Run 9. Outer Step 16. Inner Step 9.

config.json

@@ -2,170 +2,170 @@
   "_name_or_path": "distributed/llama-1b",
   "all_reduce_scores": {
     "0": "NOT_ALIVE",
-    "1": "NON_PARTICIPATING",
-    "10": "FAIL",
+    "1": "FAIL",
+    "10": "NON_PARTICIPATING",
     "100": "NOT_ALIVE",
     "101": "NOT_ALIVE",
     "102": "NOT_ALIVE",
-    "103": "NON_PARTICIPATING",
-    "104": "FAIL",
+    "103": "FAIL",
+    "104": "NON_PARTICIPATING",
     "105": "FAIL",
-    "106": "NOT_ALIVE",
-    "107": "NOT_ALIVE",
+    "106": "FAIL",
+    "107": "NON_PARTICIPATING",
     "108": "NOT_ALIVE",
-    "109": "FAIL",
+    "109": "NON_PARTICIPATING",
     "11": "NON_PARTICIPATING",
-    "110": "NON_PARTICIPATING",
+    "110": "NOT_ALIVE",
     "111": "NON_PARTICIPATING",
-    "112": "FAIL",
+    "112": "NOT_ALIVE",
     "113": "NON_PARTICIPATING",
     "114": "NON_PARTICIPATING",
     "115": "NON_PARTICIPATING",
-    "116": "NON_PARTICIPATING",
-    "117": "NON_PARTICIPATING",
+    "116": "FAIL",
+    "117": "FAIL",
     "118": "NOT_ALIVE",
-    "119": "NON_PARTICIPATING",
+    "119": "NOT_ALIVE",
     "12": "NOT_ALIVE",
-    "120": "FAIL",
+    "120": "NOT_ALIVE",
     "121": "NOT_ALIVE",
-    "122": "FAIL",
+    "122": "NON_PARTICIPATING",
     "123": "NON_PARTICIPATING",
-    "124": "NON_PARTICIPATING",
-    "125": "NON_PARTICIPATING",
-    "126": "NOT_ALIVE",
-    "127": "NOT_ALIVE",
+    "124": "FAIL",
+    "125": "NOT_ALIVE",
+    "126": "NON_PARTICIPATING",
+    "127": "FAIL",
     "128": "NON_PARTICIPATING",
-    "129": "NON_PARTICIPATING",
+    "129": "NOT_ALIVE",
     "13": "FAIL",
-    "130": "FAIL",
+    "130": "NON_PARTICIPATING",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
-    "133": "NON_PARTICIPATING",
+    "133": "NOT_ALIVE",
     "134": "NON_PARTICIPATING",
     "135": "FAIL",
     "136": "NOT_ALIVE",
-    "137": "NON_PARTICIPATING",
-    "138": "FAIL",
+    "137": "NOT_ALIVE",
+    "138": "NOT_ALIVE",
     "139": "NOT_ALIVE",
     "14": "NON_PARTICIPATING",
     "140": "NOT_ALIVE",
     "141": "NOT_ALIVE",
-    "142": "NON_PARTICIPATING",
-    "143": "NOT_ALIVE",
+    "142": "NOT_ALIVE",
+    "143": "FAIL",
     "144": "NOT_ALIVE",
-    "145": "NON_PARTICIPATING",
-    "146": "NON_PARTICIPATING",
-    "147": "NON_PARTICIPATING",
+    "145": "NOT_ALIVE",
+    "146": "NOT_ALIVE",
+    "147": "NOT_ALIVE",
     "148": "NON_PARTICIPATING",
-    "149": "NOT_ALIVE",
-    "15": "NON_PARTICIPATING",
-    "150": "NON_PARTICIPATING",
+    "149": "NON_PARTICIPATING",
+    "15": "NOT_ALIVE",
+    "150": "FAIL",
     "151": "NOT_ALIVE",
-    "152": "NON_PARTICIPATING",
+    "152": "NOT_ALIVE",
     "153": "NON_PARTICIPATING",
-    "154": "FAIL",
+    "154": "NON_PARTICIPATING",
     "155": "NON_PARTICIPATING",
     "156": "FAIL",
-    "157": "NOT_ALIVE",
-    "158": "NON_PARTICIPATING",
