Run 9. Outer Step 18. Inner Step 5.

config.json

@@ -2,56 +2,56 @@
   "_name_or_path": "distributed/llama-1b",
   "all_reduce_scores": {
     "0": "NOT_ALIVE",
-    "1": "NON_PARTICIPATING",
-    "10": "NOT_ALIVE",
+    "1": "FAIL",
+    "10": "NON_PARTICIPATING",
     "100": "NOT_ALIVE",
     "101": "NOT_ALIVE",
     "102": "NOT_ALIVE",
     "103": "FAIL",
-    "104": "NOT_ALIVE",
+    "104": "NON_PARTICIPATING",
     "105": "FAIL",
-    "106": "NON_PARTICIPATING",
-    "107": "NON_PARTICIPATING",
+    "106": "FAIL",
+    "107": "NOT_ALIVE",
     "108": "NOT_ALIVE",
-    "109": "FAIL",
+    "109": "NON_PARTICIPATING",
     "11": "NON_PARTICIPATING",
     "110": "NOT_ALIVE",
     "111": "NON_PARTICIPATING",
-    "112": "FAIL",
+    "112": "NON_PARTICIPATING",
     "113": "NON_PARTICIPATING",
     "114": "NON_PARTICIPATING",
     "115": "NON_PARTICIPATING",
-    "116": "NON_PARTICIPATING",
+    "116": "NOT_ALIVE",
     "117": "NON_PARTICIPATING",
-    "118": "NON_PARTICIPATING",
+    "118": "FAIL",
     "119": "FAIL",
     "12": "NOT_ALIVE",
     "120": "NOT_ALIVE",
     "121": "NOT_ALIVE",
     "122": "NON_PARTICIPATING",
     "123": "FAIL",
-    "124": "NOT_ALIVE",
+    "124": "NON_PARTICIPATING",
     "125": "NON_PARTICIPATING",
-    "126": "FAIL",
+    "126": "NON_PARTICIPATING",
     "127": "NOT_ALIVE",
-    "128": "NON_PARTICIPATING",
+    "128": "FAIL",
     "129": "NOT_ALIVE",
-    "13": "NON_PARTICIPATING",
+    "13": "FAIL",
     "130": "NON_PARTICIPATING",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
     "133": "NON_PARTICIPATING",
     "134": "FAIL",
-    "135": "NON_PARTICIPATING",
+    "135": "NOT_ALIVE",
     "136": "NOT_ALIVE",
     "137": "NOT_ALIVE",
     "138": "NOT_ALIVE",
     "139": "NOT_ALIVE",
     "14": "NOT_ALIVE",
     "140": "NOT_ALIVE",
-    "141": "NON_PARTICIPATING",
+    "141": "NOT_ALIVE",
     "142": "NON_PARTICIPATING",
-    "143": "NON_PARTICIPATING",
+    "143": "FAIL",
     "144": "NOT_ALIVE",
     "145": "NOT_ALIVE",
     "146": "FAIL",
@@ -59,173 +59,173 @@
     "148": "NOT_ALIVE",
     "149": "NON_PARTICIPATING",
     "15": "NOT_ALIVE",
-    "150": "NON_PARTICIPATING",
+    "150": "NOT_ALIVE",
     "151": "NOT_ALIVE",
     "152": "NOT_ALIVE",
-    "153": "FAIL",
+    "153": "NON_PARTICIPATING",
     "154": "NON_PARTICIPATING",
-    "155": "NON_PARTICIPATING",
-    "156": "NON_PARTICIPATING",
-    "157": "FAIL",
+    "155": "NOT_ALIVE",
+    "156": "FAIL",
+    "157": "NOT_ALIVE",
     "158": "NON_PARTICIPATING",
-    "159": "NON_PARTICIPATING",
-    "16": "NON_PARTICIPATING",
+    "159": "FAIL",
+    "16": "FAIL",
     "160": "NOT_ALIVE",
     "161": "NOT_ALIVE",
-    "162": "NOT_ALIVE",
