Run 10. Outer Step 17. Inner Step 5.

config.json

@@ -2,26 +2,26 @@
   "_name_or_path": "distributed/llama-1b",
   "all_reduce_scores": {
     "0": "NOT_ALIVE",
-    "1": "NOT_ALIVE",
+    "1": "SUCCESS",
     "10": "NON_PARTICIPATING",
     "100": "NOT_ALIVE",
     "101": "NOT_ALIVE",
-    "102": "FAIL",
-    "103": "FAIL",
+    "102": "SUCCESS",
+    "103": "SUCCESS",
     "104": "NOT_ALIVE",
     "105": "FAIL",
     "106": "NON_PARTICIPATING",
     "107": "NON_PARTICIPATING",
     "108": "NOT_ALIVE",
     "109": "NOT_ALIVE",
-    "11": "NON_PARTICIPATING",
+    "11": "FAIL",
     "110": "NOT_ALIVE",
-    "111": "FAIL",
-    "112": "FAIL",
-    "113": "NOT_ALIVE",
-    "114": "NOT_ALIVE",
-    "115": "NOT_ALIVE",
-    "116": "FAIL",
+    "111": "SUCCESS",
+    "112": "SUCCESS",
+    "113": "SUCCESS",
+    "114": "SUCCESS",
+    "115": "NON_PARTICIPATING",
+    "116": "NOT_ALIVE",
     "117": "NON_PARTICIPATING",
     "118": "NON_PARTICIPATING",
     "119": "FAIL",
@@ -29,19 +29,19 @@
     "120": "NOT_ALIVE",
     "121": "NOT_ALIVE",
     "122": "NON_PARTICIPATING",
-    "123": "FAIL",
-    "124": "NOT_ALIVE",
-    "125": "NON_PARTICIPATING",
-    "126": "FAIL",
-    "127": "NON_PARTICIPATING",
+    "123": "NON_PARTICIPATING",
+    "124": "NON_PARTICIPATING",
+    "125": "SUCCESS",
+    "126": "SUCCESS",
+    "127": "NOT_ALIVE",
     "128": "NON_PARTICIPATING",
     "129": "NOT_ALIVE",
-    "13": "FAIL",
+    "13": "NON_PARTICIPATING",
     "130": "NOT_ALIVE",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
     "133": "NOT_ALIVE",
-    "134": "NOT_ALIVE",
+    "134": "NON_PARTICIPATING",
     "135": "NON_PARTICIPATING",
     "136": "NOT_ALIVE",
     "137": "NOT_ALIVE",
@@ -50,22 +50,22 @@
     "14": "NOT_ALIVE",
     "140": "NOT_ALIVE",
     "141": "NOT_ALIVE",
-    "142": "NON_PARTICIPATING",
+    "142": "NOT_ALIVE",
     "143": "NOT_ALIVE",
     "144": "NON_PARTICIPATING",
     "145": "NOT_ALIVE",
-    "146": "FAIL",
+    "146": "SUCCESS",
     "147": "NON_PARTICIPATING",
     "148": "NON_PARTICIPATING",
-    "149": "NOT_ALIVE",
+    "149": "NON_PARTICIPATING",
     "15": "NOT_ALIVE",
-    "150": "FAIL",
+    "150": "SUCCESS",
     "151": "NOT_ALIVE",
     "152": "NOT_ALIVE",
-    "153": "FAIL",
-    "154": "FAIL",
+    "153": "SUCCESS",
+    "154": "NOT_ALIVE",
     "155": "NOT_ALIVE",
-    "156": "FAIL",
+    "156": "SUCCESS",
     "157": "NON_PARTICIPATING",
     "158": "NON_PARTICIPATING",
     "159": "NON_PARTICIPATING",
@@ -73,98 +73,98 @@
     "160": "NOT_ALIVE",
     "161": "NOT_ALIVE",
     "162": "NON_PARTICIPATING",
-    "163": "NOT_ALIVE",
+    "163": "SUCCESS",