-    "159": "NON_PARTICIPATING",
-    "16": "NON_PARTICIPATING",
-    "160": "NON_PARTICIPATING",
+    "157": "NON_PARTICIPATING",
+    "158": "FAIL",
+    "159": "NOT_ALIVE",
+    "16": "NOT_ALIVE",
+    "160": "NOT_ALIVE",
     "161": "NOT_ALIVE",
     "162": "NON_PARTICIPATING",
-    "163": "FAIL",
-    "164": "NON_PARTICIPATING",
-    "165": "NON_PARTICIPATING",
+    "163": "NOT_ALIVE",
+    "164": "NOT_ALIVE",
+    "165": "NOT_ALIVE",
     "166": "NON_PARTICIPATING",
-    "167": "NOT_ALIVE",
+    "167": "NON_PARTICIPATING",
     "168": "NON_PARTICIPATING",
-    "169": "FAIL",
-    "17": "NON_PARTICIPATING",
-    "170": "NON_PARTICIPATING",
+    "169": "NOT_ALIVE",
+    "17": "NOT_ALIVE",
+    "170": "NOT_ALIVE",
     "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
     "173": "FAIL",
-    "174": "NON_PARTICIPATING",
+    "174": "NOT_ALIVE",
     "175": "NON_PARTICIPATING",
     "176": "NOT_ALIVE",
     "177": "FAIL",
     "178": "NOT_ALIVE",
-    "179": "NON_PARTICIPATING",
+    "179": "NOT_ALIVE",
     "18": "NON_PARTICIPATING",
-    "180": "FAIL",
+    "180": "NON_PARTICIPATING",
     "181": "NON_PARTICIPATING",
     "182": "NOT_ALIVE",
     "183": "NOT_ALIVE",
     "184": "NON_PARTICIPATING",
-    "185": "NOT_ALIVE",
+    "185": "FAIL",
     "186": "NON_PARTICIPATING",
-    "187": "NOT_ALIVE",
+    "187": "NON_PARTICIPATING",
     "188": "NON_PARTICIPATING",
-    "189": "FAIL",
+    "189": "NOT_ALIVE",
     "19": "NON_PARTICIPATING",
-    "190": "NON_PARTICIPATING",
+    "190": "NOT_ALIVE",
     "191": "NON_PARTICIPATING",
-    "192": "NON_PARTICIPATING",
-    "193": "NON_PARTICIPATING",
-    "194": "NON_PARTICIPATING",
+    "192": "NOT_ALIVE",
+    "193": "FAIL",
+    "194": "NOT_ALIVE",
     "195": "FAIL",
     "196": "NON_PARTICIPATING",
-    "197": "NOT_ALIVE",
+    "197": "NON_PARTICIPATING",
     "198": "NOT_ALIVE",
-    "199": "FAIL",
+    "199": "NON_PARTICIPATING",
     "2": "NON_PARTICIPATING",
-    "20": "NOT_ALIVE",
+    "20": "FAIL",
     "200": "NON_PARTICIPATING",
-    "201": "FAIL",
+    "201": "NON_PARTICIPATING",
     "202": "FAIL",
-    "203": "NOT_ALIVE",
+    "203": "NON_PARTICIPATING",
     "204": "NOT_ALIVE",
-    "205": "NOT_ALIVE",
-    "206": "NOT_ALIVE",
+    "205": "FAIL",
+    "206": "NON_PARTICIPATING",
     "207": "NOT_ALIVE",
-    "208": "FAIL",
-    "209": "NON_PARTICIPATING",
-    "21": "FAIL",
-    "210": "NOT_ALIVE",
-    "211": "FAIL",
-    "212": "FAIL",
+    "208": "NOT_ALIVE",
+    "209": "FAIL",
+    "21": "NON_PARTICIPATING",
+    "210": "NON_PARTICIPATING",
+    "211": "NON_PARTICIPATING",
+    "212": "NON_PARTICIPATING",
     "213": "NOT_ALIVE",
     "214": "NON_PARTICIPATING",
-    "215": "FAIL",
+    "215": "NOT_ALIVE",
     "216": "FAIL",
     "217": "NON_PARTICIPATING",
-    "218": "NON_PARTICIPATING",
-    "219": "NON_PARTICIPATING",
