Run 10. Outer Step 18. Inner Step 5.

config.json

@@ -2,260 +2,260 @@
   "_name_or_path": "distributed/llama-1b",
   "all_reduce_scores": {
     "0": "NOT_ALIVE",
-    "1": "FAIL",
-    "10": "NON_PARTICIPATING",
+    "1": "SUCCESS",
+    "10": "SUCCESS",
     "100": "NOT_ALIVE",
     "101": "NOT_ALIVE",
-    "102": "NON_PARTICIPATING",
-    "103": "FAIL",
+    "102": "SUCCESS",
+    "103": "SUCCESS",
     "104": "NOT_ALIVE",
-    "105": "FAIL",
+    "105": "NON_PARTICIPATING",
     "106": "NON_PARTICIPATING",
-    "107": "NOT_ALIVE",
+    "107": "NON_PARTICIPATING",
     "108": "NOT_ALIVE",
-    "109": "FAIL",
-    "11": "NON_PARTICIPATING",
-    "110": "NON_PARTICIPATING",
-    "111": "FAIL",
-    "112": "NOT_ALIVE",
-    "113": "NOT_ALIVE",
-    "114": "FAIL",
-    "115": "NOT_ALIVE",
-    "116": "NOT_ALIVE",
+    "109": "SUCCESS",
+    "11": "NOT_ALIVE",
+    "110": "NOT_ALIVE",
+    "111": "SUCCESS",
+    "112": "SUCCESS",
+    "113": "SUCCESS",
+    "114": "SUCCESS",
+    "115": "NON_PARTICIPATING",
+    "116": "NON_PARTICIPATING",
     "117": "NON_PARTICIPATING",
     "118": "NON_PARTICIPATING",
-    "119": "NON_PARTICIPATING",
+    "119": "SUCCESS",
     "12": "NOT_ALIVE",
-    "120": "NON_PARTICIPATING",
+    "120": "NOT_ALIVE",
     "121": "NOT_ALIVE",
     "122": "NON_PARTICIPATING",
-    "123": "NOT_ALIVE",
-    "124": "NON_PARTICIPATING",
-    "125": "NOT_ALIVE",
-    "126": "FAIL",
+    "123": "SUCCESS",
+    "124": "NOT_ALIVE",
+    "125": "NON_PARTICIPATING",
+    "126": "SUCCESS",
     "127": "NON_PARTICIPATING",
     "128": "NON_PARTICIPATING",
-    "129": "NON_PARTICIPATING",
-    "13": "FAIL",
+    "129": "NOT_ALIVE",
+    "13": "SUCCESS",
     "130": "NOT_ALIVE",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
-    "133": "NON_PARTICIPATING",
-    "134": "NOT_ALIVE",
+    "133": "NOT_ALIVE",
+    "134": "NON_PARTICIPATING",
     "135": "NON_PARTICIPATING",
     "136": "NOT_ALIVE",
     "137": "NOT_ALIVE",
-    "138": "NON_PARTICIPATING",
+    "138": "NOT_ALIVE",
     "139": "NOT_ALIVE",
-    "14": "NON_PARTICIPATING",
+    "14": "NOT_ALIVE",
     "140": "NOT_ALIVE",
     "141": "NOT_ALIVE",
     "142": "FAIL",
-    "143": "NON_PARTICIPATING",
+    "143": "NOT_ALIVE",
     "144": "NON_PARTICIPATING",
     "145": "NOT_ALIVE",
-    "146": "FAIL",
+    "146": "NOT_ALIVE",
     "147": "NON_PARTICIPATING",
-    "148": "NON_PARTICIPATING",
-    "149": "FAIL",
-    "15": "FAIL",
+    "148": "NOT_ALIVE",
+    "149": "NON_PARTICIPATING",
+    "15": "NOT_ALIVE",
     "150": "FAIL",
     "151": "NOT_ALIVE",
     "152": "NOT_ALIVE",
-    "153": "FAIL",
-    "154": "FAIL",
-    "155": "NON_PARTICIPATING",
-    "156": "FAIL",
+    "153": "SUCCESS",