-    "163": "NOT_ALIVE",
+    "162": "NON_PARTICIPATING",
+    "163": "FAIL",
     "164": "NOT_ALIVE",
     "165": "NOT_ALIVE",
-    "166": "FAIL",
+    "166": "NON_PARTICIPATING",
     "167": "NON_PARTICIPATING",
     "168": "NON_PARTICIPATING",
-    "169": "NOT_ALIVE",
+    "169": "FAIL",
     "17": "NON_PARTICIPATING",
-    "170": "FAIL",
+    "170": "NON_PARTICIPATING",
     "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
-    "173": "NON_PARTICIPATING",
-    "174": "NON_PARTICIPATING",
-    "175": "NON_PARTICIPATING",
+    "173": "FAIL",
+    "174": "NOT_ALIVE",
+    "175": "FAIL",
     "176": "NOT_ALIVE",
     "177": "NON_PARTICIPATING",
     "178": "NOT_ALIVE",
     "179": "NOT_ALIVE",
-    "18": "FAIL",
+    "18": "NON_PARTICIPATING",
     "180": "NON_PARTICIPATING",
     "181": "NON_PARTICIPATING",
-    "182": "NON_PARTICIPATING",
-    "183": "NOT_ALIVE",
+    "182": "FAIL",
+    "183": "NON_PARTICIPATING",
     "184": "NON_PARTICIPATING",
-    "185": "FAIL",
-    "186": "NON_PARTICIPATING",
+    "185": "NON_PARTICIPATING",
+    "186": "FAIL",
     "187": "NON_PARTICIPATING",
     "188": "NON_PARTICIPATING",
     "189": "NON_PARTICIPATING",
-    "19": "FAIL",
+    "19": "NON_PARTICIPATING",
     "190": "NON_PARTICIPATING",
     "191": "FAIL",
-    "192": "NOT_ALIVE",
-    "193": "NON_PARTICIPATING",
+    "192": "NON_PARTICIPATING",
+    "193": "NOT_ALIVE",
     "194": "NON_PARTICIPATING",
     "195": "FAIL",
     "196": "NON_PARTICIPATING",
     "197": "NON_PARTICIPATING",
-    "198": "FAIL",
+    "198": "NON_PARTICIPATING",
     "199": "NON_PARTICIPATING",
-    "2": "FAIL",
-    "20": "FAIL",
+    "2": "NON_PARTICIPATING",
+    "20": "NON_PARTICIPATING",
     "200": "NON_PARTICIPATING",
     "201": "NON_PARTICIPATING",
     "202": "NON_PARTICIPATING",
-    "203": "NON_PARTICIPATING",
+    "203": "NOT_ALIVE",
     "204": "NOT_ALIVE",
     "205": "NON_PARTICIPATING",
-    "206": "FAIL",
+    "206": "NON_PARTICIPATING",
     "207": "NOT_ALIVE",
     "208": "NOT_ALIVE",
     "209": "FAIL",
     "21": "NON_PARTICIPATING",
     "210": "NON_PARTICIPATING",
-    "211": "NON_PARTICIPATING",
+    "211": "NOT_ALIVE",
     "212": "FAIL",
     "213": "NOT_ALIVE",
-    "214": "FAIL",
+    "214": "NOT_ALIVE",
     "215": "FAIL",
     "216": "NON_PARTICIPATING",
     "217": "NON_PARTICIPATING",
     "218": "NON_PARTICIPATING",
-    "219": "FAIL",
+    "219": "NON_PARTICIPATING",
     "22": "NOT_ALIVE",
     "220": "NOT_ALIVE",
-    "221": "NON_PARTICIPATING",
+    "221": "NOT_ALIVE",
     "222": "NON_PARTICIPATING",
     "223": "FAIL",
     "224": "NOT_ALIVE",
-    "225": "NON_PARTICIPATING",
-    "226": "FAIL",
-    "227": "NON_PARTICIPATING",