     "164": "NOT_ALIVE",
     "165": "NOT_ALIVE",
     "166": "NOT_ALIVE",
-    "167": "NOT_ALIVE",
-    "168": "FAIL",
-    "169": "FAIL",
+    "167": "NON_PARTICIPATING",
+    "168": "SUCCESS",
+    "169": "SUCCESS",
     "17": "NON_PARTICIPATING",
     "170": "FAIL",
     "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
-    "173": "FAIL",
+    "173": "SUCCESS",
     "174": "NOT_ALIVE",
     "175": "NON_PARTICIPATING",
     "176": "NOT_ALIVE",
-    "177": "FAIL",
+    "177": "SUCCESS",
     "178": "NOT_ALIVE",
     "179": "NOT_ALIVE",
-    "18": "NON_PARTICIPATING",
+    "18": "SUCCESS",
     "180": "NON_PARTICIPATING",
     "181": "NOT_ALIVE",
     "182": "NON_PARTICIPATING",
-    "183": "NOT_ALIVE",
-    "184": "FAIL",
-    "185": "FAIL",
-    "186": "NOT_ALIVE",
+    "183": "NON_PARTICIPATING",
+    "184": "NOT_ALIVE",
+    "185": "NON_PARTICIPATING",
+    "186": "FAIL",
     "187": "NOT_ALIVE",
-    "188": "FAIL",
+    "188": "NON_PARTICIPATING",
     "189": "NOT_ALIVE",
     "19": "NON_PARTICIPATING",
     "190": "NON_PARTICIPATING",
     "191": "NOT_ALIVE",
-    "192": "NOT_ALIVE",
-    "193": "FAIL",
+    "192": "NON_PARTICIPATING",
+    "193": "NON_PARTICIPATING",
     "194": "NOT_ALIVE",
-    "195": "NOT_ALIVE",
-    "196": "FAIL",
-    "197": "NON_PARTICIPATING",
-    "198": "FAIL",
+    "195": "NON_PARTICIPATING",
+    "196": "SUCCESS",
+    "197": "NOT_ALIVE",
+    "198": "SUCCESS",
     "199": "NON_PARTICIPATING",
-    "2": "NON_PARTICIPATING",
-    "20": "FAIL",
-    "200": "NON_PARTICIPATING",
+    "2": "NOT_ALIVE",
+    "20": "SUCCESS",
+    "200": "SUCCESS",
     "201": "NOT_ALIVE",
     "202": "NOT_ALIVE",
     "203": "NOT_ALIVE",
     "204": "NOT_ALIVE",
-    "205": "FAIL",
-    "206": "NOT_ALIVE",
+    "205": "NON_PARTICIPATING",
+    "206": "SUCCESS",
     "207": "NOT_ALIVE",
     "208": "NON_PARTICIPATING",
-    "209": "NON_PARTICIPATING",
-    "21": "NON_PARTICIPATING",
+    "209": "FAIL",
+    "21": "SUCCESS",
     "210": "NON_PARTICIPATING",
-    "211": "NON_PARTICIPATING",
-    "212": "NOT_ALIVE",
+    "211": "FAIL",
+    "212": "SUCCESS",
     "213": "NON_PARTICIPATING",
-    "214": "FAIL",
+    "214": "SUCCESS",
     "215": "NOT_ALIVE",
     "216": "NOT_ALIVE",
     "217": "NOT_ALIVE",
-    "218": "NOT_ALIVE",
-    "219": "FAIL",
+    "218": "NON_PARTICIPATING",
+    "219": "SUCCESS",
     "22": "NOT_ALIVE",
     "220": "NON_PARTICIPATING",
-    "221": "FAIL",
+    "221": "NON_PARTICIPATING",
     "222": "FAIL",
-    "223": "NOT_ALIVE",
+    "223": "NON_PARTICIPATING",
     "224": "NOT_ALIVE",
-    "225": "NOT_ALIVE",
-    "226": "NOT_ALIVE",
-    "227": "FAIL",