+    "218": "FAIL",
+    "219": "FAIL",
     "22": "NOT_ALIVE",
-    "220": "NON_PARTICIPATING",
-    "221": "NON_PARTICIPATING",
+    "220": "FAIL",
+    "221": "FAIL",
     "222": "NON_PARTICIPATING",
-    "223": "NON_PARTICIPATING",
+    "223": "FAIL",
     "224": "NOT_ALIVE",
     "225": "FAIL",
     "226": "NON_PARTICIPATING",
     "227": "FAIL",
-    "228": "NON_PARTICIPATING",
-    "229": "NON_PARTICIPATING",
-    "23": "NON_PARTICIPATING",
+    "228": "NOT_ALIVE",
+    "229": "NOT_ALIVE",
+    "23": "FAIL",
     "230": "NOT_ALIVE",
     "231": "NOT_ALIVE",
     "232": "NON_PARTICIPATING",
-    "233": "NON_PARTICIPATING",
+    "233": "NOT_ALIVE",
     "234": "NON_PARTICIPATING",
-    "235": "NON_PARTICIPATING",
-    "236": "NON_PARTICIPATING",
-    "237": "FAIL",
-    "238": "FAIL",
+    "235": "FAIL",
+    "236": "FAIL",
+    "237": "NON_PARTICIPATING",
+    "238": "NON_PARTICIPATING",
     "239": "NON_PARTICIPATING",
-    "24": "FAIL",
+    "24": "NOT_ALIVE",
     "240": "NOT_ALIVE",
     "241": "NON_PARTICIPATING",
-    "242": "NON_PARTICIPATING",
-    "243": "NON_PARTICIPATING",
+    "242": "NOT_ALIVE",
+    "243": "NOT_ALIVE",
     "244": "NOT_ALIVE",
-    "245": "NON_PARTICIPATING",
-    "246": "NON_PARTICIPATING",
+    "245": "FAIL",
+    "246": "FAIL",
     "247": "FAIL",
     "248": "NON_PARTICIPATING",
     "249": "NON_PARTICIPATING",
@@ -174,88 +174,88 @@
     "251": "NOT_ALIVE",
     "252": "NON_PARTICIPATING",
     "253": "NON_PARTICIPATING",
-    "254": "NOT_ALIVE",
+    "254": "NON_PARTICIPATING",
     "255": "FAIL",
     "26": "NON_PARTICIPATING",
-    "27": "NON_PARTICIPATING",
-    "28": "NON_PARTICIPATING",
-    "29": "FAIL",
-    "3": "NOT_ALIVE",
+    "27": "NOT_ALIVE",
+    "28": "NOT_ALIVE",
+    "29": "NON_PARTICIPATING",
+    "3": "NON_PARTICIPATING",
     "30": "FAIL",
-    "31": "FAIL",
-    "32": "NON_PARTICIPATING",
+    "31": "NOT_ALIVE",
+    "32": "NOT_ALIVE",
     "33": "NOT_ALIVE",
-    "34": "NON_PARTICIPATING",
+    "34": "NOT_ALIVE",
     "35": "NON_PARTICIPATING",
     "36": "NON_PARTICIPATING",
     "37": "NOT_ALIVE",
     "38": "FAIL",
     "39": "NON_PARTICIPATING",
     "4": "NON_PARTICIPATING",
-    "40": "FAIL",
+    "40": "NON_PARTICIPATING",
     "41": "NOT_ALIVE",
-    "42": "NON_PARTICIPATING",
-    "43": "NOT_ALIVE",
+    "42": "NOT_ALIVE",
+    "43": "NON_PARTICIPATING",
     "44": "NON_PARTICIPATING",
-    "45": "NON_PARTICIPATING",
-    "46": "NOT_ALIVE",
+    "45": "NOT_ALIVE",
+    "46": "NON_PARTICIPATING",
     "47": "NON_PARTICIPATING",
     "48": "NON_PARTICIPATING",
     "49": "NOT_ALIVE",
-    "5": "FAIL",
+    "5": "NON_PARTICIPATING",
     "50": "NON_PARTICIPATING",
     "51": "NON_PARTICIPATING",
     "52": "NON_PARTICIPATING",
     "53": "FAIL",
-    "54": "NON_PARTICIPATING",
-    "55": "NON_PARTICIPATING",
-    "56": "NOT_ALIVE",
