+import functools
+import sys
+import math
+import warnings
+
+import numpy as np
+from .._utils import set_module
+import numpy.core.numeric as _nx
+from numpy.core.numeric import ScalarType, array
+from numpy.core.numerictypes import issubdtype
+
+import numpy.matrixlib as matrixlib
+from .function_base import diff
+from numpy.core.multiarray import ravel_multi_index, unravel_index
+from numpy.core import overrides, linspace
+from numpy.lib.stride_tricks import as_strided
+
+
+array_function_dispatch = functools.partial(
+    overrides.array_function_dispatch, module='numpy')
+
+
+__all__ = [
+    'ravel_multi_index', 'unravel_index', 'mgrid', 'ogrid', 'r_', 'c_',
+    's_', 'index_exp', 'ix_', 'ndenumerate', 'ndindex', 'fill_diagonal',
+    'diag_indices', 'diag_indices_from'
+]
+
+
+def _ix__dispatcher(*args):
+    return args
+
+
+@array_function_dispatch(_ix__dispatcher)
+def ix_(*args):
+    """
+    Construct an open mesh from multiple sequences.
+
+    This function takes N 1-D sequences and returns N outputs with N
+    dimensions each, such that the shape is 1 in all but one dimension
+    and the dimension with the non-unit shape value cycles through all
+    N dimensions.
+
+    Using `ix_` one can quickly construct index arrays that will index
+    the cross product. ``a[np.ix_([1,3],[2,5])]`` returns the array
+    ``[[a[1,2] a[1,5]], [a[3,2] a[3,5]]]``.
+
+    Parameters
+    ----------
+    args : 1-D sequences
+        Each sequence should be of integer or boolean type.
+        Boolean sequences will be interpreted as boolean masks for the
+        corresponding dimension (equivalent to passing in
+        ``np.nonzero(boolean_sequence)``).
+
+    Returns
+    -------
+    out : tuple of ndarrays
+        N arrays with N dimensions each, with N the number of input
+        sequences. Together these arrays form an open mesh.
+
+    See Also
+    --------
+    ogrid, mgrid, meshgrid
+
+    Examples
+    --------
+    >>> a = np.arange(10).reshape(2, 5)
+    >>> a
+    array([[0, 1, 2, 3, 4],
+           [5, 6, 7, 8, 9]])
+    >>> ixgrid = np.ix_([0, 1], [2, 4])
+    >>> ixgrid
+    (array([[0],
+           [1]]), array([[2, 4]]))
+    >>> ixgrid[0].shape, ixgrid[1].shape
+    ((2, 1), (1, 2))
+    >>> a[ixgrid]
+    array([[2, 4],
+           [7, 9]])
+
+    >>> ixgrid = np.ix_([True, True], [2, 4])
+    >>> a[ixgrid]
+    array([[2, 4],
+           [7, 9]])
+    >>> ixgrid = np.ix_([True, True], [False, False, True, False, True])
+    >>> a[ixgrid]
+    array([[2, 4],
+           [7, 9]])
+
+    """
+    out = []
+    nd = len(args)
+    for k, new in enumerate(args):
+        if not isinstance(new, _nx.ndarray):
+            new = np.asarray(new)
+            if new.size == 0:
+                # Explicitly type empty arrays to avoid float default
+                new = new.astype(_nx.intp)
+        if new.ndim != 1:
+            raise ValueError("Cross index must be 1 dimensional")
+        if issubdtype(new.dtype, _nx.bool_):
+            new, = new.nonzero()
+        new = new.reshape((1,)*k + (new.size,) + (1,)*(nd-k-1))
+        out.append(new)
+    return tuple(out)
+
+
+class nd_grid:
+    """
+    Construct a multi-dimensional "meshgrid".
+
+    ``grid = nd_grid()`` creates an instance which will return a mesh-grid
+    when indexed.  The dimension and number of the output arrays are equal
+    to the number of indexing dimensions.  If the step length is not a
+    complex number, then the stop is not inclusive.
+
+    However, if the step length is a **complex number** (e.g. 5j), then the
+    integer part of its magnitude is interpreted as specifying the
+    number of points to create between the start and stop values, where
+    the stop value **is inclusive**.
+
+    If instantiated with an argument of ``sparse=True``, the mesh-grid is
+    open (or not fleshed out) so that only one-dimension of each returned
+    argument is greater than 1.
+
+    Parameters
+    ----------
+    sparse : bool, optional
+        Whether the grid is sparse or not. Default is False.
+
+    Notes
+    -----
+    Two instances of `nd_grid` are made available in the NumPy namespace,
+    `mgrid` and `ogrid`, approximately defined as::
+
+        mgrid = nd_grid(sparse=False)
+        ogrid = nd_grid(sparse=True)
+
+    Users should use these pre-defined instances instead of using `nd_grid`
+    directly.
+    """
+
+    def __init__(self, sparse=False):
+        self.sparse = sparse
+
+    def __getitem__(self, key):
+        try:
+            size = []
+            # Mimic the behavior of `np.arange` and use a data type
+            # which is at least as large as `np.int_`
+            num_list = [0]
+            for k in range(len(key)):
+                step = key[k].step
+                start = key[k].start
+                stop = key[k].stop
+                if start is None:
+                    start = 0
+                if step is None:
+                    step = 1
+                if isinstance(step, (_nx.complexfloating, complex)):
+                    step = abs(step)
+                    size.append(int(step))
+                else:
+                    size.append(
+                        int(math.ceil((stop - start) / (step*1.0))))
+                num_list += [start, stop, step]
+            typ = _nx.result_type(*num_list)
+            if self.sparse:
+                nn = [_nx.arange(_x, dtype=_t)
+                      for _x, _t in zip(size, (typ,)*len(size))]
+            else:
+                nn = _nx.indices(size, typ)
+            for k, kk in enumerate(key):
+                step = kk.step
+                start = kk.start
+                if start is None:
+                    start = 0
+                if step is None:
+                    step = 1
+                if isinstance(step, (_nx.complexfloating, complex)):
+                    step = int(abs(step))
+                    if step != 1:
+                        step = (kk.stop - start) / float(step - 1)
+                nn[k] = (nn[k]*step+start)
+            if self.sparse:
+                slobj = [_nx.newaxis]*len(size)
+                for k in range(len(size)):
+                    slobj[k] = slice(None, None)
+                    nn[k] = nn[k][tuple(slobj)]
+                    slobj[k] = _nx.newaxis
+            return nn
+        except (IndexError, TypeError):
+            step = key.step
+            stop = key.stop
+            start = key.start
+            if start is None:
+                start = 0
+            if isinstance(step, (_nx.complexfloating, complex)):
+                # Prevent the (potential) creation of integer arrays
+                step_float = abs(step)
+                step = length = int(step_float)
+                if step != 1:
+                    step = (key.stop-start)/float(step-1)
+                typ = _nx.result_type(start, stop, step_float)
+                return _nx.arange(0, length, 1, dtype=typ)*step + start
+            else:
+                return _nx.arange(start, stop, step)
+
+
+class MGridClass(nd_grid):
+    """
+    An instance which returns a dense multi-dimensional "meshgrid".
+
+    An instance which returns a dense (or fleshed out) mesh-grid
+    when indexed, so that each returned argument has the same shape.
+    The dimensions and number of the output arrays are equal to the
+    number of indexing dimensions.  If the step length is not a complex
+    number, then the stop is not inclusive.
+
+    However, if the step length is a **complex number** (e.g. 5j), then
+    the integer part of its magnitude is interpreted as specifying the
+    number of points to create between the start and stop values, where
+    the stop value **is inclusive**.
+
+    Returns
+    -------
+    mesh-grid `ndarrays` all of the same dimensions
+
+    See Also
+    --------
+    ogrid : like `mgrid` but returns open (not fleshed out) mesh grids
+    meshgrid: return coordinate matrices from coordinate vectors
+    r_ : array concatenator
+    :ref:`how-to-partition`
+
+    Examples
+    --------
+    >>> np.mgrid[0:5, 0:5]
+    array([[[0, 0, 0, 0, 0],
+            [1, 1, 1, 1, 1],
+            [2, 2, 2, 2, 2],
+            [3, 3, 3, 3, 3],
+            [4, 4, 4, 4, 4]],
+           [[0, 1, 2, 3, 4],
+            [0, 1, 2, 3, 4],
+            [0, 1, 2, 3, 4],
+            [0, 1, 2, 3, 4],
+            [0, 1, 2, 3, 4]]])
+    >>> np.mgrid[-1:1:5j]
+    array([-1. , -0.5,  0. ,  0.5,  1. ])
+
+    """
+
+    def __init__(self):
+        super().__init__(sparse=False)
+
+
+mgrid = MGridClass()
+
+
+class OGridClass(nd_grid):
+    """
+    An instance which returns an open multi-dimensional "meshgrid".
+
+    An instance which returns an open (i.e. not fleshed out) mesh-grid
+    when indexed, so that only one dimension of each returned array is
+    greater than 1.  The dimension and number of the output arrays are
+    equal to the number of indexing dimensions.  If the step length is
+    not a complex number, then the stop is not inclusive.
+
+    However, if the step length is a **complex number** (e.g. 5j), then
+    the integer part of its magnitude is interpreted as specifying the
+    number of points to create between the start and stop values, where
+    the stop value **is inclusive**.
+
+    Returns
+    -------
+    mesh-grid
+        `ndarrays` with only one dimension not equal to 1
+
+    See Also
+    --------
+    mgrid : like `ogrid` but returns dense (or fleshed out) mesh grids
+    meshgrid: return coordinate matrices from coordinate vectors
+    r_ : array concatenator
+    :ref:`how-to-partition`
+
+    Examples
+    --------
+    >>> from numpy import ogrid
+    >>> ogrid[-1:1:5j]
+    array([-1. , -0.5,  0. ,  0.5,  1. ])
+    >>> ogrid[0:5,0:5]
+    [array([[0],
+            [1],
+            [2],
+            [3],
+            [4]]), array([[0, 1, 2, 3, 4]])]
+
+    """
+
+    def __init__(self):
+        super().__init__(sparse=True)
+
+
+ogrid = OGridClass()
+
+
+class AxisConcatenator:
+    """
+    Translates slice objects to concatenation along an axis.
+
+    For detailed documentation on usage, see `r_`.
+    """
+    # allow ma.mr_ to override this
+    concatenate = staticmethod(_nx.concatenate)
+    makemat = staticmethod(matrixlib.matrix)
+
+    def __init__(self, axis=0, matrix=False, ndmin=1, trans1d=-1):
+        self.axis = axis
+        self.matrix = matrix
+        self.trans1d = trans1d
+        self.ndmin = ndmin
+
+    def __getitem__(self, key):
+        # handle matrix builder syntax
+        if isinstance(key, str):
+            frame = sys._getframe().f_back
+            mymat = matrixlib.bmat(key, frame.f_globals, frame.f_locals)
+            return mymat
+
+        if not isinstance(key, tuple):
+            key = (key,)
+
+        # copy attributes, since they can be overridden in the first argument
+        trans1d = self.trans1d
+        ndmin = self.ndmin
+        matrix = self.matrix
+        axis = self.axis
+
+        objs = []
+        # dtypes or scalars for weak scalar handling in result_type
+        result_type_objs = []
+
+        for k, item in enumerate(key):
+            scalar = False
+            if isinstance(item, slice):
+                step = item.step
+                start = item.start
+                stop = item.stop
+                if start is None:
+                    start = 0
+                if step is None:
+                    step = 1
+                if isinstance(step, (_nx.complexfloating, complex)):
+                    size = int(abs(step))
+                    newobj = linspace(start, stop, num=size)
+                else:
+                    newobj = _nx.arange(start, stop, step)
+                if ndmin > 1:
+                    newobj = array(newobj, copy=False, ndmin=ndmin)
+                    if trans1d != -1:
+                        newobj = newobj.swapaxes(-1, trans1d)
+            elif isinstance(item, str):
+                if k != 0:
+                    raise ValueError("special directives must be the "
+                                     "first entry.")
+                if item in ('r', 'c'):
+                    matrix = True
+                    col = (item == 'c')
+                    continue
+                if ',' in item:
+                    vec = item.split(',')
+                    try:
+                        axis, ndmin = [int(x) for x in vec[:2]]
+                        if len(vec) == 3:
+                            trans1d = int(vec[2])
+                        continue
+                    except Exception as e:
+                        raise ValueError(
+                            "unknown special directive {!r}".format(item)
+                        ) from e
+                try:
+                    axis = int(item)
+                    continue
+                except (ValueError, TypeError) as e:
+                    raise ValueError("unknown special directive") from e
+            elif type(item) in ScalarType:
+                scalar = True
+                newobj = item
+            else:
+                item_ndim = np.ndim(item)
+                newobj = array(item, copy=False, subok=True, ndmin=ndmin)
+                if trans1d != -1 and item_ndim < ndmin:
+                    k2 = ndmin - item_ndim
+                    k1 = trans1d
+                    if k1 < 0:
+                        k1 += k2 + 1
+                    defaxes = list(range(ndmin))
+                    axes = defaxes[:k1] + defaxes[k2:] + defaxes[k1:k2]
+                    newobj = newobj.transpose(axes)
+
+            objs.append(newobj)
+            if scalar:
+                result_type_objs.append(item)
+            else:
+                result_type_objs.append(newobj.dtype)
+
+        # Ensure that scalars won't up-cast unless warranted, for 0, drops
+        # through to error in concatenate.
+        if len(result_type_objs) != 0:
+            final_dtype = _nx.result_type(*result_type_objs)
+            # concatenate could do cast, but that can be overriden:
+            objs = [array(obj, copy=False, subok=True,
+                          ndmin=ndmin, dtype=final_dtype) for obj in objs]
+
+        res = self.concatenate(tuple(objs), axis=axis)
+
+        if matrix:
+            oldndim = res.ndim
+            res = self.makemat(res)
+            if oldndim == 1 and col:
+                res = res.T
+        return res
+
+    def __len__(self):
+        return 0
+
+# separate classes are used here instead of just making r_ = concatentor(0),
+# etc. because otherwise we couldn't get the doc string to come out right
+# in help(r_)
+
+
+class RClass(AxisConcatenator):
+    """
+    Translates slice objects to concatenation along the first axis.
+
+    This is a simple way to build up arrays quickly. There are two use cases.
+
+    1. If the index expression contains comma separated arrays, then stack
+       them along their first axis.
+    2. If the index expression contains slice notation or scalars then create
+       a 1-D array with a range indicated by the slice notation.
+
+    If slice notation is used, the syntax ``start:stop:step`` is equivalent
+    to ``np.arange(start, stop, step)`` inside of the brackets. However, if
+    ``step`` is an imaginary number (i.e. 100j) then its integer portion is
+    interpreted as a number-of-points desired and the start and stop are
+    inclusive. In other words ``start:stop:stepj`` is interpreted as
+    ``np.linspace(start, stop, step, endpoint=1)`` inside of the brackets.
+    After expansion of slice notation, all comma separated sequences are
+    concatenated together.
+
+    Optional character strings placed as the first element of the index
+    expression can be used to change the output. The strings 'r' or 'c' result
+    in matrix output. If the result is 1-D and 'r' is specified a 1 x N (row)
+    matrix is produced. If the result is 1-D and 'c' is specified, then a N x 1
+    (column) matrix is produced. If the result is 2-D then both provide the
+    same matrix result.
+
+    A string integer specifies which axis to stack multiple comma separated
+    arrays along. A string of two comma-separated integers allows indication
+    of the minimum number of dimensions to force each entry into as the
+    second integer (the axis to concatenate along is still the first integer).
+
+    A string with three comma-separated integers allows specification of the
+    axis to concatenate along, the minimum number of dimensions to force the
+    entries to, and which axis should contain the start of the arrays which
+    are less than the specified number of dimensions. In other words the third
+    integer allows you to specify where the 1's should be placed in the shape
+    of the arrays that have their shapes upgraded. By default, they are placed
+    in the front of the shape tuple. The third argument allows you to specify
+    where the start of the array should be instead. Thus, a third argument of
+    '0' would place the 1's at the end of the array shape. Negative integers
+    specify where in the new shape tuple the last dimension of upgraded arrays
+    should be placed, so the default is '-1'.
+
+    Parameters
+    ----------
+    Not a function, so takes no parameters
+
+
+    Returns
+    -------
+    A concatenated ndarray or matrix.
+
+    See Also
+    --------
+    concatenate : Join a sequence of arrays along an existing axis.
+    c_ : Translates slice objects to concatenation along the second axis.
+
+    Examples
+    --------
+    >>> np.r_[np.array([1,2,3]), 0, 0, np.array([4,5,6])]
+    array([1, 2, 3, ..., 4, 5, 6])
+    >>> np.r_[-1:1:6j, [0]*3, 5, 6]
+    array([-1. , -0.6, -0.2,  0.2,  0.6,  1. ,  0. ,  0. ,  0. ,  5. ,  6. ])
+
+    String integers specify the axis to concatenate along or the minimum
+    number of dimensions to force entries into.
+
+    >>> a = np.array([[0, 1, 2], [3, 4, 5]])
+    >>> np.r_['-1', a, a] # concatenate along last axis
+    array([[0, 1, 2, 0, 1, 2],
+           [3, 4, 5, 3, 4, 5]])
+    >>> np.r_['0,2', [1,2,3], [4,5,6]] # concatenate along first axis, dim>=2
+    array([[1, 2, 3],
+           [4, 5, 6]])
+
+    >>> np.r_['0,2,0', [1,2,3], [4,5,6]]
+    array([[1],
+           [2],
+           [3],
+           [4],
+           [5],
+           [6]])
+    >>> np.r_['1,2,0', [1,2,3], [4,5,6]]
+    array([[1, 4],
+           [2, 5],
+           [3, 6]])
+
+    Using 'r' or 'c' as a first string argument creates a matrix.
+
+    >>> np.r_['r',[1,2,3], [4,5,6]]
+    matrix([[1, 2, 3, 4, 5, 6]])
+
+    """
+
+    def __init__(self):
+        AxisConcatenator.__init__(self, 0)
+
+
+r_ = RClass()
+
+
+class CClass(AxisConcatenator):
+    """
+    Translates slice objects to concatenation along the second axis.
+
+    This is short-hand for ``np.r_['-1,2,0', index expression]``, which is
+    useful because of its common occurrence. In particular, arrays will be
+    stacked along their last axis after being upgraded to at least 2-D with
+    1's post-pended to the shape (column vectors made out of 1-D arrays).
+
+    See Also
+    --------
+    column_stack : Stack 1-D arrays as columns into a 2-D array.
+    r_ : For more detailed documentation.
+
+    Examples
+    --------
+    >>> np.c_[np.array([1,2,3]), np.array([4,5,6])]
+    array([[1, 4],
+           [2, 5],
+           [3, 6]])
+    >>> np.c_[np.array([[1,2,3]]), 0, 0, np.array([[4,5,6]])]
+    array([[1, 2, 3, ..., 4, 5, 6]])
+
+    """
+
+    def __init__(self):
+        AxisConcatenator.__init__(self, -1, ndmin=2, trans1d=0)
+
+
+c_ = CClass()
+
+
+@set_module('numpy')
+class ndenumerate:
+    """
+    Multidimensional index iterator.
+
+    Return an iterator yielding pairs of array coordinates and values.
+
+    Parameters
+    ----------
+    arr : ndarray
+      Input array.
+
+    See Also
+    --------
+    ndindex, flatiter
+
+    Examples
+    --------
+    >>> a = np.array([[1, 2], [3, 4]])
+    >>> for index, x in np.ndenumerate(a):
+    ...     print(index, x)
+    (0, 0) 1
+    (0, 1) 2
+    (1, 0) 3
+    (1, 1) 4
+
+    """
+
+    def __init__(self, arr):
+        self.iter = np.asarray(arr).flat
+
+    def __next__(self):
+        """
+        Standard iterator method, returns the index tuple and array value.
+
+        Returns
+        -------
+        coords : tuple of ints
+            The indices of the current iteration.
+        val : scalar
+            The array element of the current iteration.
+
+        """
+        return self.iter.coords, next(self.iter)
+
+    def __iter__(self):
+        return self
+
+
+@set_module('numpy')
+class ndindex:
+    """
+    An N-dimensional iterator object to index arrays.
+
+    Given the shape of an array, an `ndindex` instance iterates over
+    the N-dimensional index of the array. At each iteration a tuple
+    of indices is returned, the last dimension is iterated over first.
+
+    Parameters
+    ----------
+    shape : ints, or a single tuple of ints
+        The size of each dimension of the array can be passed as
+        individual parameters or as the elements of a tuple.
+
+    See Also
+    --------
+    ndenumerate, flatiter
+
+    Examples
+    --------
+    Dimensions as individual arguments
+
+    >>> for index in np.ndindex(3, 2, 1):
+    ...     print(index)
+    (0, 0, 0)
+    (0, 1, 0)
+    (1, 0, 0)
+    (1, 1, 0)
+    (2, 0, 0)
+    (2, 1, 0)
+
+    Same dimensions - but in a tuple ``(3, 2, 1)``
+
+    >>> for index in np.ndindex((3, 2, 1)):
+    ...     print(index)
+    (0, 0, 0)
+    (0, 1, 0)
+    (1, 0, 0)
+    (1, 1, 0)
+    (2, 0, 0)
+    (2, 1, 0)
+
+    """
+
+    def __init__(self, *shape):
+        if len(shape) == 1 and isinstance(shape[0], tuple):
+            shape = shape[0]
+        x = as_strided(_nx.zeros(1), shape=shape,
+                       strides=_nx.zeros_like(shape))
+        self._it = _nx.nditer(x, flags=['multi_index', 'zerosize_ok'],
+                              order='C')
+
+    def __iter__(self):
+        return self
+
+    def ndincr(self):
+        """
+        Increment the multi-dimensional index by one.
+
+        This method is for backward compatibility only: do not use.
+
+        .. deprecated:: 1.20.0
+            This method has been advised against since numpy 1.8.0, but only
+            started emitting DeprecationWarning as of this version.
+        """
+        # NumPy 1.20.0, 2020-09-08
+        warnings.warn(
+            "`ndindex.ndincr()` is deprecated, use `next(ndindex)` instead",
+            DeprecationWarning, stacklevel=2)
+        next(self)
+
+    def __next__(self):
+        """
+        Standard iterator method, updates the index and returns the index
+        tuple.
+
+        Returns
+        -------
+        val : tuple of ints
+            Returns a tuple containing the indices of the current
+            iteration.
+
+        """
+        next(self._it)
+        return self._it.multi_index
+
+
+# You can do all this with slice() plus a few special objects,
+# but there's a lot to remember. This version is simpler because
+# it uses the standard array indexing syntax.
+#
+# Written by Konrad Hinsen <hinsen@cnrs-orleans.fr>
+# last revision: 1999-7-23
+#
+# Cosmetic changes by T. Oliphant 2001
+#
+#
+
+class IndexExpression:
+    """
+    A nicer way to build up index tuples for arrays.
+
+    .. note::
+       Use one of the two predefined instances `index_exp` or `s_`
+       rather than directly using `IndexExpression`.
+
+    For any index combination, including slicing and axis insertion,
+    ``a[indices]`` is the same as ``a[np.index_exp[indices]]`` for any
+    array `a`. However, ``np.index_exp[indices]`` can be used anywhere
+    in Python code and returns a tuple of slice objects that can be
+    used in the construction of complex index expressions.
+
+    Parameters
+    ----------
+    maketuple : bool
+        If True, always returns a tuple.
+
+    See Also
+    --------
+    index_exp : Predefined instance that always returns a tuple:
+       `index_exp = IndexExpression(maketuple=True)`.
+    s_ : Predefined instance without tuple conversion:
+       `s_ = IndexExpression(maketuple=False)`.
+
+    Notes
+    -----
+    You can do all this with `slice()` plus a few special objects,
+    but there's a lot to remember and this version is simpler because
+    it uses the standard array indexing syntax.
