Run 10. Outer Step 20. Inner Step 4.

config.json

@@ -2,47 +2,47 @@
   "_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",
+    "102": "FAIL",
     "103": "FAIL",
     "104": "NOT_ALIVE",
     "105": "NOT_ALIVE",
     "106": "NON_PARTICIPATING",
     "107": "NON_PARTICIPATING",
     "108": "NOT_ALIVE",
-    "109": "NON_PARTICIPATING",
-    "11": "FAIL",
+    "109": "FAIL",
+    "11": "NON_PARTICIPATING",
     "110": "NOT_ALIVE",
-    "111": "NOT_ALIVE",
-    "112": "FAIL",
+    "111": "SUCCESS",
+    "112": "NON_PARTICIPATING",
     "113": "NOT_ALIVE",
-    "114": "SUCCESS",
+    "114": "FAIL",
     "115": "NON_PARTICIPATING",
-    "116": "SUCCESS",
+    "116": "NON_PARTICIPATING",
     "117": "NON_PARTICIPATING",
-    "118": "NOT_ALIVE",
+    "118": "NON_PARTICIPATING",
     "119": "NON_PARTICIPATING",
     "12": "NOT_ALIVE",
-    "120": "NOT_ALIVE",
+    "120": "NON_PARTICIPATING",
     "121": "NOT_ALIVE",
-    "122": "NON_PARTICIPATING",
+    "122": "NOT_ALIVE",
     "123": "FAIL",
     "124": "NOT_ALIVE",
     "125": "NOT_ALIVE",
-    "126": "NOT_ALIVE",
+    "126": "FAIL",
     "127": "NON_PARTICIPATING",
-    "128": "NON_PARTICIPATING",
+    "128": "NOT_ALIVE",
     "129": "NON_PARTICIPATING",
-    "13": "NON_PARTICIPATING",
+    "13": "NOT_ALIVE",
     "130": "NOT_ALIVE",
     "131": "NOT_ALIVE",
     "132": "NOT_ALIVE",
     "133": "NON_PARTICIPATING",
-    "134": "SUCCESS",
-    "135": "NON_PARTICIPATING",
+    "134": "NOT_ALIVE",
+    "135": "NOT_ALIVE",
     "136": "NOT_ALIVE",
     "137": "NOT_ALIVE",
     "138": "NON_PARTICIPATING",
@@ -50,186 +50,186 @@
     "14": "NON_PARTICIPATING",
     "140": "NOT_ALIVE",
     "141": "NOT_ALIVE",
-    "142": "NOT_ALIVE",
-    "143": "FAIL",
-    "144": "NOT_ALIVE",
+    "142": "FAIL",
+    "143": "NON_PARTICIPATING",
+    "144": "NON_PARTICIPATING",
     "145": "NOT_ALIVE",
-    "146": "NON_PARTICIPATING",
-    "147": "NON_PARTICIPATING",
+    "146": "FAIL",
+    "147": "NOT_ALIVE",
     "148": "NON_PARTICIPATING",
-    "149": "NOT_ALIVE",
-    "15": "NON_PARTICIPATING",
+    "149": "NON_PARTICIPATING",
+    "15": "FAIL",
     "150": "NON_PARTICIPATING",
     "151": "NOT_ALIVE",
     "152": "NON_PARTICIPATING",
     "153": "FAIL",
-    "154": "SUCCESS",
-    "155": "FAIL",
-    "156": "NON_PARTICIPATING",
+    "154": "NON_PARTICIPATING",
+    "155": "NOT_ALIVE",
+    "156": "FAIL",
     "157": "NON_PARTICIPATING",
-    "158": "NON_PARTICIPATING",
+    "158": "NOT_ALIVE",
     "159": "NOT_ALIVE",
     "16": "NOT_ALIVE",
-    "160": "FAIL",
+    "160": "NON_PARTICIPATING",
     "161": "NOT_ALIVE",
-    "162": "NOT_ALIVE",
-    "163": "SUCCESS",
+    "162": "NON_PARTICIPATING",
