AbhishekGhosh1998 commited on
Commit
a981508
·
verified ·
1 Parent(s): 0e03f9a

granularity results granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb @ 2026-09-21T07:04:04Z from sp-aiinfra-interactive-aghosh-8-mb99k-8xb200-moe-cg-0

Browse files
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/mem_samples.csv CHANGED
@@ -150,3 +150,75 @@
150
  5, 77239
151
  6, 77239
152
  7, 77239
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
150
  5, 77239
151
  6, 77239
152
  7, 77239
153
+ 0, 77497
154
+ 1, 77241
155
+ 2, 77241
156
+ 3, 77241
157
+ 4, 77241
158
+ 5, 77241
159
+ 6, 77241
160
+ 7, 77241
161
+ 0, 81847
162
+ 1, 81551
163
+ 2, 81625
164
+ 3, 81549
165
+ 4, 81589
166
+ 5, 81627
167
+ 6, 81609
168
+ 7, 81625
169
+ 0, 82815
170
+ 1, 82585
171
+ 2, 82547
172
+ 3, 82545
173
+ 4, 82543
174
+ 5, 82569
175
+ 6, 82555
176
+ 7, 82529
177
+ 0, 83831
178
+ 1, 83591
179
+ 2, 83599
180
+ 3, 83631
181
+ 4, 83633
182
+ 5, 83633
183
+ 6, 83597
184
+ 7, 83637
185
+ 0, 97849
186
+ 1, 97509
187
+ 2, 97481
188
+ 3, 97505
189
+ 4, 97509
190
+ 5, 97509
191
+ 6, 97509
192
+ 7, 97509
193
+ 0, 97849
194
+ 1, 97509
195
+ 2, 97481
196
+ 3, 97505
197
+ 4, 97509
198
+ 5, 97509
199
+ 6, 97509
200
+ 7, 97509
201
+ 0, 95413
202
+ 1, 95267
203
+ 2, 95301
204
+ 3, 94617
205
+ 4, 94621
206
+ 5, 94621
207
+ 6, 95009
208
+ 7, 94621
209
+ 0, 93345
210
+ 1, 93587
211
+ 2, 93233
212
+ 3, 92937
213
+ 4, 92935
214
+ 5, 93005
215
+ 6, 93587
216
+ 7, 93261
217
+ 0, 68
218
+ 1, 76425
219
+ 2, 76397
220
+ 3, 324
221
+ 4, 4
222
+ 5, 68
223
+ 6, 76425
224
+ 7, 324
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2/train.log CHANGED
@@ -838,3 +838,66 @@ INFO:moonep.api:MoonEP comm buffer sizes:
838
  local_temp_buffer_size : 1.27 MiB
839
  total_size : 563.27 MiB
840
  [MoonEP selfcheck 1/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [768, 640, 640, 1024, 640, 640, 640, 512, 1280, 640, 896, 640, 1536, 2304, 1536, 896, 1408, 384, 1280, 1664, 1152, 640, 640, 1152, 1280, 1536, 768, 512, 1664, 384, 384, 896, 2432, 1792, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[215, 197, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [194, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [205, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [78, 91, 86, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [203, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [50, 32, 34, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
838
  local_temp_buffer_size : 1.27 MiB
839
  total_size : 563.27 MiB
840
  [MoonEP selfcheck 1/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [768, 640, 640, 1024, 640, 640, 640, 512, 1280, 640, 896, 640, 1536, 2304, 1536, 896, 1408, 384, 1280, 1664, 1152, 640, 640, 1152, 1280, 1536, 768, 512, 1664, 384, 384, 896, 2432, 1792, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[215, 197, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [194, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [205, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [78, 91, 86, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [203, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [50, 32, 34, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
841
+ [2026-09-21 07:02:37] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 23843.9 | throughput per GPU (TFLOP/s/GPU): 3.1 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.226155E+01 | loss scale: 1.0 | grad norm: 15.011 | number of skipped iterations: 0 | number of nan iterations: 0 |
842
+ Number of parameters in transformer block in billions: 13.26
843
+ Number of parameters in embedding layers in billions: 0.72
844
+ Total number of parameters in billions: 13.98
845
+ Number of parameters in most loaded shard in billions: 13.9771