+
+    Examples
+    --------
+    >>> np.s_[2::2]
+    slice(2, None, 2)
+    >>> np.index_exp[2::2]
+    (slice(2, None, 2),)
+
+    >>> np.array([0, 1, 2, 3, 4])[np.s_[2::2]]
+    array([2, 4])
+
+    """
+
+    def __init__(self, maketuple):
+        self.maketuple = maketuple
+
+    def __getitem__(self, item):
+        if self.maketuple and not isinstance(item, tuple):
+            return (item,)
+        else:
+            return item
+
+
+index_exp = IndexExpression(maketuple=True)
+s_ = IndexExpression(maketuple=False)
+
+# End contribution from Konrad.
+
+
+# The following functions complement those in twodim_base, but are
+# applicable to N-dimensions.
+
+
+def _fill_diagonal_dispatcher(a, val, wrap=None):
+    return (a,)
+
+
+@array_function_dispatch(_fill_diagonal_dispatcher)
+def fill_diagonal(a, val, wrap=False):
+    """Fill the main diagonal of the given array of any dimensionality.
+
+    For an array `a` with ``a.ndim >= 2``, the diagonal is the list of
+    locations with indices ``a[i, ..., i]`` all identical. This function
+    modifies the input array in-place, it does not return a value.
+
+    Parameters
+    ----------
+    a : array, at least 2-D.
+      Array whose diagonal is to be filled, it gets modified in-place.
+
+    val : scalar or array_like
+      Value(s) to write on the diagonal. If `val` is scalar, the value is
+      written along the diagonal. If array-like, the flattened `val` is
+      written along the diagonal, repeating if necessary to fill all
+      diagonal entries.
+
+    wrap : bool
+      For tall matrices in NumPy version up to 1.6.2, the
+      diagonal "wrapped" after N columns. You can have this behavior
+      with this option. This affects only tall matrices.
+
+    See also
+    --------
+    diag_indices, diag_indices_from
+
+    Notes
+    -----
+    .. versionadded:: 1.4.0
+
+    This functionality can be obtained via `diag_indices`, but internally
+    this version uses a much faster implementation that never constructs the
+    indices and uses simple slicing.
+
+    Examples
+    --------
+    >>> a = np.zeros((3, 3), int)
+    >>> np.fill_diagonal(a, 5)
+    >>> a
+    array([[5, 0, 0],
+           [0, 5, 0],
+           [0, 0, 5]])
+
+    The same function can operate on a 4-D array:
+
+    >>> a = np.zeros((3, 3, 3, 3), int)
+    >>> np.fill_diagonal(a, 4)
+
+    We only show a few blocks for clarity:
+
+    >>> a[0, 0]
+    array([[4, 0, 0],
+           [0, 0, 0],
+           [0, 0, 0]])
+    >>> a[1, 1]
+    array([[0, 0, 0],
+           [0, 4, 0],
+           [0, 0, 0]])
+    >>> a[2, 2]
+    array([[0, 0, 0],
+           [0, 0, 0],
+           [0, 0, 4]])
+
+    The wrap option affects only tall matrices:
+
+    >>> # tall matrices no wrap
+    >>> a = np.zeros((5, 3), int)
+    >>> np.fill_diagonal(a, 4)
+    >>> a
+    array([[4, 0, 0],
+           [0, 4, 0],
+           [0, 0, 4],
+           [0, 0, 0],
+           [0, 0, 0]])
+
+    >>> # tall matrices wrap
+    >>> a = np.zeros((5, 3), int)
+    >>> np.fill_diagonal(a, 4, wrap=True)
+    >>> a
+    array([[4, 0, 0],
+           [0, 4, 0],
+           [0, 0, 4],
+           [0, 0, 0],
+           [4, 0, 0]])
+
+    >>> # wide matrices
+    >>> a = np.zeros((3, 5), int)
+    >>> np.fill_diagonal(a, 4, wrap=True)
+    >>> a
+    array([[4, 0, 0, 0, 0],
+           [0, 4, 0, 0, 0],
+           [0, 0, 4, 0, 0]])
+
+    The anti-diagonal can be filled by reversing the order of elements
+    using either `numpy.flipud` or `numpy.fliplr`.
+
+    >>> a = np.zeros((3, 3), int);
+    >>> np.fill_diagonal(np.fliplr(a), [1,2,3])  # Horizontal flip
+    >>> a
+    array([[0, 0, 1],
+           [0, 2, 0],
+           [3, 0, 0]])
+    >>> np.fill_diagonal(np.flipud(a), [1,2,3])  # Vertical flip
+    >>> a
+    array([[0, 0, 3],
+           [0, 2, 0],
+           [1, 0, 0]])
+
+    Note that the order in which the diagonal is filled varies depending
+    on the flip function.
+    """
+    if a.ndim < 2:
+        raise ValueError("array must be at least 2-d")
+    end = None
+    if a.ndim == 2:
+        # Explicit, fast formula for the common case.  For 2-d arrays, we
+        # accept rectangular ones.
+        step = a.shape[1] + 1
+        # This is needed to don't have tall matrix have the diagonal wrap.
+        if not wrap:
+            end = a.shape[1] * a.shape[1]
+    else:
+        # For more than d=2, the strided formula is only valid for arrays with
+        # all dimensions equal, so we check first.
+        if not np.all(diff(a.shape) == 0):
+            raise ValueError("All dimensions of input must be of equal length")
+        step = 1 + (np.cumprod(a.shape[:-1])).sum()
+
+    # Write the value out into the diagonal.
+    a.flat[:end:step] = val
+
+
+@set_module('numpy')
+def diag_indices(n, ndim=2):
+    """
+    Return the indices to access the main diagonal of an array.
+
+    This returns a tuple of indices that can be used to access the main
+    diagonal of an array `a` with ``a.ndim >= 2`` dimensions and shape
+    (n, n, ..., n). For ``a.ndim = 2`` this is the usual diagonal, for
+    ``a.ndim > 2`` this is the set of indices to access ``a[i, i, ..., i]``
+    for ``i = [0..n-1]``.
+
+    Parameters
+    ----------
+    n : int
+      The size, along each dimension, of the arrays for which the returned
+      indices can be used.
+
+    ndim : int, optional
+      The number of dimensions.
+
+    See Also
+    --------
+    diag_indices_from
+
+    Notes
+    -----
+    .. versionadded:: 1.4.0
+
+    Examples
+    --------
+    Create a set of indices to access the diagonal of a (4, 4) array:
+
+    >>> di = np.diag_indices(4)
+    >>> di
+    (array([0, 1, 2, 3]), array([0, 1, 2, 3]))
+    >>> a = np.arange(16).reshape(4, 4)
+    >>> a
+    array([[ 0,  1,  2,  3],
+           [ 4,  5,  6,  7],
+           [ 8,  9, 10, 11],
+           [12, 13, 14, 15]])
+    >>> a[di] = 100
+    >>> a
+    array([[100,   1,   2,   3],
+           [  4, 100,   6,   7],
+           [  8,   9, 100,  11],
+           [ 12,  13,  14, 100]])
+
+    Now, we create indices to manipulate a 3-D array:
+
+    >>> d3 = np.diag_indices(2, 3)
+    >>> d3
+    (array([0, 1]), array([0, 1]), array([0, 1]))
+
+    And use it to set the diagonal of an array of zeros to 1:
+
+    >>> a = np.zeros((2, 2, 2), dtype=int)
+    >>> a[d3] = 1
+    >>> a
+    array([[[1, 0],
+            [0, 0]],
+           [[0, 0],
+            [0, 1]]])
+
+    """
+    idx = np.arange(n)
+    return (idx,) * ndim
+
+
+def _diag_indices_from(arr):
+    return (arr,)
+
+
+@array_function_dispatch(_diag_indices_from)
+def diag_indices_from(arr):
+    """
+    Return the indices to access the main diagonal of an n-dimensional array.
+
+    See `diag_indices` for full details.
+
+    Parameters
+    ----------
+    arr : array, at least 2-D
+
+    See Also
+    --------
+    diag_indices
+
+    Notes
+    -----
+    .. versionadded:: 1.4.0
+
+    Examples
+    --------
+    
+    Create a 4 by 4 array.
+
+    >>> a = np.arange(16).reshape(4, 4)
+    >>> a
+    array([[ 0,  1,  2,  3],
+           [ 4,  5,  6,  7],
+           [ 8,  9, 10, 11],
+           [12, 13, 14, 15]])
+    
+    Get the indices of the diagonal elements.
+
+    >>> di = np.diag_indices_from(a)
+    >>> di
+    (array([0, 1, 2, 3]), array([0, 1, 2, 3]))
+
+    >>> a[di]
+    array([ 0,  5, 10, 15])
+
+    This is simply syntactic sugar for diag_indices.
+
+    >>> np.diag_indices(a.shape[0])
+    (array([0, 1, 2, 3]), array([0, 1, 2, 3]))
+
+    """
+
+    if not arr.ndim >= 2:
+        raise ValueError("input array must be at least 2-d")
+    # For more than d=2, the strided formula is only valid for arrays with
+    # all dimensions equal, so we check first.
+    if not np.all(diff(arr.shape) == 0):
+        raise ValueError("All dimensions of input must be of equal length")
+
+    return diag_indices(arr.shape[0], arr.ndim)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/mixins.pyi

@@ -0,0 +1,74 @@
+from abc import ABCMeta, abstractmethod
+from typing import Literal as L, Any
+
+from numpy import ufunc
+
+__all__: list[str]
+
+# NOTE: `NDArrayOperatorsMixin` is not formally an abstract baseclass,
+# even though it's reliant on subclasses implementing `__array_ufunc__`
+
+# NOTE: The accepted input- and output-types of the various dunders are
+# completely dependent on how `__array_ufunc__` is implemented.
+# As such, only little type safety can be provided here.
+
+class NDArrayOperatorsMixin(metaclass=ABCMeta):
+    @abstractmethod
+    def __array_ufunc__(
+        self,
+        ufunc: ufunc,
+        method: L["__call__", "reduce", "reduceat", "accumulate", "outer", "inner"],
+        *inputs: Any,
+        **kwargs: Any,
+    ) -> Any: ...
+    def __lt__(self, other: Any) -> Any: ...
+    def __le__(self, other: Any) -> Any: ...
+    def __eq__(self, other: Any) -> Any: ...
+    def __ne__(self, other: Any) -> Any: ...
+    def __gt__(self, other: Any) -> Any: ...
+    def __ge__(self, other: Any) -> Any: ...
+    def __add__(self, other: Any) -> Any: ...
+    def __radd__(self, other: Any) -> Any: ...
+    def __iadd__(self, other: Any) -> Any: ...
+    def __sub__(self, other: Any) -> Any: ...
+    def __rsub__(self, other: Any) -> Any: ...
+    def __isub__(self, other: Any) -> Any: ...
+    def __mul__(self, other: Any) -> Any: ...
+    def __rmul__(self, other: Any) -> Any: ...
+    def __imul__(self, other: Any) -> Any: ...
+    def __matmul__(self, other: Any) -> Any: ...
+    def __rmatmul__(self, other: Any) -> Any: ...
+    def __imatmul__(self, other: Any) -> Any: ...
+    def __truediv__(self, other: Any) -> Any: ...
+    def __rtruediv__(self, other: Any) -> Any: ...
+    def __itruediv__(self, other: Any) -> Any: ...
+    def __floordiv__(self, other: Any) -> Any: ...
+    def __rfloordiv__(self, other: Any) -> Any: ...
+    def __ifloordiv__(self, other: Any) -> Any: ...
+    def __mod__(self, other: Any) -> Any: ...
+    def __rmod__(self, other: Any) -> Any: ...
+    def __imod__(self, other: Any) -> Any: ...
+    def __divmod__(self, other: Any) -> Any: ...
+    def __rdivmod__(self, other: Any) -> Any: ...
+    def __pow__(self, other: Any) -> Any: ...
+    def __rpow__(self, other: Any) -> Any: ...
+    def __ipow__(self, other: Any) -> Any: ...
+    def __lshift__(self, other: Any) -> Any: ...
+    def __rlshift__(self, other: Any) -> Any: ...
+    def __ilshift__(self, other: Any) -> Any: ...
+    def __rshift__(self, other: Any) -> Any: ...
+    def __rrshift__(self, other: Any) -> Any: ...
+    def __irshift__(self, other: Any) -> Any: ...
+    def __and__(self, other: Any) -> Any: ...
+    def __rand__(self, other: Any) -> Any: ...
+    def __iand__(self, other: Any) -> Any: ...
+    def __xor__(self, other: Any) -> Any: ...
+    def __rxor__(self, other: Any) -> Any: ...
+    def __ixor__(self, other: Any) -> Any: ...
+    def __or__(self, other: Any) -> Any: ...
+    def __ror__(self, other: Any) -> Any: ...
+    def __ior__(self, other: Any) -> Any: ...
+    def __neg__(self) -> Any: ...
+    def __pos__(self) -> Any: ...
+    def __abs__(self) -> Any: ...
+    def __invert__(self) -> Any: ...

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/polynomial.pyi

@@ -0,0 +1,303 @@
+from typing import (
+    Literal as L,
+    overload,
+    Any,
+    SupportsInt,
+    SupportsIndex,
+    TypeVar,
+    NoReturn,
+)
+
+from numpy import (
+    RankWarning as RankWarning,
+    poly1d as poly1d,
+    unsignedinteger,
+    signedinteger,
+    floating,
+    complexfloating,
+    bool_,
+    int32,
+    int64,
+    float64,
+    complex128,
+    object_,
+)
+
+from numpy._typing import (
+    NDArray,
+    ArrayLike,
+    _ArrayLikeBool_co,
+    _ArrayLikeUInt_co,
+    _ArrayLikeInt_co,
+    _ArrayLikeFloat_co,
+    _ArrayLikeComplex_co,
+    _ArrayLikeObject_co,
+)
+
+_T = TypeVar("_T")
+
+_2Tup = tuple[_T, _T]
+_5Tup = tuple[
+    _T,
+    NDArray[float64],
+    NDArray[int32],
+    NDArray[float64],
+    NDArray[float64],
+]
+
+__all__: list[str]
+
+def poly(seq_of_zeros: ArrayLike) -> NDArray[floating[Any]]: ...
+
+# Returns either a float or complex array depending on the input values.
+# See `np.linalg.eigvals`.
+def roots(p: ArrayLike) -> NDArray[complexfloating[Any, Any]] | NDArray[floating[Any]]: ...
+
+@overload
+def polyint(
+    p: poly1d,
+    m: SupportsInt | SupportsIndex = ...,
+    k: None | _ArrayLikeComplex_co | _ArrayLikeObject_co = ...,
+) -> poly1d: ...
+@overload
+def polyint(
+    p: _ArrayLikeFloat_co,
+    m: SupportsInt | SupportsIndex = ...,
+    k: None | _ArrayLikeFloat_co = ...,
+) -> NDArray[floating[Any]]: ...
+@overload
+def polyint(
+    p: _ArrayLikeComplex_co,
+    m: SupportsInt | SupportsIndex = ...,
+    k: None | _ArrayLikeComplex_co = ...,
+) -> NDArray[complexfloating[Any, Any]]: ...
+@overload
+def polyint(
+    p: _ArrayLikeObject_co,
+    m: SupportsInt | SupportsIndex = ...,
+    k: None | _ArrayLikeObject_co = ...,
+) -> NDArray[object_]: ...
+
+@overload
+def polyder(
+    p: poly1d,
+    m: SupportsInt | SupportsIndex = ...,
+) -> poly1d: ...
+@overload
+def polyder(
+    p: _ArrayLikeFloat_co,
+    m: SupportsInt | SupportsIndex = ...,
+) -> NDArray[floating[Any]]: ...
+@overload
+def polyder(
+    p: _ArrayLikeComplex_co,
+    m: SupportsInt | SupportsIndex = ...,
+) -> NDArray[complexfloating[Any, Any]]: ...
+@overload
+def polyder(
+    p: _ArrayLikeObject_co,
+    m: SupportsInt | SupportsIndex = ...,
+) -> NDArray[object_]: ...
+
+@overload
+def polyfit(
+    x: _ArrayLikeFloat_co,
+    y: _ArrayLikeFloat_co,
+    deg: SupportsIndex | SupportsInt,
+    rcond: None | float = ...,
+    full: L[False] = ...,
+    w: None | _ArrayLikeFloat_co = ...,
+    cov: L[False] = ...,
+) -> NDArray[float64]: ...
+@overload
+def polyfit(
+    x: _ArrayLikeComplex_co,
+    y: _ArrayLikeComplex_co,
+    deg: SupportsIndex | SupportsInt,
+    rcond: None | float = ...,
+    full: L[False] = ...,
+    w: None | _ArrayLikeFloat_co = ...,
+    cov: L[False] = ...,
+) -> NDArray[complex128]: ...
+@overload
+def polyfit(
+    x: _ArrayLikeFloat_co,
+    y: _ArrayLikeFloat_co,
+    deg: SupportsIndex | SupportsInt,
+    rcond: None | float = ...,
+    full: L[False] = ...,
+    w: None | _ArrayLikeFloat_co = ...,
+    cov: L[True, "unscaled"] = ...,
+) -> _2Tup[NDArray[float64]]: ...
+@overload
+def polyfit(
+    x: _ArrayLikeComplex_co,
+    y: _ArrayLikeComplex_co,
+    deg: SupportsIndex | SupportsInt,
+    rcond: None | float = ...,
+    full: L[False] = ...,
+    w: None | _ArrayLikeFloat_co = ...,
+    cov: L[True, "unscaled"] = ...,
+) -> _2Tup[NDArray[complex128]]: ...
+@overload
+def polyfit(
+    x: _ArrayLikeFloat_co,
+    y: _ArrayLikeFloat_co,
+    deg: SupportsIndex | SupportsInt,
+    rcond: None | float = ...,
+    full: L[True] = ...,
+    w: None | _ArrayLikeFloat_co = ...,
+    cov: bool | L["unscaled"] = ...,
+) -> _5Tup[NDArray[float64]]: ...
+@overload
+def polyfit(
+    x: _ArrayLikeComplex_co,
+    y: _ArrayLikeComplex_co,
+    deg: SupportsIndex | SupportsInt,
+    rcond: None | float = ...,
+    full: L[True] = ...,
+    w: None | _ArrayLikeFloat_co = ...,
+    cov: bool | L["unscaled"] = ...,
+) -> _5Tup[NDArray[complex128]]: ...
+
+@overload
+def polyval(
+    p: _ArrayLikeBool_co,
+    x: _ArrayLikeBool_co,
+) -> NDArray[int64]: ...
+@overload
+def polyval(
+    p: _ArrayLikeUInt_co,
+    x: _ArrayLikeUInt_co,
+) -> NDArray[unsignedinteger[Any]]: ...
+@overload
+def polyval(
+    p: _ArrayLikeInt_co,
+    x: _ArrayLikeInt_co,
+) -> NDArray[signedinteger[Any]]: ...
+@overload
+def polyval(
+    p: _ArrayLikeFloat_co,
+    x: _ArrayLikeFloat_co,
+) -> NDArray[floating[Any]]: ...
+@overload
+def polyval(
+    p: _ArrayLikeComplex_co,
+    x: _ArrayLikeComplex_co,
+) -> NDArray[complexfloating[Any, Any]]: ...
+@overload
+def polyval(
+    p: _ArrayLikeObject_co,
+    x: _ArrayLikeObject_co,
+) -> NDArray[object_]: ...
+
+@overload
+def polyadd(
+    a1: poly1d,
+    a2: _ArrayLikeComplex_co | _ArrayLikeObject_co,
+) -> poly1d: ...
+@overload
+def polyadd(
+    a1: _ArrayLikeComplex_co | _ArrayLikeObject_co,
+    a2: poly1d,
+) -> poly1d: ...
+@overload
+def polyadd(
+    a1: _ArrayLikeBool_co,
+    a2: _ArrayLikeBool_co,
+) -> NDArray[bool_]: ...
+@overload
+def polyadd(
+    a1: _ArrayLikeUInt_co,
+    a2: _ArrayLikeUInt_co,
+) -> NDArray[unsignedinteger[Any]]: ...
+@overload
+def polyadd(
+    a1: _ArrayLikeInt_co,
+    a2: _ArrayLikeInt_co,
+) -> NDArray[signedinteger[Any]]: ...
+@overload
+def polyadd(
+    a1: _ArrayLikeFloat_co,
+    a2: _ArrayLikeFloat_co,
+) -> NDArray[floating[Any]]: ...
+@overload
+def polyadd(
+    a1: _ArrayLikeComplex_co,
+    a2: _ArrayLikeComplex_co,
+) -> NDArray[complexfloating[Any, Any]]: ...
+@overload
+def polyadd(
+    a1: _ArrayLikeObject_co,
+    a2: _ArrayLikeObject_co,
+) -> NDArray[object_]: ...
+
+@overload
+def polysub(
+    a1: poly1d,
+    a2: _ArrayLikeComplex_co | _ArrayLikeObject_co,
+) -> poly1d: ...
+@overload
+def polysub(
+    a1: _ArrayLikeComplex_co | _ArrayLikeObject_co,
+    a2: poly1d,
+) -> poly1d: ...
+@overload
+def polysub(
+    a1: _ArrayLikeBool_co,
+    a2: _ArrayLikeBool_co,
+) -> NoReturn: ...
+@overload
+def polysub(
+    a1: _ArrayLikeUInt_co,
+    a2: _ArrayLikeUInt_co,
+) -> NDArray[unsignedinteger[Any]]: ...
+@overload
+def polysub(
+    a1: _ArrayLikeInt_co,
+    a2: _ArrayLikeInt_co,
+) -> NDArray[signedinteger[Any]]: ...
+@overload
+def polysub(
+    a1: _ArrayLikeFloat_co,
+    a2: _ArrayLikeFloat_co,
+) -> NDArray[floating[Any]]: ...
+@overload
+def polysub(
+    a1: _ArrayLikeComplex_co,
+    a2: _ArrayLikeComplex_co,
+) -> NDArray[complexfloating[Any, Any]]: ...
+@overload
+def polysub(
+    a1: _ArrayLikeObject_co,
+    a2: _ArrayLikeObject_co,
+) -> NDArray[object_]: ...
+
+# NOTE: Not an alias, but they do have the same signature (that we can reuse)
+polymul = polyadd
+
+@overload
+def polydiv(
+    u: poly1d,
+    v: _ArrayLikeComplex_co | _ArrayLikeObject_co,
+) -> _2Tup[poly1d]: ...
+@overload
+def polydiv(
+    u: _ArrayLikeComplex_co | _ArrayLikeObject_co,
+    v: poly1d,
+) -> _2Tup[poly1d]: ...
+@overload
+def polydiv(
+    u: _ArrayLikeFloat_co,
+    v: _ArrayLikeFloat_co,
+) -> _2Tup[NDArray[floating[Any]]]: ...
+@overload
+def polydiv(
+    u: _ArrayLikeComplex_co,
+    v: _ArrayLikeComplex_co,
+) -> _2Tup[NDArray[complexfloating[Any, Any]]]: ...
+@overload
+def polydiv(
+    u: _ArrayLikeObject_co,
+    v: _ArrayLikeObject_co,
+) -> _2Tup[NDArray[Any]]: ...

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/recfunctions.py

@@ -0,0 +1,1673 @@
+"""
+Collection of utilities to manipulate structured arrays.
+
+Most of these functions were initially implemented by John Hunter for
+matplotlib.  They have been rewritten and extended for convenience.
+
+"""
+import itertools
+import numpy as np
+import numpy.ma as ma
+from numpy import ndarray, recarray
+from numpy.ma import MaskedArray
+from numpy.ma.mrecords import MaskedRecords
+from numpy.core.overrides import array_function_dispatch
+from numpy.lib._iotools import _is_string_like
+
+_check_fill_value = np.ma.core._check_fill_value
+
+
+__all__ = [
+    'append_fields', 'apply_along_fields', 'assign_fields_by_name',
+    'drop_fields', 'find_duplicates', 'flatten_descr',
+    'get_fieldstructure', 'get_names', 'get_names_flat',
+    'join_by', 'merge_arrays', 'rec_append_fields',
+    'rec_drop_fields', 'rec_join', 'recursive_fill_fields',
+    'rename_fields', 'repack_fields', 'require_fields',
+    'stack_arrays', 'structured_to_unstructured', 'unstructured_to_structured',
+    ]
+
+
+def _recursive_fill_fields_dispatcher(input, output):
+    return (input, output)
+
+
+@array_function_dispatch(_recursive_fill_fields_dispatcher)
+def recursive_fill_fields(input, output):
+    """
+    Fills fields from output with fields from input,
+    with support for nested structures.