+    "163": "FAIL",
     "164": "NOT_ALIVE",
-    "165": "NOT_ALIVE",
+    "165": "NON_PARTICIPATING",
     "166": "NOT_ALIVE",
-    "167": "FAIL",
+    "167": "NON_PARTICIPATING",
     "168": "FAIL",
-    "169": "NOT_ALIVE",
-    "17": "NON_PARTICIPATING",
-    "170": "NON_PARTICIPATING",
+    "169": "FAIL",
+    "17": "NOT_ALIVE",
+    "170": "FAIL",
     "171": "NOT_ALIVE",
     "172": "NOT_ALIVE",
     "173": "NON_PARTICIPATING",
     "174": "NON_PARTICIPATING",
-    "175": "FAIL",
+    "175": "NON_PARTICIPATING",
     "176": "NOT_ALIVE",
-    "177": "NOT_ALIVE",
+    "177": "NON_PARTICIPATING",
     "178": "NOT_ALIVE",
     "179": "NON_PARTICIPATING",
-    "18": "FAIL",
+    "18": "NON_PARTICIPATING",
     "180": "NON_PARTICIPATING",
     "181": "NOT_ALIVE",
-    "182": "NON_PARTICIPATING",
-    "183": "NON_PARTICIPATING",
-    "184": "NON_PARTICIPATING",
+    "182": "NOT_ALIVE",
+    "183": "FAIL",
+    "184": "FAIL",
     "185": "NOT_ALIVE",
-    "186": "FAIL",
+    "186": "SUCCESS",
     "187": "NOT_ALIVE",
-    "188": "FAIL",
+    "188": "NOT_ALIVE",
     "189": "NOT_ALIVE",
     "19": "NON_PARTICIPATING",
-    "190": "NON_PARTICIPATING",
+    "190": "FAIL",
     "191": "NOT_ALIVE",
     "192": "NON_PARTICIPATING",
-    "193": "NON_PARTICIPATING",
+    "193": "NOT_ALIVE",
     "194": "NOT_ALIVE",
-    "195": "FAIL",
-    "196": "NON_PARTICIPATING",
+    "195": "NON_PARTICIPATING",
+    "196": "FAIL",
     "197": "NON_PARTICIPATING",
-    "198": "NON_PARTICIPATING",
+    "198": "FAIL",
     "199": "NON_PARTICIPATING",
-    "2": "NON_PARTICIPATING",
-    "20": "NON_PARTICIPATING",
-    "200": "NON_PARTICIPATING",
-    "201": "FAIL",
+    "2": "FAIL",
+    "20": "FAIL",
+    "200": "FAIL",
+    "201": "NON_PARTICIPATING",
     "202": "NOT_ALIVE",
-    "203": "NOT_ALIVE",
+    "203": "NON_PARTICIPATING",
     "204": "NOT_ALIVE",
-    "205": "NON_PARTICIPATING",
+    "205": "FAIL",
     "206": "FAIL",
     "207": "NOT_ALIVE",
-    "208": "NOT_ALIVE",
-    "209": "FAIL",
-    "21": "NON_PARTICIPATING",
+    "208": "NON_PARTICIPATING",
+    "209": "NON_PARTICIPATING",
+    "21": "NOT_ALIVE",
     "210": "NON_PARTICIPATING",
-    "211": "FAIL",
-    "212": "NON_PARTICIPATING",
+    "211": "NON_PARTICIPATING",
+    "212": "FAIL",
     "213": "SUCCESS",
-    "214": "NON_PARTICIPATING",
-    "215": "NOT_ALIVE",
+    "214": "NOT_ALIVE",
+    "215": "NON_PARTICIPATING",
     "216": "NOT_ALIVE",
     "217": "NOT_ALIVE",
-    "218": "FAIL",
-    "219": "NOT_ALIVE",
-    "22": "FAIL",
+    "218": "NON_PARTICIPATING",
+    "219": "FAIL",
+    "22": "NON_PARTICIPATING",
     "220": "NOT_ALIVE",
-    "221": "NON_PARTICIPATING",
