Run 10. Outer Step 9. Inner Step 5.

config.json

@@ -2,70 +2,70 @@
   "_name_or_path": "distributed/llama-1b",
   "all_reduce_scores": {
     "0": "NOT_ALIVE",
-    "1": "NOT_ALIVE",
-    "10": "NON_PARTICIPATING",
+    "1": "FAIL",
+    "10": "NOT_ALIVE",
     "100": "NOT_ALIVE",
     "101": "NOT_ALIVE",
-    "102": "FAIL",
+    "102": "SUCCESS",
     "103": "NOT_ALIVE",
     "104": "NOT_ALIVE",
-    "105": "FAIL",
-    "106": "FAIL",
-    "107": "NOT_ALIVE",
+    "105": "NOT_ALIVE",
+    "106": "NOT_ALIVE",
+    "107": "SUCCESS",
     "108": "NOT_ALIVE",
     "109": "NOT_ALIVE",
     "11": "NON_PARTICIPATING",
     "110": "NON_PARTICIPATING",
-    "111": "NON_PARTICIPATING",
-    "112": "NON_PARTICIPATING",
-    "113": "FAIL",
+    "111": "NOT_ALIVE",
+    "112": "NOT_ALIVE",
+    "113": "SUCCESS",
     "114": "NOT_ALIVE",
-    "115": "FAIL",
-    "116": "NON_PARTICIPATING",
-    "117": "NON_PARTICIPATING",
+    "115": "NON_PARTICIPATING",
+    "116": "SUCCESS",
+    "117": "FAIL",
     "118": "NON_PARTICIPATING",
-    "119": "FAIL",
+    "119": "NOT_ALIVE",
     "12": "NOT_ALIVE",
     "120": "NON_PARTICIPATING",
     "121": "NOT_ALIVE",
     "122": "NON_PARTICIPATING",
     "123": "FAIL",
-    "124": "FAIL",
-    "125": "NON_PARTICIPATING",
-    "126": "NON_PARTICIPATING",
-    "127": "FAIL",
-    "128": "NON_PARTICIPATING",
-    "129": "NOT_ALIVE",
-    "13": "FAIL",
+    "124": "NON_PARTICIPATING",
+    "125": "SUCCESS",
+    "126": "NOT_ALIVE",
+    "127": "NON_PARTICIPATING",
+    "128": "NOT_ALIVE",
+    "129": "NON_PARTICIPATING",
+    "13": "NON_PARTICIPATING",
     "130": "NON_PARTICIPATING",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
-    "133": "SUCCESS",
-    "134": "NON_PARTICIPATING",
-    "135": "FAIL",
+    "133": "NOT_ALIVE",
+    "134": "SUCCESS",
+    "135": "NON_PARTICIPATING",
     "136": "NOT_ALIVE",
     "137": "NOT_ALIVE",
     "138": "NON_PARTICIPATING",
     "139": "NOT_ALIVE",
-    "14": "FAIL",
+    "14": "NOT_ALIVE",
     "140": "NOT_ALIVE",
     "141": "NOT_ALIVE",
-    "142": "NON_PARTICIPATING",
+    "142": "NOT_ALIVE",
     "143": "NOT_ALIVE",
-    "144": "NON_PARTICIPATING",
+    "144": "SUCCESS",
     "145": "NOT_ALIVE",
-    "146": "NON_PARTICIPATING",
-    "147": "FAIL",
-    "148": "FAIL",
-    "149": "NON_PARTICIPATING",
+    "146": "NOT_ALIVE",
+    "147": "SUCCESS",
+    "148": "SUCCESS",
+    "149": "SUCCESS",
     "15": "NOT_ALIVE",
-    "150": "NOT_ALIVE",
+    "150": "NON_PARTICIPATING",
     "151": "NOT_ALIVE",
     "152": "NOT_ALIVE",
-    "153": "NON_PARTICIPATING",
+    "153": "FAIL",
     "154": "NOT_ALIVE",
-    "155": "NON_PARTICIPATING",
-    "156": "NON_PARTICIPATING",
+    "155": "NOT_ALIVE",
+    "156": "NOT_ALIVE",
     "157": "NON_PARTICIPATING",