+
+    Parameters
+    ----------
+    input : ndarray
+        Input array.
+    output : ndarray
+        Output array.
+
+    Notes
+    -----
+    * `output` should be at least the same size as `input`
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> a = np.array([(1, 10.), (2, 20.)], dtype=[('A', np.int64), ('B', np.float64)])
+    >>> b = np.zeros((3,), dtype=a.dtype)
+    >>> rfn.recursive_fill_fields(a, b)
+    array([(1, 10.), (2, 20.), (0,  0.)], dtype=[('A', '<i8'), ('B', '<f8')])
+
+    """
+    newdtype = output.dtype
+    for field in newdtype.names:
+        try:
+            current = input[field]
+        except ValueError:
+            continue
+        if current.dtype.names is not None:
+            recursive_fill_fields(current, output[field])
+        else:
+            output[field][:len(current)] = current
+    return output
+
+
+def _get_fieldspec(dtype):
+    """
+    Produce a list of name/dtype pairs corresponding to the dtype fields
+
+    Similar to dtype.descr, but the second item of each tuple is a dtype, not a
+    string. As a result, this handles subarray dtypes
+
+    Can be passed to the dtype constructor to reconstruct the dtype, noting that
+    this (deliberately) discards field offsets.
+
+    Examples
+    --------
+    >>> dt = np.dtype([(('a', 'A'), np.int64), ('b', np.double, 3)])
+    >>> dt.descr
+    [(('a', 'A'), '<i8'), ('b', '<f8', (3,))]
+    >>> _get_fieldspec(dt)
+    [(('a', 'A'), dtype('int64')), ('b', dtype(('<f8', (3,))))]
+
+    """
+    if dtype.names is None:
+        # .descr returns a nameless field, so we should too
+        return [('', dtype)]
+    else:
+        fields = ((name, dtype.fields[name]) for name in dtype.names)
+        # keep any titles, if present
+        return [
+            (name if len(f) == 2 else (f[2], name), f[0])
+            for name, f in fields
+        ]
+
+
+def get_names(adtype):
+    """
+    Returns the field names of the input datatype as a tuple. Input datatype
+    must have fields otherwise error is raised.
+
+    Parameters
+    ----------
+    adtype : dtype
+        Input datatype
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> rfn.get_names(np.empty((1,), dtype=[('A', int)]).dtype)
+    ('A',)
+    >>> rfn.get_names(np.empty((1,), dtype=[('A',int), ('B', float)]).dtype)
+    ('A', 'B')
+    >>> adtype = np.dtype([('a', int), ('b', [('ba', int), ('bb', int)])])
+    >>> rfn.get_names(adtype)
+    ('a', ('b', ('ba', 'bb')))
+    """
+    listnames = []
+    names = adtype.names
+    for name in names:
+        current = adtype[name]
+        if current.names is not None:
+            listnames.append((name, tuple(get_names(current))))
+        else:
+            listnames.append(name)
+    return tuple(listnames)
+
+
+def get_names_flat(adtype):
+    """
+    Returns the field names of the input datatype as a tuple. Input datatype
+    must have fields otherwise error is raised.
+    Nested structure are flattened beforehand.
+
+    Parameters
+    ----------
+    adtype : dtype
+        Input datatype
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> rfn.get_names_flat(np.empty((1,), dtype=[('A', int)]).dtype) is None
+    False
+    >>> rfn.get_names_flat(np.empty((1,), dtype=[('A',int), ('B', str)]).dtype)
+    ('A', 'B')
+    >>> adtype = np.dtype([('a', int), ('b', [('ba', int), ('bb', int)])])
+    >>> rfn.get_names_flat(adtype)
+    ('a', 'b', 'ba', 'bb')
+    """
+    listnames = []
+    names = adtype.names
+    for name in names:
+        listnames.append(name)
+        current = adtype[name]
+        if current.names is not None:
+            listnames.extend(get_names_flat(current))
+    return tuple(listnames)
+
+
+def flatten_descr(ndtype):
+    """
+    Flatten a structured data-type description.
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> ndtype = np.dtype([('a', '<i4'), ('b', [('ba', '<f8'), ('bb', '<i4')])])
+    >>> rfn.flatten_descr(ndtype)
+    (('a', dtype('int32')), ('ba', dtype('float64')), ('bb', dtype('int32')))
+
+    """
+    names = ndtype.names
+    if names is None:
+        return (('', ndtype),)
+    else:
+        descr = []
+        for field in names:
+            (typ, _) = ndtype.fields[field]
+            if typ.names is not None:
+                descr.extend(flatten_descr(typ))
+            else:
+                descr.append((field, typ))
+        return tuple(descr)
+
+
+def _zip_dtype(seqarrays, flatten=False):
+    newdtype = []
+    if flatten:
+        for a in seqarrays:
+            newdtype.extend(flatten_descr(a.dtype))
+    else:
+        for a in seqarrays:
+            current = a.dtype
+            if current.names is not None and len(current.names) == 1:
+                # special case - dtypes of 1 field are flattened
+                newdtype.extend(_get_fieldspec(current))
+            else:
+                newdtype.append(('', current))
+    return np.dtype(newdtype)
+
+
+def _zip_descr(seqarrays, flatten=False):
+    """
+    Combine the dtype description of a series of arrays.
+
+    Parameters
+    ----------
+    seqarrays : sequence of arrays
+        Sequence of arrays
+    flatten : {boolean}, optional
+        Whether to collapse nested descriptions.
+    """
+    return _zip_dtype(seqarrays, flatten=flatten).descr
+
+
+def get_fieldstructure(adtype, lastname=None, parents=None,):
+    """
+    Returns a dictionary with fields indexing lists of their parent fields.
+
+    This function is used to simplify access to fields nested in other fields.
+
+    Parameters
+    ----------
+    adtype : np.dtype
+        Input datatype
+    lastname : optional
+        Last processed field name (used internally during recursion).
+    parents : dictionary
+        Dictionary of parent fields (used interbally during recursion).
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> ndtype =  np.dtype([('A', int),
+    ...                     ('B', [('BA', int),
+    ...                            ('BB', [('BBA', int), ('BBB', int)])])])
+    >>> rfn.get_fieldstructure(ndtype)
+    ... # XXX: possible regression, order of BBA and BBB is swapped
+    {'A': [], 'B': [], 'BA': ['B'], 'BB': ['B'], 'BBA': ['B', 'BB'], 'BBB': ['B', 'BB']}
+
+    """
+    if parents is None:
+        parents = {}
+    names = adtype.names
+    for name in names:
+        current = adtype[name]
+        if current.names is not None:
+            if lastname:
+                parents[name] = [lastname, ]
+            else:
+                parents[name] = []
+            parents.update(get_fieldstructure(current, name, parents))
+        else:
+            lastparent = [_ for _ in (parents.get(lastname, []) or [])]
+            if lastparent:
+                lastparent.append(lastname)
+            elif lastname:
+                lastparent = [lastname, ]
+            parents[name] = lastparent or []
+    return parents
+
+
+def _izip_fields_flat(iterable):
+    """
+    Returns an iterator of concatenated fields from a sequence of arrays,
+    collapsing any nested structure.
+
+    """
+    for element in iterable:
+        if isinstance(element, np.void):
+            yield from _izip_fields_flat(tuple(element))
+        else:
+            yield element
+
+
+def _izip_fields(iterable):
+    """
+    Returns an iterator of concatenated fields from a sequence of arrays.
+
+    """
+    for element in iterable:
+        if (hasattr(element, '__iter__') and
+                not isinstance(element, str)):
+            yield from _izip_fields(element)
+        elif isinstance(element, np.void) and len(tuple(element)) == 1:
+            # this statement is the same from the previous expression
+            yield from _izip_fields(element)
+        else:
+            yield element
+
+
+def _izip_records(seqarrays, fill_value=None, flatten=True):
+    """
+    Returns an iterator of concatenated items from a sequence of arrays.
+
+    Parameters
+    ----------
+    seqarrays : sequence of arrays
+        Sequence of arrays.
+    fill_value : {None, integer}
+        Value used to pad shorter iterables.
+    flatten : {True, False},
+        Whether to
+    """
+
+    # Should we flatten the items, or just use a nested approach
+    if flatten:
+        zipfunc = _izip_fields_flat
+    else:
+        zipfunc = _izip_fields
+
+    for tup in itertools.zip_longest(*seqarrays, fillvalue=fill_value):
+        yield tuple(zipfunc(tup))
+
+
+def _fix_output(output, usemask=True, asrecarray=False):
+    """
+    Private function: return a recarray, a ndarray, a MaskedArray
+    or a MaskedRecords depending on the input parameters
+    """
+    if not isinstance(output, MaskedArray):
+        usemask = False
+    if usemask:
+        if asrecarray:
+            output = output.view(MaskedRecords)
+    else:
+        output = ma.filled(output)
+        if asrecarray:
+            output = output.view(recarray)
+    return output
+
+
+def _fix_defaults(output, defaults=None):
+    """
+    Update the fill_value and masked data of `output`
+    from the default given in a dictionary defaults.
+    """
+    names = output.dtype.names
+    (data, mask, fill_value) = (output.data, output.mask, output.fill_value)
+    for (k, v) in (defaults or {}).items():
+        if k in names:
+            fill_value[k] = v
+            data[k][mask[k]] = v
+    return output
+
+
+def _merge_arrays_dispatcher(seqarrays, fill_value=None, flatten=None,
+                             usemask=None, asrecarray=None):
+    return seqarrays
+
+
+@array_function_dispatch(_merge_arrays_dispatcher)
+def merge_arrays(seqarrays, fill_value=-1, flatten=False,
+                 usemask=False, asrecarray=False):
+    """
+    Merge arrays field by field.
+
+    Parameters
+    ----------
+    seqarrays : sequence of ndarrays
+        Sequence of arrays
+    fill_value : {float}, optional
+        Filling value used to pad missing data on the shorter arrays.
+    flatten : {False, True}, optional
+        Whether to collapse nested fields.
+    usemask : {False, True}, optional
+        Whether to return a masked array or not.
+    asrecarray : {False, True}, optional
+        Whether to return a recarray (MaskedRecords) or not.
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> rfn.merge_arrays((np.array([1, 2]), np.array([10., 20., 30.])))
+    array([( 1, 10.), ( 2, 20.), (-1, 30.)],
+          dtype=[('f0', '<i8'), ('f1', '<f8')])
+
+    >>> rfn.merge_arrays((np.array([1, 2], dtype=np.int64),
+    ...         np.array([10., 20., 30.])), usemask=False)
+     array([(1, 10.0), (2, 20.0), (-1, 30.0)],
+             dtype=[('f0', '<i8'), ('f1', '<f8')])
+    >>> rfn.merge_arrays((np.array([1, 2]).view([('a', np.int64)]),
+    ...               np.array([10., 20., 30.])),
+    ...              usemask=False, asrecarray=True)
+    rec.array([( 1, 10.), ( 2, 20.), (-1, 30.)],
+              dtype=[('a', '<i8'), ('f1', '<f8')])
+
+    Notes
+    -----
+    * Without a mask, the missing value will be filled with something,
+      depending on what its corresponding type:
+
+      * ``-1``      for integers
+      * ``-1.0``    for floating point numbers
+      * ``'-'``     for characters
+      * ``'-1'``    for strings
+      * ``True``    for boolean values
+    * XXX: I just obtained these values empirically
+    """
+    # Only one item in the input sequence ?
+    if (len(seqarrays) == 1):
+        seqarrays = np.asanyarray(seqarrays[0])
+    # Do we have a single ndarray as input ?
+    if isinstance(seqarrays, (ndarray, np.void)):
+        seqdtype = seqarrays.dtype
+        # Make sure we have named fields
+        if seqdtype.names is None:
+            seqdtype = np.dtype([('', seqdtype)])
+        if not flatten or _zip_dtype((seqarrays,), flatten=True) == seqdtype:
+            # Minimal processing needed: just make sure everything's a-ok
+            seqarrays = seqarrays.ravel()
+            # Find what type of array we must return
+            if usemask:
+                if asrecarray:
+                    seqtype = MaskedRecords
+                else:
+                    seqtype = MaskedArray
+            elif asrecarray:
+                seqtype = recarray
+            else:
+                seqtype = ndarray
+            return seqarrays.view(dtype=seqdtype, type=seqtype)
+        else:
+            seqarrays = (seqarrays,)
+    else:
+        # Make sure we have arrays in the input sequence
+        seqarrays = [np.asanyarray(_m) for _m in seqarrays]
+    # Find the sizes of the inputs and their maximum
+    sizes = tuple(a.size for a in seqarrays)
+    maxlength = max(sizes)
+    # Get the dtype of the output (flattening if needed)
+    newdtype = _zip_dtype(seqarrays, flatten=flatten)
+    # Initialize the sequences for data and mask
+    seqdata = []
+    seqmask = []
+    # If we expect some kind of MaskedArray, make a special loop.
+    if usemask:
+        for (a, n) in zip(seqarrays, sizes):
+            nbmissing = (maxlength - n)
+            # Get the data and mask
+            data = a.ravel().__array__()
+            mask = ma.getmaskarray(a).ravel()
+            # Get the filling value (if needed)
+            if nbmissing:
+                fval = _check_fill_value(fill_value, a.dtype)
+                if isinstance(fval, (ndarray, np.void)):
+                    if len(fval.dtype) == 1:
+                        fval = fval.item()[0]
+                        fmsk = True
+                    else:
+                        fval = np.array(fval, dtype=a.dtype, ndmin=1)
+                        fmsk = np.ones((1,), dtype=mask.dtype)
+            else:
+                fval = None
+                fmsk = True
+            # Store an iterator padding the input to the expected length
+            seqdata.append(itertools.chain(data, [fval] * nbmissing))
+            seqmask.append(itertools.chain(mask, [fmsk] * nbmissing))
+        # Create an iterator for the data
+        data = tuple(_izip_records(seqdata, flatten=flatten))
+        output = ma.array(np.fromiter(data, dtype=newdtype, count=maxlength),
+                          mask=list(_izip_records(seqmask, flatten=flatten)))
+        if asrecarray:
+            output = output.view(MaskedRecords)
+    else:
+        # Same as before, without the mask we don't need...
+        for (a, n) in zip(seqarrays, sizes):
+            nbmissing = (maxlength - n)
+            data = a.ravel().__array__()
+            if nbmissing:
+                fval = _check_fill_value(fill_value, a.dtype)
+                if isinstance(fval, (ndarray, np.void)):
+                    if len(fval.dtype) == 1:
+                        fval = fval.item()[0]
+                    else:
+                        fval = np.array(fval, dtype=a.dtype, ndmin=1)
+            else:
+                fval = None
+            seqdata.append(itertools.chain(data, [fval] * nbmissing))
+        output = np.fromiter(tuple(_izip_records(seqdata, flatten=flatten)),
+                             dtype=newdtype, count=maxlength)
+        if asrecarray:
+            output = output.view(recarray)
+    # And we're done...
+    return output
+
+
+def _drop_fields_dispatcher(base, drop_names, usemask=None, asrecarray=None):
+    return (base,)
+
+
+@array_function_dispatch(_drop_fields_dispatcher)
+def drop_fields(base, drop_names, usemask=True, asrecarray=False):
+    """
+    Return a new array with fields in `drop_names` dropped.
+
+    Nested fields are supported.
+
+    .. versionchanged:: 1.18.0
+        `drop_fields` returns an array with 0 fields if all fields are dropped,
+        rather than returning ``None`` as it did previously.
+
+    Parameters
+    ----------
+    base : array
+        Input array
+    drop_names : string or sequence
+        String or sequence of strings corresponding to the names of the
+        fields to drop.
+    usemask : {False, True}, optional
+        Whether to return a masked array or not.
+    asrecarray : string or sequence, optional
+        Whether to return a recarray or a mrecarray (`asrecarray=True`) or
+        a plain ndarray or masked array with flexible dtype. The default
+        is False.
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> a = np.array([(1, (2, 3.0)), (4, (5, 6.0))],
+    ...   dtype=[('a', np.int64), ('b', [('ba', np.double), ('bb', np.int64)])])
+    >>> rfn.drop_fields(a, 'a')
+    array([((2., 3),), ((5., 6),)],
+          dtype=[('b', [('ba', '<f8'), ('bb', '<i8')])])
+    >>> rfn.drop_fields(a, 'ba')
+    array([(1, (3,)), (4, (6,))], dtype=[('a', '<i8'), ('b', [('bb', '<i8')])])
+    >>> rfn.drop_fields(a, ['ba', 'bb'])
+    array([(1,), (4,)], dtype=[('a', '<i8')])
+    """
+    if _is_string_like(drop_names):
+        drop_names = [drop_names]
+    else:
+        drop_names = set(drop_names)
+
+    def _drop_descr(ndtype, drop_names):
+        names = ndtype.names
+        newdtype = []
+        for name in names:
+            current = ndtype[name]
+            if name in drop_names:
+                continue
+            if current.names is not None:
+                descr = _drop_descr(current, drop_names)
+                if descr:
+                    newdtype.append((name, descr))
+            else:
+                newdtype.append((name, current))
+        return newdtype
+
+    newdtype = _drop_descr(base.dtype, drop_names)
+
+    output = np.empty(base.shape, dtype=newdtype)
+    output = recursive_fill_fields(base, output)
+    return _fix_output(output, usemask=usemask, asrecarray=asrecarray)
+
+
+def _keep_fields(base, keep_names, usemask=True, asrecarray=False):
+    """
+    Return a new array keeping only the fields in `keep_names`,
+    and preserving the order of those fields.
+
+    Parameters
+    ----------
+    base : array
+        Input array
+    keep_names : string or sequence
+        String or sequence of strings corresponding to the names of the
+        fields to keep. Order of the names will be preserved.
+    usemask : {False, True}, optional
+        Whether to return a masked array or not.
+    asrecarray : string or sequence, optional
+        Whether to return a recarray or a mrecarray (`asrecarray=True`) or
+        a plain ndarray or masked array with flexible dtype. The default
+        is False.
+    """
+    newdtype = [(n, base.dtype[n]) for n in keep_names]
+    output = np.empty(base.shape, dtype=newdtype)
+    output = recursive_fill_fields(base, output)
+    return _fix_output(output, usemask=usemask, asrecarray=asrecarray)
+
+
+def _rec_drop_fields_dispatcher(base, drop_names):
+    return (base,)
+
+
+@array_function_dispatch(_rec_drop_fields_dispatcher)
+def rec_drop_fields(base, drop_names):
+    """
+    Returns a new numpy.recarray with fields in `drop_names` dropped.
+    """
+    return drop_fields(base, drop_names, usemask=False, asrecarray=True)
+
+
+def _rename_fields_dispatcher(base, namemapper):
+    return (base,)
+
+
+@array_function_dispatch(_rename_fields_dispatcher)
+def rename_fields(base, namemapper):
+    """
+    Rename the fields from a flexible-datatype ndarray or recarray.
+
+    Nested fields are supported.
+
+    Parameters
+    ----------
+    base : ndarray
+        Input array whose fields must be modified.
+    namemapper : dictionary
+        Dictionary mapping old field names to their new version.
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> a = np.array([(1, (2, [3.0, 30.])), (4, (5, [6.0, 60.]))],
+    ...   dtype=[('a', int),('b', [('ba', float), ('bb', (float, 2))])])
+    >>> rfn.rename_fields(a, {'a':'A', 'bb':'BB'})
+    array([(1, (2., [ 3., 30.])), (4, (5., [ 6., 60.]))],
+          dtype=[('A', '<i8'), ('b', [('ba', '<f8'), ('BB', '<f8', (2,))])])
+
+    """
+    def _recursive_rename_fields(ndtype, namemapper):
+        newdtype = []
+        for name in ndtype.names:
+            newname = namemapper.get(name, name)
+            current = ndtype[name]
+            if current.names is not None:
+                newdtype.append(
+                    (newname, _recursive_rename_fields(current, namemapper))
+                    )
+            else:
+                newdtype.append((newname, current))
+        return newdtype
+    newdtype = _recursive_rename_fields(base.dtype, namemapper)
+    return base.view(newdtype)
+
+
+def _append_fields_dispatcher(base, names, data, dtypes=None,
+                              fill_value=None, usemask=None, asrecarray=None):
+    yield base
+    yield from data
+
+
+@array_function_dispatch(_append_fields_dispatcher)
+def append_fields(base, names, data, dtypes=None,
+                  fill_value=-1, usemask=True, asrecarray=False):
+    """
+    Add new fields to an existing array.
+
+    The names of the fields are given with the `names` arguments,
+    the corresponding values with the `data` arguments.
+    If a single field is appended, `names`, `data` and `dtypes` do not have
+    to be lists but just values.
+
+    Parameters
+    ----------
+    base : array
+        Input array to extend.
+    names : string, sequence
+        String or sequence of strings corresponding to the names
+        of the new fields.
+    data : array or sequence of arrays
+        Array or sequence of arrays storing the fields to add to the base.
+    dtypes : sequence of datatypes, optional
+        Datatype or sequence of datatypes.
+        If None, the datatypes are estimated from the `data`.
+    fill_value : {float}, optional
+        Filling value used to pad missing data on the shorter arrays.
+    usemask : {False, True}, optional
+        Whether to return a masked array or not.
+    asrecarray : {False, True}, optional
+        Whether to return a recarray (MaskedRecords) or not.
+
+    """
+    # Check the names
+    if isinstance(names, (tuple, list)):
+        if len(names) != len(data):
+            msg = "The number of arrays does not match the number of names"
+            raise ValueError(msg)
+    elif isinstance(names, str):
+        names = [names, ]
+        data = [data, ]
+    #
+    if dtypes is None:
+        data = [np.array(a, copy=False, subok=True) for a in data]
+        data = [a.view([(name, a.dtype)]) for (name, a) in zip(names, data)]
+    else:
+        if not isinstance(dtypes, (tuple, list)):
+            dtypes = [dtypes, ]
+        if len(data) != len(dtypes):
+            if len(dtypes) == 1:
+                dtypes = dtypes * len(data)
+            else:
+                msg = "The dtypes argument must be None, a dtype, or a list."
+                raise ValueError(msg)
+        data = [np.array(a, copy=False, subok=True, dtype=d).view([(n, d)])
+                for (a, n, d) in zip(data, names, dtypes)]
+    #
+    base = merge_arrays(base, usemask=usemask, fill_value=fill_value)
+    if len(data) > 1:
+        data = merge_arrays(data, flatten=True, usemask=usemask,
+                            fill_value=fill_value)
+    else:
+        data = data.pop()
+    #
+    output = ma.masked_all(
+        max(len(base), len(data)),
+        dtype=_get_fieldspec(base.dtype) + _get_fieldspec(data.dtype))
+    output = recursive_fill_fields(base, output)
+    output = recursive_fill_fields(data, output)
+    #
+    return _fix_output(output, usemask=usemask, asrecarray=asrecarray)
+
+
+def _rec_append_fields_dispatcher(base, names, data, dtypes=None):
+    yield base
+    yield from data
+
+
+@array_function_dispatch(_rec_append_fields_dispatcher)
+def rec_append_fields(base, names, data, dtypes=None):
+    """
+    Add new fields to an existing array.
+
+    The names of the fields are given with the `names` arguments,
+    the corresponding values with the `data` arguments.
+    If a single field is appended, `names`, `data` and `dtypes` do not have
+    to be lists but just values.
+
+    Parameters
+    ----------
+    base : array
+        Input array to extend.
+    names : string, sequence
+        String or sequence of strings corresponding to the names
+        of the new fields.
+    data : array or sequence of arrays
+        Array or sequence of arrays storing the fields to add to the base.
+    dtypes : sequence of datatypes, optional
+        Datatype or sequence of datatypes.
+        If None, the datatypes are estimated from the `data`.
+
+    See Also
+    --------
+    append_fields
+
+    Returns
+    -------
+    appended_array : np.recarray
+    """
+    return append_fields(base, names, data=data, dtypes=dtypes,
+                         asrecarray=True, usemask=False)
+
+
+def _repack_fields_dispatcher(a, align=None, recurse=None):
+    return (a,)
+
+
+@array_function_dispatch(_repack_fields_dispatcher)
+def repack_fields(a, align=False, recurse=False):
+    """
+    Re-pack the fields of a structured array or dtype in memory.