-    "228": "NON_PARTICIPATING",
+    "225": "FAIL",
+    "226": "NOT_ALIVE",
+    "227": "FAIL",
+    "228": "NOT_ALIVE",
     "229": "NOT_ALIVE",
-    "23": "NOT_ALIVE",
+    "23": "NON_PARTICIPATING",
     "230": "NOT_ALIVE",
     "231": "NOT_ALIVE",
-    "232": "FAIL",
+    "232": "NON_PARTICIPATING",
     "233": "FAIL",
-    "234": "NON_PARTICIPATING",
-    "235": "FAIL",
-    "236": "FAIL",
+    "234": "FAIL",
+    "235": "NOT_ALIVE",
+    "236": "NON_PARTICIPATING",
     "237": "NON_PARTICIPATING",
-    "238": "FAIL",
-    "239": "FAIL",
+    "238": "NON_PARTICIPATING",
+    "239": "NON_PARTICIPATING",
     "24": "NON_PARTICIPATING",
     "240": "NOT_ALIVE",
-    "241": "NON_PARTICIPATING",
+    "241": "NOT_ALIVE",
     "242": "NOT_ALIVE",
     "243": "NOT_ALIVE",
     "244": "NOT_ALIVE",
-    "245": "NON_PARTICIPATING",
-    "246": "NON_PARTICIPATING",
-    "247": "FAIL",
+    "245": "FAIL",
+    "246": "FAIL",
+    "247": "NON_PARTICIPATING",
     "248": "NON_PARTICIPATING",
     "249": "FAIL",
     "25": "SUCCESS",
-    "250": "FAIL",
+    "250": "NON_PARTICIPATING",
     "251": "NOT_ALIVE",
-    "252": "FAIL",
-    "253": "NON_PARTICIPATING",
+    "252": "NON_PARTICIPATING",
+    "253": "NOT_ALIVE",
     "254": "NON_PARTICIPATING",
     "255": "NON_PARTICIPATING",
-    "26": "FAIL",
+    "26": "NON_PARTICIPATING",
     "27": "NON_PARTICIPATING",
     "28": "FAIL",
-    "29": "FAIL",
+    "29": "NOT_ALIVE",
     "3": "NON_PARTICIPATING",
-    "30": "FAIL",
-    "31": "NON_PARTICIPATING",
+    "30": "NON_PARTICIPATING",
+    "31": "FAIL",
     "32": "NOT_ALIVE",
     "33": "NOT_ALIVE",
     "34": "NOT_ALIVE",
     "35": "FAIL",
     "36": "FAIL",
     "37": "NOT_ALIVE",
-    "38": "NON_PARTICIPATING",
-    "39": "FAIL",
-    "4": "FAIL",
+    "38": "FAIL",
+    "39": "NON_PARTICIPATING",
+    "4": "NON_PARTICIPATING",
     "40": "FAIL",
     "41": "NOT_ALIVE",
-    "42": "NON_PARTICIPATING",
-    "43": "NON_PARTICIPATING",
-    "44": "NOT_ALIVE",
-    "45": "NON_PARTICIPATING",
+    "42": "FAIL",
+    "43": "NOT_ALIVE",
+    "44": "NON_PARTICIPATING",
+    "45": "NOT_ALIVE",
     "46": "NOT_ALIVE",
     "47": "NON_PARTICIPATING",
     "48": "NON_PARTICIPATING",
     "49": "NOT_ALIVE",
-    "5": "NOT_ALIVE",
-    "50": "NON_PARTICIPATING",
-    "51": "NON_PARTICIPATING",
+    "5": "NON_PARTICIPATING",
+    "50": "NOT_ALIVE",
+    "51": "FAIL",
     "52": "NON_PARTICIPATING",
-    "53": "FAIL",
+    "53": "NON_PARTICIPATING",
     "54": "NOT_ALIVE",
     "55": "NOT_ALIVE",
-    "56": "NOT_ALIVE",
+    "56": "FAIL",
     "57": "NON_PARTICIPATING",
     "58": "NON_PARTICIPATING",
     "59": "NOT_ALIVE",
-    "6": "NOT_ALIVE",