846
+ Theoretical memory footprints: weight and optimizer=99972.23 MB
847
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 46904.38818359375 | max allocated: 57624.96875 | reserved: 60652.0 | max reserved: 60652.0
848
+ [MoonEP selfcheck 2/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [768, 768, 768, 640, 512, 640, 768, 512, 1408, 640, 896, 768, 1408, 2560, 1408, 1024, 1280, 512, 1664, 1664, 1152, 640, 640, 1024, 1280, 1408, 640, 512, 1536, 384, 384, 896, 2176, 1664, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[205, 197, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [158, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [151, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, 78, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [50, 32, 34, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [91, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [215, 194, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
849
+ [2026-09-21 07:02:37] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 200.6 | throughput per GPU (TFLOP/s/GPU): 371.2 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.227770E+01 | loss scale: 1.0 | grad norm: 14.853 | number of skipped iterations: 0 | number of nan iterations: 0 |
850
+ [MoonEP selfcheck 3/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [384, 640, 1152, 512, 384, 768, 896, 256, 896, 512, 1024, 768, 1664, 1920, 1024, 896, 1408, 256, 1152, 1792, 256, 768, 384, 896, 768, 1024, 384, 256, 640, 384, 384, 1024, 5248, 3200, 1280, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[50, 34, 62, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [66, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [151, 158, 137, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [81, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, 78, 91, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [239, 252, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [52, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
851
+ [2026-09-21 07:02:38] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 197.8 | throughput per GPU (TFLOP/s/GPU): 376.4 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.070648E+01 | loss scale: 1.0 | grad norm: 18.850 | number of skipped iterations: 0 | number of nan iterations: 0 |
852
+ Capture CUDA graph for training!!!
853
+ CUDA graph capture done!!!
854
+ [2026-09-21 07:02:38] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 729.7 | throughput per GPU (TFLOP/s/GPU): 102.1 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.557410E+01 | loss scale: 1.0 | grad norm: 97.887 | number of skipped iterations: 0 | number of nan iterations: 0 |
855
+ [2026-09-21 07:02:38] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 193.8 | throughput per GPU (TFLOP/s/GPU): 384.3 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.447998E+01 | loss scale: 1.0 | grad norm: 97.902 | number of skipped iterations: 0 | number of nan iterations: 0 |
856
+ [2026-09-21 07:02:39] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 192.5 | throughput per GPU (TFLOP/s/GPU): 386.9 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.293705E+01 | loss scale: 1.0 | grad norm: 16.034 | number of skipped iterations: 0 | number of nan iterations: 0 |
857
+ [2026-09-21 07:02:39] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 193.3 | throughput per GPU (TFLOP/s/GPU): 385.1 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.689716E+01 | loss scale: 1.0 | grad norm: 100.878 | number of skipped iterations: 0 | number of nan iterations: 0 |
858
+ [2026-09-21 07:02:39] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 192.6 | throughput per GPU (TFLOP/s/GPU): 386.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.098987E+01 | loss scale: 1.0 | grad norm: 18.714 | number of skipped iterations: 0 | number of nan iterations: 0 |