+
+    The memory layout of structured datatypes allows fields at arbitrary
+    byte offsets. This means the fields can be separated by padding bytes,
+    their offsets can be non-monotonically increasing, and they can overlap.
+
+    This method removes any overlaps and reorders the fields in memory so they
+    have increasing byte offsets, and adds or removes padding bytes depending
+    on the `align` option, which behaves like the `align` option to
+    `numpy.dtype`.
+
+    If `align=False`, this method produces a "packed" memory layout in which
+    each field starts at the byte the previous field ended, and any padding
+    bytes are removed.
+
+    If `align=True`, this methods produces an "aligned" memory layout in which
+    each field's offset is a multiple of its alignment, and the total itemsize
+    is a multiple of the largest alignment, by adding padding bytes as needed.
+
+    Parameters
+    ----------
+    a : ndarray or dtype
+       array or dtype for which to repack the fields.
+    align : boolean
+       If true, use an "aligned" memory layout, otherwise use a "packed" layout.
+    recurse : boolean
+       If True, also repack nested structures.
+
+    Returns
+    -------
+    repacked : ndarray or dtype
+       Copy of `a` with fields repacked, or `a` itself if no repacking was
+       needed.
+
+    Examples
+    --------
+
+    >>> from numpy.lib import recfunctions as rfn
+    >>> def print_offsets(d):
+    ...     print("offsets:", [d.fields[name][1] for name in d.names])
+    ...     print("itemsize:", d.itemsize)
+    ...
+    >>> dt = np.dtype('u1, <i8, <f8', align=True)
+    >>> dt
+    dtype({'names': ['f0', 'f1', 'f2'], 'formats': ['u1', '<i8', '<f8'], \
+'offsets': [0, 8, 16], 'itemsize': 24}, align=True)
+    >>> print_offsets(dt)
+    offsets: [0, 8, 16]
+    itemsize: 24
+    >>> packed_dt = rfn.repack_fields(dt)
+    >>> packed_dt
+    dtype([('f0', 'u1'), ('f1', '<i8'), ('f2', '<f8')])
+    >>> print_offsets(packed_dt)
+    offsets: [0, 1, 9]
+    itemsize: 17
+
+    """
+    if not isinstance(a, np.dtype):
+        dt = repack_fields(a.dtype, align=align, recurse=recurse)
+        return a.astype(dt, copy=False)
+
+    if a.names is None:
+        return a
+
+    fieldinfo = []
+    for name in a.names:
+        tup = a.fields[name]
+        if recurse:
+            fmt = repack_fields(tup[0], align=align, recurse=True)
+        else:
+            fmt = tup[0]
+
+        if len(tup) == 3:
+            name = (tup[2], name)
+
+        fieldinfo.append((name, fmt))
+
+    dt = np.dtype(fieldinfo, align=align)
+    return np.dtype((a.type, dt))
+
+def _get_fields_and_offsets(dt, offset=0):
+    """
+    Returns a flat list of (dtype, count, offset) tuples of all the
+    scalar fields in the dtype "dt", including nested fields, in left
+    to right order.
+    """
+
+    # counts up elements in subarrays, including nested subarrays, and returns
+    # base dtype and count
+    def count_elem(dt):
+        count = 1
+        while dt.shape != ():
+            for size in dt.shape:
+                count *= size
+            dt = dt.base
+        return dt, count
+
+    fields = []
+    for name in dt.names:
+        field = dt.fields[name]
+        f_dt, f_offset = field[0], field[1]
+        f_dt, n = count_elem(f_dt)
+
+        if f_dt.names is None:
+            fields.append((np.dtype((f_dt, (n,))), n, f_offset + offset))
+        else:
+            subfields = _get_fields_and_offsets(f_dt, f_offset + offset)
+            size = f_dt.itemsize
+
+            for i in range(n):
+                if i == 0:
+                    # optimization: avoid list comprehension if no subarray
+                    fields.extend(subfields)
+                else:
+                    fields.extend([(d, c, o + i*size) for d, c, o in subfields])
+    return fields
+
+def _common_stride(offsets, counts, itemsize):
+    """
+    Returns the stride between the fields, or None if the stride is not
+    constant. The values in "counts" designate the lengths of
+    subarrays. Subarrays are treated as many contiguous fields, with
+    always positive stride.
+    """
+    if len(offsets) <= 1:
+        return itemsize
+
+    negative = offsets[1] < offsets[0]  # negative stride
+    if negative:
+        # reverse, so offsets will be ascending
+        it = zip(reversed(offsets), reversed(counts))
+    else:
+        it = zip(offsets, counts)
+
+    prev_offset = None
+    stride = None
+    for offset, count in it:
+        if count != 1:  # subarray: always c-contiguous
+            if negative:
+                return None  # subarrays can never have a negative stride
+            if stride is None:
+                stride = itemsize
+            if stride != itemsize:
+                return None
+            end_offset = offset + (count - 1) * itemsize
+        else:
+            end_offset = offset
+
+        if prev_offset is not None:
+            new_stride = offset - prev_offset
+            if stride is None:
+                stride = new_stride
+            if stride != new_stride:
+                return None
+
+        prev_offset = end_offset
+
+    if negative:
+        return -stride
+    return stride
+
+
+def _structured_to_unstructured_dispatcher(arr, dtype=None, copy=None,
+                                           casting=None):
+    return (arr,)
+
+@array_function_dispatch(_structured_to_unstructured_dispatcher)
+def structured_to_unstructured(arr, dtype=None, copy=False, casting='unsafe'):
+    """
+    Converts an n-D structured array into an (n+1)-D unstructured array.
+
+    The new array will have a new last dimension equal in size to the
+    number of field-elements of the input array. If not supplied, the output
+    datatype is determined from the numpy type promotion rules applied to all
+    the field datatypes.
+
+    Nested fields, as well as each element of any subarray fields, all count
+    as a single field-elements.
+
+    Parameters
+    ----------
+    arr : ndarray
+       Structured array or dtype to convert. Cannot contain object datatype.
+    dtype : dtype, optional
+       The dtype of the output unstructured array.
+    copy : bool, optional
+        If true, always return a copy. If false, a view is returned if
+        possible, such as when the `dtype` and strides of the fields are
+        suitable and the array subtype is one of `np.ndarray`, `np.recarray`
+        or `np.memmap`.
+
+        .. versionchanged:: 1.25.0
+            A view can now be returned if the fields are separated by a
+            uniform stride.
+
+    casting : {'no', 'equiv', 'safe', 'same_kind', 'unsafe'}, optional
+        See casting argument of `numpy.ndarray.astype`. Controls what kind of
+        data casting may occur.
+
+    Returns
+    -------
+    unstructured : ndarray
+       Unstructured array with one more dimension.
+
+    Examples
+    --------
+
+    >>> from numpy.lib import recfunctions as rfn
+    >>> a = np.zeros(4, dtype=[('a', 'i4'), ('b', 'f4,u2'), ('c', 'f4', 2)])
+    >>> a
+    array([(0, (0., 0), [0., 0.]), (0, (0., 0), [0., 0.]),
+           (0, (0., 0), [0., 0.]), (0, (0., 0), [0., 0.])],
+          dtype=[('a', '<i4'), ('b', [('f0', '<f4'), ('f1', '<u2')]), ('c', '<f4', (2,))])
+    >>> rfn.structured_to_unstructured(a)
+    array([[0., 0., 0., 0., 0.],
+           [0., 0., 0., 0., 0.],
+           [0., 0., 0., 0., 0.],
+           [0., 0., 0., 0., 0.]])
+
+    >>> b = np.array([(1, 2, 5), (4, 5, 7), (7, 8 ,11), (10, 11, 12)],
+    ...              dtype=[('x', 'i4'), ('y', 'f4'), ('z', 'f8')])
+    >>> np.mean(rfn.structured_to_unstructured(b[['x', 'z']]), axis=-1)
+    array([ 3. ,  5.5,  9. , 11. ])
+
+    """
+    if arr.dtype.names is None:
+        raise ValueError('arr must be a structured array')
+
+    fields = _get_fields_and_offsets(arr.dtype)
+    n_fields = len(fields)
+    if n_fields == 0 and dtype is None:
+        raise ValueError("arr has no fields. Unable to guess dtype")
+    elif n_fields == 0:
+        # too many bugs elsewhere for this to work now
+        raise NotImplementedError("arr with no fields is not supported")
+
+    dts, counts, offsets = zip(*fields)
+    names = ['f{}'.format(n) for n in range(n_fields)]
+
+    if dtype is None:
+        out_dtype = np.result_type(*[dt.base for dt in dts])
+    else:
+        out_dtype = np.dtype(dtype)
+
+    # Use a series of views and casts to convert to an unstructured array:
+
+    # first view using flattened fields (doesn't work for object arrays)
+    # Note: dts may include a shape for subarrays
+    flattened_fields = np.dtype({'names': names,
+                                 'formats': dts,
+                                 'offsets': offsets,
+                                 'itemsize': arr.dtype.itemsize})
+    arr = arr.view(flattened_fields)
+
+    # we only allow a few types to be unstructured by manipulating the
+    # strides, because we know it won't work with, for example, np.matrix nor
+    # np.ma.MaskedArray.
+    can_view = type(arr) in (np.ndarray, np.recarray, np.memmap)
+    if (not copy) and can_view and all(dt.base == out_dtype for dt in dts):
+        # all elements have the right dtype already; if they have a common
+        # stride, we can just return a view
+        common_stride = _common_stride(offsets, counts, out_dtype.itemsize)
+        if common_stride is not None:
+            wrap = arr.__array_wrap__
+
+            new_shape = arr.shape + (sum(counts), out_dtype.itemsize)
+            new_strides = arr.strides + (abs(common_stride), 1)
+
+            arr = arr[..., np.newaxis].view(np.uint8)  # view as bytes
+            arr = arr[..., min(offsets):]  # remove the leading unused data
+            arr = np.lib.stride_tricks.as_strided(arr,
+                                                  new_shape,
+                                                  new_strides,
+                                                  subok=True)
+
+            # cast and drop the last dimension again
+            arr = arr.view(out_dtype)[..., 0]
+
+            if common_stride < 0:
+                arr = arr[..., ::-1]  # reverse, if the stride was negative
+            if type(arr) is not type(wrap.__self__):
+                # Some types (e.g. recarray) turn into an ndarray along the
+                # way, so we have to wrap it again in order to match the
+                # behavior with copy=True.
+                arr = wrap(arr)
+            return arr
+
+    # next cast to a packed format with all fields converted to new dtype
+    packed_fields = np.dtype({'names': names,
+                              'formats': [(out_dtype, dt.shape) for dt in dts]})
+    arr = arr.astype(packed_fields, copy=copy, casting=casting)
+
+    # finally is it safe to view the packed fields as the unstructured type
+    return arr.view((out_dtype, (sum(counts),)))
+
+
+def _unstructured_to_structured_dispatcher(arr, dtype=None, names=None,
+                                           align=None, copy=None, casting=None):
+    return (arr,)
+
+@array_function_dispatch(_unstructured_to_structured_dispatcher)
+def unstructured_to_structured(arr, dtype=None, names=None, align=False,
+                               copy=False, casting='unsafe'):
+    """
+    Converts an n-D unstructured array into an (n-1)-D structured array.
+
+    The last dimension of the input array is converted into a structure, with
+    number of field-elements equal to the size of the last dimension of the
+    input array. By default all output fields have the input array's dtype, but
+    an output structured dtype with an equal number of fields-elements can be
+    supplied instead.
+
+    Nested fields, as well as each element of any subarray fields, all count
+    towards the number of field-elements.
+
+    Parameters
+    ----------
+    arr : ndarray
+       Unstructured array or dtype to convert.
+    dtype : dtype, optional
+       The structured dtype of the output array
+    names : list of strings, optional
+       If dtype is not supplied, this specifies the field names for the output
+       dtype, in order. The field dtypes will be the same as the input array.
+    align : boolean, optional
+       Whether to create an aligned memory layout.
+    copy : bool, optional
+        See copy argument to `numpy.ndarray.astype`. If true, always return a
+        copy. If false, and `dtype` requirements are satisfied, a view is
+        returned.
+    casting : {'no', 'equiv', 'safe', 'same_kind', 'unsafe'}, optional
+        See casting argument of `numpy.ndarray.astype`. Controls what kind of
+        data casting may occur.
+
+    Returns
+    -------
+    structured : ndarray
+       Structured array with fewer dimensions.
+
+    Examples
+    --------
+
+    >>> from numpy.lib import recfunctions as rfn
+    >>> dt = np.dtype([('a', 'i4'), ('b', 'f4,u2'), ('c', 'f4', 2)])
+    >>> a = np.arange(20).reshape((4,5))
+    >>> a
+    array([[ 0,  1,  2,  3,  4],
+           [ 5,  6,  7,  8,  9],
+           [10, 11, 12, 13, 14],
+           [15, 16, 17, 18, 19]])
+    >>> rfn.unstructured_to_structured(a, dt)
+    array([( 0, ( 1.,  2), [ 3.,  4.]), ( 5, ( 6.,  7), [ 8.,  9.]),
+           (10, (11., 12), [13., 14.]), (15, (16., 17), [18., 19.])],
+          dtype=[('a', '<i4'), ('b', [('f0', '<f4'), ('f1', '<u2')]), ('c', '<f4', (2,))])
+
+    """
+    if arr.shape == ():
+        raise ValueError('arr must have at least one dimension')
+    n_elem = arr.shape[-1]
+    if n_elem == 0:
+        # too many bugs elsewhere for this to work now
+        raise NotImplementedError("last axis with size 0 is not supported")
+
+    if dtype is None:
+        if names is None:
+            names = ['f{}'.format(n) for n in range(n_elem)]
+        out_dtype = np.dtype([(n, arr.dtype) for n in names], align=align)
+        fields = _get_fields_and_offsets(out_dtype)
+        dts, counts, offsets = zip(*fields)
+    else:
+        if names is not None:
+            raise ValueError("don't supply both dtype and names")
+        # if dtype is the args of np.dtype, construct it
+        dtype = np.dtype(dtype)
+        # sanity check of the input dtype
+        fields = _get_fields_and_offsets(dtype)
+        if len(fields) == 0:
+            dts, counts, offsets = [], [], []
+        else:
+            dts, counts, offsets = zip(*fields)
+
+        if n_elem != sum(counts):
+            raise ValueError('The length of the last dimension of arr must '
+                             'be equal to the number of fields in dtype')
+        out_dtype = dtype
+        if align and not out_dtype.isalignedstruct:
+            raise ValueError("align was True but dtype is not aligned")
+
+    names = ['f{}'.format(n) for n in range(len(fields))]
+
+    # Use a series of views and casts to convert to a structured array:
+
+    # first view as a packed structured array of one dtype
+    packed_fields = np.dtype({'names': names,
+                              'formats': [(arr.dtype, dt.shape) for dt in dts]})
+    arr = np.ascontiguousarray(arr).view(packed_fields)
+
+    # next cast to an unpacked but flattened format with varied dtypes
+    flattened_fields = np.dtype({'names': names,
+                                 'formats': dts,
+                                 'offsets': offsets,
+                                 'itemsize': out_dtype.itemsize})
+    arr = arr.astype(flattened_fields, copy=copy, casting=casting)
+
+    # finally view as the final nested dtype and remove the last axis
+    return arr.view(out_dtype)[..., 0]
+
+def _apply_along_fields_dispatcher(func, arr):
+    return (arr,)
+
+@array_function_dispatch(_apply_along_fields_dispatcher)
+def apply_along_fields(func, arr):
+    """
+    Apply function 'func' as a reduction across fields of a structured array.
+
+    This is similar to `apply_along_axis`, but treats the fields of a
+    structured array as an extra axis. The fields are all first cast to a
+    common type following the type-promotion rules from `numpy.result_type`
+    applied to the field's dtypes.
+
+    Parameters
+    ----------
+    func : function
+       Function to apply on the "field" dimension. This function must
+       support an `axis` argument, like np.mean, np.sum, etc.
+    arr : ndarray
+       Structured array for which to apply func.
+
+    Returns
+    -------
+    out : ndarray
+       Result of the recution operation
+
+    Examples
+    --------
+
+    >>> from numpy.lib import recfunctions as rfn
+    >>> b = np.array([(1, 2, 5), (4, 5, 7), (7, 8 ,11), (10, 11, 12)],
+    ...              dtype=[('x', 'i4'), ('y', 'f4'), ('z', 'f8')])
+    >>> rfn.apply_along_fields(np.mean, b)
+    array([ 2.66666667,  5.33333333,  8.66666667, 11.        ])
+    >>> rfn.apply_along_fields(np.mean, b[['x', 'z']])
+    array([ 3. ,  5.5,  9. , 11. ])
+
+    """
+    if arr.dtype.names is None:
+        raise ValueError('arr must be a structured array')
+
+    uarr = structured_to_unstructured(arr)
+    return func(uarr, axis=-1)
+    # works and avoids axis requirement, but very, very slow:
+    #return np.apply_along_axis(func, -1, uarr)
+
+def _assign_fields_by_name_dispatcher(dst, src, zero_unassigned=None):
+    return dst, src
+
+@array_function_dispatch(_assign_fields_by_name_dispatcher)
+def assign_fields_by_name(dst, src, zero_unassigned=True):
+    """
+    Assigns values from one structured array to another by field name.
+
+    Normally in numpy >= 1.14, assignment of one structured array to another
+    copies fields "by position", meaning that the first field from the src is
+    copied to the first field of the dst, and so on, regardless of field name.
+
+    This function instead copies "by field name", such that fields in the dst
+    are assigned from the identically named field in the src. This applies
+    recursively for nested structures. This is how structure assignment worked
+    in numpy >= 1.6 to <= 1.13.
+
+    Parameters
+    ----------
+    dst : ndarray
+    src : ndarray
+        The source and destination arrays during assignment.
+    zero_unassigned : bool, optional
+        If True, fields in the dst for which there was no matching
+        field in the src are filled with the value 0 (zero). This
+        was the behavior of numpy <= 1.13. If False, those fields
+        are not modified.
+    """
+
+    if dst.dtype.names is None:
+        dst[...] = src
+        return
+
+    for name in dst.dtype.names:
+        if name not in src.dtype.names:
+            if zero_unassigned:
+                dst[name] = 0
+        else:
+            assign_fields_by_name(dst[name], src[name],
+                                  zero_unassigned)
+
+def _require_fields_dispatcher(array, required_dtype):
+    return (array,)
+
+@array_function_dispatch(_require_fields_dispatcher)
+def require_fields(array, required_dtype):
+    """
+    Casts a structured array to a new dtype using assignment by field-name.
+
+    This function assigns from the old to the new array by name, so the
+    value of a field in the output array is the value of the field with the
+    same name in the source array. This has the effect of creating a new
+    ndarray containing only the fields "required" by the required_dtype.
+
+    If a field name in the required_dtype does not exist in the
+    input array, that field is created and set to 0 in the output array.
+
+    Parameters
+    ----------
+    a : ndarray
+       array to cast
+    required_dtype : dtype
+       datatype for output array
+
+    Returns
+    -------
+    out : ndarray
+        array with the new dtype, with field values copied from the fields in
+        the input array with the same name
+
+    Examples
+    --------
+
+    >>> from numpy.lib import recfunctions as rfn
+    >>> a = np.ones(4, dtype=[('a', 'i4'), ('b', 'f8'), ('c', 'u1')])
+    >>> rfn.require_fields(a, [('b', 'f4'), ('c', 'u1')])
+    array([(1., 1), (1., 1), (1., 1), (1., 1)],
+      dtype=[('b', '<f4'), ('c', 'u1')])
+    >>> rfn.require_fields(a, [('b', 'f4'), ('newf', 'u1')])
+    array([(1., 0), (1., 0), (1., 0), (1., 0)],
+      dtype=[('b', '<f4'), ('newf', 'u1')])
+
+    """
+    out = np.empty(array.shape, dtype=required_dtype)
+    assign_fields_by_name(out, array)
+    return out
+
+
+def _stack_arrays_dispatcher(arrays, defaults=None, usemask=None,
+                             asrecarray=None, autoconvert=None):
+    return arrays
+
+
+@array_function_dispatch(_stack_arrays_dispatcher)
+def stack_arrays(arrays, defaults=None, usemask=True, asrecarray=False,
+                 autoconvert=False):
+    """
+    Superposes arrays fields by fields
+
+    Parameters
+    ----------
+    arrays : array or sequence
+        Sequence of input arrays.
+    defaults : dictionary, optional
+        Dictionary mapping field names to the corresponding default values.
+    usemask : {True, False}, optional
+        Whether to return a MaskedArray (or MaskedRecords is
+        `asrecarray==True`) or a ndarray.
+    asrecarray : {False, True}, optional
+        Whether to return a recarray (or MaskedRecords if `usemask==True`)
+        or just a flexible-type ndarray.
+    autoconvert : {False, True}, optional
+        Whether automatically cast the type of the field to the maximum.
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> x = np.array([1, 2,])
+    >>> rfn.stack_arrays(x) is x
+    True
+    >>> z = np.array([('A', 1), ('B', 2)], dtype=[('A', '|S3'), ('B', float)])
+    >>> zz = np.array([('a', 10., 100.), ('b', 20., 200.), ('c', 30., 300.)],
+    ...   dtype=[('A', '|S3'), ('B', np.double), ('C', np.double)])
+    >>> test = rfn.stack_arrays((z,zz))
+    >>> test
+    masked_array(data=[(b'A', 1.0, --), (b'B', 2.0, --), (b'a', 10.0, 100.0),
+                       (b'b', 20.0, 200.0), (b'c', 30.0, 300.0)],
+                 mask=[(False, False,  True), (False, False,  True),
+                       (False, False, False), (False, False, False),
+                       (False, False, False)],
+           fill_value=(b'N/A', 1.e+20, 1.e+20),
+                dtype=[('A', 'S3'), ('B', '<f8'), ('C', '<f8')])
+
+    """
+    if isinstance(arrays, ndarray):
+        return arrays
+    elif len(arrays) == 1:
+        return arrays[0]
+    seqarrays = [np.asanyarray(a).ravel() for a in arrays]
+    nrecords = [len(a) for a in seqarrays]
+    ndtype = [a.dtype for a in seqarrays]
+    fldnames = [d.names for d in ndtype]
+    #
+    dtype_l = ndtype[0]
+    newdescr = _get_fieldspec(dtype_l)
+    names = [n for n, d in newdescr]
+    for dtype_n in ndtype[1:]:
+        for fname, fdtype in _get_fieldspec(dtype_n):
+            if fname not in names:
+                newdescr.append((fname, fdtype))
+                names.append(fname)
+            else:
+                nameidx = names.index(fname)
+                _, cdtype = newdescr[nameidx]
+                if autoconvert:
+                    newdescr[nameidx] = (fname, max(fdtype, cdtype))
+                elif fdtype != cdtype:
+                    raise TypeError("Incompatible type '%s' <> '%s'" %
+                                    (cdtype, fdtype))
+    # Only one field: use concatenate
+    if len(newdescr) == 1:
+        output = ma.concatenate(seqarrays)
+    else:
+        #
+        output = ma.masked_all((np.sum(nrecords),), newdescr)
+        offset = np.cumsum(np.r_[0, nrecords])
+        seen = []
+        for (a, n, i, j) in zip(seqarrays, fldnames, offset[:-1], offset[1:]):
+            names = a.dtype.names
+            if names is None:
+                output['f%i' % len(seen)][i:j] = a
+            else:
+                for name in n:
+                    output[name][i:j] = a[name]
+                    if name not in seen:
+                        seen.append(name)
+    #
+    return _fix_output(_fix_defaults(output, defaults),
+                       usemask=usemask, asrecarray=asrecarray)
+
+
+def _find_duplicates_dispatcher(
+        a, key=None, ignoremask=None, return_index=None):
+    return (a,)
+
+
+@array_function_dispatch(_find_duplicates_dispatcher)
+def find_duplicates(a, key=None, ignoremask=True, return_index=False):
+    """
+    Find the duplicates in a structured array along a given key
+
+    Parameters
+    ----------
+    a : array-like
+        Input array
+    key : {string, None}, optional
+        Name of the fields along which to check the duplicates.
