Run 9. Outer Step 12. Inner Step 9.

config.json

@@ -2,42 +2,42 @@
   "_name_or_path": "distributed/llama-1b",
   "all_reduce_scores": {
     "0": "NOT_ALIVE",
-    "1": "NON_PARTICIPATING",
-    "10": "NON_PARTICIPATING",
+    "1": "NOT_ALIVE",
+    "10": "FAIL",
     "100": "NOT_ALIVE",
     "101": "NOT_ALIVE",
     "102": "NOT_ALIVE",
     "103": "NON_PARTICIPATING",
-    "104": "NON_PARTICIPATING",
-    "105": "NON_PARTICIPATING",
+    "104": "FAIL",
+    "105": "FAIL",
     "106": "FAIL",
     "107": "NON_PARTICIPATING",
     "108": "NOT_ALIVE",
-    "109": "NON_PARTICIPATING",
-    "11": "FAIL",
-    "110": "NOT_ALIVE",
+    "109": "NOT_ALIVE",
+    "11": "NON_PARTICIPATING",
+    "110": "FAIL",
     "111": "NON_PARTICIPATING",
-    "112": "NOT_ALIVE",
-    "113": "NON_PARTICIPATING",
+    "112": "FAIL",
+    "113": "FAIL",
     "114": "NON_PARTICIPATING",
     "115": "NON_PARTICIPATING",
-    "116": "FAIL",
-    "117": "NOT_ALIVE",
+    "116": "NON_PARTICIPATING",
+    "117": "NON_PARTICIPATING",
     "118": "NON_PARTICIPATING",
-    "119": "NOT_ALIVE",
-    "12": "NON_PARTICIPATING",
+    "119": "NON_PARTICIPATING",
+    "12": "NOT_ALIVE",
     "120": "NOT_ALIVE",
     "121": "NOT_ALIVE",
     "122": "NON_PARTICIPATING",
-    "123": "FAIL",
+    "123": "NOT_ALIVE",
     "124": "NON_PARTICIPATING",
-    "125": "FAIL",
+    "125": "NOT_ALIVE",
     "126": "NON_PARTICIPATING",
-    "127": "NON_PARTICIPATING",
+    "127": "NOT_ALIVE",
     "128": "FAIL",
-    "129": "FAIL",
+    "129": "NON_PARTICIPATING",
     "13": "FAIL",
-    "130": "NOT_ALIVE",
+    "130": "NON_PARTICIPATING",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
     "133": "NON_PARTICIPATING",
@@ -45,13 +45,13 @@
     "135": "NOT_ALIVE",
     "136": "NOT_ALIVE",
     "137": "NOT_ALIVE",
-    "138": "NON_PARTICIPATING",
+    "138": "NOT_ALIVE",
     "139": "NOT_ALIVE",
     "14": "NOT_ALIVE",
     "140": "NOT_ALIVE",
     "141": "NOT_ALIVE",
     "142": "NON_PARTICIPATING",
-    "143": "NON_PARTICIPATING",
+    "143": "NOT_ALIVE",
     "144": "NOT_ALIVE",
     "145": "NON_PARTICIPATING",
     "146": "NON_PARTICIPATING",
@@ -64,23 +64,23 @@
     "152": "NOT_ALIVE",
     "153": "FAIL",
     "154": "NON_PARTICIPATING",
-    "155": "NON_PARTICIPATING",
+    "155": "NOT_ALIVE",
     "156": "NON_PARTICIPATING",
-    "157": "FAIL",
+    "157": "NON_PARTICIPATING",
     "158": "NON_PARTICIPATING",
     "159": "NON_PARTICIPATING",
     "16": "NON_PARTICIPATING",
     "160": "NOT_ALIVE",
     "161": "NOT_ALIVE",
     "162": "NON_PARTICIPATING",
-    "163": "NON_PARTICIPATING",
+    "163": "FAIL",
     "164": "NOT_ALIVE",
-    "165": "FAIL",
+    "165": "NON_PARTICIPATING",
     "166": "NON_PARTICIPATING",
     "167": "NON_PARTICIPATING",
     "168": "NON_PARTICIPATING",
-    "169": "NON_PARTICIPATING",
-    "17": "NON_PARTICIPATING",
+    "169": "FAIL",
+    "17": "FAIL",
     "170": "NON_PARTICIPATING",
     "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
@@ -88,105 +88,105 @@
     "174": "NOT_ALIVE",
     "175": "NON_PARTICIPATING",
     "176": "NOT_ALIVE",
-    "177": "NON_PARTICIPATING",
+    "177": "NOT_ALIVE",
     "178": "NOT_ALIVE",
-    "179": "FAIL",
-    "18": "NON_PARTICIPATING",
+    "179": "NON_PARTICIPATING",
+    "18": "FAIL",
     "180": "NON_PARTICIPATING",
     "181": "NON_PARTICIPATING",
     "182": "NOT_ALIVE",
-    "183": "NOT_ALIVE",
+    "183": "NON_PARTICIPATING",
     "184": "FAIL",