859
+ [2026-09-21 07:02:39] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 192.1 | throughput per GPU (TFLOP/s/GPU): 387.6 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.048015E+01 | loss scale: 1.0 | grad norm: 16.061 | number of skipped iterations: 0 | number of nan iterations: 0 |
860
+ [2026-09-21 07:02:39] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 194.6 | throughput per GPU (TFLOP/s/GPU): 382.7 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.017142E+01 | loss scale: 1.0 | grad norm: 11.421 | number of skipped iterations: 0 | number of nan iterations: 0 |
861
+ [2026-09-21 07:02:40] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 194.4 | throughput per GPU (TFLOP/s/GPU): 383.0 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.971597E+00 | loss scale: 1.0 | grad norm: 11.756 | number of skipped iterations: 0 | number of nan iterations: 0 |
862
+ [2026-09-21 07:02:40] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 193.4 | throughput per GPU (TFLOP/s/GPU): 384.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.694415E+00 | loss scale: 1.0 | grad norm: 17.317 | number of skipped iterations: 0 | number of nan iterations: 0 |
863
+ [after training is done] datetime: 2026-09-21 07:02:40
864
+ [rank0]:[W921 07:02:41.482150415 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
865
+ [rank3]:[W921 07:02:42.511431902 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
866
+ [rank5]:[W921 07:02:42.512477698 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
867
+ [rank7]:[W921 07:02:42.610224663 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
868
+ [rank1]:[W921 07:02:42.700460172 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
869
+ [rank6]:[W921 07:02:43.106989150 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
870
+ [rank4]:[W921 07:02:43.187799343 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
871
+ [rank2]:[W921 07:02:43.207712693 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
872
+ W0921 07:02:48.113000 362905 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
873
+ W0921 07:02:48.114000 362905 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 363104 closing signal SIGTERM
874
+ W0921 07:02:48.115000 362905 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 363105 closing signal SIGTERM
875
+ W0921 07:02:48.115000 362905 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 363109 closing signal SIGTERM
876
+ Traceback (most recent call last):
877
+ File "/usr/local/bin/torchrun", line 8, in <module>
878
+ sys.exit(main())
879
+ ^^^^^^
880
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
881
+ return f(*args, **kwargs)
882
+ ^^^^^^^^^^^^^^^^^^
883
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
884
+ run(args)
885
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
886
+ elastic_launch(
887
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
888
+ return launch_agent(self._config, self._entrypoint, list(args))
889
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
890
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
891
+ result = agent.run()
892
+ ^^^^^^^^^^^
893
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
894
+ result = f(*args, **kwargs)
895
+ ^^^^^^^^^^^^^^^^^^
896
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
897
+ result = self._invoke_run(role)
898
+ ^^^^^^^^^^^^^^^^^^^^^^
899
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
900
+ time.sleep(monitor_interval)
901
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
902
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
903