-    "222": "NON_PARTICIPATING",
-    "223": "FAIL",
+    "221": "FAIL",
+    "222": "FAIL",
+    "223": "NON_PARTICIPATING",
     "224": "NOT_ALIVE",
     "225": "NON_PARTICIPATING",
     "226": "NON_PARTICIPATING",
-    "227": "FAIL",
+    "227": "SUCCESS",
     "228": "NOT_ALIVE",
-    "229": "NON_PARTICIPATING",
-    "23": "NOT_ALIVE",
+    "229": "NOT_ALIVE",
+    "23": "NON_PARTICIPATING",
     "230": "NOT_ALIVE",
     "231": "NOT_ALIVE",
     "232": "NON_PARTICIPATING",
     "233": "FAIL",
-    "234": "NOT_ALIVE",
-    "235": "SUCCESS",
+    "234": "NON_PARTICIPATING",
+    "235": "FAIL",
     "236": "FAIL",
     "237": "FAIL",
-    "238": "NON_PARTICIPATING",
+    "238": "FAIL",
     "239": "NON_PARTICIPATING",
-    "24": "NON_PARTICIPATING",
+    "24": "FAIL",
     "240": "NOT_ALIVE",
-    "241": "SUCCESS",
-    "242": "FAIL",
+    "241": "FAIL",
+    "242": "NON_PARTICIPATING",
     "243": "NOT_ALIVE",
     "244": "NOT_ALIVE",
-    "245": "NON_PARTICIPATING",
-    "246": "FAIL",
+    "245": "FAIL",
+    "246": "NON_PARTICIPATING",
     "247": "NOT_ALIVE",
-    "248": "FAIL",
-    "249": "FAIL",
+    "248": "NON_PARTICIPATING",
+    "249": "NON_PARTICIPATING",
     "25": "SUCCESS",
     "250": "NOT_ALIVE",
     "251": "NOT_ALIVE",
     "252": "FAIL",
     "253": "NOT_ALIVE",
-    "254": "NOT_ALIVE",
-    "255": "NOT_ALIVE",
-    "26": "NOT_ALIVE",
+    "254": "FAIL",
+    "255": "NON_PARTICIPATING",
+    "26": "NON_PARTICIPATING",
     "27": "NON_PARTICIPATING",
-    "28": "FAIL",
+    "28": "NON_PARTICIPATING",
     "29": "FAIL",
-    "3": "FAIL",
+    "3": "NON_PARTICIPATING",
     "30": "NON_PARTICIPATING",
     "31": "NOT_ALIVE",
-    "32": "FAIL",
+    "32": "NOT_ALIVE",
     "33": "NOT_ALIVE",
     "34": "NON_PARTICIPATING",
     "35": "FAIL",
-    "36": "NON_PARTICIPATING",
+    "36": "FAIL",
     "37": "NOT_ALIVE",
-    "38": "FAIL",
-    "39": "NON_PARTICIPATING",
-    "4": "NON_PARTICIPATING",
-    "40": "FAIL",
+    "38": "NON_PARTICIPATING",
+    "39": "FAIL",
+    "4": "FAIL",
+    "40": "NON_PARTICIPATING",
     "41": "NOT_ALIVE",
     "42": "NOT_ALIVE",
     "43": "NOT_ALIVE",
-    "44": "NON_PARTICIPATING",
-    "45": "FAIL",
-    "46": "NOT_ALIVE",
-    "47": "FAIL",
+    "44": "FAIL",
+    "45": "NOT_ALIVE",
+    "46": "FAIL",
+    "47": "NOT_ALIVE",
     "48": "NON_PARTICIPATING",
-    "49": "NOT_ALIVE",
+    "49": "NON_PARTICIPATING",
     "5": "NON_PARTICIPATING",
     "50": "NON_PARTICIPATING",
     "51": "NON_PARTICIPATING",
     "52": "NON_PARTICIPATING",
-    "53": "NON_PARTICIPATING",
+    "53": "FAIL",
     "54": "NON_PARTICIPATING",
     "55": "NON_PARTICIPATING",