     "185": "NOT_ALIVE",
-    "186": "NOT_ALIVE",
+    "186": "NON_PARTICIPATING",
     "187": "NON_PARTICIPATING",
     "188": "NON_PARTICIPATING",
     "189": "NON_PARTICIPATING",
-    "19": "FAIL",
+    "19": "NON_PARTICIPATING",
     "190": "NON_PARTICIPATING",
-    "191": "NOT_ALIVE",
+    "191": "NON_PARTICIPATING",
     "192": "NON_PARTICIPATING",
     "193": "NON_PARTICIPATING",
-    "194": "FAIL",
-    "195": "NOT_ALIVE",
+    "194": "NON_PARTICIPATING",
+    "195": "NON_PARTICIPATING",
     "196": "FAIL",
     "197": "NON_PARTICIPATING",
-    "198": "NOT_ALIVE",
-    "199": "FAIL",
-    "2": "NON_PARTICIPATING",
-    "20": "NOT_ALIVE",
-    "200": "NOT_ALIVE",
-    "201": "NON_PARTICIPATING",
-    "202": "FAIL",
-    "203": "FAIL",
+    "198": "FAIL",
+    "199": "NON_PARTICIPATING",
+    "2": "NOT_ALIVE",
+    "20": "FAIL",
+    "200": "NON_PARTICIPATING",
+    "201": "FAIL",
+    "202": "NON_PARTICIPATING",
+    "203": "NON_PARTICIPATING",
     "204": "NOT_ALIVE",
-    "205": "NOT_ALIVE",
+    "205": "FAIL",
     "206": "NON_PARTICIPATING",
     "207": "NOT_ALIVE",
     "208": "FAIL",
-    "209": "NON_PARTICIPATING",
-    "21": "NOT_ALIVE",
+    "209": "FAIL",
+    "21": "FAIL",
     "210": "NON_PARTICIPATING",
     "211": "FAIL",
-    "212": "FAIL",
+    "212": "NON_PARTICIPATING",
     "213": "NOT_ALIVE",
-    "214": "NOT_ALIVE",
-    "215": "FAIL",
+    "214": "FAIL",
+    "215": "NON_PARTICIPATING",
     "216": "NOT_ALIVE",
-    "217": "FAIL",
-    "218": "FAIL",
+    "217": "NON_PARTICIPATING",
+    "218": "NOT_ALIVE",
     "219": "FAIL",
     "22": "NOT_ALIVE",
     "220": "FAIL",
-    "221": "NON_PARTICIPATING",
-    "222": "NON_PARTICIPATING",
-    "223": "FAIL",
+    "221": "FAIL",
+    "222": "FAIL",
+    "223": "NON_PARTICIPATING",
     "224": "NOT_ALIVE",
     "225": "FAIL",
     "226": "NON_PARTICIPATING",
-    "227": "FAIL",
+    "227": "NON_PARTICIPATING",
     "228": "NON_PARTICIPATING",
     "229": "NOT_ALIVE",
     "23": "NON_PARTICIPATING",
     "230": "NOT_ALIVE",
     "231": "NOT_ALIVE",
     "232": "NON_PARTICIPATING",
-    "233": "NON_PARTICIPATING",
-    "234": "NON_PARTICIPATING",
+    "233": "FAIL",
+    "234": "FAIL",
     "235": "NON_PARTICIPATING",
-    "236": "FAIL",
+    "236": "NOT_ALIVE",
     "237": "NON_PARTICIPATING",
-    "238": "NOT_ALIVE",
-    "239": "NOT_ALIVE",
+    "238": "NON_PARTICIPATING",
+    "239": "NON_PARTICIPATING",
     "24": "FAIL",
     "240": "NOT_ALIVE",
     "241": "NON_PARTICIPATING",
     "242": "NOT_ALIVE",
-    "243": "NOT_ALIVE",
+    "243": "NON_PARTICIPATING",
     "244": "NOT_ALIVE",
-    "245": "NON_PARTICIPATING",
-    "246": "NON_PARTICIPATING",
-    "247": "NOT_ALIVE",
-    "248": "NON_PARTICIPATING",
+    "245": "FAIL",
+    "246": "FAIL",
+    "247": "FAIL",
+    "248": "NOT_ALIVE",
     "249": "NOT_ALIVE",
     "25": "SUCCESS",
     "250": "FAIL",
     "251": "NOT_ALIVE",
-    "252": "FAIL",
+    "252": "NON_PARTICIPATING",
     "253": "NON_PARTICIPATING",
     "254": "NON_PARTICIPATING",
     "255": "NON_PARTICIPATING",
     "26": "NOT_ALIVE",
-    "27": "NON_PARTICIPATING",
+    "27": "FAIL",
     "28": "NON_PARTICIPATING",
-    "29": "NOT_ALIVE",
+    "29": "FAIL",
     "3": "NON_PARTICIPATING",
-    "30": "FAIL",
-    "31": "NON_PARTICIPATING",
+    "30": "NOT_ALIVE",
+    "31": "NOT_ALIVE",
     "32": "NOT_ALIVE",
     "33": "NOT_ALIVE",
-    "34": "FAIL",
-    "35": "NOT_ALIVE",
+    "34": "NON_PARTICIPATING",