+    "154": "SUCCESS",
+    "155": "NOT_ALIVE",
+    "156": "SUCCESS",
     "157": "NON_PARTICIPATING",
     "158": "NON_PARTICIPATING",
     "159": "NON_PARTICIPATING",
     "16": "NOT_ALIVE",
     "160": "NOT_ALIVE",
     "161": "NOT_ALIVE",
-    "162": "NON_PARTICIPATING",
-    "163": "FAIL",
+    "162": "NOT_ALIVE",
+    "163": "SUCCESS",
     "164": "NOT_ALIVE",
-    "165": "NON_PARTICIPATING",
+    "165": "NOT_ALIVE",
     "166": "NOT_ALIVE",
     "167": "NON_PARTICIPATING",
-    "168": "FAIL",
-    "169": "FAIL",
+    "168": "SUCCESS",
+    "169": "SUCCESS",
     "17": "NON_PARTICIPATING",
-    "170": "FAIL",
+    "170": "NOT_ALIVE",
     "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
-    "173": "FAIL",
-    "174": "NON_PARTICIPATING",
+    "173": "SUCCESS",
+    "174": "NOT_ALIVE",
     "175": "NON_PARTICIPATING",
     "176": "NOT_ALIVE",
     "177": "FAIL",
     "178": "NOT_ALIVE",
     "179": "NOT_ALIVE",
-    "18": "NOT_ALIVE",
+    "18": "NON_PARTICIPATING",
     "180": "NON_PARTICIPATING",
     "181": "NOT_ALIVE",
     "182": "NON_PARTICIPATING",
-    "183": "NOT_ALIVE",
-    "184": "NOT_ALIVE",
-    "185": "FAIL",
-    "186": "FAIL",
+    "183": "SUCCESS",
+    "184": "SUCCESS",
+    "185": "NOT_ALIVE",
+    "186": "SUCCESS",
     "187": "NOT_ALIVE",
-    "188": "FAIL",
+    "188": "SUCCESS",
     "189": "NOT_ALIVE",
     "19": "NOT_ALIVE",
     "190": "NON_PARTICIPATING",
     "191": "NOT_ALIVE",
     "192": "NON_PARTICIPATING",
-    "193": "NON_PARTICIPATING",
+    "193": "SUCCESS",
     "194": "NON_PARTICIPATING",
-    "195": "NON_PARTICIPATING",
-    "196": "FAIL",
+    "195": "NOT_ALIVE",
+    "196": "SUCCESS",
     "197": "NON_PARTICIPATING",
-    "198": "NOT_ALIVE",
+    "198": "SUCCESS",
     "199": "NON_PARTICIPATING",
     "2": "NON_PARTICIPATING",
-    "20": "FAIL",
-    "200": "FAIL",
+    "20": "SUCCESS",
+    "200": "SUCCESS",
     "201": "NON_PARTICIPATING",
     "202": "NOT_ALIVE",
     "203": "NOT_ALIVE",
     "204": "NOT_ALIVE",
-    "205": "FAIL",
-    "206": "FAIL",
+    "205": "NOT_ALIVE",
+    "206": "SUCCESS",
     "207": "NOT_ALIVE",
     "208": "NON_PARTICIPATING",
-    "209": "FAIL",
+    "209": "NON_PARTICIPATING",
     "21": "NON_PARTICIPATING",
     "210": "NON_PARTICIPATING",
-    "211": "FAIL",
-    "212": "FAIL",
+    "211": "NON_PARTICIPATING",
+    "212": "NOT_ALIVE",
     "213": "NON_PARTICIPATING",
-    "214": "NON_PARTICIPATING",
+    "214": "NOT_ALIVE",
     "215": "NOT_ALIVE",
     "216": "NOT_ALIVE",
     "217": "NOT_ALIVE",
-    "218": "NOT_ALIVE",
+    "218": "NON_PARTICIPATING",
     "219": "FAIL",
-    "22": "NOT_ALIVE",
+    "22": "NON_PARTICIPATING",
     "220": "NOT_ALIVE",
-    "221": "FAIL",