-    "56": "FAIL",
-    "57": "NON_PARTICIPATING",
-    "58": "NON_PARTICIPATING",
+    "56": "NOT_ALIVE",
+    "57": "FAIL",
+    "58": "NOT_ALIVE",
     "59": "NON_PARTICIPATING",
     "6": "NOT_ALIVE",
-    "60": "NON_PARTICIPATING",
+    "60": "FAIL",
     "61": "NOT_ALIVE",
-    "62": "SUCCESS",
+    "62": "NON_PARTICIPATING",
     "63": "NON_PARTICIPATING",
     "64": "NOT_ALIVE",
     "65": "NOT_ALIVE",
-    "66": "NOT_ALIVE",
+    "66": "NON_PARTICIPATING",
     "67": "NON_PARTICIPATING",
     "68": "NOT_ALIVE",
     "69": "FAIL",
-    "7": "SUCCESS",
+    "7": "FAIL",
     "70": "NOT_ALIVE",
     "71": "NOT_ALIVE",
     "72": "NON_PARTICIPATING",
-    "73": "NOT_ALIVE",
-    "74": "NOT_ALIVE",
+    "73": "NON_PARTICIPATING",
+    "74": "FAIL",
     "75": "NON_PARTICIPATING",
     "76": "NON_PARTICIPATING",
     "77": "NON_PARTICIPATING",
@@ -239,19 +239,19 @@
     "80": "NOT_ALIVE",
     "81": "NOT_ALIVE",
     "82": "NOT_ALIVE",
-    "83": "NON_PARTICIPATING",
-    "84": "NOT_ALIVE",
-    "85": "NOT_ALIVE",
-    "86": "FAIL",
-    "87": "SUCCESS",
-    "88": "FAIL",
+    "83": "NOT_ALIVE",
+    "84": "FAIL",
+    "85": "NON_PARTICIPATING",
+    "86": "NON_PARTICIPATING",
+    "87": "FAIL",
+    "88": "NON_PARTICIPATING",
     "89": "NOT_ALIVE",
-    "9": "NOT_ALIVE",
+    "9": "SUCCESS",
     "90": "FAIL",
-    "91": "SUCCESS",
-    "92": "FAIL",
+    "91": "FAIL",
+    "92": "NON_PARTICIPATING",
     "93": "NOT_ALIVE",
-    "94": "NON_PARTICIPATING",
+    "94": "NOT_ALIVE",
     "95": "NOT_ALIVE",
     "96": "NOT_ALIVE",
     "97": "NON_PARTICIPATING",
@@ -264,17 +264,17 @@
   "attention_bias": false,
   "attention_dropout": 0.0,
   "block_list": [
-    6478757,
-    6478769
+    6478908,
+    6478925
   ],
   "bos_token_id": 1,
   "eos_token_id": 2,
   "hidden_act": "silu",
   "hidden_size": 2048,
   "initializer_range": 0.02,
-  "inner_step": 39,
+  "inner_step": 4,
   "intermediate_size": 5632,
-  "last_allreduce_block": 6478480,
+  "last_allreduce_block": 6478819,
   "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:f58fbe7d16bd236a62befa1bc791da1083ed431b574ffc1bd757f39bfd1e8f90
+oid sha256:a890374cdb11e0b064e14975b2b86ceb5cb82ff019ca22b726a91705bd278745
 size 88389626

inner_optimizer.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1a8949b2ee4fd88954305a39d8e390e05d971b2d6b95b1e1b635cbb0c6f57c7e
+oid sha256:c79337eccb4e138de046bbf4dafa5cb18228473c2d3225f976d1f560dbf86903
 size 4924926906

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fed6690369f39aa26b74c5b105f6908f4ef77dd32d5d5929acd2a5d4970bca03
+oid sha256:239780ce986e9ea2233f38a546db68924611361310a14a2f727878014ee865f0
 size 4400216536