+    "54": "NOT_ALIVE",
+    "55": "NOT_ALIVE",
+    "56": "FAIL",
     "57": "FAIL",
-    "58": "NOT_ALIVE",
+    "58": "NON_PARTICIPATING",
     "59": "NOT_ALIVE",
     "6": "NON_PARTICIPATING",
     "60": "NOT_ALIVE",
-    "61": "FAIL",
-    "62": "FAIL",
-    "63": "NON_PARTICIPATING",
+    "61": "NON_PARTICIPATING",
+    "62": "NON_PARTICIPATING",
+    "63": "FAIL",
     "64": "NOT_ALIVE",
     "65": "FAIL",
     "66": "FAIL",
     "67": "NON_PARTICIPATING",
     "68": "NOT_ALIVE",
     "69": "FAIL",
-    "7": "FAIL",
+    "7": "NOT_ALIVE",
     "70": "FAIL",
     "71": "NOT_ALIVE",
-    "72": "FAIL",
-    "73": "NOT_ALIVE",
-    "74": "FAIL",
-    "75": "NON_PARTICIPATING",
-    "76": "NON_PARTICIPATING",
-    "77": "NON_PARTICIPATING",
-    "78": "NON_PARTICIPATING",
+    "72": "NON_PARTICIPATING",
+    "73": "NON_PARTICIPATING",
+    "74": "NOT_ALIVE",
+    "75": "NOT_ALIVE",
+    "76": "NOT_ALIVE",
+    "77": "NOT_ALIVE",
+    "78": "NOT_ALIVE",
     "79": "NON_PARTICIPATING",
-    "8": "NOT_ALIVE",
-    "80": "NOT_ALIVE",
+    "8": "NON_PARTICIPATING",
+    "80": "NON_PARTICIPATING",
     "81": "NOT_ALIVE",
     "82": "NOT_ALIVE",
     "83": "NON_PARTICIPATING",
-    "84": "NON_PARTICIPATING",
-    "85": "NON_PARTICIPATING",
-    "86": "NON_PARTICIPATING",
+    "84": "FAIL",
+    "85": "FAIL",
+    "86": "FAIL",
     "87": "NON_PARTICIPATING",
     "88": "FAIL",
-    "89": "NOT_ALIVE",
+    "89": "FAIL",
     "9": "FAIL",
     "90": "FAIL",
-    "91": "FAIL",
+    "91": "NOT_ALIVE",
     "92": "FAIL",
     "93": "NOT_ALIVE",
-    "94": "NON_PARTICIPATING",
-    "95": "NON_PARTICIPATING",
-    "96": "NOT_ALIVE",
+    "94": "FAIL",
+    "95": "NOT_ALIVE",
+    "96": "FAIL",
     "97": "NON_PARTICIPATING",
-    "98": "NOT_ALIVE",
+    "98": "NON_PARTICIPATING",
     "99": "NON_PARTICIPATING"
   },
   "architectures": [
@@ -264,17 +264,17 @@
   "attention_bias": false,
   "attention_dropout": 0.0,
   "block_list": [
-    6459346,
-    6459384
+    6459520,
+    6459553
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 72,
+  "inner_step": 9,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6458233,
+  "last_allreduce_block": 6459469,
   "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 sha256:180e6bb3b57b42f36d055f51a1d46382a22961073489791adbde3f325d0c0fc2
+oid sha256:24c16c25149230e74be64947f99094cfcb7b14cbb917d437ffa9d49ed2201f19
 size 88389626

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:18528982494f1a9175e68dad4ec31831addd09b776b7ec87c0d1f6c48ebe96bd
+oid sha256:c22bff077746ede47f3a277a109ae17965ab4fb73aa8b9724aa4ca42178e51cf
 size 4924926906

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e3d37fc2ebf5ba0e36bf5eaee62726d3e44c3441d133d01c585e4a6e2df66a10
+oid sha256:3f87beb99dc5c85c7b25bb30065c4eee5321ea716dfde1e480a97eed65b4a425
 size 4400216536