+    "225": "NON_PARTICIPATING",
+    "226": "NON_PARTICIPATING",
+    "227": "SUCCESS",
     "228": "NOT_ALIVE",
-    "229": "NON_PARTICIPATING",
+    "229": "SUCCESS",
     "23": "NON_PARTICIPATING",
     "230": "NOT_ALIVE",
     "231": "NOT_ALIVE",
-    "232": "NON_PARTICIPATING",
-    "233": "FAIL",
+    "232": "SUCCESS",
+    "233": "NON_PARTICIPATING",
     "234": "NON_PARTICIPATING",
-    "235": "NOT_ALIVE",
-    "236": "NON_PARTICIPATING",
-    "237": "FAIL",
-    "238": "NON_PARTICIPATING",
+    "235": "NON_PARTICIPATING",
+    "236": "SUCCESS",
+    "237": "SUCCESS",
+    "238": "NOT_ALIVE",
     "239": "NON_PARTICIPATING",
     "24": "NOT_ALIVE",
     "240": "NOT_ALIVE",
-    "241": "FAIL",
+    "241": "NOT_ALIVE",
     "242": "NON_PARTICIPATING",
     "243": "NOT_ALIVE",
     "244": "NOT_ALIVE",
-    "245": "FAIL",
+    "245": "SUCCESS",
     "246": "NON_PARTICIPATING",
     "247": "NOT_ALIVE",
     "248": "NON_PARTICIPATING",
@@ -172,91 +172,91 @@
     "25": "SUCCESS",
     "250": "NON_PARTICIPATING",
     "251": "NOT_ALIVE",
-    "252": "NON_PARTICIPATING",
+    "252": "SUCCESS",
     "253": "NOT_ALIVE",
-    "254": "NON_PARTICIPATING",
+    "254": "SUCCESS",
     "255": "NON_PARTICIPATING",
-    "26": "FAIL",
+    "26": "NON_PARTICIPATING",
     "27": "NOT_ALIVE",
-    "28": "FAIL",
-    "29": "FAIL",
-    "3": "NON_PARTICIPATING",
+    "28": "NON_PARTICIPATING",
+    "29": "SUCCESS",
+    "3": "FAIL",
     "30": "NON_PARTICIPATING",
     "31": "NOT_ALIVE",
     "32": "NOT_ALIVE",
     "33": "NOT_ALIVE",
     "34": "NOT_ALIVE",
-    "35": "FAIL",
-    "36": "FAIL",
+    "35": "SUCCESS",
+    "36": "NON_PARTICIPATING",
     "37": "NOT_ALIVE",
     "38": "NON_PARTICIPATING",
-    "39": "FAIL",
-    "4": "FAIL",
-    "40": "FAIL",
+    "39": "NON_PARTICIPATING",
+    "4": "NON_PARTICIPATING",
+    "40": "NON_PARTICIPATING",
     "41": "NOT_ALIVE",
-    "42": "NOT_ALIVE",
+    "42": "NON_PARTICIPATING",
     "43": "NOT_ALIVE",
-    "44": "FAIL",
-    "45": "NON_PARTICIPATING",
-    "46": "FAIL",
-    "47": "FAIL",
+    "44": "NON_PARTICIPATING",
+    "45": "FAIL",
+    "46": "SUCCESS",
+    "47": "NOT_ALIVE",
     "48": "NON_PARTICIPATING",
     "49": "NOT_ALIVE",
     "5": "NON_PARTICIPATING",
     "50": "NON_PARTICIPATING",
-    "51": "NON_PARTICIPATING",
+    "51": "NOT_ALIVE",
     "52": "NON_PARTICIPATING",
-    "53": "NOT_ALIVE",
+    "53": "NON_PARTICIPATING",
     "54": "NOT_ALIVE",
     "55": "NOT_ALIVE",
-    "56": "NON_PARTICIPATING",
-    "57": "FAIL",
+    "56": "SUCCESS",
+    "57": "SUCCESS",
     "58": "NON_PARTICIPATING",
-    "59": "NON_PARTICIPATING",
+    "59": "NOT_ALIVE",
