Run 10. Outer Step 69. Inner Step 5.

config.json

@@ -2,23 +2,23 @@
   "_name_or_path": "distributed/llama-1b",
   "all_reduce_scores": {
     "0": "NOT_ALIVE",
-    "1": "NON_PARTICIPATING",
+    "1": "SUCCESS",
     "10": "NON_PARTICIPATING",
     "100": "NOT_ALIVE",
     "101": "NON_PARTICIPATING",
-    "102": "FAIL",
+    "102": "SUCCESS",
     "103": "NON_PARTICIPATING",
     "104": "NOT_ALIVE",
     "105": "SUCCESS",
-    "106": "NOT_ALIVE",
+    "106": "NON_PARTICIPATING",
     "107": "NON_PARTICIPATING",
     "108": "NON_PARTICIPATING",
-    "109": "SUCCESS",
+    "109": "NON_PARTICIPATING",
     "11": "SUCCESS",
-    "110": "NON_PARTICIPATING",
+    "110": "NOT_ALIVE",
     "111": "SUCCESS",
     "112": "NON_PARTICIPATING",
-    "113": "SUCCESS",
+    "113": "NOT_ALIVE",
     "114": "SUCCESS",
     "115": "NON_PARTICIPATING",
     "116": "SUCCESS",
@@ -27,26 +27,26 @@
     "119": "NON_PARTICIPATING",
     "12": "NON_PARTICIPATING",
     "120": "NOT_ALIVE",
-    "121": "NOT_ALIVE",
-    "122": "NOT_ALIVE",
+    "121": "SUCCESS",
+    "122": "NON_PARTICIPATING",
     "123": "SUCCESS",
-    "124": "NOT_ALIVE",
-    "125": "NOT_ALIVE",
-    "126": "NOT_ALIVE",
-    "127": "NOT_ALIVE",
-    "128": "SUCCESS",
-    "129": "SUCCESS",
-    "13": "SUCCESS",
+    "124": "SUCCESS",
+    "125": "SUCCESS",
+    "126": "SUCCESS",
+    "127": "NON_PARTICIPATING",
+    "128": "NOT_ALIVE",
+    "129": "NON_PARTICIPATING",
+    "13": "NON_PARTICIPATING",
     "130": "NOT_ALIVE",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
     "133": "NOT_ALIVE",
-    "134": "NON_PARTICIPATING",
+    "134": "NOT_ALIVE",
     "135": "NON_PARTICIPATING",
     "136": "NON_PARTICIPATING",
     "137": "NOT_ALIVE",
     "138": "NON_PARTICIPATING",
-    "139": "NON_PARTICIPATING",
+    "139": "NOT_ALIVE",
     "14": "NOT_ALIVE",
     "140": "NON_PARTICIPATING",
     "141": "NON_PARTICIPATING",
@@ -56,118 +56,118 @@
     "145": "NOT_ALIVE",
     "146": "SUCCESS",
     "147": "NON_PARTICIPATING",
-    "148": "NON_PARTICIPATING",
+    "148": "NOT_ALIVE",
     "149": "NON_PARTICIPATING",
     "15": "NON_PARTICIPATING",
     "150": "NON_PARTICIPATING",
     "151": "NOT_ALIVE",
-    "152": "NON_PARTICIPATING",
+    "152": "NOT_ALIVE",
     "153": "SUCCESS",
     "154": "SUCCESS",
-    "155": "SUCCESS",
+    "155": "NOT_ALIVE",
     "156": "NON_PARTICIPATING",
     "157": "NON_PARTICIPATING",
     "158": "NON_PARTICIPATING",
     "159": "SUCCESS",
     "16": "NOT_ALIVE",
     "160": "NOT_ALIVE",
-    "161": "NON_PARTICIPATING",
-    "162": "NON_PARTICIPATING",
-    "163": "SUCCESS",
+    "161": "SUCCESS",
+    "162": "NOT_ALIVE",
+    "163": "NOT_ALIVE",
     "164": "NOT_ALIVE",
     "165": "NOT_ALIVE",
     "166": "NON_PARTICIPATING",
     "167": "NON_PARTICIPATING",
     "168": "SUCCESS",
-    "169": "SUCCESS",
+    "169": "NON_PARTICIPATING",
     "17": "NON_PARTICIPATING",
     "170": "SUCCESS",
-    "171": "NON_PARTICIPATING",
+    "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
-    "173": "NOT_ALIVE",
+    "173": "NON_PARTICIPATING",
     "174": "NOT_ALIVE",
     "175": "SUCCESS",
     "176": "NOT_ALIVE",
     "177": "NOT_ALIVE",
-    "178": "NON_PARTICIPATING",
+    "178": "NOT_ALIVE",
     "179": "NOT_ALIVE",
-    "18": "NON_PARTICIPATING",
+    "18": "NOT_ALIVE",
     "180": "NOT_ALIVE",
     "181": "NON_PARTICIPATING",
     "182": "NON_PARTICIPATING",
     "183": "NOT_ALIVE",
     "184": "SUCCESS",
