Run 10. Outer Step 19. Inner Step 5.

config.json

@@ -2,258 +2,258 @@
   "_name_or_path": "distributed/llama-1b",
   "all_reduce_scores": {
     "0": "NOT_ALIVE",
-    "1": "SUCCESS",
-    "10": "SUCCESS",
+    "1": "NON_PARTICIPATING",
+    "10": "NOT_ALIVE",
     "100": "NOT_ALIVE",
     "101": "NOT_ALIVE",
-    "102": "SUCCESS",
-    "103": "SUCCESS",
+    "102": "NOT_ALIVE",
+    "103": "FAIL",
     "104": "NOT_ALIVE",
-    "105": "NON_PARTICIPATING",
+    "105": "NOT_ALIVE",
     "106": "NON_PARTICIPATING",
     "107": "NON_PARTICIPATING",
     "108": "NOT_ALIVE",
-    "109": "SUCCESS",
-    "11": "NOT_ALIVE",
+    "109": "NON_PARTICIPATING",
+    "11": "FAIL",
     "110": "NOT_ALIVE",
-    "111": "SUCCESS",
-    "112": "SUCCESS",
-    "113": "SUCCESS",
+    "111": "NOT_ALIVE",
+    "112": "FAIL",
+    "113": "NOT_ALIVE",
     "114": "SUCCESS",
     "115": "NON_PARTICIPATING",
-    "116": "NON_PARTICIPATING",
+    "116": "SUCCESS",
     "117": "NON_PARTICIPATING",
-    "118": "NON_PARTICIPATING",
-    "119": "SUCCESS",
+    "118": "NOT_ALIVE",
+    "119": "NON_PARTICIPATING",
     "12": "NOT_ALIVE",
     "120": "NOT_ALIVE",
     "121": "NOT_ALIVE",
     "122": "NON_PARTICIPATING",
-    "123": "SUCCESS",
+    "123": "FAIL",
     "124": "NOT_ALIVE",
-    "125": "NON_PARTICIPATING",
-    "126": "SUCCESS",
+    "125": "NOT_ALIVE",
+    "126": "NOT_ALIVE",
     "127": "NON_PARTICIPATING",
     "128": "NON_PARTICIPATING",
-    "129": "NOT_ALIVE",
-    "13": "SUCCESS",
+    "129": "NON_PARTICIPATING",
+    "13": "NON_PARTICIPATING",
     "130": "NOT_ALIVE",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
-    "133": "NOT_ALIVE",
-    "134": "NON_PARTICIPATING",
+    "133": "NON_PARTICIPATING",
+    "134": "SUCCESS",
     "135": "NON_PARTICIPATING",
     "136": "NOT_ALIVE",
     "137": "NOT_ALIVE",
-    "138": "NOT_ALIVE",
+    "138": "NON_PARTICIPATING",
     "139": "NOT_ALIVE",
-    "14": "NOT_ALIVE",
+    "14": "NON_PARTICIPATING",
     "140": "NOT_ALIVE",
     "141": "NOT_ALIVE",
-    "142": "FAIL",
-    "143": "NOT_ALIVE",
-    "144": "NON_PARTICIPATING",
+    "142": "NOT_ALIVE",
+    "143": "FAIL",
+    "144": "NOT_ALIVE",
     "145": "NOT_ALIVE",
-    "146": "NOT_ALIVE",
+    "146": "NON_PARTICIPATING",
     "147": "NON_PARTICIPATING",
-    "148": "NOT_ALIVE",
-    "149": "NON_PARTICIPATING",
-    "15": "NOT_ALIVE",
-    "150": "FAIL",
+    "148": "NON_PARTICIPATING",
+    "149": "NOT_ALIVE",
+    "15": "NON_PARTICIPATING",
+    "150": "NON_PARTICIPATING",
     "151": "NOT_ALIVE",
-    "152": "NOT_ALIVE",
-    "153": "SUCCESS",
+    "152": "NON_PARTICIPATING",
+    "153": "FAIL",
     "154": "SUCCESS",
-    "155": "NOT_ALIVE",
-    "156": "SUCCESS",
+    "155": "FAIL",
+    "156": "NON_PARTICIPATING",
     "157": "NON_PARTICIPATING",