     "6": "NOT_ALIVE",
-    "60": "NON_PARTICIPATING",
+    "60": "SUCCESS",
     "61": "NOT_ALIVE",
-    "62": "NON_PARTICIPATING",
+    "62": "SUCCESS",
     "63": "NON_PARTICIPATING",
-    "64": "FAIL",
+    "64": "NOT_ALIVE",
     "65": "NOT_ALIVE",
     "66": "NOT_ALIVE",
     "67": "NON_PARTICIPATING",
     "68": "NOT_ALIVE",
-    "69": "FAIL",
-    "7": "FAIL",
+    "69": "SUCCESS",
+    "7": "SUCCESS",
     "70": "NOT_ALIVE",
     "71": "NOT_ALIVE",
     "72": "NON_PARTICIPATING",
     "73": "NOT_ALIVE",
-    "74": "FAIL",
+    "74": "NON_PARTICIPATING",
     "75": "NOT_ALIVE",
     "76": "NON_PARTICIPATING",
     "77": "NOT_ALIVE",
-    "78": "NON_PARTICIPATING",
+    "78": "NOT_ALIVE",
     "79": "NON_PARTICIPATING",
-    "8": "FAIL",
+    "8": "SUCCESS",
     "80": "NOT_ALIVE",
     "81": "NOT_ALIVE",
     "82": "NOT_ALIVE",
-    "83": "NOT_ALIVE",
-    "84": "FAIL",
-    "85": "NON_PARTICIPATING",
-    "86": "FAIL",
-    "87": "FAIL",
+    "83": "NON_PARTICIPATING",
+    "84": "SUCCESS",
+    "85": "NOT_ALIVE",
+    "86": "NON_PARTICIPATING",
+    "87": "NOT_ALIVE",
     "88": "NON_PARTICIPATING",
     "89": "NOT_ALIVE",
-    "9": "FAIL",
-    "90": "FAIL",
-    "91": "FAIL",
-    "92": "NON_PARTICIPATING",
+    "9": "SUCCESS",
+    "90": "NON_PARTICIPATING",
+    "91": "SUCCESS",
+    "92": "SUCCESS",
     "93": "NOT_ALIVE",
     "94": "NON_PARTICIPATING",
     "95": "NOT_ALIVE",
-    "96": "FAIL",
-    "97": "NOT_ALIVE",
+    "96": "SUCCESS",
+    "97": "NON_PARTICIPATING",
     "98": "NON_PARTICIPATING",
-    "99": "NOT_ALIVE"
+    "99": "NON_PARTICIPATING"
   },
   "architectures": [
     "LlamaForCausalLM"
@@ -264,17 +264,18 @@
   "attention_bias": false,
   "attention_dropout": 0.0,
   "block_list": [
-    6476444,
-    6476444
+    6476500,
+    6476518,
+    6476520
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 98,
+  "inner_step": 4,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6476005,
+  "last_allreduce_block": 6476411,
   "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:a451b816838840c301007b036b3c340f5254b6de55e887c38519f1b2bee6a8a8
+oid sha256:2dfd0caf342cc7ce3a191469bc7c823b9ed4694104d93a605410bd3263a27231
 size 88390091

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ebbfccead8c32d0b5449c645daafeb857e0f6cefb6edb5ea29048f63e14274ca
+oid sha256:01090723268bfe6a0634327ff35b24ee85c8635f6cf6d8efccb6130a9ed884ea
 size 4924928111

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:966571a3bee65ce5060bea4534b00c0c4a01f47cd9d700fa0e3228e65c65dfe8
+oid sha256:ca1ddd1a8eff47768e6f8f607fe2343749ec1d4fb9a25bae20211ab46b83de85
 size 4400216536