-    "185": "SUCCESS",
-    "186": "SUCCESS",
+    "185": "NOT_ALIVE",
+    "186": "NON_PARTICIPATING",
     "187": "SUCCESS",
-    "188": "NON_PARTICIPATING",
+    "188": "NOT_ALIVE",
     "189": "NOT_ALIVE",
     "19": "NON_PARTICIPATING",
     "190": "NON_PARTICIPATING",
     "191": "NOT_ALIVE",
     "192": "NON_PARTICIPATING",
-    "193": "NON_PARTICIPATING",
+    "193": "SUCCESS",
     "194": "NOT_ALIVE",
     "195": "NOT_ALIVE",
     "196": "SUCCESS",
     "197": "NON_PARTICIPATING",
     "198": "SUCCESS",
-    "199": "NON_PARTICIPATING",
-    "2": "NON_PARTICIPATING",
-    "20": "NON_PARTICIPATING",
+    "199": "SUCCESS",
+    "2": "NOT_ALIVE",
+    "20": "SUCCESS",
     "200": "NON_PARTICIPATING",
-    "201": "NON_PARTICIPATING",
+    "201": "NOT_ALIVE",
     "202": "NON_PARTICIPATING",
     "203": "NON_PARTICIPATING",
     "204": "NOT_ALIVE",
-    "205": "SUCCESS",
+    "205": "NOT_ALIVE",
     "206": "SUCCESS",
     "207": "NOT_ALIVE",
     "208": "NON_PARTICIPATING",
     "209": "NON_PARTICIPATING",
-    "21": "NON_PARTICIPATING",
+    "21": "NOT_ALIVE",
     "210": "NON_PARTICIPATING",
     "211": "NON_PARTICIPATING",
-    "212": "NON_PARTICIPATING",
+    "212": "SUCCESS",
     "213": "NOT_ALIVE",
     "214": "SUCCESS",
     "215": "NOT_ALIVE",
     "216": "NOT_ALIVE",
     "217": "NOT_ALIVE",
     "218": "NOT_ALIVE",
-    "219": "SUCCESS",
+    "219": "NON_PARTICIPATING",
     "22": "NOT_ALIVE",
-    "220": "NON_PARTICIPATING",
+    "220": "NOT_ALIVE",
     "221": "SUCCESS",
     "222": "SUCCESS",
     "223": "NON_PARTICIPATING",
     "224": "NOT_ALIVE",
-    "225": "NOT_ALIVE",
+    "225": "NON_PARTICIPATING",
     "226": "NON_PARTICIPATING",
-    "227": "NON_PARTICIPATING",
-    "228": "SUCCESS",
-    "229": "SUCCESS",
-    "23": "NON_PARTICIPATING",
+    "227": "SUCCESS",
+    "228": "NON_PARTICIPATING",
+    "229": "NON_PARTICIPATING",
+    "23": "NOT_ALIVE",
     "230": "NON_PARTICIPATING",
     "231": "NON_PARTICIPATING",
     "232": "NON_PARTICIPATING",
     "233": "NON_PARTICIPATING",
-    "234": "NOT_ALIVE",
-    "235": "SUCCESS",
+    "234": "SUCCESS",
+    "235": "NON_PARTICIPATING",
     "236": "NON_PARTICIPATING",
     "237": "NON_PARTICIPATING",
-    "238": "NON_PARTICIPATING",
+    "238": "SUCCESS",
     "239": "NON_PARTICIPATING",
     "24": "NON_PARTICIPATING",
     "240": "NON_PARTICIPATING",
     "241": "NON_PARTICIPATING",
-    "242": "NON_PARTICIPATING",
+    "242": "NOT_ALIVE",
     "243": "NOT_ALIVE",
     "244": "NOT_ALIVE",
     "245": "NON_PARTICIPATING",
     "246": "NON_PARTICIPATING",
     "247": "NOT_ALIVE",
-    "248": "NON_PARTICIPATING",
+    "248": "SUCCESS",
     "249": "NOT_ALIVE",
     "25": "NOT_ALIVE",
     "250": "SUCCESS",
@@ -176,27 +176,27 @@
     "253": "NOT_ALIVE",
     "254": "NON_PARTICIPATING",
     "255": "SUCCESS",
-    "26": "NON_PARTICIPATING",
+    "26": "SUCCESS",
     "27": "NOT_ALIVE",
-    "28": "SUCCESS",
+    "28": "NOT_ALIVE",
     "29": "NON_PARTICIPATING",
     "3": "SUCCESS",
-    "30": "NOT_ALIVE",
+    "30": "NON_PARTICIPATING",
     "31": "SUCCESS",
-    "32": "NON_PARTICIPATING",
+    "32": "NOT_ALIVE",
     "33": "NOT_ALIVE",
-    "34": "NOT_ALIVE",
-    "35": "NON_PARTICIPATING",
-    "36": "NOT_ALIVE",
+    "34": "NON_PARTICIPATING",
+    "35": "SUCCESS",
+    "36": "NON_PARTICIPATING",
     "37": "NON_PARTICIPATING",
     "38": "NOT_ALIVE",
     "39": "NON_PARTICIPATING",
-    "4": "FAIL",
+    "4": "SUCCESS",
     "40": "NOT_ALIVE",