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 362905 got signal: 15
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep3/mem_samples.csv ADDED
@@ -0,0 +1,216 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 0, 0
2
+ 1, 0
3
+ 2, 0
4
+ 3, 0
5
+ 4, 0
6
+ 5, 0
7
+ 6, 0
8
+ 7, 0
9
+ 0, 4
10
+ 1, 4
11
+ 2, 1
12
+ 3, 1
13
+ 4, 1
14
+ 5, 1
15
+ 6, 1
16
+ 7, 1
17
+ 0, 4
18
+ 1, 4
19
+ 2, 4
20
+ 3, 4
21
+ 4, 4
22
+ 5, 4
23
+ 6, 4
24
+ 7, 1
25
+ 0, 0
26
+ 1, 0
27
+ 2, 0
28
+ 3, 0
29
+ 4, 0
30
+ 5, 0
31
+ 6, 0
32
+ 7, 0
33
+ 0, 0
34
+ 1, 0
35
+ 2, 0
36
+ 3, 0
37
+ 4, 0
38
+ 5, 0
39
+ 6, 0
40
+ 7, 0
41
+ 0, 4
42
+ 1, 1
43
+ 2, 1
44
+ 3, 1
45
+ 4, 1
46
+ 5, 1
47
+ 6, 1
48
+ 7, 1
49
+ 0, 4
50
+ 1, 4
51
+ 2, 4
52
+ 3, 4
53
+ 4, 4
54
+ 5, 1
55
+ 6, 1
56
+ 7, 1
57
+ 0, 4
58
+ 1, 4
59
+ 2, 4
60
+ 3, 4
61
+ 4, 4
62
+ 5, 4
63
+ 6, 4
64
+ 7, 4
65
+ 0, 4
66
+ 1, 4
67
+ 2, 4
68
+ 3, 45
69
+ 4, 4
70
+ 5, 4
71
+ 6, 4
72
+ 7, 4
73
+ 0, 821
74
+ 1, 821
75
+ 2, 821
76
+ 3, 821
77
+ 4, 821
78
+ 5, 821
79
+ 6, 821
80
+ 7, 821
81
+ 0, 1927
82
+ 1, 1927
83
+ 2, 1927
84
+ 3, 1927
85
+ 4, 1927
86
+ 5, 1927
87
+ 6, 1927
88
+ 7, 1927
89
+ 0, 37691
90
+ 1, 37691
91
+ 2, 37691
92
+ 3, 37691
93
+ 4, 37691
94
+ 5, 37691
95
+ 6, 37691
96
+ 7, 37691
97
+ 0, 44939
98
+ 1, 45033
99
+ 2, 45069
100
+ 3, 45107
101
+ 4, 45077
102
+ 5, 45061
103
+ 6, 45095
104
+ 7, 45057
105
+ 0, 75511
106
+ 1, 75813
107
+ 2, 75839
108
+ 3, 75811
109
+ 4, 75817
110
+ 5, 75827
111
+ 6, 75849
112
+ 7, 75843
113
+ 0, 76023
114
+ 1, 76329
115
+ 2, 76329
116
+ 3, 76329
117
+ 4, 76329
118
+ 5, 76329
119
+ 6, 76329
120
+ 7, 76329
121
+ 0, 76025
122
+ 1, 76329
123
+ 2, 76329
124
+ 3, 76329
125
+ 4, 76329
126
+ 5, 76329
127
+ 6, 76329
128
+ 7, 76329
129
+ 0, 76223
130
+ 1, 76527
131
+ 2, 76527
132
+ 3, 76527
133
+ 4, 76527
134
+ 5, 76527
135
+ 6, 76527
136
+ 7, 76527
137
+ 0, 76483
138
+ 1, 76797
139
+ 2, 76799
140
+ 3, 76813
141
+ 4, 76843
142
+ 5, 76811
143
+ 6, 76799
144
+ 7, 76829
145
+ 0, 77497
146
+ 1, 77241
147
+ 2, 77241
148
+ 3, 77241
149
+ 4, 77241
150
+ 5, 77241
151
+ 6, 77241
152
+ 7, 77241
153
+ 0, 81765
154
+ 1, 81547
155
+ 2, 81575
156
+ 3, 81623
157
+ 4, 81583
158
+ 5, 81609
159
+ 6, 81623
160
+ 7, 81611
161
+ 0, 82755
162
+ 1, 82511
163
+ 2, 82497
164
+ 3, 82507
165
+ 4, 82537
166
+ 5, 82523
167
+ 6, 82501
168
+ 7, 82505
169
+ 0, 83791
170
+ 1, 83569
171
+ 2, 83585
172
+ 3, 83593
173
+ 4, 83603
174
+ 5, 83593
175
+ 6, 83595
176
+ 7, 83615
177
+ 0, 97849
178
+ 1, 97509
179
+ 2, 97481
180
+ 3, 97505
181
+ 4, 97509
182
+ 5, 97509
183
+ 6, 97509
184
+ 7, 97509
185
+ 0, 97849
186
+ 1, 97509
187
+ 2, 97481
188
+ 3, 97505
189
+ 4, 97509
190
+ 5, 97509
191
+ 6, 97509
192
+ 7, 97509
193
+ 0, 94961
194
+ 1, 95009
195
+ 2, 94593
196
+ 3, 94617
197
+ 4, 94621
198
+ 5, 94759
199
+ 6, 95009
200
+ 7, 95009
201
+ 0, 93563
202
+ 1, 93211
203
+ 2, 92919
204
+ 3, 92937
205
+ 4, 92725
206
+ 5, 92941
207
+ 6, 93011
208
+ 7, 93791
209
+ 0, 65
210
+ 1, 76425
211
+ 2, 324
212
+ 3, 4
213
+ 4, 4
214
+ 5, 4
215
+ 6, 4
216
+ 7, 4
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep3/run.env ADDED
@@ -0,0 +1,105 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ backend=moonep
2
+ config=DEEPSEEK_V3_L4
3
+ family=tierB
4
+ source_model=DeepSeek-V3
5
+ standard_num_layers=61
6
+ base_rung=
7
+ repo=/workspace/Megatron-MoE-Echo
8
+ branch=moonep_backend_async_grad_reduce
9
+ commit=2e68c8ad
10
+ moonep_source_dir=/workspace/MoonEP
11
+ moonep_source_branch=megatron-ddp-symmetric-params
12
+ moonep_source_commit=d8bc666
13
+ num_layers=4
14
+ hidden=7168
15
+ ffn_hidden=18432
16
+ moe_ffn_hidden=2048
17
+ num_experts=256
18
+ topk=8
19
+ experts_per_rank=32
20
+ expert_params=44040192
21
+ heads=128
22
+ kv_groups=128
23
+ kv_channels=128
24
+ qk_layernorm=1
25