     "158": "NON_PARTICIPATING",
-    "159": "NON_PARTICIPATING",
+    "159": "NOT_ALIVE",
     "16": "NOT_ALIVE",
-    "160": "NOT_ALIVE",
+    "160": "FAIL",
     "161": "NOT_ALIVE",
     "162": "NOT_ALIVE",
     "163": "SUCCESS",
     "164": "NOT_ALIVE",
     "165": "NOT_ALIVE",
     "166": "NOT_ALIVE",
-    "167": "NON_PARTICIPATING",
-    "168": "SUCCESS",
-    "169": "SUCCESS",
+    "167": "FAIL",
+    "168": "FAIL",
+    "169": "NOT_ALIVE",
     "17": "NON_PARTICIPATING",
-    "170": "NOT_ALIVE",
+    "170": "NON_PARTICIPATING",
     "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
-    "173": "SUCCESS",
-    "174": "NOT_ALIVE",
-    "175": "NON_PARTICIPATING",
+    "173": "NON_PARTICIPATING",
+    "174": "NON_PARTICIPATING",
+    "175": "FAIL",
     "176": "NOT_ALIVE",
-    "177": "FAIL",
+    "177": "NOT_ALIVE",
     "178": "NOT_ALIVE",
-    "179": "NOT_ALIVE",
-    "18": "NON_PARTICIPATING",
+    "179": "NON_PARTICIPATING",
+    "18": "FAIL",
     "180": "NON_PARTICIPATING",
     "181": "NOT_ALIVE",
     "182": "NON_PARTICIPATING",
-    "183": "SUCCESS",
-    "184": "SUCCESS",
+    "183": "NON_PARTICIPATING",
+    "184": "NON_PARTICIPATING",
     "185": "NOT_ALIVE",
-    "186": "SUCCESS",
+    "186": "FAIL",
     "187": "NOT_ALIVE",
-    "188": "SUCCESS",
+    "188": "FAIL",
     "189": "NOT_ALIVE",
-    "19": "NOT_ALIVE",
+    "19": "NON_PARTICIPATING",
     "190": "NON_PARTICIPATING",
     "191": "NOT_ALIVE",
     "192": "NON_PARTICIPATING",
-    "193": "SUCCESS",
-    "194": "NON_PARTICIPATING",
-    "195": "NOT_ALIVE",
-    "196": "SUCCESS",
+    "193": "NON_PARTICIPATING",
+    "194": "NOT_ALIVE",
+    "195": "FAIL",
+    "196": "NON_PARTICIPATING",
     "197": "NON_PARTICIPATING",
-    "198": "SUCCESS",
+    "198": "NON_PARTICIPATING",
     "199": "NON_PARTICIPATING",
     "2": "NON_PARTICIPATING",
-    "20": "SUCCESS",
-    "200": "SUCCESS",
-    "201": "NON_PARTICIPATING",
+    "20": "NON_PARTICIPATING",
+    "200": "NON_PARTICIPATING",
+    "201": "FAIL",
     "202": "NOT_ALIVE",
     "203": "NOT_ALIVE",
     "204": "NOT_ALIVE",
-    "205": "NOT_ALIVE",
-    "206": "SUCCESS",
+    "205": "NON_PARTICIPATING",
+    "206": "FAIL",
     "207": "NOT_ALIVE",
-    "208": "NON_PARTICIPATING",
-    "209": "NON_PARTICIPATING",
+    "208": "NOT_ALIVE",
+    "209": "FAIL",
     "21": "NON_PARTICIPATING",
     "210": "NON_PARTICIPATING",
-    "211": "NON_PARTICIPATING",
-    "212": "NOT_ALIVE",
-    "213": "NON_PARTICIPATING",
-    "214": "NOT_ALIVE",
+    "211": "FAIL",
+    "212": "NON_PARTICIPATING",
+    "213": "SUCCESS",
+    "214": "NON_PARTICIPATING",
     "215": "NOT_ALIVE",
     "216": "NOT_ALIVE",
     "217": "NOT_ALIVE",
-    "218": "NON_PARTICIPATING",
-    "219": "FAIL",
-    "22": "NON_PARTICIPATING",
+    "218": "FAIL",
+    "219": "NOT_ALIVE",