     "158": "NON_PARTICIPATING",
     "159": "NON_PARTICIPATING",
@@ -73,90 +73,90 @@
     "160": "NOT_ALIVE",
     "161": "NOT_ALIVE",
     "162": "NON_PARTICIPATING",
-    "163": "NON_PARTICIPATING",
+    "163": "SUCCESS",
     "164": "NOT_ALIVE",
     "165": "NON_PARTICIPATING",
     "166": "NOT_ALIVE",
-    "167": "NON_PARTICIPATING",
-    "168": "NON_PARTICIPATING",
-    "169": "NON_PARTICIPATING",
-    "17": "NON_PARTICIPATING",
-    "170": "FAIL",
+    "167": "NOT_ALIVE",
+    "168": "SUCCESS",
+    "169": "NOT_ALIVE",
+    "17": "NOT_ALIVE",
+    "170": "SUCCESS",
     "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
-    "173": "NOT_ALIVE",
-    "174": "NON_PARTICIPATING",
+    "173": "SUCCESS",
+    "174": "NOT_ALIVE",
     "175": "NON_PARTICIPATING",
     "176": "NOT_ALIVE",
     "177": "FAIL",
     "178": "NOT_ALIVE",
-    "179": "NOT_ALIVE",
-    "18": "NON_PARTICIPATING",
-    "180": "FAIL",
+    "179": "NON_PARTICIPATING",
+    "18": "FAIL",
+    "180": "NON_PARTICIPATING",
     "181": "NOT_ALIVE",
     "182": "NON_PARTICIPATING",
-    "183": "NOT_ALIVE",
-    "184": "FAIL",
-    "185": "FAIL",
+    "183": "SUCCESS",
+    "184": "SUCCESS",
+    "185": "SUCCESS",
     "186": "SUCCESS",
     "187": "NOT_ALIVE",
-    "188": "FAIL",
+    "188": "SUCCESS",
     "189": "NOT_ALIVE",
     "19": "NON_PARTICIPATING",
-    "190": "FAIL",
-    "191": "NON_PARTICIPATING",
+    "190": "NON_PARTICIPATING",
+    "191": "NOT_ALIVE",
     "192": "NON_PARTICIPATING",
     "193": "NON_PARTICIPATING",
-    "194": "NOT_ALIVE",
-    "195": "NON_PARTICIPATING",
-    "196": "FAIL",
+    "194": "NON_PARTICIPATING",
+    "195": "NOT_ALIVE",
+    "196": "SUCCESS",
     "197": "NON_PARTICIPATING",
-    "198": "NON_PARTICIPATING",
-    "199": "FAIL",
-    "2": "FAIL",
-    "20": "NON_PARTICIPATING",
-    "200": "FAIL",
+    "198": "NOT_ALIVE",
+    "199": "NOT_ALIVE",
+    "2": "NON_PARTICIPATING",
+    "20": "NOT_ALIVE",
+    "200": "SUCCESS",
     "201": "NON_PARTICIPATING",
     "202": "NOT_ALIVE",
     "203": "NOT_ALIVE",
     "204": "NOT_ALIVE",
-    "205": "NOT_ALIVE",
-    "206": "FAIL",
+    "205": "SUCCESS",
+    "206": "SUCCESS",
     "207": "NOT_ALIVE",
-    "208": "NON_PARTICIPATING",
-    "209": "FAIL",
-    "21": "FAIL",
+    "208": "SUCCESS",
+    "209": "SUCCESS",
+    "21": "SUCCESS",
     "210": "NON_PARTICIPATING",
-    "211": "NON_PARTICIPATING",
-    "212": "FAIL",
-    "213": "NON_PARTICIPATING",
+    "211": "NOT_ALIVE",
+    "212": "SUCCESS",
+    "213": "NOT_ALIVE",
     "214": "NOT_ALIVE",
     "215": "NOT_ALIVE",
     "216": "NOT_ALIVE",
     "217": "NOT_ALIVE",
-    "218": "NOT_ALIVE",
-    "219": "SUCCESS",
+    "218": "NON_PARTICIPATING",
+    "219": "NOT_ALIVE",
     "22": "NOT_ALIVE",
     "220": "NOT_ALIVE",