+ shared_ffn=2048
26
+ shared_expert_overlap=0
27
+ moe_token_dispatcher=flex
28
+ mla=1
29
+ q_lora_rank=1536
30
+ kv_lora_rank=512
31
+ qk_head_dim=128
32
+ qk_pos_emb_head_dim=64
33
+ v_head_dim=128
34
+ dense_first_layers=3
35
+ moe_layers=1
36
+ rotary_base=10000
37
+ rotary_percent=1.0
38
+ rotary_scaling=40
39
+ mscale=1.0
40
+ mscale_all_dim=1.0
41
+ qkv_bias=0
42
+ router_balancing=seq_aux_loss
43
+ aux_loss_coeff=1e-4
44
+ router_pre_softmax=0
45
+ router_fusion=0
46
+ router_dtype=fp32
47
+ router_groups=8
48
+ router_group_topk=4
49
+ router_score_function=sigmoid
50
+ router_expert_bias=1
51
+ router_bias_update_rate=1e-3
52
+ router_topk_scaling_factor=2.5
53
+ shared_expert_gate=0
54
+ apply_layernorm_1p=0
55
+ attention_output_gate=0
56
+ linear_attention_type=
57
+ linear_attention_freq=
58
+ linear_conv_kernel_dim=
59
+ linear_key_head_dim=
60
+ linear_value_head_dim=
61
+ linear_num_key_heads=
62
+ linear_num_value_heads=
63
+ mtp_num_layers=0
64
+ mtp_loss_scaling_factor=0.1
65
+ no_weight_decay_cond_type=
66
+ max_position_embeddings=4096
67
+ init_method_std=0.02
68
+ device_initiated_grouped_gemm=1
69
+ norm_epsilon=1e-6
70
+ seq_length=4096
71
+ micro_batch_size=1
72
+ global_batch_size=8
73
+ num_microbatches=1
74
+ assignments_per_rank=32768
75
+ mean_tokens_per_expert=1024
76
+ slot_policy=matched
77
+ slots_per_rank=32
78
+ capacity_mode=none
79
+ capacity=none
80
+ cuda_graph=1
81
+ probe_mode=0
82
+ async_grad_reduce=1
83
+ branch_role=timing
84
+ memory_dump_mode=none
85
+ memory_snapshot_path=
86
+ cuda_graph_warmup_steps=3
87
+ recompute=none
88
+ force_balance=0
89
+ train_iters=12
90
+ ep_size=8
91
+ echo_dispatch_overlap=1
92
+ echo_dispatch_recompute=0
93
+ ultraep_quota_min_tokens=
94
+ ultraep_weight_sync_join=
95
+ profile_enable=0
96
+ kernel_trace=0
97
+ fp8_mode=mxfp8
98
+ fp8_recipe=mxfp8
99
+ training_seed=1234
100
+ learning_rate=1e-4
101
+ min_learning_rate=1.0e-5
102
+ lr_warmup_iters=5
103
+ state_dump_enable=0
104
+ state_dump_dir=
105
+ moonep_assert_no_pending_reduce=1
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep3/train.log ADDED
@@ -0,0 +1,904 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [07:02:58] WARN: moonep_backend_async_grad_reduce lacks --moe-router-fusion; DEEPSEEK_V3_L4 runs unfused (timing only, routing identical)
2
+ ==============================================================
3
+ DEEPSEEK_V3_L4 / moonep (slots/rank=32 policy=matched)
4
+ E=256 K=8 H'=2048 m=32
5
+ L=4 seq=4096 gbs=8 (1 microbatches)
6
+ cap_mode=none cap=none cg=1 recompute=none
7
+ memory_dump=none snapshots=/workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep3/memory_snapshots/snapshot.pickle
8
+ repo=/workspace/Megatron-MoE-Echo @ 2e68c8ad
9
+ ==============================================================
10
+ W0921 07:02:59.870000 365649 torch/distributed/run.py:766]
11
+ W0921 07:02:59.870000 365649 torch/distributed/run.py:766] *****************************************
12
+ W0921 07:02:59.870000 365649 torch/distributed/run.py:766] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
13
+ W0921 07:02:59.870000 365649 torch/distributed/run.py:766] *****************************************
14
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
15
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
16
+ using world size: 8, data-parallel size: 8, context-parallel size: 1, hierarchical context-parallel sizes: None, tensor-model-parallel size: 1, pipeline-model-parallel size: 1
17
+ WARNING: overriding default arguments for tokenizer_type:GPT2BPETokenizer with tokenizer_type:GPT2BPETokenizer
18
+ Number of virtual stages per pipeline stage: None
19
+ accumulate and all-reduce gradients in fp32 for bfloat16 data type.