+        If None, the search is performed by records
+    ignoremask : {True, False}, optional
+        Whether masked data should be discarded or considered as duplicates.
+    return_index : {False, True}, optional
+        Whether to return the indices of the duplicated values.
+
+    Examples
+    --------
+    >>> from numpy.lib import recfunctions as rfn
+    >>> ndtype = [('a', int)]
+    >>> a = np.ma.array([1, 1, 1, 2, 2, 3, 3],
+    ...         mask=[0, 0, 1, 0, 0, 0, 1]).view(ndtype)
+    >>> rfn.find_duplicates(a, ignoremask=True, return_index=True)
+    (masked_array(data=[(1,), (1,), (2,), (2,)],
+                 mask=[(False,), (False,), (False,), (False,)],
+           fill_value=(999999,),
+                dtype=[('a', '<i8')]), array([0, 1, 3, 4]))
+    """
+    a = np.asanyarray(a).ravel()
+    # Get a dictionary of fields
+    fields = get_fieldstructure(a.dtype)
+    # Get the sorting data (by selecting the corresponding field)
+    base = a
+    if key:
+        for f in fields[key]:
+            base = base[f]
+        base = base[key]
+    # Get the sorting indices and the sorted data
+    sortidx = base.argsort()
+    sortedbase = base[sortidx]
+    sorteddata = sortedbase.filled()
+    # Compare the sorting data
+    flag = (sorteddata[:-1] == sorteddata[1:])
+    # If masked data must be ignored, set the flag to false where needed
+    if ignoremask:
+        sortedmask = sortedbase.recordmask
+        flag[sortedmask[1:]] = False
+    flag = np.concatenate(([False], flag))
+    # We need to take the point on the left as well (else we're missing it)
+    flag[:-1] = flag[:-1] + flag[1:]
+    duplicates = a[sortidx][flag]
+    if return_index:
+        return (duplicates, sortidx[flag])
+    else:
+        return duplicates
+
+
+def _join_by_dispatcher(
+        key, r1, r2, jointype=None, r1postfix=None, r2postfix=None,
+        defaults=None, usemask=None, asrecarray=None):
+    return (r1, r2)
+
+
+@array_function_dispatch(_join_by_dispatcher)
+def join_by(key, r1, r2, jointype='inner', r1postfix='1', r2postfix='2',
+            defaults=None, usemask=True, asrecarray=False):
+    """
+    Join arrays `r1` and `r2` on key `key`.
+
+    The key should be either a string or a sequence of string corresponding
+    to the fields used to join the array.  An exception is raised if the
+    `key` field cannot be found in the two input arrays.  Neither `r1` nor
+    `r2` should have any duplicates along `key`: the presence of duplicates
+    will make the output quite unreliable. Note that duplicates are not
+    looked for by the algorithm.
+
+    Parameters
+    ----------
+    key : {string, sequence}
+        A string or a sequence of strings corresponding to the fields used
+        for comparison.
+    r1, r2 : arrays
+        Structured arrays.
+    jointype : {'inner', 'outer', 'leftouter'}, optional
+        If 'inner', returns the elements common to both r1 and r2.
+        If 'outer', returns the common elements as well as the elements of
+        r1 not in r2 and the elements of not in r2.
+        If 'leftouter', returns the common elements and the elements of r1
+        not in r2.
+    r1postfix : string, optional
+        String appended to the names of the fields of r1 that are present
+        in r2 but absent of the key.
+    r2postfix : string, optional
+        String appended to the names of the fields of r2 that are present
+        in r1 but absent of the key.
+    defaults : {dictionary}, optional
+        Dictionary mapping field names to the corresponding default values.
+    usemask : {True, False}, optional
+        Whether to return a MaskedArray (or MaskedRecords is
+        `asrecarray==True`) or a ndarray.
+    asrecarray : {False, True}, optional
+        Whether to return a recarray (or MaskedRecords if `usemask==True`)
+        or just a flexible-type ndarray.
+
+    Notes
+    -----
+    * The output is sorted along the key.
+    * A temporary array is formed by dropping the fields not in the key for
+      the two arrays and concatenating the result. This array is then
+      sorted, and the common entries selected. The output is constructed by
+      filling the fields with the selected entries. Matching is not
+      preserved if there are some duplicates...
+
+    """
+    # Check jointype
+    if jointype not in ('inner', 'outer', 'leftouter'):
+        raise ValueError(
+                "The 'jointype' argument should be in 'inner', "
+                "'outer' or 'leftouter' (got '%s' instead)" % jointype
+                )
+    # If we have a single key, put it in a tuple
+    if isinstance(key, str):
+        key = (key,)
+
+    # Check the keys
+    if len(set(key)) != len(key):
+        dup = next(x for n,x in enumerate(key) if x in key[n+1:])
+        raise ValueError("duplicate join key %r" % dup)
+    for name in key:
+        if name not in r1.dtype.names:
+            raise ValueError('r1 does not have key field %r' % name)
+        if name not in r2.dtype.names:
+            raise ValueError('r2 does not have key field %r' % name)
+
+    # Make sure we work with ravelled arrays
+    r1 = r1.ravel()
+    r2 = r2.ravel()
+    # Fixme: nb2 below is never used. Commenting out for pyflakes.
+    # (nb1, nb2) = (len(r1), len(r2))
+    nb1 = len(r1)
+    (r1names, r2names) = (r1.dtype.names, r2.dtype.names)
+
+    # Check the names for collision
+    collisions = (set(r1names) & set(r2names)) - set(key)
+    if collisions and not (r1postfix or r2postfix):
+        msg = "r1 and r2 contain common names, r1postfix and r2postfix "
+        msg += "can't both be empty"
+        raise ValueError(msg)
+
+    # Make temporary arrays of just the keys
+    #  (use order of keys in `r1` for back-compatibility)
+    key1 = [ n for n in r1names if n in key ]
+    r1k = _keep_fields(r1, key1)
+    r2k = _keep_fields(r2, key1)
+
+    # Concatenate the two arrays for comparison
+    aux = ma.concatenate((r1k, r2k))
+    idx_sort = aux.argsort(order=key)
+    aux = aux[idx_sort]
+    #
+    # Get the common keys
+    flag_in = ma.concatenate(([False], aux[1:] == aux[:-1]))
+    flag_in[:-1] = flag_in[1:] + flag_in[:-1]
+    idx_in = idx_sort[flag_in]
+    idx_1 = idx_in[(idx_in < nb1)]
+    idx_2 = idx_in[(idx_in >= nb1)] - nb1
+    (r1cmn, r2cmn) = (len(idx_1), len(idx_2))
+    if jointype == 'inner':
+        (r1spc, r2spc) = (0, 0)
+    elif jointype == 'outer':
+        idx_out = idx_sort[~flag_in]
+        idx_1 = np.concatenate((idx_1, idx_out[(idx_out < nb1)]))
+        idx_2 = np.concatenate((idx_2, idx_out[(idx_out >= nb1)] - nb1))
+        (r1spc, r2spc) = (len(idx_1) - r1cmn, len(idx_2) - r2cmn)
+    elif jointype == 'leftouter':
+        idx_out = idx_sort[~flag_in]
+        idx_1 = np.concatenate((idx_1, idx_out[(idx_out < nb1)]))
+        (r1spc, r2spc) = (len(idx_1) - r1cmn, 0)
+    # Select the entries from each input
+    (s1, s2) = (r1[idx_1], r2[idx_2])
+    #
+    # Build the new description of the output array .......
+    # Start with the key fields
+    ndtype = _get_fieldspec(r1k.dtype)
+
+    # Add the fields from r1
+    for fname, fdtype in _get_fieldspec(r1.dtype):
+        if fname not in key:
+            ndtype.append((fname, fdtype))
+
+    # Add the fields from r2
+    for fname, fdtype in _get_fieldspec(r2.dtype):
+        # Have we seen the current name already ?
+        # we need to rebuild this list every time
+        names = list(name for name, dtype in ndtype)
+        try:
+            nameidx = names.index(fname)
+        except ValueError:
+            #... we haven't: just add the description to the current list
+            ndtype.append((fname, fdtype))
+        else:
+            # collision
+            _, cdtype = ndtype[nameidx]
+            if fname in key:
+                # The current field is part of the key: take the largest dtype
+                ndtype[nameidx] = (fname, max(fdtype, cdtype))
+            else:
+                # The current field is not part of the key: add the suffixes,
+                # and place the new field adjacent to the old one
+                ndtype[nameidx:nameidx + 1] = [
+                    (fname + r1postfix, cdtype),
+                    (fname + r2postfix, fdtype)
+                ]
+    # Rebuild a dtype from the new fields
+    ndtype = np.dtype(ndtype)
+    # Find the largest nb of common fields :
+    # r1cmn and r2cmn should be equal, but...
+    cmn = max(r1cmn, r2cmn)
+    # Construct an empty array
+    output = ma.masked_all((cmn + r1spc + r2spc,), dtype=ndtype)
+    names = output.dtype.names
+    for f in r1names:
+        selected = s1[f]
+        if f not in names or (f in r2names and not r2postfix and f not in key):
+            f += r1postfix
+        current = output[f]
+        current[:r1cmn] = selected[:r1cmn]
+        if jointype in ('outer', 'leftouter'):
+            current[cmn:cmn + r1spc] = selected[r1cmn:]
+    for f in r2names:
+        selected = s2[f]
+        if f not in names or (f in r1names and not r1postfix and f not in key):
+            f += r2postfix
+        current = output[f]
+        current[:r2cmn] = selected[:r2cmn]
+        if (jointype == 'outer') and r2spc:
+            current[-r2spc:] = selected[r2cmn:]
+    # Sort and finalize the output
+    output.sort(order=key)
+    kwargs = dict(usemask=usemask, asrecarray=asrecarray)
+    return _fix_output(_fix_defaults(output, defaults), **kwargs)
+
+
+def _rec_join_dispatcher(
+        key, r1, r2, jointype=None, r1postfix=None, r2postfix=None,
+        defaults=None):
+    return (r1, r2)
+
+
+@array_function_dispatch(_rec_join_dispatcher)
+def rec_join(key, r1, r2, jointype='inner', r1postfix='1', r2postfix='2',
+             defaults=None):
+    """
+    Join arrays `r1` and `r2` on keys.
+    Alternative to join_by, that always returns a np.recarray.
+
+    See Also
+    --------
+    join_by : equivalent function
+    """
+    kwargs = dict(jointype=jointype, r1postfix=r1postfix, r2postfix=r2postfix,
+                  defaults=defaults, usemask=False, asrecarray=True)
+    return join_by(key, r1, r2, **kwargs)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/scimath.py

@@ -0,0 +1,625 @@
+"""
+Wrapper functions to more user-friendly calling of certain math functions
+whose output data-type is different than the input data-type in certain
+domains of the input.
+
+For example, for functions like `log` with branch cuts, the versions in this
+module provide the mathematically valid answers in the complex plane::
+
+  >>> import math
+  >>> np.emath.log(-math.exp(1)) == (1+1j*math.pi)
+  True
+
+Similarly, `sqrt`, other base logarithms, `power` and trig functions are
+correctly handled.  See their respective docstrings for specific examples.
+
+Functions
+---------
+
+.. autosummary::
+   :toctree: generated/
+
+   sqrt
+   log
+   log2
+   logn
+   log10
+   power
+   arccos
+   arcsin
+   arctanh
+
+"""
+import numpy.core.numeric as nx
+import numpy.core.numerictypes as nt
+from numpy.core.numeric import asarray, any
+from numpy.core.overrides import array_function_dispatch
+from numpy.lib.type_check import isreal
+
+
+__all__ = [
+    'sqrt', 'log', 'log2', 'logn', 'log10', 'power', 'arccos', 'arcsin',
+    'arctanh'
+    ]
+
+
+_ln2 = nx.log(2.0)
+
+
+def _tocomplex(arr):
+    """Convert its input `arr` to a complex array.
+
+    The input is returned as a complex array of the smallest type that will fit
+    the original data: types like single, byte, short, etc. become csingle,
+    while others become cdouble.
+
+    A copy of the input is always made.
+
+    Parameters
+    ----------
+    arr : array
+
+    Returns
+    -------
+    array
+        An array with the same input data as the input but in complex form.
+
+    Examples
+    --------
+
+    First, consider an input of type short:
+
+    >>> a = np.array([1,2,3],np.short)
+
+    >>> ac = np.lib.scimath._tocomplex(a); ac
+    array([1.+0.j, 2.+0.j, 3.+0.j], dtype=complex64)
+
+    >>> ac.dtype
+    dtype('complex64')
+
+    If the input is of type double, the output is correspondingly of the
+    complex double type as well:
+
+    >>> b = np.array([1,2,3],np.double)
+
+    >>> bc = np.lib.scimath._tocomplex(b); bc
+    array([1.+0.j, 2.+0.j, 3.+0.j])
+
+    >>> bc.dtype
+    dtype('complex128')
+
+    Note that even if the input was complex to begin with, a copy is still
+    made, since the astype() method always copies:
+
+    >>> c = np.array([1,2,3],np.csingle)
+
+    >>> cc = np.lib.scimath._tocomplex(c); cc
+    array([1.+0.j,  2.+0.j,  3.+0.j], dtype=complex64)
+
+    >>> c *= 2; c
+    array([2.+0.j,  4.+0.j,  6.+0.j], dtype=complex64)
+
+    >>> cc
+    array([1.+0.j,  2.+0.j,  3.+0.j], dtype=complex64)
+    """
+    if issubclass(arr.dtype.type, (nt.single, nt.byte, nt.short, nt.ubyte,
+                                   nt.ushort, nt.csingle)):
+        return arr.astype(nt.csingle)
+    else:
+        return arr.astype(nt.cdouble)
+
+
+def _fix_real_lt_zero(x):
+    """Convert `x` to complex if it has real, negative components.
+
+    Otherwise, output is just the array version of the input (via asarray).
+
+    Parameters
+    ----------
+    x : array_like
+
+    Returns
+    -------
+    array
+
+    Examples
+    --------
+    >>> np.lib.scimath._fix_real_lt_zero([1,2])
+    array([1, 2])
+
+    >>> np.lib.scimath._fix_real_lt_zero([-1,2])
+    array([-1.+0.j,  2.+0.j])
+
+    """
+    x = asarray(x)
+    if any(isreal(x) & (x < 0)):
+        x = _tocomplex(x)
+    return x
+
+
+def _fix_int_lt_zero(x):
+    """Convert `x` to double if it has real, negative components.
+
+    Otherwise, output is just the array version of the input (via asarray).
+
+    Parameters
+    ----------
+    x : array_like
+
+    Returns
+    -------
+    array
+
+    Examples
+    --------
+    >>> np.lib.scimath._fix_int_lt_zero([1,2])
+    array([1, 2])
+
+    >>> np.lib.scimath._fix_int_lt_zero([-1,2])
+    array([-1.,  2.])
+    """
+    x = asarray(x)
+    if any(isreal(x) & (x < 0)):
+        x = x * 1.0
+    return x
+
+
+def _fix_real_abs_gt_1(x):
+    """Convert `x` to complex if it has real components x_i with abs(x_i)>1.
+
+    Otherwise, output is just the array version of the input (via asarray).
+
+    Parameters
+    ----------
+    x : array_like
+
+    Returns
+    -------
+    array
+
+    Examples
+    --------
+    >>> np.lib.scimath._fix_real_abs_gt_1([0,1])
+    array([0, 1])
+
+    >>> np.lib.scimath._fix_real_abs_gt_1([0,2])
+    array([0.+0.j, 2.+0.j])
+    """
+    x = asarray(x)
+    if any(isreal(x) & (abs(x) > 1)):
+        x = _tocomplex(x)
+    return x
+
+
+def _unary_dispatcher(x):
+    return (x,)
+
+
+@array_function_dispatch(_unary_dispatcher)
+def sqrt(x):
+    """
+    Compute the square root of x.
+
+    For negative input elements, a complex value is returned
+    (unlike `numpy.sqrt` which returns NaN).
+
+    Parameters
+    ----------
+    x : array_like
+       The input value(s).
+
+    Returns
+    -------
+    out : ndarray or scalar
+       The square root of `x`. If `x` was a scalar, so is `out`,
+       otherwise an array is returned.
+
+    See Also
+    --------
+    numpy.sqrt
+
+    Examples
+    --------
+    For real, non-negative inputs this works just like `numpy.sqrt`:
+
+    >>> np.emath.sqrt(1)
+    1.0
+    >>> np.emath.sqrt([1, 4])
+    array([1.,  2.])
+
+    But it automatically handles negative inputs:
+
+    >>> np.emath.sqrt(-1)
+    1j
+    >>> np.emath.sqrt([-1,4])
+    array([0.+1.j, 2.+0.j])
+
+    Different results are expected because:
+    floating point 0.0 and -0.0 are distinct.
+
+    For more control, explicitly use complex() as follows:
+
+    >>> np.emath.sqrt(complex(-4.0, 0.0))
+    2j
+    >>> np.emath.sqrt(complex(-4.0, -0.0))
+    -2j
+    """
+    x = _fix_real_lt_zero(x)
+    return nx.sqrt(x)
+
+
+@array_function_dispatch(_unary_dispatcher)
+def log(x):
+    """
+    Compute the natural logarithm of `x`.
+
+    Return the "principal value" (for a description of this, see `numpy.log`)
+    of :math:`log_e(x)`. For real `x > 0`, this is a real number (``log(0)``
+    returns ``-inf`` and ``log(np.inf)`` returns ``inf``). Otherwise, the
+    complex principle value is returned.
+
+    Parameters
+    ----------
+    x : array_like
+       The value(s) whose log is (are) required.
+
+    Returns
+    -------
+    out : ndarray or scalar
+       The log of the `x` value(s). If `x` was a scalar, so is `out`,
+       otherwise an array is returned.
+
+    See Also
+    --------
+    numpy.log
+
+    Notes
+    -----
+    For a log() that returns ``NAN`` when real `x < 0`, use `numpy.log`
+    (note, however, that otherwise `numpy.log` and this `log` are identical,
+    i.e., both return ``-inf`` for `x = 0`, ``inf`` for `x = inf`, and,
+    notably, the complex principle value if ``x.imag != 0``).
+
+    Examples
+    --------
+    >>> np.emath.log(np.exp(1))
+    1.0
+
+    Negative arguments are handled "correctly" (recall that
+    ``exp(log(x)) == x`` does *not* hold for real ``x < 0``):
+
+    >>> np.emath.log(-np.exp(1)) == (1 + np.pi * 1j)
+    True
+
+    """
+    x = _fix_real_lt_zero(x)
+    return nx.log(x)
+
+
+@array_function_dispatch(_unary_dispatcher)
+def log10(x):
+    """
+    Compute the logarithm base 10 of `x`.
+
+    Return the "principal value" (for a description of this, see
+    `numpy.log10`) of :math:`log_{10}(x)`. For real `x > 0`, this
+    is a real number (``log10(0)`` returns ``-inf`` and ``log10(np.inf)``
+    returns ``inf``). Otherwise, the complex principle value is returned.
+
+    Parameters
+    ----------
+    x : array_like or scalar
+       The value(s) whose log base 10 is (are) required.
+
+    Returns
+    -------
+    out : ndarray or scalar
+       The log base 10 of the `x` value(s). If `x` was a scalar, so is `out`,
+       otherwise an array object is returned.
+
+    See Also
+    --------
+    numpy.log10
+
+    Notes
+    -----
+    For a log10() that returns ``NAN`` when real `x < 0`, use `numpy.log10`
+    (note, however, that otherwise `numpy.log10` and this `log10` are
+    identical, i.e., both return ``-inf`` for `x = 0`, ``inf`` for `x = inf`,
+    and, notably, the complex principle value if ``x.imag != 0``).
+
+    Examples
+    --------
+
+    (We set the printing precision so the example can be auto-tested)
+
+    >>> np.set_printoptions(precision=4)
+
+    >>> np.emath.log10(10**1)
+    1.0
+
+    >>> np.emath.log10([-10**1, -10**2, 10**2])
+    array([1.+1.3644j, 2.+1.3644j, 2.+0.j    ])
+
+    """
+    x = _fix_real_lt_zero(x)
+    return nx.log10(x)
+
+
+def _logn_dispatcher(n, x):
+    return (n, x,)
+
+
+@array_function_dispatch(_logn_dispatcher)
+def logn(n, x):
+    """
+    Take log base n of x.
+
+    If `x` contains negative inputs, the answer is computed and returned in the
+    complex domain.
+
+    Parameters
+    ----------
+    n : array_like
+       The integer base(s) in which the log is taken.
+    x : array_like
+       The value(s) whose log base `n` is (are) required.
+
+    Returns
+    -------
+    out : ndarray or scalar
+       The log base `n` of the `x` value(s). If `x` was a scalar, so is
+       `out`, otherwise an array is returned.
+
+    Examples
+    --------
+    >>> np.set_printoptions(precision=4)
+
+    >>> np.emath.logn(2, [4, 8])
+    array([2., 3.])
+    >>> np.emath.logn(2, [-4, -8, 8])
+    array([2.+4.5324j, 3.+4.5324j, 3.+0.j    ])
+
+    """
+    x = _fix_real_lt_zero(x)
+    n = _fix_real_lt_zero(n)
+    return nx.log(x)/nx.log(n)
+
+
+@array_function_dispatch(_unary_dispatcher)
+def log2(x):
+    """
+    Compute the logarithm base 2 of `x`.
+
+    Return the "principal value" (for a description of this, see
+    `numpy.log2`) of :math:`log_2(x)`. For real `x > 0`, this is
+    a real number (``log2(0)`` returns ``-inf`` and ``log2(np.inf)`` returns
+    ``inf``). Otherwise, the complex principle value is returned.
+
+    Parameters
+    ----------
+    x : array_like
+       The value(s) whose log base 2 is (are) required.
+
+    Returns
+    -------
+    out : ndarray or scalar
+       The log base 2 of the `x` value(s). If `x` was a scalar, so is `out`,
+       otherwise an array is returned.
+
+    See Also
+    --------
+    numpy.log2
+
+    Notes
+    -----
+    For a log2() that returns ``NAN`` when real `x < 0`, use `numpy.log2`
+    (note, however, that otherwise `numpy.log2` and this `log2` are
+    identical, i.e., both return ``-inf`` for `x = 0`, ``inf`` for `x = inf`,
+    and, notably, the complex principle value if ``x.imag != 0``).
+
+    Examples
+    --------
+    We set the printing precision so the example can be auto-tested:
+
+    >>> np.set_printoptions(precision=4)
+
+    >>> np.emath.log2(8)
+    3.0
+    >>> np.emath.log2([-4, -8, 8])
+    array([2.+4.5324j, 3.+4.5324j, 3.+0.j    ])
+
+    """
+    x = _fix_real_lt_zero(x)
+    return nx.log2(x)
+
+
+def _power_dispatcher(x, p):
+    return (x, p)
+
+
+@array_function_dispatch(_power_dispatcher)
+def power(x, p):
+    """
+    Return x to the power p, (x**p).
+
+    If `x` contains negative values, the output is converted to the
+    complex domain.
+
+    Parameters
+    ----------
+    x : array_like
+        The input value(s).
+    p : array_like of ints
+        The power(s) to which `x` is raised. If `x` contains multiple values,
+        `p` has to either be a scalar, or contain the same number of values
+        as `x`. In the latter case, the result is
+        ``x[0]**p[0], x[1]**p[1], ...``.
+
+    Returns
+    -------
+    out : ndarray or scalar
+        The result of ``x**p``. If `x` and `p` are scalars, so is `out`,
+        otherwise an array is returned.
+
+    See Also
+    --------
+    numpy.power
+
+    Examples
+    --------
+    >>> np.set_printoptions(precision=4)
+
+    >>> np.emath.power([2, 4], 2)
+    array([ 4, 16])
+    >>> np.emath.power([2, 4], -2)
+    array([0.25  ,  0.0625])
+    >>> np.emath.power([-2, 4], 2)
+    array([ 4.-0.j, 16.+0.j])
+
+    """
+    x = _fix_real_lt_zero(x)
+    p = _fix_int_lt_zero(p)
+    return nx.power(x, p)
+
+
+@array_function_dispatch(_unary_dispatcher)
+def arccos(x):
+    """
+    Compute the inverse cosine of x.