+    "22": "FAIL",
     "220": "NOT_ALIVE",
-    "221": "SUCCESS",
-    "222": "FAIL",
-    "223": "NOT_ALIVE",
+    "221": "NON_PARTICIPATING",
+    "222": "NON_PARTICIPATING",
+    "223": "FAIL",
     "224": "NOT_ALIVE",
     "225": "NON_PARTICIPATING",
-    "226": "NOT_ALIVE",
-    "227": "SUCCESS",
+    "226": "NON_PARTICIPATING",
+    "227": "FAIL",
     "228": "NOT_ALIVE",
     "229": "NON_PARTICIPATING",
-    "23": "NON_PARTICIPATING",
+    "23": "NOT_ALIVE",
     "230": "NOT_ALIVE",
     "231": "NOT_ALIVE",
-    "232": "SUCCESS",
-    "233": "NOT_ALIVE",
-    "234": "NON_PARTICIPATING",
-    "235": "NON_PARTICIPATING",
+    "232": "NON_PARTICIPATING",
+    "233": "FAIL",
+    "234": "NOT_ALIVE",
+    "235": "SUCCESS",
     "236": "FAIL",
     "237": "FAIL",
     "238": "NON_PARTICIPATING",
     "239": "NON_PARTICIPATING",
-    "24": "NOT_ALIVE",
+    "24": "NON_PARTICIPATING",
     "240": "NOT_ALIVE",
-    "241": "NON_PARTICIPATING",
-    "242": "NON_PARTICIPATING",
+    "241": "SUCCESS",
+    "242": "FAIL",
     "243": "NOT_ALIVE",
     "244": "NOT_ALIVE",
     "245": "NON_PARTICIPATING",
-    "246": "NON_PARTICIPATING",
+    "246": "FAIL",
     "247": "NOT_ALIVE",
-    "248": "NON_PARTICIPATING",
-    "249": "NOT_ALIVE",
+    "248": "FAIL",
+    "249": "FAIL",
     "25": "SUCCESS",
     "250": "NOT_ALIVE",
     "251": "NOT_ALIVE",
-    "252": "NON_PARTICIPATING",
+    "252": "FAIL",
     "253": "NOT_ALIVE",
     "254": "NOT_ALIVE",
-    "255": "NON_PARTICIPATING",
-    "26": "SUCCESS",
-    "27": "NOT_ALIVE",
-    "28": "NON_PARTICIPATING",
-    "29": "SUCCESS",
-    "3": "NON_PARTICIPATING",
+    "255": "NOT_ALIVE",
+    "26": "NOT_ALIVE",
+    "27": "NON_PARTICIPATING",
+    "28": "FAIL",
+    "29": "FAIL",
+    "3": "FAIL",
     "30": "NON_PARTICIPATING",
     "31": "NOT_ALIVE",
-    "32": "NOT_ALIVE",
+    "32": "FAIL",
     "33": "NOT_ALIVE",
-    "34": "NOT_ALIVE",
-    "35": "SUCCESS",
-    "36": "SUCCESS",
+    "34": "NON_PARTICIPATING",
+    "35": "FAIL",
+    "36": "NON_PARTICIPATING",
     "37": "NOT_ALIVE",
-    "38": "NON_PARTICIPATING",
-    "39": "SUCCESS",
-    "4": "SUCCESS",
-    "40": "NOT_ALIVE",
+    "38": "FAIL",
+    "39": "NON_PARTICIPATING",
+    "4": "NON_PARTICIPATING",
+    "40": "FAIL",
     "41": "NOT_ALIVE",
     "42": "NOT_ALIVE",
     "43": "NOT_ALIVE",
-    "44": "SUCCESS",
-    "45": "NON_PARTICIPATING",
-    "46": "SUCCESS",
-    "47": "SUCCESS",
+    "44": "NON_PARTICIPATING",
+    "45": "FAIL",
+    "46": "NOT_ALIVE",
+    "47": "FAIL",
     "48": "NON_PARTICIPATING",
     "49": "NOT_ALIVE",
     "5": "NON_PARTICIPATING",
     "50": "NON_PARTICIPATING",
     "51": "NON_PARTICIPATING",
     "52": "NON_PARTICIPATING",
-    "53": "SUCCESS",
-    "54": "NOT_ALIVE",
-    "55": "NOT_ALIVE",
-    "56": "NON_PARTICIPATING",