20
+ using torch.bfloat16 for parameters ...
21
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
22
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
23
+ ------------------------ arguments ------------------------
24
+ account_for_embedding_in_pipeline_split ......... False
25
+ account_for_loss_in_pipeline_split .............. False
26
+ accumulate_allreduce_grads_in_fp32 .............. True
27
+ adam_beta1 ...................................... 0.9
28
+ adam_beta2 ...................................... 0.999
29
+ adam_eps ........................................ 1e-08
30
+ add_bias_linear ................................. False
31
+ add_position_embedding .......................... False
32
+ add_qkv_bias .................................... False
33
+ adlr_autoresume ................................. False
34
+ adlr_autoresume_interval ........................ 1000
35
+ align_grad_reduce ............................... True
36
+ align_param_gather .............................. False
37
+ app_tag_run_name ................................ None
38
+ app_tag_run_version ............................. 0.0.0
39
+ apply_layernorm_1p .............................. False
40
+ apply_query_key_layer_scaling ................... False
41
+ apply_residual_connection_post_layernorm ........ False
42
+ apply_rope_fusion ............................... False
43
+ async_save ...................................... None
44
+ async_tensor_model_parallel_allreduce ........... True
45
+ attention_backend ............................... AttnBackend.auto
46
+ attention_dropout ............................... 0.0
47
+ attention_softmax_in_fp32 ....................... False
48
+ auto_detect_ckpt_format ......................... False
49
+ barrier_with_L1_time ............................ True
50
+ bert_binary_head ................................ True
51
+ bert_embedder_type .............................. megatron
52
+ bert_load ....................................... None
53
+ bf16 ............................................ True
54
+ bias_dropout_fusion ............................. True
55
+ bias_gelu_fusion ................................ False
56
+ bias_swiglu_fusion .............................. True
57
+ biencoder_projection_dim ........................ 0
58
+ biencoder_shared_query_context_model ............ False
59
+ block_data_path ................................. None
60
+ calc_ft_timeouts ................................ False
61
+ calculate_per_token_loss ........................ False
62
+ check_for_large_grads ........................... False
63
+ check_for_nan_in_loss_and_grad .................. False
64
+ check_for_spiky_loss ............................ False
65
+ check_weight_hash_across_dp_replicas_interval ... None
66
+ ckpt_assume_constant_structure .................. False
67
+ ckpt_convert_format ............................. None
68
+ ckpt_convert_save ............................... None
69
+ ckpt_convert_update_legacy_dist_opt_format ...... False
70
+ ckpt_format ..................................... torch_dist
71
+ ckpt_fully_parallel_load ........................ False
72
+ ckpt_fully_parallel_save ........................ True
73
+ ckpt_fully_parallel_save_deprecated ............. False
74
+ ckpt_step ....................................... None
75
+ classes_fraction ................................ 1.0
76
+ clip_grad ....................................... 1.0
77
+ clone_scatter_output_in_embedding ............... True
78
+ config_logger_dir ...............................