-    "60": "NON_PARTICIPATING",
-    "61": "NOT_ALIVE",
+    "6": "NON_PARTICIPATING",
+    "60": "FAIL",
+    "61": "NON_PARTICIPATING",
     "62": "NON_PARTICIPATING",
-    "63": "NON_PARTICIPATING",
-    "64": "NOT_ALIVE",
-    "65": "NON_PARTICIPATING",
-    "66": "NON_PARTICIPATING",
+    "63": "NOT_ALIVE",
+    "64": "FAIL",
+    "65": "NOT_ALIVE",
+    "66": "NOT_ALIVE",
     "67": "NON_PARTICIPATING",
     "68": "NON_PARTICIPATING",
-    "69": "NOT_ALIVE",
-    "7": "NON_PARTICIPATING",
-    "70": "NON_PARTICIPATING",
+    "69": "NON_PARTICIPATING",
+    "7": "FAIL",
+    "70": "NOT_ALIVE",
     "71": "NOT_ALIVE",
     "72": "NON_PARTICIPATING",
     "73": "NON_PARTICIPATING",
@@ -233,30 +233,30 @@
     "75": "NOT_ALIVE",
     "76": "NON_PARTICIPATING",
     "77": "NOT_ALIVE",
-    "78": "NON_PARTICIPATING",
+    "78": "FAIL",
     "79": "NON_PARTICIPATING",
-    "8": "NOT_ALIVE",
+    "8": "FAIL",
     "80": "FAIL",
     "81": "NOT_ALIVE",
     "82": "NOT_ALIVE",
     "83": "NOT_ALIVE",
-    "84": "FAIL",
-    "85": "NOT_ALIVE",
+    "84": "NOT_ALIVE",
+    "85": "NON_PARTICIPATING",
     "86": "NOT_ALIVE",
     "87": "NON_PARTICIPATING",
-    "88": "NON_PARTICIPATING",
-    "89": "NON_PARTICIPATING",
-    "9": "NON_PARTICIPATING",
+    "88": "NOT_ALIVE",
+    "89": "FAIL",
+    "9": "NOT_ALIVE",
     "90": "FAIL",
     "91": "FAIL",
-    "92": "NON_PARTICIPATING",
+    "92": "FAIL",
     "93": "NOT_ALIVE",
     "94": "FAIL",
     "95": "NOT_ALIVE",
-    "96": "NON_PARTICIPATING",
+    "96": "NOT_ALIVE",
     "97": "NOT_ALIVE",
     "98": "NOT_ALIVE",
-    "99": "FAIL"
+    "99": "NON_PARTICIPATING"
   },
   "architectures": [
     "LlamaForCausalLM"
@@ -264,18 +264,18 @@
   "attention_bias": false,
   "attention_dropout": 0.0,
   "block_list": [
-    6460741,
-    6460755,
-    6460769
+    6460812,
+    6460828,
+    6460844
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 67,
+  "inner_step": 4,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6460132,
+  "last_allreduce_block": 6460778,
   "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:dd0cb1bbe0651b0594349ce04d94ddc42ac0c67ff5f9d93705b3def5c8e618ad
+oid sha256:8994430f5779d82c8e6a761c160eb2d9d269842ab7ec60ce6ff569c569384147
 size 88389626

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5700754ef45adf87daa862b92b9caa3f47fcee16386569d8bec8f2bbd24832d2
+oid sha256:c1b4e0f242de93461e87bfb0607707833dde9b54136c53c5578de93c395f101c
 size 4924926906

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ff8cec7ebbe32618410a8ba81b3b424cb7182c626d831f2408c2253f09c8759a
+oid sha256:dd49d45ee3d4d2b56b3c88cecf004f4ce6b768b89af224ded8a94a822848a3f5
 size 4400216536