-    "57": "SUCCESS",
+    "53": "NON_PARTICIPATING",
+    "54": "NON_PARTICIPATING",
+    "55": "NON_PARTICIPATING",
+    "56": "FAIL",
+    "57": "NON_PARTICIPATING",
     "58": "NON_PARTICIPATING",
     "59": "NON_PARTICIPATING",
     "6": "NOT_ALIVE",
-    "60": "SUCCESS",
+    "60": "NON_PARTICIPATING",
     "61": "NOT_ALIVE",
-    "62": "NON_PARTICIPATING",
+    "62": "SUCCESS",
     "63": "NON_PARTICIPATING",
     "64": "NOT_ALIVE",
     "65": "NOT_ALIVE",
-    "66": "NON_PARTICIPATING",
+    "66": "NOT_ALIVE",
     "67": "NON_PARTICIPATING",
     "68": "NOT_ALIVE",
-    "69": "SUCCESS",
+    "69": "FAIL",
     "7": "SUCCESS",
     "70": "NOT_ALIVE",
     "71": "NOT_ALIVE",
     "72": "NON_PARTICIPATING",
     "73": "NOT_ALIVE",
-    "74": "SUCCESS",
-    "75": "NOT_ALIVE",
+    "74": "NOT_ALIVE",
+    "75": "NON_PARTICIPATING",
     "76": "NON_PARTICIPATING",
-    "77": "NOT_ALIVE",
+    "77": "NON_PARTICIPATING",
     "78": "NON_PARTICIPATING",
     "79": "NON_PARTICIPATING",
-    "8": "SUCCESS",
+    "8": "NON_PARTICIPATING",
     "80": "NOT_ALIVE",
     "81": "NOT_ALIVE",
     "82": "NOT_ALIVE",
     "83": "NON_PARTICIPATING",
-    "84": "SUCCESS",
-    "85": "NON_PARTICIPATING",
-    "86": "NOT_ALIVE",
-    "87": "NON_PARTICIPATING",
-    "88": "NON_PARTICIPATING",
+    "84": "NOT_ALIVE",
+    "85": "NOT_ALIVE",
+    "86": "FAIL",
+    "87": "SUCCESS",
+    "88": "FAIL",
     "89": "NOT_ALIVE",
-    "9": "SUCCESS",
-    "90": "NON_PARTICIPATING",
+    "9": "NOT_ALIVE",
+    "90": "FAIL",
     "91": "SUCCESS",
-    "92": "SUCCESS",
+    "92": "FAIL",
     "93": "NOT_ALIVE",
     "94": "NON_PARTICIPATING",
     "95": "NOT_ALIVE",
-    "96": "SUCCESS",
+    "96": "NOT_ALIVE",
     "97": "NON_PARTICIPATING",
     "98": "NON_PARTICIPATING",
     "99": "NON_PARTICIPATING"
@@ -264,18 +264,17 @@
   "attention_bias": false,
   "attention_dropout": 0.0,
   "block_list": [
-    6478404,
-    6478420,
-    6478435
+    6478508,
+    6478524
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 81,
+  "inner_step": 4,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6477264,
+  "last_allreduce_block": 6478480,
   "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:bbbe96a17c312e5fef32505e79e4f75283bd394df0777a89bf16ec0cda4006ab
+oid sha256:4b940eb804d3edf40f08ccbbf8e2356233df4f5fe9edecf122c04c66bf1938a1
 size 88389626

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8f0a3f43d5de1cecc1cc8f1d99a01a3d10aff8f55bfadbc3ba5b516a7628644b
+oid sha256:961803ec44ff2dcd73d6ac079a01291557c8d5f63f20bdd11d7653df5c4be566
 size 4924926906

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fb56d33fbe4f1b379f1155a6179ff506fbd9e0c44c7dfbe90c8c2d6f304c4c58
+oid sha256:f049c314e8ad147a3482da2a3e28bafa169343066be0e28c4ba0f8749970412c
 size 4400216536