79
+ consumed_train_samples .......................... 0
80
+ consumed_valid_samples .......................... 0
81
+ context_parallel_size ........................... 1
82
+ cp_comm_type .................................... ['p2p']
83
+ create_attention_mask_in_dataloader ............. True
84
+ cross_entropy_fusion_impl ....................... native
85
+ cross_entropy_loss_fusion ....................... False
86
+ cuda_graph_scope ................................ full_iteration
87
+ cuda_graph_warmup_steps ......................... 3
88
+ data_args_path .................................. None
89
+ data_cache_path ................................. None
90
+ data_parallel_random_init ....................... False
91
+ data_parallel_sharding_strategy ................. no_shard
92
+ data_parallel_size .............................. 8
93
+ data_path ....................................... ['/workspace/data/wikitext/wikitext103_gpt2_text_document']
94
+ data_per_class_fraction ......................... 1.0
95
+ data_sharding ................................... True
96
+ dataloader_type ................................. single
97
+ ddp_average_in_collective ....................... False
98
+ ddp_bucket_size ................................. None
99
+ ddp_num_buckets ................................. None
100
+ ddp_pad_buckets_for_high_nccl_busbw ............. False
101
+ decoder_first_pipeline_num_layers ............... None
102
+ decoder_last_pipeline_num_layers ................ None
103
+ decoder_num_layers .............................. None
104
+ decoder_seq_length .............................. None
105
+ decoupled_lr .................................... None
106
+ decoupled_min_lr ................................ None
107
+ decrease_batch_size_if_needed ................... False
108
+ defer_embedding_wgrad_compute ................... False
109
+ delay_wgrad_compute ............................. False
110
+ deprecated_use_mcore_models ..................... True
111
+ deterministic_mode .............................. False
112
+ dino_bottleneck_size ............................ 256
113
+ dino_freeze_last_layer .......................... 1
114
+ dino_head_hidden_size ........................... 2048
115
+ dino_local_crops_number ......................... 10
116
+ dino_local_img_size ............................. 96
117
+ dino_norm_last_layer ............................ False
118
+ dino_teacher_temp ............................... 0.07
119
+ dino_warmup_teacher_temp ........................ 0.04
120
+ dino_warmup_teacher_temp_epochs ................. 30
121
+ disable_bf16_reduced_precision_matmul ........... False
122
+ disable_mamba_mem_eff_path ...................... False
123
+ disable_straggler_on_startup .................... False
124
+ dist_ckpt_format_deprecated ..................... None
125
+ dist_ckpt_strictness ............................ assume_ok_unexpected
126
+ distribute_saved_activations .................... False
127
+ distributed_backend ............................. nccl
128
+ distributed_timeout_minutes ..................... 10
129
+ embedding_init_method_std ....................... None
130
+ embedding_path .................................. None
131
+ empty_unused_memory_level ....................... 0
132
+ enable_cuda_graph ............................... True
133
+ enable_experimental ............................. False
134
+ enable_ft_package ............................... False
135
+ enable_gloo_process_groups ...................... True
136
+ enable_msc ...................................... True
137
+ enable_one_logger ............................... True
138
+ encoder_num_layers .............................. 4
139
+ encoder_seq_length .............................. 4096
140
+ end_weight_decay ................................ 0.1
141
+ eod_mask_loss ................................... False
142
+ error_injection_rate ............................ 0
143
+ error_injection_type ............................ transient_error
144
+ eval_interval ................................... 1000000
145
+ eval_iters ...................................... 0
146
+ evidence_data_path .............................. None
147
+ exit_duration_in_mins ........................... None
148
+ exit_interval ................................... None
149
+ exit_on_missing_checkpoint ...................... False
150
+ exit_signal_handler ............................. False
151
+ exp_avg_dtype ................................... torch.float32
152
+ exp_avg_sq_dtype ................................ torch.float32
153
+ expert_model_parallel_size ...................... 8
154
+ expert_tensor_parallel_size ..................... 1
155
+ export_force_local_attention .................... False
156
+ export_kd_cfg ................................... None
157
+ export_kd_teacher_ckpt_format ................... None
158
+ export_kd_teacher_load .......................... None
159
+ export_kv_cache_quant ........................... False
160
+ export_legacy_megatron .......................... False
161