+
+    Return the "principal value" (for a description of this, see
+    `numpy.arccos`) of the inverse cosine of `x`. For real `x` such that
+    `abs(x) <= 1`, this is a real number in the closed interval
+    :math:`[0, \\pi]`.  Otherwise, the complex principle value is returned.
+
+    Parameters
+    ----------
+    x : array_like or scalar
+       The value(s) whose arccos is (are) required.
+
+    Returns
+    -------
+    out : ndarray or scalar
+       The inverse cosine(s) of the `x` value(s). If `x` was a scalar, so
+       is `out`, otherwise an array object is returned.
+
+    See Also
+    --------
+    numpy.arccos
+
+    Notes
+    -----
+    For an arccos() that returns ``NAN`` when real `x` is not in the
+    interval ``[-1,1]``, use `numpy.arccos`.
+
+    Examples
+    --------
+    >>> np.set_printoptions(precision=4)
+
+    >>> np.emath.arccos(1) # a scalar is returned
+    0.0
+
+    >>> np.emath.arccos([1,2])
+    array([0.-0.j   , 0.-1.317j])
+
+    """
+    x = _fix_real_abs_gt_1(x)
+    return nx.arccos(x)
+
+
+@array_function_dispatch(_unary_dispatcher)
+def arcsin(x):
+    """
+    Compute the inverse sine of x.
+
+    Return the "principal value" (for a description of this, see
+    `numpy.arcsin`) of the inverse sine of `x`. For real `x` such that
+    `abs(x) <= 1`, this is a real number in the closed interval
+    :math:`[-\\pi/2, \\pi/2]`.  Otherwise, the complex principle value is
+    returned.
+
+    Parameters
+    ----------
+    x : array_like or scalar
+       The value(s) whose arcsin is (are) required.
+
+    Returns
+    -------
+    out : ndarray or scalar
+       The inverse sine(s) of the `x` value(s). If `x` was a scalar, so
+       is `out`, otherwise an array object is returned.
+
+    See Also
+    --------
+    numpy.arcsin
+
+    Notes
+    -----
+    For an arcsin() that returns ``NAN`` when real `x` is not in the
+    interval ``[-1,1]``, use `numpy.arcsin`.
+
+    Examples
+    --------
+    >>> np.set_printoptions(precision=4)
+
+    >>> np.emath.arcsin(0)
+    0.0
+
+    >>> np.emath.arcsin([0,1])
+    array([0.    , 1.5708])
+
+    """
+    x = _fix_real_abs_gt_1(x)
+    return nx.arcsin(x)
+
+
+@array_function_dispatch(_unary_dispatcher)
+def arctanh(x):
+    """
+    Compute the inverse hyperbolic tangent of `x`.
+
+    Return the "principal value" (for a description of this, see
+    `numpy.arctanh`) of ``arctanh(x)``. For real `x` such that
+    ``abs(x) < 1``, this is a real number.  If `abs(x) > 1`, or if `x` is
+    complex, the result is complex. Finally, `x = 1` returns``inf`` and
+    ``x=-1`` returns ``-inf``.
+
+    Parameters
+    ----------
+    x : array_like
+       The value(s) whose arctanh is (are) required.
+
+    Returns
+    -------
+    out : ndarray or scalar
+       The inverse hyperbolic tangent(s) of the `x` value(s). If `x` was
+       a scalar so is `out`, otherwise an array is returned.
+
+
+    See Also
+    --------
+    numpy.arctanh
+
+    Notes
+    -----
+    For an arctanh() that returns ``NAN`` when real `x` is not in the
+    interval ``(-1,1)``, use `numpy.arctanh` (this latter, however, does
+    return +/-inf for ``x = +/-1``).
+
+    Examples
+    --------
+    >>> np.set_printoptions(precision=4)
+
+    >>> from numpy.testing import suppress_warnings
+    >>> with suppress_warnings() as sup:
+    ...     sup.filter(RuntimeWarning)
+    ...     np.emath.arctanh(np.eye(2))
+    array([[inf,  0.],
+           [ 0., inf]])
+    >>> np.emath.arctanh([1j])
+    array([0.+0.7854j])
+
+    """
+    x = _fix_real_abs_gt_1(x)
+    return nx.arctanh(x)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/shape_base.py

@@ -0,0 +1,1274 @@
+import functools
+
+import numpy.core.numeric as _nx
+from numpy.core.numeric import asarray, zeros, array, asanyarray
+from numpy.core.fromnumeric import reshape, transpose
+from numpy.core.multiarray import normalize_axis_index
+from numpy.core import overrides
+from numpy.core import vstack, atleast_3d
+from numpy.core.numeric import normalize_axis_tuple
+from numpy.core.shape_base import _arrays_for_stack_dispatcher
+from numpy.lib.index_tricks import ndindex
+from numpy.matrixlib.defmatrix import matrix  # this raises all the right alarm bells
+
+
+__all__ = [
+    'column_stack', 'row_stack', 'dstack', 'array_split', 'split',
+    'hsplit', 'vsplit', 'dsplit', 'apply_over_axes', 'expand_dims',
+    'apply_along_axis', 'kron', 'tile', 'get_array_wrap', 'take_along_axis',
+    'put_along_axis'
+    ]
+
+
+array_function_dispatch = functools.partial(
+    overrides.array_function_dispatch, module='numpy')
+
+
+def _make_along_axis_idx(arr_shape, indices, axis):
+    # compute dimensions to iterate over
+    if not _nx.issubdtype(indices.dtype, _nx.integer):
+        raise IndexError('`indices` must be an integer array')
+    if len(arr_shape) != indices.ndim:
+        raise ValueError(
+            "`indices` and `arr` must have the same number of dimensions")
+    shape_ones = (1,) * indices.ndim
+    dest_dims = list(range(axis)) + [None] + list(range(axis+1, indices.ndim))
+
+    # build a fancy index, consisting of orthogonal aranges, with the
+    # requested index inserted at the right location
+    fancy_index = []
+    for dim, n in zip(dest_dims, arr_shape):
+        if dim is None:
+            fancy_index.append(indices)
+        else:
+            ind_shape = shape_ones[:dim] + (-1,) + shape_ones[dim+1:]
+            fancy_index.append(_nx.arange(n).reshape(ind_shape))
+
+    return tuple(fancy_index)
+
+
+def _take_along_axis_dispatcher(arr, indices, axis):
+    return (arr, indices)
+
+
+@array_function_dispatch(_take_along_axis_dispatcher)
+def take_along_axis(arr, indices, axis):
+    """
+    Take values from the input array by matching 1d index and data slices.
+
+    This iterates over matching 1d slices oriented along the specified axis in
+    the index and data arrays, and uses the former to look up values in the
+    latter. These slices can be different lengths.
+
+    Functions returning an index along an axis, like `argsort` and
+    `argpartition`, produce suitable indices for this function.
+
+    .. versionadded:: 1.15.0
+
+    Parameters
+    ----------
+    arr : ndarray (Ni..., M, Nk...)
+        Source array
+    indices : ndarray (Ni..., J, Nk...)
+        Indices to take along each 1d slice of `arr`. This must match the
+        dimension of arr, but dimensions Ni and Nj only need to broadcast
+        against `arr`.
+    axis : int
+        The axis to take 1d slices along. If axis is None, the input array is
+        treated as if it had first been flattened to 1d, for consistency with
+        `sort` and `argsort`.
+
+    Returns
+    -------
+    out: ndarray (Ni..., J, Nk...)
+        The indexed result.
+
+    Notes
+    -----
+    This is equivalent to (but faster than) the following use of `ndindex` and
+    `s_`, which sets each of ``ii`` and ``kk`` to a tuple of indices::
+
+        Ni, M, Nk = a.shape[:axis], a.shape[axis], a.shape[axis+1:]
+        J = indices.shape[axis]  # Need not equal M
+        out = np.empty(Ni + (J,) + Nk)
+
+        for ii in ndindex(Ni):
+            for kk in ndindex(Nk):
+                a_1d       = a      [ii + s_[:,] + kk]
+                indices_1d = indices[ii + s_[:,] + kk]
+                out_1d     = out    [ii + s_[:,] + kk]
+                for j in range(J):
+                    out_1d[j] = a_1d[indices_1d[j]]
+
+    Equivalently, eliminating the inner loop, the last two lines would be::
+
+                out_1d[:] = a_1d[indices_1d]
+
+    See Also
+    --------
+    take : Take along an axis, using the same indices for every 1d slice
+    put_along_axis :
+        Put values into the destination array by matching 1d index and data slices
+
+    Examples
+    --------
+
+    For this sample array
+
+    >>> a = np.array([[10, 30, 20], [60, 40, 50]])
+
+    We can sort either by using sort directly, or argsort and this function
+
+    >>> np.sort(a, axis=1)
+    array([[10, 20, 30],
+           [40, 50, 60]])
+    >>> ai = np.argsort(a, axis=1)
+    >>> ai
+    array([[0, 2, 1],
+           [1, 2, 0]])
+    >>> np.take_along_axis(a, ai, axis=1)
+    array([[10, 20, 30],
+           [40, 50, 60]])
+
+    The same works for max and min, if you maintain the trivial dimension
+    with ``keepdims``:
+
+    >>> np.max(a, axis=1, keepdims=True)
+    array([[30],
+           [60]])
+    >>> ai = np.argmax(a, axis=1, keepdims=True)
+    >>> ai
+    array([[1],
+           [0]])
+    >>> np.take_along_axis(a, ai, axis=1)
+    array([[30],
+           [60]])
+
+    If we want to get the max and min at the same time, we can stack the
+    indices first
+
+    >>> ai_min = np.argmin(a, axis=1, keepdims=True)
+    >>> ai_max = np.argmax(a, axis=1, keepdims=True)
+    >>> ai = np.concatenate([ai_min, ai_max], axis=1)
+    >>> ai
+    array([[0, 1],
+           [1, 0]])
+    >>> np.take_along_axis(a, ai, axis=1)
+    array([[10, 30],
+           [40, 60]])
+    """
+    # normalize inputs
+    if axis is None:
+        arr = arr.flat
+        arr_shape = (len(arr),)  # flatiter has no .shape
+        axis = 0
+    else:
+        axis = normalize_axis_index(axis, arr.ndim)
+        arr_shape = arr.shape
+
+    # use the fancy index
+    return arr[_make_along_axis_idx(arr_shape, indices, axis)]
+
+
+def _put_along_axis_dispatcher(arr, indices, values, axis):
+    return (arr, indices, values)
+
+
+@array_function_dispatch(_put_along_axis_dispatcher)
+def put_along_axis(arr, indices, values, axis):
+    """
+    Put values into the destination array by matching 1d index and data slices.
+
+    This iterates over matching 1d slices oriented along the specified axis in
+    the index and data arrays, and uses the former to place values into the
+    latter. These slices can be different lengths.
+
+    Functions returning an index along an axis, like `argsort` and
+    `argpartition`, produce suitable indices for this function.
+
+    .. versionadded:: 1.15.0
+
+    Parameters
+    ----------
+    arr : ndarray (Ni..., M, Nk...)
+        Destination array.
+    indices : ndarray (Ni..., J, Nk...)
+        Indices to change along each 1d slice of `arr`. This must match the
+        dimension of arr, but dimensions in Ni and Nj may be 1 to broadcast
+        against `arr`.
+    values : array_like (Ni..., J, Nk...)
+        values to insert at those indices. Its shape and dimension are
+        broadcast to match that of `indices`.
+    axis : int
+        The axis to take 1d slices along. If axis is None, the destination
+        array is treated as if a flattened 1d view had been created of it.
+
+    Notes
+    -----
+    This is equivalent to (but faster than) the following use of `ndindex` and
+    `s_`, which sets each of ``ii`` and ``kk`` to a tuple of indices::
+
+        Ni, M, Nk = a.shape[:axis], a.shape[axis], a.shape[axis+1:]
+        J = indices.shape[axis]  # Need not equal M
+
+        for ii in ndindex(Ni):
+            for kk in ndindex(Nk):
+                a_1d       = a      [ii + s_[:,] + kk]
+                indices_1d = indices[ii + s_[:,] + kk]
+                values_1d  = values [ii + s_[:,] + kk]
+                for j in range(J):
+                    a_1d[indices_1d[j]] = values_1d[j]
+
+    Equivalently, eliminating the inner loop, the last two lines would be::
+
+                a_1d[indices_1d] = values_1d
+
+    See Also
+    --------
+    take_along_axis :
+        Take values from the input array by matching 1d index and data slices
+
+    Examples
+    --------
+
+    For this sample array
+
+    >>> a = np.array([[10, 30, 20], [60, 40, 50]])
+
+    We can replace the maximum values with:
+
+    >>> ai = np.argmax(a, axis=1, keepdims=True)
+    >>> ai
+    array([[1],
+           [0]])
+    >>> np.put_along_axis(a, ai, 99, axis=1)
+    >>> a
+    array([[10, 99, 20],
+           [99, 40, 50]])
+
+    """
+    # normalize inputs
+    if axis is None:
+        arr = arr.flat
+        axis = 0
+        arr_shape = (len(arr),)  # flatiter has no .shape
+    else:
+        axis = normalize_axis_index(axis, arr.ndim)
+        arr_shape = arr.shape
+
+    # use the fancy index
+    arr[_make_along_axis_idx(arr_shape, indices, axis)] = values
+
+
+def _apply_along_axis_dispatcher(func1d, axis, arr, *args, **kwargs):
+    return (arr,)
+
+
+@array_function_dispatch(_apply_along_axis_dispatcher)
+def apply_along_axis(func1d, axis, arr, *args, **kwargs):
+    """
+    Apply a function to 1-D slices along the given axis.
+
+    Execute `func1d(a, *args, **kwargs)` where `func1d` operates on 1-D arrays
+    and `a` is a 1-D slice of `arr` along `axis`.
+
+    This is equivalent to (but faster than) the following use of `ndindex` and
+    `s_`, which sets each of ``ii``, ``jj``, and ``kk`` to a tuple of indices::
+
+        Ni, Nk = a.shape[:axis], a.shape[axis+1:]
+        for ii in ndindex(Ni):
+            for kk in ndindex(Nk):
+                f = func1d(arr[ii + s_[:,] + kk])
+                Nj = f.shape
+                for jj in ndindex(Nj):
+                    out[ii + jj + kk] = f[jj]
+
+    Equivalently, eliminating the inner loop, this can be expressed as::
+
+        Ni, Nk = a.shape[:axis], a.shape[axis+1:]
+        for ii in ndindex(Ni):
+            for kk in ndindex(Nk):
+                out[ii + s_[...,] + kk] = func1d(arr[ii + s_[:,] + kk])
+
+    Parameters
+    ----------
+    func1d : function (M,) -> (Nj...)
+        This function should accept 1-D arrays. It is applied to 1-D
+        slices of `arr` along the specified axis.
+    axis : integer
+        Axis along which `arr` is sliced.
+    arr : ndarray (Ni..., M, Nk...)
+        Input array.
+    args : any
+        Additional arguments to `func1d`.
+    kwargs : any
+        Additional named arguments to `func1d`.
+
+        .. versionadded:: 1.9.0
+
+
+    Returns
+    -------
+    out : ndarray  (Ni..., Nj..., Nk...)
+        The output array. The shape of `out` is identical to the shape of
+        `arr`, except along the `axis` dimension. This axis is removed, and
+        replaced with new dimensions equal to the shape of the return value
+        of `func1d`. So if `func1d` returns a scalar `out` will have one
+        fewer dimensions than `arr`.
+
+    See Also
+    --------
+    apply_over_axes : Apply a function repeatedly over multiple axes.
+
+    Examples
+    --------
+    >>> def my_func(a):
+    ...     \"\"\"Average first and last element of a 1-D array\"\"\"
+    ...     return (a[0] + a[-1]) * 0.5
+    >>> b = np.array([[1,2,3], [4,5,6], [7,8,9]])
+    >>> np.apply_along_axis(my_func, 0, b)
+    array([4., 5., 6.])
+    >>> np.apply_along_axis(my_func, 1, b)
+    array([2.,  5.,  8.])
+
+    For a function that returns a 1D array, the number of dimensions in
+    `outarr` is the same as `arr`.
+
+    >>> b = np.array([[8,1,7], [4,3,9], [5,2,6]])
+    >>> np.apply_along_axis(sorted, 1, b)
+    array([[1, 7, 8],
+           [3, 4, 9],
+           [2, 5, 6]])
+
+    For a function that returns a higher dimensional array, those dimensions
+    are inserted in place of the `axis` dimension.
+
+    >>> b = np.array([[1,2,3], [4,5,6], [7,8,9]])
+    >>> np.apply_along_axis(np.diag, -1, b)
+    array([[[1, 0, 0],
+            [0, 2, 0],
+            [0, 0, 3]],
+           [[4, 0, 0],
+            [0, 5, 0],
+            [0, 0, 6]],
+           [[7, 0, 0],
+            [0, 8, 0],
+            [0, 0, 9]]])
+    """
+    # handle negative axes
+    arr = asanyarray(arr)
+    nd = arr.ndim
+    axis = normalize_axis_index(axis, nd)
+
+    # arr, with the iteration axis at the end
+    in_dims = list(range(nd))
+    inarr_view = transpose(arr, in_dims[:axis] + in_dims[axis+1:] + [axis])
+
+    # compute indices for the iteration axes, and append a trailing ellipsis to
+    # prevent 0d arrays decaying to scalars, which fixes gh-8642
+    inds = ndindex(inarr_view.shape[:-1])
+    inds = (ind + (Ellipsis,) for ind in inds)
+
+    # invoke the function on the first item
+    try:
+        ind0 = next(inds)
+    except StopIteration:
+        raise ValueError(
+            'Cannot apply_along_axis when any iteration dimensions are 0'
+        ) from None
+    res = asanyarray(func1d(inarr_view[ind0], *args, **kwargs))
+
+    # build a buffer for storing evaluations of func1d.
+    # remove the requested axis, and add the new ones on the end.
+    # laid out so that each write is contiguous.
+    # for a tuple index inds, buff[inds] = func1d(inarr_view[inds])
+    buff = zeros(inarr_view.shape[:-1] + res.shape, res.dtype)
+
+    # permutation of axes such that out = buff.transpose(buff_permute)
+    buff_dims = list(range(buff.ndim))
+    buff_permute = (
+        buff_dims[0 : axis] +
+        buff_dims[buff.ndim-res.ndim : buff.ndim] +
+        buff_dims[axis : buff.ndim-res.ndim]
+    )
+
+    # matrices have a nasty __array_prepare__ and __array_wrap__
+    if not isinstance(res, matrix):
+        buff = res.__array_prepare__(buff)
+
+    # save the first result, then compute and save all remaining results
+    buff[ind0] = res
+    for ind in inds:
+        buff[ind] = asanyarray(func1d(inarr_view[ind], *args, **kwargs))
+
+    if not isinstance(res, matrix):
+        # wrap the array, to preserve subclasses
+        buff = res.__array_wrap__(buff)
+
+        # finally, rotate the inserted axes back to where they belong
+        return transpose(buff, buff_permute)
+
+    else:
+        # matrices have to be transposed first, because they collapse dimensions!
+        out_arr = transpose(buff, buff_permute)
+        return res.__array_wrap__(out_arr)
+
+
+def _apply_over_axes_dispatcher(func, a, axes):
+    return (a,)
+
+
+@array_function_dispatch(_apply_over_axes_dispatcher)
+def apply_over_axes(func, a, axes):
+    """
+    Apply a function repeatedly over multiple axes.
+
+    `func` is called as `res = func(a, axis)`, where `axis` is the first
+    element of `axes`.  The result `res` of the function call must have
+    either the same dimensions as `a` or one less dimension.  If `res`
+    has one less dimension than `a`, a dimension is inserted before
+    `axis`.  The call to `func` is then repeated for each axis in `axes`,
+    with `res` as the first argument.
+
+    Parameters
+    ----------
+    func : function
+        This function must take two arguments, `func(a, axis)`.
+    a : array_like
+        Input array.
+    axes : array_like
+        Axes over which `func` is applied; the elements must be integers.
+
+    Returns
+    -------
+    apply_over_axis : ndarray
+        The output array.  The number of dimensions is the same as `a`,
+        but the shape can be different.  This depends on whether `func`
+        changes the shape of its output with respect to its input.
+
+    See Also
+    --------
+    apply_along_axis :
+        Apply a function to 1-D slices of an array along the given axis.
+
+    Notes
+    -----
+    This function is equivalent to tuple axis arguments to reorderable ufuncs
+    with keepdims=True. Tuple axis arguments to ufuncs have been available since
+    version 1.7.0.
+
+    Examples
+    --------
+    >>> a = np.arange(24).reshape(2,3,4)
+    >>> a
+    array([[[ 0,  1,  2,  3],
+            [ 4,  5,  6,  7],
+            [ 8,  9, 10, 11]],
+           [[12, 13, 14, 15],
+            [16, 17, 18, 19],
+            [20, 21, 22, 23]]])
+
+    Sum over axes 0 and 2. The result has same number of dimensions
+    as the original array:
+
+    >>> np.apply_over_axes(np.sum, a, [0,2])
+    array([[[ 60],
+            [ 92],
+            [124]]])
+
+    Tuple axis arguments to ufuncs are equivalent:
+
+    >>> np.sum(a, axis=(0,2), keepdims=True)
+    array([[[ 60],
+            [ 92],
+            [124]]])
+
+    """
+    val = asarray(a)
+    N = a.ndim
+    if array(axes).ndim == 0:
+        axes = (axes,)
+    for axis in axes:
+        if axis < 0:
+            axis = N + axis
+        args = (val, axis)
+        res = func(*args)
+        if res.ndim == val.ndim:
+            val = res
+        else:
+            res = expand_dims(res, axis)
+            if res.ndim == val.ndim:
+                val = res
+            else:
+                raise ValueError("function is not returning "
+                                 "an array of the correct shape")
+    return val
+
+
+def _expand_dims_dispatcher(a, axis):
+    return (a,)
+
+
+@array_function_dispatch(_expand_dims_dispatcher)
+def expand_dims(a, axis):
+    """
+    Expand the shape of an array.
+
+    Insert a new axis that will appear at the `axis` position in the expanded
+    array shape.
+
+    Parameters
+    ----------
+    a : array_like
+        Input array.
+    axis : int or tuple of ints
+        Position in the expanded axes where the new axis (or axes) is placed.
+
+        .. deprecated:: 1.13.0
+            Passing an axis where ``axis > a.ndim`` will be treated as
+            ``axis == a.ndim``, and passing ``axis < -a.ndim - 1`` will
+            be treated as ``axis == 0``. This behavior is deprecated.
+
+        .. versionchanged:: 1.18.0
+            A tuple of axes is now supported.  Out of range axes as
+            described above are now forbidden and raise an `AxisError`.
+
+    Returns
+    -------
+    result : ndarray
+        View of `a` with the number of dimensions increased.
+
+    See Also
+    --------
+    squeeze : The inverse operation, removing singleton dimensions
+    reshape : Insert, remove, and combine dimensions, and resize existing ones
+    doc.indexing, atleast_1d, atleast_2d, atleast_3d
+
+    Examples
+    --------
+    >>> x = np.array([1, 2])
+    >>> x.shape
+    (2,)
+
+    The following is equivalent to ``x[np.newaxis, :]`` or ``x[np.newaxis]``:
+
+    >>> y = np.expand_dims(x, axis=0)
+    >>> y
+    array([[1, 2]])
+    >>> y.shape
+    (1, 2)
+
+    The following is equivalent to ``x[:, np.newaxis]``:
+
+    >>> y = np.expand_dims(x, axis=1)
+    >>> y
+    array([[1],
+           [2]])
+    >>> y.shape
+    (2, 1)
+
+    ``axis`` may also be a tuple:
+
+    >>> y = np.expand_dims(x, axis=(0, 1))
+    >>> y
+    array([[[1, 2]]])
+
+    >>> y = np.expand_dims(x, axis=(2, 0))
+    >>> y
+    array([[[1],
+            [2]]])
+
+    Note that some examples may use ``None`` instead of ``np.newaxis``.  These
+    are the same objects:
+
+    >>> np.newaxis is None
+    True
+
+    """
+    if isinstance(a, matrix):
+        a = asarray(a)
+    else:
+        a = asanyarray(a)
+
+    if type(axis) not in (tuple, list):
+        axis = (axis,)
+
+    out_ndim = len(axis) + a.ndim
+    axis = normalize_axis_tuple(axis, out_ndim)
+
+    shape_it = iter(a.shape)
+    shape = [1 if ax in axis else next(shape_it) for ax in range(out_ndim)]
+
+    return a.reshape(shape)
+
+
+row_stack = vstack
+
+
+def _column_stack_dispatcher(tup):
+    return _arrays_for_stack_dispatcher(tup)
+
+
+@array_function_dispatch(_column_stack_dispatcher)
+def column_stack(tup):
+    """
+    Stack 1-D arrays as columns into a 2-D array.