+    "35": "NON_PARTICIPATING",
     "36": "NON_PARTICIPATING",
     "37": "NOT_ALIVE",
     "38": "FAIL",
@@ -194,49 +194,49 @@
     "4": "NON_PARTICIPATING",
     "40": "FAIL",
     "41": "NOT_ALIVE",
-    "42": "NOT_ALIVE",
+    "42": "FAIL",
     "43": "NON_PARTICIPATING",
-    "44": "NOT_ALIVE",
-    "45": "NON_PARTICIPATING",
+    "44": "NON_PARTICIPATING",
+    "45": "FAIL",
     "46": "FAIL",
-    "47": "NON_PARTICIPATING",
+    "47": "FAIL",
     "48": "NON_PARTICIPATING",
     "49": "NOT_ALIVE",
-    "5": "FAIL",
+    "5": "NON_PARTICIPATING",
     "50": "FAIL",
-    "51": "NON_PARTICIPATING",
+    "51": "FAIL",
     "52": "NON_PARTICIPATING",
     "53": "FAIL",
     "54": "NOT_ALIVE",
     "55": "NON_PARTICIPATING",
-    "56": "NON_PARTICIPATING",
-    "57": "FAIL",
+    "56": "FAIL",
+    "57": "NON_PARTICIPATING",
     "58": "NOT_ALIVE",
     "59": "NON_PARTICIPATING",
     "6": "NOT_ALIVE",
-    "60": "FAIL",
+    "60": "NON_PARTICIPATING",
     "61": "NON_PARTICIPATING",
-    "62": "NON_PARTICIPATING",
+    "62": "FAIL",
     "63": "NON_PARTICIPATING",
     "64": "NOT_ALIVE",
-    "65": "NOT_ALIVE",
-    "66": "FAIL",
-    "67": "NON_PARTICIPATING",
+    "65": "FAIL",
+    "66": "NON_PARTICIPATING",
+    "67": "NOT_ALIVE",
     "68": "NOT_ALIVE",
-    "69": "FAIL",
+    "69": "NON_PARTICIPATING",
     "7": "FAIL",
     "70": "FAIL",
     "71": "NOT_ALIVE",
     "72": "NON_PARTICIPATING",
     "73": "NON_PARTICIPATING",
-    "74": "NOT_ALIVE",
+    "74": "NON_PARTICIPATING",
     "75": "NON_PARTICIPATING",
-    "76": "FAIL",
+    "76": "NON_PARTICIPATING",
     "77": "NOT_ALIVE",
     "78": "NON_PARTICIPATING",
-    "79": "FAIL",
+    "79": "NON_PARTICIPATING",
     "8": "NON_PARTICIPATING",
-    "80": "NOT_ALIVE",
+    "80": "FAIL",
     "81": "NOT_ALIVE",
     "82": "NOT_ALIVE",
     "83": "NON_PARTICIPATING",
@@ -245,18 +245,18 @@
     "86": "FAIL",
     "87": "NON_PARTICIPATING",
     "88": "FAIL",
-    "89": "FAIL",
+    "89": "NON_PARTICIPATING",
     "9": "NON_PARTICIPATING",
     "90": "NON_PARTICIPATING",
-    "91": "NON_PARTICIPATING",
-    "92": "FAIL",
-    "93": "FAIL",
-    "94": "NOT_ALIVE",
+    "91": "FAIL",
+    "92": "NOT_ALIVE",
+    "93": "NON_PARTICIPATING",
+    "94": "NON_PARTICIPATING",
     "95": "NOT_ALIVE",
     "96": "NON_PARTICIPATING",
     "97": "NON_PARTICIPATING",
     "98": "NON_PARTICIPATING",
-    "99": "FAIL"
+    "99": "NON_PARTICIPATING"
   },
   "architectures": [
     "LlamaForCausalLM"
@@ -264,17 +264,17 @@
   "attention_bias": false,
   "attention_dropout": 0.0,
   "block_list": [
-    6456840,
-    6456879
+    6457024,
+    6457059
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 67,
+  "inner_step": 9,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6456333,
+  "last_allreduce_block": 6456970,
   "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:b71ad934b106ed393755911f65ad0323dcfbefe634441fe70efbf290c4e0f8ec
+oid sha256:897befa84d9502b35a707f9ef675c9516c70f8c0a76597cde8b0fe087e387573
 size 88389626

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0d13e9d39d61a113345ca1fcc326863eb72680f0dbc79430a4ec45971278e69d
+oid sha256:3f09846ac7283ac1411ec4b792ebee271ca4dea003d2fd7e3dbe418f7e80e2c7
 size 4924926906

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e5d7ed340ed9c1f147f178fa0e02cd160334aa8bce48ca8ac950c0391c8a60ac
+oid sha256:ad83de96374e2eab1f6dc1bbc0cf201cd99b9653b9fae792d9966e7dfb9da930
 size 4400216536