+    "221": "SUCCESS",
     "222": "FAIL",
-    "223": "NON_PARTICIPATING",
+    "223": "NOT_ALIVE",
     "224": "NOT_ALIVE",
     "225": "NON_PARTICIPATING",
-    "226": "NON_PARTICIPATING",
-    "227": "FAIL",
+    "226": "NOT_ALIVE",
+    "227": "SUCCESS",
     "228": "NOT_ALIVE",
     "229": "NON_PARTICIPATING",
     "23": "NON_PARTICIPATING",
     "230": "NOT_ALIVE",
     "231": "NOT_ALIVE",
-    "232": "FAIL",
-    "233": "NON_PARTICIPATING",
+    "232": "SUCCESS",
+    "233": "NOT_ALIVE",
     "234": "NON_PARTICIPATING",
-    "235": "NOT_ALIVE",
-    "236": "NON_PARTICIPATING",
+    "235": "NON_PARTICIPATING",
+    "236": "FAIL",
     "237": "FAIL",
-    "238": "FAIL",
-    "239": "NOT_ALIVE",
-    "24": "NON_PARTICIPATING",
+    "238": "NON_PARTICIPATING",
+    "239": "NON_PARTICIPATING",
+    "24": "NOT_ALIVE",
     "240": "NOT_ALIVE",
     "241": "NON_PARTICIPATING",
     "242": "NON_PARTICIPATING",
     "243": "NOT_ALIVE",
     "244": "NOT_ALIVE",
-    "245": "FAIL",
-    "246": "NOT_ALIVE",
-    "247": "NON_PARTICIPATING",
+    "245": "NON_PARTICIPATING",
+    "246": "NON_PARTICIPATING",
+    "247": "NOT_ALIVE",
     "248": "NON_PARTICIPATING",
     "249": "NOT_ALIVE",
     "25": "SUCCESS",
-    "250": "NON_PARTICIPATING",
+    "250": "NOT_ALIVE",
     "251": "NOT_ALIVE",
     "252": "NON_PARTICIPATING",
     "253": "NOT_ALIVE",
-    "254": "FAIL",
+    "254": "NOT_ALIVE",
     "255": "NON_PARTICIPATING",
-    "26": "NOT_ALIVE",
+    "26": "SUCCESS",
     "27": "NOT_ALIVE",
-    "28": "FAIL",
-    "29": "FAIL",
+    "28": "NON_PARTICIPATING",
+    "29": "SUCCESS",
     "3": "NON_PARTICIPATING",
     "30": "NON_PARTICIPATING",
     "31": "NOT_ALIVE",
-    "32": "NON_PARTICIPATING",
+    "32": "NOT_ALIVE",
     "33": "NOT_ALIVE",
-    "34": "NON_PARTICIPATING",
-    "35": "FAIL",
-    "36": "NON_PARTICIPATING",
+    "34": "NOT_ALIVE",
+    "35": "SUCCESS",
+    "36": "SUCCESS",
     "37": "NOT_ALIVE",
     "38": "NON_PARTICIPATING",
-    "39": "FAIL",
-    "4": "NOT_ALIVE",
-    "40": "FAIL",
-    "41": "NON_PARTICIPATING",
-    "42": "NON_PARTICIPATING",
+    "39": "SUCCESS",
+    "4": "SUCCESS",
+    "40": "NOT_ALIVE",
+    "41": "NOT_ALIVE",
+    "42": "NOT_ALIVE",
     "43": "NOT_ALIVE",
-    "44": "FAIL",
+    "44": "SUCCESS",
     "45": "NON_PARTICIPATING",
-    "46": "FAIL",
-    "47": "FAIL",
+    "46": "SUCCESS",
+    "47": "SUCCESS",
     "48": "NON_PARTICIPATING",
     "49": "NOT_ALIVE",
     "5": "NON_PARTICIPATING",
-    "50": "NOT_ALIVE",
+    "50": "NON_PARTICIPATING",
     "51": "NON_PARTICIPATING",
-    "52": "NOT_ALIVE",
-    "53": "NON_PARTICIPATING",
+    "52": "NON_PARTICIPATING",
+    "53": "SUCCESS",
     "54": "NOT_ALIVE",