+
+    Take a sequence of 1-D arrays and stack them as columns
+    to make a single 2-D array. 2-D arrays are stacked as-is,
+    just like with `hstack`.  1-D arrays are turned into 2-D columns
+    first.
+
+    Parameters
+    ----------
+    tup : sequence of 1-D or 2-D arrays.
+        Arrays to stack. All of them must have the same first dimension.
+
+    Returns
+    -------
+    stacked : 2-D array
+        The array formed by stacking the given arrays.
+
+    See Also
+    --------
+    stack, hstack, vstack, concatenate
+
+    Examples
+    --------
+    >>> a = np.array((1,2,3))
+    >>> b = np.array((2,3,4))
+    >>> np.column_stack((a,b))
+    array([[1, 2],
+           [2, 3],
+           [3, 4]])
+
+    """
+    arrays = []
+    for v in tup:
+        arr = asanyarray(v)
+        if arr.ndim < 2:
+            arr = array(arr, copy=False, subok=True, ndmin=2).T
+        arrays.append(arr)
+    return _nx.concatenate(arrays, 1)
+
+
+def _dstack_dispatcher(tup):
+    return _arrays_for_stack_dispatcher(tup)
+
+
+@array_function_dispatch(_dstack_dispatcher)
+def dstack(tup):
+    """
+    Stack arrays in sequence depth wise (along third axis).
+
+    This is equivalent to concatenation along the third axis after 2-D arrays
+    of shape `(M,N)` have been reshaped to `(M,N,1)` and 1-D arrays of shape
+    `(N,)` have been reshaped to `(1,N,1)`. Rebuilds arrays divided by
+    `dsplit`.
+
+    This function makes most sense for arrays with up to 3 dimensions. For
+    instance, for pixel-data with a height (first axis), width (second axis),
+    and r/g/b channels (third axis). The functions `concatenate`, `stack` and
+    `block` provide more general stacking and concatenation operations.
+
+    Parameters
+    ----------
+    tup : sequence of arrays
+        The arrays must have the same shape along all but the third axis.
+        1-D or 2-D arrays must have the same shape.
+
+    Returns
+    -------
+    stacked : ndarray
+        The array formed by stacking the given arrays, will be at least 3-D.
+
+    See Also
+    --------
+    concatenate : Join a sequence of arrays along an existing axis.
+    stack : Join a sequence of arrays along a new axis.
+    block : Assemble an nd-array from nested lists of blocks.
+    vstack : Stack arrays in sequence vertically (row wise).
+    hstack : Stack arrays in sequence horizontally (column wise).
+    column_stack : Stack 1-D arrays as columns into a 2-D array.
+    dsplit : Split array along third axis.
+
+    Examples
+    --------
+    >>> a = np.array((1,2,3))
+    >>> b = np.array((2,3,4))
+    >>> np.dstack((a,b))
+    array([[[1, 2],
+            [2, 3],
+            [3, 4]]])
+
+    >>> a = np.array([[1],[2],[3]])
+    >>> b = np.array([[2],[3],[4]])
+    >>> np.dstack((a,b))
+    array([[[1, 2]],
+           [[2, 3]],
+           [[3, 4]]])
+
+    """
+    arrs = atleast_3d(*tup)
+    if not isinstance(arrs, list):
+        arrs = [arrs]
+    return _nx.concatenate(arrs, 2)
+
+
+def _replace_zero_by_x_arrays(sub_arys):
+    for i in range(len(sub_arys)):
+        if _nx.ndim(sub_arys[i]) == 0:
+            sub_arys[i] = _nx.empty(0, dtype=sub_arys[i].dtype)
+        elif _nx.sometrue(_nx.equal(_nx.shape(sub_arys[i]), 0)):
+            sub_arys[i] = _nx.empty(0, dtype=sub_arys[i].dtype)
+    return sub_arys
+
+
+def _array_split_dispatcher(ary, indices_or_sections, axis=None):
+    return (ary, indices_or_sections)
+
+
+@array_function_dispatch(_array_split_dispatcher)
+def array_split(ary, indices_or_sections, axis=0):
+    """
+    Split an array into multiple sub-arrays.
+
+    Please refer to the ``split`` documentation.  The only difference
+    between these functions is that ``array_split`` allows
+    `indices_or_sections` to be an integer that does *not* equally
+    divide the axis. For an array of length l that should be split
+    into n sections, it returns l % n sub-arrays of size l//n + 1
+    and the rest of size l//n.
+
+    See Also
+    --------
+    split : Split array into multiple sub-arrays of equal size.
+
+    Examples
+    --------
+    >>> x = np.arange(8.0)
+    >>> np.array_split(x, 3)
+    [array([0.,  1.,  2.]), array([3.,  4.,  5.]), array([6.,  7.])]
+
+    >>> x = np.arange(9)
+    >>> np.array_split(x, 4)
+    [array([0, 1, 2]), array([3, 4]), array([5, 6]), array([7, 8])]
+
+    """
+    try:
+        Ntotal = ary.shape[axis]
+    except AttributeError:
+        Ntotal = len(ary)
+    try:
+        # handle array case.
+        Nsections = len(indices_or_sections) + 1
+        div_points = [0] + list(indices_or_sections) + [Ntotal]
+    except TypeError:
+        # indices_or_sections is a scalar, not an array.
+        Nsections = int(indices_or_sections)
+        if Nsections <= 0:
+            raise ValueError('number sections must be larger than 0.') from None
+        Neach_section, extras = divmod(Ntotal, Nsections)
+        section_sizes = ([0] +
+                         extras * [Neach_section+1] +
+                         (Nsections-extras) * [Neach_section])
+        div_points = _nx.array(section_sizes, dtype=_nx.intp).cumsum()
+
+    sub_arys = []
+    sary = _nx.swapaxes(ary, axis, 0)
+    for i in range(Nsections):
+        st = div_points[i]
+        end = div_points[i + 1]
+        sub_arys.append(_nx.swapaxes(sary[st:end], axis, 0))
+
+    return sub_arys
+
+
+def _split_dispatcher(ary, indices_or_sections, axis=None):
+    return (ary, indices_or_sections)
+
+
+@array_function_dispatch(_split_dispatcher)
+def split(ary, indices_or_sections, axis=0):
+    """
+    Split an array into multiple sub-arrays as views into `ary`.
+
+    Parameters
+    ----------
+    ary : ndarray
+        Array to be divided into sub-arrays.
+    indices_or_sections : int or 1-D array
+        If `indices_or_sections` is an integer, N, the array will be divided
+        into N equal arrays along `axis`.  If such a split is not possible,
+        an error is raised.
+
+        If `indices_or_sections` is a 1-D array of sorted integers, the entries
+        indicate where along `axis` the array is split.  For example,
+        ``[2, 3]`` would, for ``axis=0``, result in
+
+          - ary[:2]
+          - ary[2:3]
+          - ary[3:]
+
+        If an index exceeds the dimension of the array along `axis`,
+        an empty sub-array is returned correspondingly.
+    axis : int, optional
+        The axis along which to split, default is 0.
+
+    Returns
+    -------
+    sub-arrays : list of ndarrays
+        A list of sub-arrays as views into `ary`.
+
+    Raises
+    ------
+    ValueError
+        If `indices_or_sections` is given as an integer, but
+        a split does not result in equal division.
+
+    See Also
+    --------
+    array_split : Split an array into multiple sub-arrays of equal or
+                  near-equal size.  Does not raise an exception if
+                  an equal division cannot be made.
+    hsplit : Split array into multiple sub-arrays horizontally (column-wise).
+    vsplit : Split array into multiple sub-arrays vertically (row wise).
+    dsplit : Split array into multiple sub-arrays along the 3rd axis (depth).
+    concatenate : Join a sequence of arrays along an existing axis.
+    stack : Join a sequence of arrays along a new axis.
+    hstack : Stack arrays in sequence horizontally (column wise).
+    vstack : Stack arrays in sequence vertically (row wise).
+    dstack : Stack arrays in sequence depth wise (along third dimension).
+
+    Examples
+    --------
+    >>> x = np.arange(9.0)
+    >>> np.split(x, 3)
+    [array([0.,  1.,  2.]), array([3.,  4.,  5.]), array([6.,  7.,  8.])]
+
+    >>> x = np.arange(8.0)
+    >>> np.split(x, [3, 5, 6, 10])
+    [array([0.,  1.,  2.]),
+     array([3.,  4.]),
+     array([5.]),
+     array([6.,  7.]),
+     array([], dtype=float64)]
+
+    """
+    try:
+        len(indices_or_sections)
+    except TypeError:
+        sections = indices_or_sections
+        N = ary.shape[axis]
+        if N % sections:
+            raise ValueError(
+                'array split does not result in an equal division') from None
+    return array_split(ary, indices_or_sections, axis)
+
+
+def _hvdsplit_dispatcher(ary, indices_or_sections):
+    return (ary, indices_or_sections)
+
+
+@array_function_dispatch(_hvdsplit_dispatcher)
+def hsplit(ary, indices_or_sections):
+    """
+    Split an array into multiple sub-arrays horizontally (column-wise).
+
+    Please refer to the `split` documentation.  `hsplit` is equivalent
+    to `split` with ``axis=1``, the array is always split along the second
+    axis except for 1-D arrays, where it is split at ``axis=0``.
+
+    See Also
+    --------
+    split : Split an array into multiple sub-arrays of equal size.
+
+    Examples
+    --------
+    >>> x = np.arange(16.0).reshape(4, 4)
+    >>> x
+    array([[ 0.,   1.,   2.,   3.],
+           [ 4.,   5.,   6.,   7.],
+           [ 8.,   9.,  10.,  11.],
+           [12.,  13.,  14.,  15.]])
+    >>> np.hsplit(x, 2)
+    [array([[  0.,   1.],
+           [  4.,   5.],
+           [  8.,   9.],
+           [12.,  13.]]),
+     array([[  2.,   3.],
+           [  6.,   7.],
+           [10.,  11.],
+           [14.,  15.]])]
+    >>> np.hsplit(x, np.array([3, 6]))
+    [array([[ 0.,   1.,   2.],
+           [ 4.,   5.,   6.],
+           [ 8.,   9.,  10.],
+           [12.,  13.,  14.]]),
+     array([[ 3.],
+           [ 7.],
+           [11.],
+           [15.]]),
+     array([], shape=(4, 0), dtype=float64)]
+
+    With a higher dimensional array the split is still along the second axis.
+
+    >>> x = np.arange(8.0).reshape(2, 2, 2)
+    >>> x
+    array([[[0.,  1.],
+            [2.,  3.]],
+           [[4.,  5.],
+            [6.,  7.]]])
+    >>> np.hsplit(x, 2)
+    [array([[[0.,  1.]],
+           [[4.,  5.]]]),
+     array([[[2.,  3.]],
+           [[6.,  7.]]])]
+
+    With a 1-D array, the split is along axis 0.
+
+    >>> x = np.array([0, 1, 2, 3, 4, 5])
+    >>> np.hsplit(x, 2)
+    [array([0, 1, 2]), array([3, 4, 5])]
+
+    """
+    if _nx.ndim(ary) == 0:
+        raise ValueError('hsplit only works on arrays of 1 or more dimensions')
+    if ary.ndim > 1:
+        return split(ary, indices_or_sections, 1)
+    else:
+        return split(ary, indices_or_sections, 0)
+
+
+@array_function_dispatch(_hvdsplit_dispatcher)
+def vsplit(ary, indices_or_sections):
+    """
+    Split an array into multiple sub-arrays vertically (row-wise).
+
+    Please refer to the ``split`` documentation.  ``vsplit`` is equivalent
+    to ``split`` with `axis=0` (default), the array is always split along the
+    first axis regardless of the array dimension.
+
+    See Also
+    --------
+    split : Split an array into multiple sub-arrays of equal size.
+
+    Examples
+    --------
+    >>> x = np.arange(16.0).reshape(4, 4)
+    >>> x
+    array([[ 0.,   1.,   2.,   3.],
+           [ 4.,   5.,   6.,   7.],
+           [ 8.,   9.,  10.,  11.],
+           [12.,  13.,  14.,  15.]])
+    >>> np.vsplit(x, 2)
+    [array([[0., 1., 2., 3.],
+           [4., 5., 6., 7.]]), array([[ 8.,  9., 10., 11.],
+           [12., 13., 14., 15.]])]
+    >>> np.vsplit(x, np.array([3, 6]))
+    [array([[ 0.,  1.,  2.,  3.],
+           [ 4.,  5.,  6.,  7.],
+           [ 8.,  9., 10., 11.]]), array([[12., 13., 14., 15.]]), array([], shape=(0, 4), dtype=float64)]
+
+    With a higher dimensional array the split is still along the first axis.
+
+    >>> x = np.arange(8.0).reshape(2, 2, 2)
+    >>> x
+    array([[[0.,  1.],
+            [2.,  3.]],
+           [[4.,  5.],
+            [6.,  7.]]])
+    >>> np.vsplit(x, 2)
+    [array([[[0., 1.],
+            [2., 3.]]]), array([[[4., 5.],
+            [6., 7.]]])]
+
+    """
+    if _nx.ndim(ary) < 2:
+        raise ValueError('vsplit only works on arrays of 2 or more dimensions')
+    return split(ary, indices_or_sections, 0)
+
+
+@array_function_dispatch(_hvdsplit_dispatcher)
+def dsplit(ary, indices_or_sections):
+    """
+    Split array into multiple sub-arrays along the 3rd axis (depth).
+
+    Please refer to the `split` documentation.  `dsplit` is equivalent
+    to `split` with ``axis=2``, the array is always split along the third
+    axis provided the array dimension is greater than or equal to 3.
+
+    See Also
+    --------
+    split : Split an array into multiple sub-arrays of equal size.
+
+    Examples
+    --------
+    >>> x = np.arange(16.0).reshape(2, 2, 4)
+    >>> x
+    array([[[ 0.,   1.,   2.,   3.],
+            [ 4.,   5.,   6.,   7.]],
+           [[ 8.,   9.,  10.,  11.],
+            [12.,  13.,  14.,  15.]]])
+    >>> np.dsplit(x, 2)
+    [array([[[ 0.,  1.],
+            [ 4.,  5.]],
+           [[ 8.,  9.],
+            [12., 13.]]]), array([[[ 2.,  3.],
+            [ 6.,  7.]],
+           [[10., 11.],
+            [14., 15.]]])]
+    >>> np.dsplit(x, np.array([3, 6]))
+    [array([[[ 0.,   1.,   2.],
+            [ 4.,   5.,   6.]],
+           [[ 8.,   9.,  10.],
+            [12.,  13.,  14.]]]),
+     array([[[ 3.],
+            [ 7.]],
+           [[11.],
+            [15.]]]),
+    array([], shape=(2, 2, 0), dtype=float64)]
+    """
+    if _nx.ndim(ary) < 3:
+        raise ValueError('dsplit only works on arrays of 3 or more dimensions')
+    return split(ary, indices_or_sections, 2)
+
+
+def get_array_prepare(*args):
+    """Find the wrapper for the array with the highest priority.
+
+    In case of ties, leftmost wins. If no wrapper is found, return None
+    """
+    wrappers = sorted((getattr(x, '__array_priority__', 0), -i,
+                 x.__array_prepare__) for i, x in enumerate(args)
+                                   if hasattr(x, '__array_prepare__'))
+    if wrappers:
+        return wrappers[-1][-1]
+    return None
+
+
+def get_array_wrap(*args):
+    """Find the wrapper for the array with the highest priority.
+
+    In case of ties, leftmost wins. If no wrapper is found, return None
+    """
+    wrappers = sorted((getattr(x, '__array_priority__', 0), -i,
+                 x.__array_wrap__) for i, x in enumerate(args)
+                                   if hasattr(x, '__array_wrap__'))
+    if wrappers:
+        return wrappers[-1][-1]
+    return None
+
+
+def _kron_dispatcher(a, b):
+    return (a, b)
+
+
+@array_function_dispatch(_kron_dispatcher)
+def kron(a, b):
+    """
+    Kronecker product of two arrays.
+
+    Computes the Kronecker product, a composite array made of blocks of the
+    second array scaled by the first.
+
+    Parameters
+    ----------
+    a, b : array_like
+
+    Returns
+    -------
+    out : ndarray
+
+    See Also
+    --------
+    outer : The outer product
+
+    Notes
+    -----
+    The function assumes that the number of dimensions of `a` and `b`
+    are the same, if necessary prepending the smallest with ones.
+    If ``a.shape = (r0,r1,..,rN)`` and ``b.shape = (s0,s1,...,sN)``,
+    the Kronecker product has shape ``(r0*s0, r1*s1, ..., rN*SN)``.
+    The elements are products of elements from `a` and `b`, organized
+    explicitly by::
+
+        kron(a,b)[k0,k1,...,kN] = a[i0,i1,...,iN] * b[j0,j1,...,jN]
+
+    where::
+
+        kt = it * st + jt,  t = 0,...,N
+
+    In the common 2-D case (N=1), the block structure can be visualized::
+
+        [[ a[0,0]*b,   a[0,1]*b,  ... , a[0,-1]*b  ],
+         [  ...                              ...   ],
+         [ a[-1,0]*b,  a[-1,1]*b, ... , a[-1,-1]*b ]]
+
+
+    Examples
+    --------
+    >>> np.kron([1,10,100], [5,6,7])
+    array([  5,   6,   7, ..., 500, 600, 700])
+    >>> np.kron([5,6,7], [1,10,100])
+    array([  5,  50, 500, ...,   7,  70, 700])
+
+    >>> np.kron(np.eye(2), np.ones((2,2)))
+    array([[1.,  1.,  0.,  0.],
+           [1.,  1.,  0.,  0.],
+           [0.,  0.,  1.,  1.],
+           [0.,  0.,  1.,  1.]])
+
+    >>> a = np.arange(100).reshape((2,5,2,5))
+    >>> b = np.arange(24).reshape((2,3,4))
+    >>> c = np.kron(a,b)
+    >>> c.shape
+    (2, 10, 6, 20)
+    >>> I = (1,3,0,2)
+    >>> J = (0,2,1)
+    >>> J1 = (0,) + J             # extend to ndim=4
+    >>> S1 = (1,) + b.shape
+    >>> K = tuple(np.array(I) * np.array(S1) + np.array(J1))
+    >>> c[K] == a[I]*b[J]
+    True
+
+    """
+    # Working:
+    # 1. Equalise the shapes by prepending smaller array with 1s
+    # 2. Expand shapes of both the arrays by adding new axes at
+    #    odd positions for 1st array and even positions for 2nd
+    # 3. Compute the product of the modified array
+    # 4. The inner most array elements now contain the rows of
+    #    the Kronecker product
+    # 5. Reshape the result to kron's shape, which is same as
+    #    product of shapes of the two arrays.
+    b = asanyarray(b)
+    a = array(a, copy=False, subok=True, ndmin=b.ndim)
+    is_any_mat = isinstance(a, matrix) or isinstance(b, matrix)
+    ndb, nda = b.ndim, a.ndim
+    nd = max(ndb, nda)
+
+    if (nda == 0 or ndb == 0):
+        return _nx.multiply(a, b)
+
+    as_ = a.shape
+    bs = b.shape
+    if not a.flags.contiguous:
+        a = reshape(a, as_)
+    if not b.flags.contiguous:
+        b = reshape(b, bs)
+
+    # Equalise the shapes by prepending smaller one with 1s
+    as_ = (1,)*max(0, ndb-nda) + as_
+    bs = (1,)*max(0, nda-ndb) + bs
+
+    # Insert empty dimensions
+    a_arr = expand_dims(a, axis=tuple(range(ndb-nda)))
+    b_arr = expand_dims(b, axis=tuple(range(nda-ndb)))
+
+    # Compute the product
+    a_arr = expand_dims(a_arr, axis=tuple(range(1, nd*2, 2)))
+    b_arr = expand_dims(b_arr, axis=tuple(range(0, nd*2, 2)))
+    # In case of `mat`, convert result to `array`
+    result = _nx.multiply(a_arr, b_arr, subok=(not is_any_mat))
+
+    # Reshape back
+    result = result.reshape(_nx.multiply(as_, bs))
+
+    return result if not is_any_mat else matrix(result, copy=False)
+
+
+def _tile_dispatcher(A, reps):
+    return (A, reps)
+
+
+@array_function_dispatch(_tile_dispatcher)
+def tile(A, reps):
+    """
+    Construct an array by repeating A the number of times given by reps.
+
+    If `reps` has length ``d``, the result will have dimension of
+    ``max(d, A.ndim)``.
+
+    If ``A.ndim < d``, `A` is promoted to be d-dimensional by prepending new
+    axes. So a shape (3,) array is promoted to (1, 3) for 2-D replication,
+    or shape (1, 1, 3) for 3-D replication. If this is not the desired
+    behavior, promote `A` to d-dimensions manually before calling this
+    function.
+
+    If ``A.ndim > d``, `reps` is promoted to `A`.ndim by pre-pending 1's to it.
+    Thus for an `A` of shape (2, 3, 4, 5), a `reps` of (2, 2) is treated as
+    (1, 1, 2, 2).
+
+    Note : Although tile may be used for broadcasting, it is strongly
+    recommended to use numpy's broadcasting operations and functions.
+
+    Parameters
+    ----------
+    A : array_like
+        The input array.
+    reps : array_like
+        The number of repetitions of `A` along each axis.
+
+    Returns
+    -------
+    c : ndarray
+        The tiled output array.
+
+    See Also
+    --------
+    repeat : Repeat elements of an array.
+    broadcast_to : Broadcast an array to a new shape
+
+    Examples
+    --------
+    >>> a = np.array([0, 1, 2])
+    >>> np.tile(a, 2)
+    array([0, 1, 2, 0, 1, 2])
+    >>> np.tile(a, (2, 2))
+    array([[0, 1, 2, 0, 1, 2],
+           [0, 1, 2, 0, 1, 2]])
+    >>> np.tile(a, (2, 1, 2))
+    array([[[0, 1, 2, 0, 1, 2]],
+           [[0, 1, 2, 0, 1, 2]]])
+
+    >>> b = np.array([[1, 2], [3, 4]])
+    >>> np.tile(b, 2)
+    array([[1, 2, 1, 2],
+           [3, 4, 3, 4]])
+    >>> np.tile(b, (2, 1))
+    array([[1, 2],
+           [3, 4],
+           [1, 2],
+           [3, 4]])
+
+    >>> c = np.array([1,2,3,4])
+    >>> np.tile(c,(4,1))
+    array([[1, 2, 3, 4],
+           [1, 2, 3, 4],
+           [1, 2, 3, 4],
+           [1, 2, 3, 4]])
+    """
+    try:
+        tup = tuple(reps)
+    except TypeError:
+        tup = (reps,)
+    d = len(tup)
+    if all(x == 1 for x in tup) and isinstance(A, _nx.ndarray):
+        # Fixes the problem that the function does not make a copy if A is a
+        # numpy array and the repetitions are 1 in all dimensions
+        return _nx.array(A, copy=True, subok=True, ndmin=d)
+    else:
+        # Note that no copy of zero-sized arrays is made. However since they
+        # have no data there is no risk of an inadvertent overwrite.