     "221": "NON_PARTICIPATING",
-    "222": "NOT_ALIVE",
-    "223": "NON_PARTICIPATING",
+    "222": "SUCCESS",
+    "223": "NOT_ALIVE",
     "224": "NOT_ALIVE",
     "225": "NON_PARTICIPATING",
-    "226": "FAIL",
-    "227": "FAIL",
+    "226": "NOT_ALIVE",
+    "227": "SUCCESS",
     "228": "NOT_ALIVE",
     "229": "NON_PARTICIPATING",
     "23": "NOT_ALIVE",
-    "230": "NON_PARTICIPATING",
+    "230": "NOT_ALIVE",
     "231": "NOT_ALIVE",
-    "232": "NOT_ALIVE",
-    "233": "NON_PARTICIPATING",
+    "232": "SUCCESS",
+    "233": "NOT_ALIVE",
     "234": "NON_PARTICIPATING",
-    "235": "NON_PARTICIPATING",
-    "236": "NON_PARTICIPATING",
-    "237": "NON_PARTICIPATING",
-    "238": "NON_PARTICIPATING",
+    "235": "SUCCESS",
+    "236": "NOT_ALIVE",
+    "237": "SUCCESS",
+    "238": "SUCCESS",
     "239": "NON_PARTICIPATING",
     "24": "NON_PARTICIPATING",
     "240": "NOT_ALIVE",
@@ -164,99 +164,99 @@
     "242": "NON_PARTICIPATING",
     "243": "NOT_ALIVE",
     "244": "NOT_ALIVE",
-    "245": "FAIL",
+    "245": "SUCCESS",
     "246": "NON_PARTICIPATING",
     "247": "NON_PARTICIPATING",
-    "248": "FAIL",
+    "248": "SUCCESS",
     "249": "NOT_ALIVE",
     "25": "SUCCESS",
-    "250": "FAIL",
+    "250": "NOT_ALIVE",
     "251": "NOT_ALIVE",
-    "252": "FAIL",
-    "253": "FAIL",
+    "252": "NON_PARTICIPATING",
+    "253": "NOT_ALIVE",
     "254": "NOT_ALIVE",
     "255": "NON_PARTICIPATING",
-    "26": "FAIL",
+    "26": "NOT_ALIVE",
     "27": "NOT_ALIVE",
-    "28": "NON_PARTICIPATING",
-    "29": "NON_PARTICIPATING",
-    "3": "NON_PARTICIPATING",
+    "28": "SUCCESS",
+    "29": "NOT_ALIVE",
+    "3": "NOT_ALIVE",
     "30": "NON_PARTICIPATING",
     "31": "NOT_ALIVE",
     "32": "NON_PARTICIPATING",
     "33": "NOT_ALIVE",
-    "34": "NON_PARTICIPATING",
-    "35": "FAIL",
-    "36": "SUCCESS",
+    "34": "NOT_ALIVE",
+    "35": "NOT_ALIVE",
+    "36": "NOT_ALIVE",
     "37": "NOT_ALIVE",
-    "38": "NOT_ALIVE",
-    "39": "NON_PARTICIPATING",
+    "38": "NON_PARTICIPATING",
+    "39": "NOT_ALIVE",
     "4": "SUCCESS",
     "40": "NON_PARTICIPATING",
-    "41": "FAIL",
-    "42": "SUCCESS",
+    "41": "SUCCESS",
+    "42": "NON_PARTICIPATING",
     "43": "NOT_ALIVE",
-    "44": "FAIL",
-    "45": "NOT_ALIVE",
+    "44": "SUCCESS",
+    "45": "SUCCESS",
     "46": "NOT_ALIVE",
-    "47": "NON_PARTICIPATING",
+    "47": "NOT_ALIVE",
     "48": "NON_PARTICIPATING",
     "49": "NOT_ALIVE",
-    "5": "NON_PARTICIPATING",
-    "50": "NOT_ALIVE",
-    "51": "FAIL",
-    "52": "NOT_ALIVE",
-    "53": "FAIL",
+    "5": "SUCCESS",
+    "50": "NON_PARTICIPATING",
+    "51": "SUCCESS",
+    "52": "NON_PARTICIPATING",
+    "53": "SUCCESS",
     "54": "NOT_ALIVE",
     "55": "NOT_ALIVE",
     "56": "NOT_ALIVE",
     "57": "FAIL",