-    "55": "NON_PARTICIPATING",
-    "56": "FAIL",
-    "57": "NOT_ALIVE",
+    "55": "NOT_ALIVE",
+    "56": "NON_PARTICIPATING",
+    "57": "SUCCESS",
     "58": "NON_PARTICIPATING",
-    "59": "NOT_ALIVE",
+    "59": "NON_PARTICIPATING",
     "6": "NOT_ALIVE",
-    "60": "FAIL",
+    "60": "SUCCESS",
     "61": "NOT_ALIVE",
     "62": "NON_PARTICIPATING",
-    "63": "NOT_ALIVE",
-    "64": "FAIL",
+    "63": "NON_PARTICIPATING",
+    "64": "NOT_ALIVE",
     "65": "NOT_ALIVE",
     "66": "NON_PARTICIPATING",
     "67": "NON_PARTICIPATING",
     "68": "NOT_ALIVE",
-    "69": "FAIL",
-    "7": "NOT_ALIVE",
-    "70": "FAIL",
+    "69": "SUCCESS",
+    "7": "SUCCESS",
+    "70": "NOT_ALIVE",
     "71": "NOT_ALIVE",
     "72": "NON_PARTICIPATING",
-    "73": "NON_PARTICIPATING",
-    "74": "FAIL",
-    "75": "NON_PARTICIPATING",
+    "73": "NOT_ALIVE",
+    "74": "SUCCESS",
+    "75": "NOT_ALIVE",
     "76": "NON_PARTICIPATING",
     "77": "NOT_ALIVE",
-    "78": "NOT_ALIVE",
+    "78": "NON_PARTICIPATING",
     "79": "NON_PARTICIPATING",
-    "8": "NOT_ALIVE",
+    "8": "SUCCESS",
     "80": "NOT_ALIVE",
     "81": "NOT_ALIVE",
     "82": "NOT_ALIVE",
     "83": "NON_PARTICIPATING",
-    "84": "FAIL",
+    "84": "SUCCESS",
     "85": "NON_PARTICIPATING",
     "86": "NOT_ALIVE",
-    "87": "NOT_ALIVE",
+    "87": "NON_PARTICIPATING",
     "88": "NON_PARTICIPATING",
-    "89": "NON_PARTICIPATING",
-    "9": "FAIL",
-    "90": "FAIL",
-    "91": "FAIL",
-    "92": "FAIL",
+    "89": "NOT_ALIVE",
+    "9": "SUCCESS",
+    "90": "NON_PARTICIPATING",
+    "91": "SUCCESS",
+    "92": "SUCCESS",
     "93": "NOT_ALIVE",
-    "94": "FAIL",
+    "94": "NON_PARTICIPATING",
     "95": "NOT_ALIVE",
-    "96": "NOT_ALIVE",
+    "96": "SUCCESS",
     "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": [
-    6473609,
-    6473623
+    6477847,
+    6477865
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 4,
+  "inner_step": 5,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6473589,
+  "last_allreduce_block": 6477264,
   "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 sha256:77b07142e3830691277432fc08bb532a228e85969b6dff194d04d601604589a4
+oid sha256:80ae094b7b2da6b7aa7a52c43282d83e0d9d2a219021407c10086092aaeaf04d
 size 88389626

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6853ad84036a739eeb398cc63c4a6e456be1aa5bb4efe8fde4c95bbb486ff166
+oid sha256:8cc28f74e0582a897e5f2005e1c34f27e20a1c0f769c003a1554cbd2ca2e2157
 size 4924926906

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:96db3a615d051c847433285aa9a022dcb9b80dc6d3ebadc8cae30db99b9f8347
+oid sha256:7f1a2fa874b66a85311d6cd8edb01335477a7e6adf9b0edd3cf8672143cb99df
 size 4400216536