+        c = _nx.array(A, copy=False, subok=True, ndmin=d)
+    if (d < c.ndim):
+        tup = (1,)*(c.ndim-d) + tup
+    shape_out = tuple(s*t for s, t in zip(c.shape, tup))
+    n = c.size
+    if n > 0:
+        for dim_in, nrep in zip(c.shape, tup):
+            if nrep != 1:
+                c = c.reshape(-1, n).repeat(nrep, 0)
+            n //= dim_in
+    return c.reshape(shape_out)

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/stride_tricks.pyi

@@ -0,0 +1,80 @@
+from collections.abc import Iterable
+from typing import Any, TypeVar, overload, SupportsIndex
+
+from numpy import generic
+from numpy._typing import (
+    NDArray,
+    ArrayLike,
+    _ShapeLike,
+    _Shape,
+    _ArrayLike
+)
+
+_SCT = TypeVar("_SCT", bound=generic)
+
+__all__: list[str]
+
+class DummyArray:
+    __array_interface__: dict[str, Any]
+    base: None | NDArray[Any]
+    def __init__(
+        self,
+        interface: dict[str, Any],
+        base: None | NDArray[Any] = ...,
+    ) -> None: ...
+
+@overload
+def as_strided(
+    x: _ArrayLike[_SCT],
+    shape: None | Iterable[int] = ...,
+    strides: None | Iterable[int] = ...,
+    subok: bool = ...,
+    writeable: bool = ...,
+) -> NDArray[_SCT]: ...
+@overload
+def as_strided(
+    x: ArrayLike,
+    shape: None | Iterable[int] = ...,
+    strides: None | Iterable[int] = ...,
+    subok: bool = ...,
+    writeable: bool = ...,
+) -> NDArray[Any]: ...
+
+@overload
+def sliding_window_view(
+    x: _ArrayLike[_SCT],
+    window_shape: int | Iterable[int],
+    axis: None | SupportsIndex = ...,
+    *,
+    subok: bool = ...,
+    writeable: bool = ...,
+) -> NDArray[_SCT]: ...
+@overload
+def sliding_window_view(
+    x: ArrayLike,
+    window_shape: int | Iterable[int],
+    axis: None | SupportsIndex = ...,
+    *,
+    subok: bool = ...,
+    writeable: bool = ...,
+) -> NDArray[Any]: ...
+
+@overload
+def broadcast_to(
+    array: _ArrayLike[_SCT],
+    shape: int | Iterable[int],
+    subok: bool = ...,
+) -> NDArray[_SCT]: ...
+@overload
+def broadcast_to(
+    array: ArrayLike,
+    shape: int | Iterable[int],
+    subok: bool = ...,
+) -> NDArray[Any]: ...
+
+def broadcast_shapes(*args: _ShapeLike) -> _Shape: ...
+
+def broadcast_arrays(
+    *args: ArrayLike,
+    subok: bool = ...,
+) -> list[NDArray[Any]]: ...

LTA_openwebtext_dualt/mini_owt_logdirichlet/.venv_qwen35_uv/lib/python3.12/site-packages/numpy/lib/type_check.py

@@ -0,0 +1,735 @@
+"""Automatically adapted for numpy Sep 19, 2005 by convertcode.py
+
+"""
+import functools
+
+__all__ = ['iscomplexobj', 'isrealobj', 'imag', 'iscomplex',
+           'isreal', 'nan_to_num', 'real', 'real_if_close',
+           'typename', 'asfarray', 'mintypecode',
+           'common_type']
+
+from .._utils import set_module
+import numpy.core.numeric as _nx
+from numpy.core.numeric import asarray, asanyarray, isnan, zeros
+from numpy.core import overrides, getlimits
+from .ufunclike import isneginf, isposinf
+
+
+array_function_dispatch = functools.partial(
+    overrides.array_function_dispatch, module='numpy')
+
+
+_typecodes_by_elsize = 'GDFgdfQqLlIiHhBb?'
+
+
+@set_module('numpy')
+def mintypecode(typechars, typeset='GDFgdf', default='d'):
+    """
+    Return the character for the minimum-size type to which given types can
+    be safely cast.
+
+    The returned type character must represent the smallest size dtype such
+    that an array of the returned type can handle the data from an array of
+    all types in `typechars` (or if `typechars` is an array, then its
+    dtype.char).
+
+    Parameters
+    ----------
+    typechars : list of str or array_like
+        If a list of strings, each string should represent a dtype.
+        If array_like, the character representation of the array dtype is used.
+    typeset : str or list of str, optional
+        The set of characters that the returned character is chosen from.
+        The default set is 'GDFgdf'.
+    default : str, optional
+        The default character, this is returned if none of the characters in
+        `typechars` matches a character in `typeset`.
+
+    Returns
+    -------
+    typechar : str
+        The character representing the minimum-size type that was found.
+
+    See Also
+    --------
+    dtype, sctype2char, maximum_sctype
+
+    Examples
+    --------
+    >>> np.mintypecode(['d', 'f', 'S'])
+    'd'
+    >>> x = np.array([1.1, 2-3.j])
+    >>> np.mintypecode(x)
+    'D'
+
+    >>> np.mintypecode('abceh', default='G')
+    'G'
+
+    """
+    typecodes = ((isinstance(t, str) and t) or asarray(t).dtype.char
+                 for t in typechars)
+    intersection = set(t for t in typecodes if t in typeset)
+    if not intersection:
+        return default
+    if 'F' in intersection and 'd' in intersection:
+        return 'D'
+    return min(intersection, key=_typecodes_by_elsize.index)
+
+
+def _asfarray_dispatcher(a, dtype=None):
+    return (a,)
+
+
+@array_function_dispatch(_asfarray_dispatcher)
+def asfarray(a, dtype=_nx.float_):
+    """
+    Return an array converted to a float type.
+
+    Parameters
+    ----------
+    a : array_like
+        The input array.
+    dtype : str or dtype object, optional
+        Float type code to coerce input array `a`.  If `dtype` is one of the
+        'int' dtypes, it is replaced with float64.
+
+    Returns
+    -------
+    out : ndarray
+        The input `a` as a float ndarray.
+
+    Examples
+    --------
+    >>> np.asfarray([2, 3])
+    array([2.,  3.])
+    >>> np.asfarray([2, 3], dtype='float')
+    array([2.,  3.])
+    >>> np.asfarray([2, 3], dtype='int8')
+    array([2.,  3.])
+
+    """
+    if not _nx.issubdtype(dtype, _nx.inexact):
+        dtype = _nx.float_
+    return asarray(a, dtype=dtype)
+
+
+def _real_dispatcher(val):
+    return (val,)
+
+
+@array_function_dispatch(_real_dispatcher)
+def real(val):
+    """
+    Return the real part of the complex argument.
+
+    Parameters
+    ----------
+    val : array_like
+        Input array.
+
+    Returns
+    -------
+    out : ndarray or scalar
+        The real component of the complex argument. If `val` is real, the type
+        of `val` is used for the output.  If `val` has complex elements, the
+        returned type is float.
+
+    See Also
+    --------
+    real_if_close, imag, angle
+
+    Examples
+    --------
+    >>> a = np.array([1+2j, 3+4j, 5+6j])
+    >>> a.real
+    array([1.,  3.,  5.])
+    >>> a.real = 9
+    >>> a
+    array([9.+2.j,  9.+4.j,  9.+6.j])
+    >>> a.real = np.array([9, 8, 7])
+    >>> a
+    array([9.+2.j,  8.+4.j,  7.+6.j])
+    >>> np.real(1 + 1j)
+    1.0
+
+    """
+    try:
+        return val.real
+    except AttributeError:
+        return asanyarray(val).real
+
+
+def _imag_dispatcher(val):
+    return (val,)
+
+
+@array_function_dispatch(_imag_dispatcher)
+def imag(val):
+    """
+    Return the imaginary part of the complex argument.
+
+    Parameters
+    ----------
+    val : array_like
+        Input array.
+
+    Returns
+    -------
+    out : ndarray or scalar
+        The imaginary component of the complex argument. If `val` is real,
+        the type of `val` is used for the output.  If `val` has complex
+        elements, the returned type is float.
+
+    See Also
+    --------
+    real, angle, real_if_close
+
+    Examples
+    --------
+    >>> a = np.array([1+2j, 3+4j, 5+6j])
+    >>> a.imag
+    array([2.,  4.,  6.])
+    >>> a.imag = np.array([8, 10, 12])
+    >>> a
+    array([1. +8.j,  3.+10.j,  5.+12.j])
+    >>> np.imag(1 + 1j)
+    1.0
+
+    """
+    try:
+        return val.imag
+    except AttributeError:
+        return asanyarray(val).imag
+
+
+def _is_type_dispatcher(x):
+    return (x,)
+
+
+@array_function_dispatch(_is_type_dispatcher)
+def iscomplex(x):
+    """
+    Returns a bool array, where True if input element is complex.
+
+    What is tested is whether the input has a non-zero imaginary part, not if
+    the input type is complex.
+
+    Parameters
+    ----------
+    x : array_like
+        Input array.
+
+    Returns
+    -------
+    out : ndarray of bools
+        Output array.
+
+    See Also
+    --------
+    isreal
+    iscomplexobj : Return True if x is a complex type or an array of complex
+                   numbers.
+
+    Examples
+    --------
+    >>> np.iscomplex([1+1j, 1+0j, 4.5, 3, 2, 2j])
+    array([ True, False, False, False, False,  True])
+
+    """
+    ax = asanyarray(x)
+    if issubclass(ax.dtype.type, _nx.complexfloating):
+        return ax.imag != 0
+    res = zeros(ax.shape, bool)
+    return res[()]   # convert to scalar if needed
+
+
+@array_function_dispatch(_is_type_dispatcher)
+def isreal(x):
+    """
+    Returns a bool array, where True if input element is real.
+
+    If element has complex type with zero complex part, the return value
+    for that element is True.
+
+    Parameters
+    ----------
+    x : array_like
+        Input array.
+
+    Returns
+    -------
+    out : ndarray, bool
+        Boolean array of same shape as `x`.
+
+    Notes
+    -----
+    `isreal` may behave unexpectedly for string or object arrays (see examples)
+
+    See Also
+    --------
+    iscomplex
+    isrealobj : Return True if x is not a complex type.
+
+    Examples
+    --------
+    >>> a = np.array([1+1j, 1+0j, 4.5, 3, 2, 2j], dtype=complex)
+    >>> np.isreal(a)
+    array([False,  True,  True,  True,  True, False])
+
+    The function does not work on string arrays.
+
+    >>> a = np.array([2j, "a"], dtype="U")
+    >>> np.isreal(a)  # Warns about non-elementwise comparison
+    False
+
+    Returns True for all elements in input array of ``dtype=object`` even if
+    any of the elements is complex.
+
+    >>> a = np.array([1, "2", 3+4j], dtype=object)
+    >>> np.isreal(a)
+    array([ True,  True,  True])
+
+    isreal should not be used with object arrays
+
+    >>> a = np.array([1+2j, 2+1j], dtype=object)
+    >>> np.isreal(a)
+    array([ True,  True])
+
+    """
+    return imag(x) == 0
+
+
+@array_function_dispatch(_is_type_dispatcher)
+def iscomplexobj(x):
+    """
+    Check for a complex type or an array of complex numbers.
+
+    The type of the input is checked, not the value. Even if the input
+    has an imaginary part equal to zero, `iscomplexobj` evaluates to True.
+
+    Parameters
+    ----------
+    x : any
+        The input can be of any type and shape.
+
+    Returns
+    -------
+    iscomplexobj : bool
+        The return value, True if `x` is of a complex type or has at least
+        one complex element.
+
+    See Also
+    --------
+    isrealobj, iscomplex
+
+    Examples
+    --------
+    >>> np.iscomplexobj(1)
+    False
+    >>> np.iscomplexobj(1+0j)
+    True
+    >>> np.iscomplexobj([3, 1+0j, True])
+    True
+
+    """
+    try:
+        dtype = x.dtype
+        type_ = dtype.type
+    except AttributeError:
+        type_ = asarray(x).dtype.type
+    return issubclass(type_, _nx.complexfloating)
+
+
+@array_function_dispatch(_is_type_dispatcher)
+def isrealobj(x):
+    """
+    Return True if x is a not complex type or an array of complex numbers.
+
+    The type of the input is checked, not the value. So even if the input
+    has an imaginary part equal to zero, `isrealobj` evaluates to False
+    if the data type is complex.
+
+    Parameters
+    ----------
+    x : any
+        The input can be of any type and shape.
+
+    Returns
+    -------
+    y : bool
+        The return value, False if `x` is of a complex type.
+
+    See Also
+    --------
+    iscomplexobj, isreal
+
+    Notes
+    -----
+    The function is only meant for arrays with numerical values but it
+    accepts all other objects. Since it assumes array input, the return
+    value of other objects may be True.
+
+    >>> np.isrealobj('A string')
+    True
+    >>> np.isrealobj(False)
+    True
+    >>> np.isrealobj(None)
+    True
+
+    Examples
+    --------
+    >>> np.isrealobj(1)
+    True
+    >>> np.isrealobj(1+0j)
+    False
+    >>> np.isrealobj([3, 1+0j, True])
+    False
+
+    """
+    return not iscomplexobj(x)
+
+#-----------------------------------------------------------------------------
+
+def _getmaxmin(t):
+    from numpy.core import getlimits
+    f = getlimits.finfo(t)
+    return f.max, f.min
+
+
+def _nan_to_num_dispatcher(x, copy=None, nan=None, posinf=None, neginf=None):
+    return (x,)
+
+
+@array_function_dispatch(_nan_to_num_dispatcher)
+def nan_to_num(x, copy=True, nan=0.0, posinf=None, neginf=None):
+    """
+    Replace NaN with zero and infinity with large finite numbers (default
+    behaviour) or with the numbers defined by the user using the `nan`,
+    `posinf` and/or `neginf` keywords.
+
+    If `x` is inexact, NaN is replaced by zero or by the user defined value in
+    `nan` keyword, infinity is replaced by the largest finite floating point
+    values representable by ``x.dtype`` or by the user defined value in
+    `posinf` keyword and -infinity is replaced by the most negative finite
+    floating point values representable by ``x.dtype`` or by the user defined
+    value in `neginf` keyword.
+
+    For complex dtypes, the above is applied to each of the real and
+    imaginary components of `x` separately.
+
+    If `x` is not inexact, then no replacements are made.
+
+    Parameters
+    ----------
+    x : scalar or array_like
+        Input data.
+    copy : bool, optional
+        Whether to create a copy of `x` (True) or to replace values
+        in-place (False). The in-place operation only occurs if
+        casting to an array does not require a copy.
+        Default is True.
+
+        .. versionadded:: 1.13
+    nan : int, float, optional
+        Value to be used to fill NaN values. If no value is passed
+        then NaN values will be replaced with 0.0.
+
+        .. versionadded:: 1.17
+    posinf : int, float, optional
+        Value to be used to fill positive infinity values. If no value is
+        passed then positive infinity values will be replaced with a very
+        large number.
+
+        .. versionadded:: 1.17
+    neginf : int, float, optional
+        Value to be used to fill negative infinity values. If no value is
+        passed then negative infinity values will be replaced with a very
+        small (or negative) number.
+
+        .. versionadded:: 1.17
+
+
+
+    Returns
+    -------
+    out : ndarray
+        `x`, with the non-finite values replaced. If `copy` is False, this may
+        be `x` itself.
+
+    See Also
+    --------
+    isinf : Shows which elements are positive or negative infinity.
+    isneginf : Shows which elements are negative infinity.
+    isposinf : Shows which elements are positive infinity.
+    isnan : Shows which elements are Not a Number (NaN).
+    isfinite : Shows which elements are finite (not NaN, not infinity)
+
+    Notes
+    -----
+    NumPy uses the IEEE Standard for Binary Floating-Point for Arithmetic
+    (IEEE 754). This means that Not a Number is not equivalent to infinity.
+
+    Examples
+    --------
+    >>> np.nan_to_num(np.inf)
+    1.7976931348623157e+308
+    >>> np.nan_to_num(-np.inf)
+    -1.7976931348623157e+308
+    >>> np.nan_to_num(np.nan)
+    0.0
+    >>> x = np.array([np.inf, -np.inf, np.nan, -128, 128])
+    >>> np.nan_to_num(x)
+    array([ 1.79769313e+308, -1.79769313e+308,  0.00000000e+000, # may vary
+           -1.28000000e+002,  1.28000000e+002])
+    >>> np.nan_to_num(x, nan=-9999, posinf=33333333, neginf=33333333)
+    array([ 3.3333333e+07,  3.3333333e+07, -9.9990000e+03,
+           -1.2800000e+02,  1.2800000e+02])
+    >>> y = np.array([complex(np.inf, np.nan), np.nan, complex(np.nan, np.inf)])
+    array([  1.79769313e+308,  -1.79769313e+308,   0.00000000e+000, # may vary
+         -1.28000000e+002,   1.28000000e+002])
+    >>> np.nan_to_num(y)
+    array([  1.79769313e+308 +0.00000000e+000j, # may vary
+             0.00000000e+000 +0.00000000e+000j,
+             0.00000000e+000 +1.79769313e+308j])
+    >>> np.nan_to_num(y, nan=111111, posinf=222222)
+    array([222222.+111111.j, 111111.     +0.j, 111111.+222222.j])
+    """
+    x = _nx.array(x, subok=True, copy=copy)
+    xtype = x.dtype.type
+
+    isscalar = (x.ndim == 0)
+
+    if not issubclass(xtype, _nx.inexact):
+        return x[()] if isscalar else x
+
+    iscomplex = issubclass(xtype, _nx.complexfloating)
+
+    dest = (x.real, x.imag) if iscomplex else (x,)
+    maxf, minf = _getmaxmin(x.real.dtype)
+    if posinf is not None:
+        maxf = posinf
+    if neginf is not None:
+        minf = neginf
+    for d in dest:
+        idx_nan = isnan(d)
+        idx_posinf = isposinf(d)
+        idx_neginf = isneginf(d)
+        _nx.copyto(d, nan, where=idx_nan)
+        _nx.copyto(d, maxf, where=idx_posinf)
+        _nx.copyto(d, minf, where=idx_neginf)
+    return x[()] if isscalar else x
+
+#-----------------------------------------------------------------------------
+
+def _real_if_close_dispatcher(a, tol=None):
+    return (a,)
+
+
+@array_function_dispatch(_real_if_close_dispatcher)
+def real_if_close(a, tol=100):
+    """
+    If input is complex with all imaginary parts close to zero, return
+    real parts.
+
+    "Close to zero" is defined as `tol` * (machine epsilon of the type for
+    `a`).
+
+    Parameters
+    ----------
+    a : array_like
+        Input array.
+    tol : float
+        Tolerance in machine epsilons for the complex part of the elements
+        in the array. If the tolerance is <=1, then the absolute tolerance
+        is used.
+
+    Returns
+    -------
+    out : ndarray
+        If `a` is real, the type of `a` is used for the output.  If `a`
+        has complex elements, the returned type is float.
+
+    See Also
+    --------
+    real, imag, angle
+
+    Notes
+    -----
+    Machine epsilon varies from machine to machine and between data types
+    but Python floats on most platforms have a machine epsilon equal to
+    2.2204460492503131e-16.  You can use 'np.finfo(float).eps' to print
+    out the machine epsilon for floats.
+
+    Examples
+    --------
+    >>> np.finfo(float).eps
+    2.2204460492503131e-16 # may vary
+
+    >>> np.real_if_close([2.1 + 4e-14j, 5.2 + 3e-15j], tol=1000)
+    array([2.1, 5.2])
+    >>> np.real_if_close([2.1 + 4e-13j, 5.2 + 3e-15j], tol=1000)
+    array([2.1+4.e-13j, 5.2 + 3e-15j])
+
+    """
+    a = asanyarray(a)
+    type_ = a.dtype.type
+    if not issubclass(type_, _nx.complexfloating):
+        return a
+    if tol > 1:
+        f = getlimits.finfo(type_)
+        tol = f.eps * tol
+    if _nx.all(_nx.absolute(a.imag) < tol):
+        a = a.real
+    return a
+
+
+#-----------------------------------------------------------------------------
+
+_namefromtype = {'S1': 'character',
+                 '?': 'bool',
+                 'b': 'signed char',
+                 'B': 'unsigned char',
+                 'h': 'short',
+                 'H': 'unsigned short',
+                 'i': 'integer',
+                 'I': 'unsigned integer',
+                 'l': 'long integer',
+                 'L': 'unsigned long integer',
+                 'q': 'long long integer',
+                 'Q': 'unsigned long long integer',
+                 'f': 'single precision',
+                 'd': 'double precision',
+                 'g': 'long precision',
+                 'F': 'complex single precision',
+                 'D': 'complex double precision',
+                 'G': 'complex long double precision',
+                 'S': 'string',
+                 'U': 'unicode',
+                 'V': 'void',
+                 'O': 'object'
+                 }
+
+@set_module('numpy')
+def typename(char):
+    """
+    Return a description for the given data type code.
+
+    Parameters
+    ----------
+    char : str
+        Data type code.
+
+    Returns
+    -------
+    out : str
+        Description of the input data type code.
+
+    See Also
+    --------
+    dtype, typecodes
+
+    Examples
+    --------
+    >>> typechars = ['S1', '?', 'B', 'D', 'G', 'F', 'I', 'H', 'L', 'O', 'Q',
+    ...              'S', 'U', 'V', 'b', 'd', 'g', 'f', 'i', 'h', 'l', 'q']
+    >>> for typechar in typechars:
+    ...     print(typechar, ' : ', np.typename(typechar))
+    ...
+    S1  :  character
+    ?  :  bool
+    B  :  unsigned char
+    D  :  complex double precision
+    G  :  complex long double precision
+    F  :  complex single precision
+    I  :  unsigned integer
+    H  :  unsigned short
+    L  :  unsigned long integer
+    O  :  object
+    Q  :  unsigned long long integer
+    S  :  string
+    U  :  unicode
+    V  :  void
+    b  :  signed char
+    d  :  double precision
+    g  :  long precision
+    f  :  single precision
+    i  :  integer
+    h  :  short
+    l  :  long integer
+    q  :  long long integer
+
+    """
+    return _namefromtype[char]
+
+#-----------------------------------------------------------------------------
+
+#determine the "minimum common type" for a group of arrays.
+array_type = [[_nx.half, _nx.single, _nx.double, _nx.longdouble],
+              [None, _nx.csingle, _nx.cdouble, _nx.clongdouble]]
+array_precision = {_nx.half: 0,
+                   _nx.single: 1,
+                   _nx.double: 2,
+                   _nx.longdouble: 3,
+                   _nx.csingle: 1,
+                   _nx.cdouble: 2,
+                   _nx.clongdouble: 3}
+
+
+def _common_type_dispatcher(*arrays):
+    return arrays
+
+
+@array_function_dispatch(_common_type_dispatcher)
+def common_type(*arrays):
+    """
+    Return a scalar type which is common to the input arrays.
+
+    The return type will always be an inexact (i.e. floating point) scalar
+    type, even if all the arrays are integer arrays. If one of the inputs is
+    an integer array, the minimum precision type that is returned is a
+    64-bit floating point dtype.
+
+    All input arrays except int64 and uint64 can be safely cast to the
+    returned dtype without loss of information.
+
+    Parameters
+    ----------
+    array1, array2, ... : ndarrays
+        Input arrays.
+
+    Returns
+    -------
+    out : data type code
+        Data type code.
+
+    See Also
+    --------
+    dtype, mintypecode
+
+    Examples
+    --------
+    >>> np.common_type(np.arange(2, dtype=np.float32))
+    <class 'numpy.float32'>
+    >>> np.common_type(np.arange(2, dtype=np.float32), np.arange(2))
+    <class 'numpy.float64'>
+    >>> np.common_type(np.arange(4), np.array([45, 6.j]), np.array([45.0]))
+    <class 'numpy.complex128'>
+
+    """
+    is_complex = False
+    precision = 0
+    for a in arrays:
+        t = a.dtype.type
+        if iscomplexobj(a):
+            is_complex = True
+        if issubclass(t, _nx.integer):
+            p = 2  # array_precision[_nx.double]
+        else:
+            p = array_precision.get(t, None)
+            if p is None:
+                raise TypeError("can't get common type for non-numeric array")
+        precision = max(precision, p)
+    if is_complex:
+        return array_type[1][precision]
+    else:
+        return array_type[0][precision]