-    "58": "NOT_ALIVE",
-    "59": "NOT_ALIVE",
+    "58": "SUCCESS",
+    "59": "NON_PARTICIPATING",
     "6": "NOT_ALIVE",
-    "60": "NOT_ALIVE",
-    "61": "NON_PARTICIPATING",
+    "60": "SUCCESS",
+    "61": "NOT_ALIVE",
     "62": "NON_PARTICIPATING",
-    "63": "NON_PARTICIPATING",
-    "64": "FAIL",
+    "63": "SUCCESS",
+    "64": "SUCCESS",
     "65": "NOT_ALIVE",
-    "66": "SUCCESS",
-    "67": "NON_PARTICIPATING",
+    "66": "NOT_ALIVE",
+    "67": "NOT_ALIVE",
     "68": "NOT_ALIVE",
-    "69": "FAIL",
-    "7": "NOT_ALIVE",
+    "69": "SUCCESS",
+    "7": "NON_PARTICIPATING",
     "70": "NOT_ALIVE",
     "71": "NOT_ALIVE",
-    "72": "NON_PARTICIPATING",
+    "72": "NOT_ALIVE",
     "73": "NON_PARTICIPATING",
-    "74": "FAIL",
+    "74": "NOT_ALIVE",
     "75": "NON_PARTICIPATING",
-    "76": "NON_PARTICIPATING",
+    "76": "NOT_ALIVE",
     "77": "NOT_ALIVE",
     "78": "NON_PARTICIPATING",
     "79": "NON_PARTICIPATING",
-    "8": "NOT_ALIVE",
-    "80": "NOT_ALIVE",
+    "8": "SUCCESS",
+    "80": "NON_PARTICIPATING",
     "81": "NOT_ALIVE",
     "82": "NOT_ALIVE",
     "83": "NON_PARTICIPATING",
-    "84": "FAIL",
+    "84": "SUCCESS",
     "85": "NON_PARTICIPATING",
     "86": "NOT_ALIVE",
-    "87": "NON_PARTICIPATING",
+    "87": "SUCCESS",
     "88": "NON_PARTICIPATING",
     "89": "FAIL",
     "9": "SUCCESS",
-    "90": "NON_PARTICIPATING",
-    "91": "NON_PARTICIPATING",
+    "90": "SUCCESS",
+    "91": "NOT_ALIVE",
     "92": "SUCCESS",
     "93": "NOT_ALIVE",
-    "94": "FAIL",
+    "94": "NOT_ALIVE",
     "95": "NOT_ALIVE",
-    "96": "NON_PARTICIPATING",
+    "96": "NOT_ALIVE",
     "97": "NON_PARTICIPATING",
     "98": "NON_PARTICIPATING",
-    "99": "NON_PARTICIPATING"
+    "99": "SUCCESS"
   },
   "architectures": [
     "LlamaForCausalLM"
@@ -264,17 +264,17 @@
   "attention_bias": false,
   "attention_dropout": 0.0,
   "block_list": [
-    6470600,
-    6470606
+    6471511,
+    6471521
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 347,
+  "inner_step": 5,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6468297,
+  "last_allreduce_block": 6471308,
   "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:b2ebee6d915c473eeb4d705fafb6cc16dc5005fa42aeaaaf97da330e7ba80ebd
+oid sha256:f624e0e639b8460de92850aeadf3d81752d870a2194a7d612c3d3bd1144912a3
 size 88389626

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cca9418594a75a5e525a2302eb0f72f22869996bf18dbf6b0111fe07fc256821
+oid sha256:52452432a3967369a31c00f7e3994e8fed9bfe64faad284319ad3b7037ce8d34
 size 4924926906

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7db94844dc21232627e72032023671980214ce0e4cef79ab4d8c22acc9a3cb26
+oid sha256:0c28418713b12548df44d2ad172f0f0c884ef316b166333881ee2111f4a0ac1b
 size 4400216536