     "41": "NON_PARTICIPATING",
-    "42": "NON_PARTICIPATING",
+    "42": "SUCCESS",
     "43": "NOT_ALIVE",
-    "44": "NOT_ALIVE",
+    "44": "NON_PARTICIPATING",
     "45": "NON_PARTICIPATING",
     "46": "NON_PARTICIPATING",
     "47": "NON_PARTICIPATING",
@@ -204,10 +204,10 @@
     "49": "NOT_ALIVE",
     "5": "NOT_ALIVE",
     "50": "NON_PARTICIPATING",
-    "51": "NOT_ALIVE",
-    "52": "SUCCESS",
+    "51": "NON_PARTICIPATING",
+    "52": "NON_PARTICIPATING",
     "53": "NON_PARTICIPATING",
-    "54": "NON_PARTICIPATING",
+    "54": "NOT_ALIVE",
     "55": "NOT_ALIVE",
     "56": "NON_PARTICIPATING",
     "57": "SUCCESS",
@@ -215,47 +215,47 @@
     "59": "SUCCESS",
     "6": "NOT_ALIVE",
     "60": "SUCCESS",
-    "61": "SUCCESS",
-    "62": "NON_PARTICIPATING",
+    "61": "NON_PARTICIPATING",
+    "62": "NOT_ALIVE",
     "63": "NON_PARTICIPATING",
-    "64": "SUCCESS",
+    "64": "NOT_ALIVE",
     "65": "NON_PARTICIPATING",
     "66": "NON_PARTICIPATING",
     "67": "NON_PARTICIPATING",
-    "68": "NON_PARTICIPATING",
+    "68": "NOT_ALIVE",
     "69": "NON_PARTICIPATING",
     "7": "NON_PARTICIPATING",
     "70": "NON_PARTICIPATING",
-    "71": "NOT_ALIVE",
+    "71": "NON_PARTICIPATING",
     "72": "NON_PARTICIPATING",
     "73": "NON_PARTICIPATING",
     "74": "SUCCESS",
-    "75": "NOT_ALIVE",
+    "75": "NON_PARTICIPATING",
     "76": "NOT_ALIVE",
-    "77": "NON_PARTICIPATING",
-    "78": "NON_PARTICIPATING",
-    "79": "NON_PARTICIPATING",
-    "8": "NOT_ALIVE",
+    "77": "NOT_ALIVE",
+    "78": "NOT_ALIVE",
+    "79": "NOT_ALIVE",
+    "8": "NON_PARTICIPATING",
     "80": "SUCCESS",
     "81": "NOT_ALIVE",
-    "82": "FAIL",
-    "83": "NON_PARTICIPATING",
+    "82": "SUCCESS",
+    "83": "NOT_ALIVE",
     "84": "SUCCESS",
     "85": "NON_PARTICIPATING",
     "86": "SUCCESS",
     "87": "SUCCESS",
-    "88": "NON_PARTICIPATING",
+    "88": "NOT_ALIVE",
     "89": "SUCCESS",
-    "9": "NON_PARTICIPATING",
-    "90": "NON_PARTICIPATING",
-    "91": "NOT_ALIVE",
-    "92": "SUCCESS",
+    "9": "NOT_ALIVE",
+    "90": "NOT_ALIVE",
+    "91": "NON_PARTICIPATING",
+    "92": "NOT_ALIVE",
     "93": "NOT_ALIVE",
     "94": "SUCCESS",
     "95": "NOT_ALIVE",
     "96": "SUCCESS",
-    "97": "NON_PARTICIPATING",
-    "98": "NOT_ALIVE",
+    "97": "NOT_ALIVE",
+    "98": "SUCCESS",
     "99": "NON_PARTICIPATING"
   },
   "architectures": [
@@ -264,17 +264,17 @@
   "attention_bias": false,
   "attention_dropout": 0.0,
   "block_list": [
-    6503948,
-    6503966
+    6504146,
+    6504153
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 43,
+  "inner_step": 4,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6502961,
+  "last_allreduce_block": 6504102,
   "max_position_embeddings": 2048,
   "mlp_bias": false,
   "model_type": "llama",

gradients.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d27948c001a1e0d774ccd1fca6e24c314e84627d07ebcf16638159315134c86e
+oid sha256:a35270f51b6fc64a5e1f5bc0e9609d936fceb9b6ba4b1be3e3739cc1df1bb45f
 size 88389626

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5730f28faf6ce952632c278e4267aaf4b5319ea7c10135ee9727a4b7c1b9f8d3
+oid sha256:ab8e3233308d38c542e6dea708bc3435c06eaf281e5f262aa303d163e68dadec
 size 4924926906

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f8924a6e1f3ffecae24ef19c03cc349b55d26d28be86a1b37cd2867be1231681
+oid sha256:35b52911fa570aefabda04149e98ff71ad2b4a8ca8c5e66d7e9a307272c26850
 size 4400216536