+ export_model_type ............................... GPTModel
162
+ export_moe_apply_probs_on_input ................. False
163
+ export_qk_l2_norm ............................... False
164
+ export_quant_cfg ................................ None
165
+ export_real_quant_cfg ........................... None
166
+ export_te_mcore_model ........................... False
167
+ external_cuda_graph ............................. False
168
+ ffn_hidden_size ................................. 18432
169
+ finetune ........................................ False
170
+ finetune_data_split ............................. train
171
+ finetune_hf_dataset ............................. None
172
+ first_last_layers_bf16 .......................... False
173
+ flash_decode .................................... False
174
+ fp16 ............................................ False
175
+ fp16_lm_cross_entropy ........................... False
176
+ fp32_residual_connection ........................ False
177
+ fp8 ............................................. e4m3
178
+ fp8_amax_compute_algo ........................... most_recent
179
+ fp8_amax_history_len ............................ 1
180
+ fp8_interval .................................... 1
181
+ fp8_margin ...................................... 0
182
+ fp8_param_gather ................................ False
183
+ fp8_recipe ...................................... mxfp8
184
+ fp8_wgrad ....................................... True
185
+ fsdp_double_buffer .............................. False
186
+ global_batch_size ............................... 8
187
+ grad_reduce_in_bf16 ............................. False
188
+ gradient_accumulation_fusion .................... True
189
+ gradient_reduce_div_fusion ...................... True
190
+ group_query_attention ........................... False
191
+ head_lr_mult .................................... 1.0
192
+ heterogeneous_layers_config_encoded_json ........ None
193
+ heterogeneous_layers_config_path ................ None
194
+ hidden_dropout .................................. 0.0
195
+ hidden_size ..................................... 7168
196
+ hierarchical_context_parallel_sizes ............. None
197
+ high_priority_stream_groups ..................... []
198
+ hybrid_attention_ratio .......................... 0.0
199
+ hybrid_mlp_ratio ................................ 0.0
200
+ hybrid_override_pattern ......................... None
201
+ hysteresis ...................................... 2
202
+ ict_head_size ................................... None
203
+ ict_load ........................................ None
204
+ img_h ........................................... 224
205
+ img_w ........................................... 224
206
+ indexer_batch_size .............................. 128
207
+ indexer_log_interval ............................ 1000
208
+ inference_batch_times_seqlen_threshold .......... -1
209
+ inference_dynamic_batching ...................... False
210
+ inference_dynamic_batching_buffer_guaranteed_fraction 0.2
211
+ inference_dynamic_batching_buffer_overflow_factor None
212
+ inference_dynamic_batching_buffer_size_gb ....... 40.0
213
+ inference_dynamic_batching_chunk_size ........... 256
214
+ inference_dynamic_batching_max_requests_override None
215
+ inference_dynamic_batching_max_tokens_override .. None
216
+ inference_dynamic_batching_num_cuda_graphs ...... 16
217
+ inference_max_batch_size ........................ 8
218
+ inference_max_seq_length ........................ 2560
219
+ inference_rng_tracker ........................... False
220
+ init_method_std ................................. 0.02
221
+ init_method_xavier_uniform ...................... False
222
+ init_model_with_meta_device ..................... False
223
+ initial_loss_scale .............................. 4294967296
224
+ inprocess_active_world_size ..................... 8
225
+ inprocess_barrier_timeout ....................... 120
226
+ inprocess_completion_timeout .................... 120
227
+ inprocess_empty_cuda_cache ...................... False
228
+ inprocess_granularity ........................... node
229
+ inprocess_hard_timeout .......................... 90
230
+ inprocess_heartbeat_interval .................... 30
231
+ inprocess_heartbeat_timeout ..................... 60
232
+ inprocess_last_call_wait ........................ 1
233
+ inprocess_max_iterations ........................ None
234
+ inprocess_monitor_process_interval .............. 1.0
235
+ inprocess_monitor_thread_interval ............... 1.0
236
+ inprocess_progress_watchdog_interval ............ 1.0
237
+ inprocess_restart ............................... False
238
+ inprocess_soft_timeout .......................... 60
239
+ inprocess_termination_grace_time ................ 1
240
+ is_hybrid_model ................................. False
241
+ iter_per_epoch .................................. 1250
242
+ iterations_to_skip .............................. []
243
+ keep_fp8_transpose_cache_when_using_custom_fsdp . False
244
+ kitchen_config_file ............................. None
245
+ kitchen_recipe_number ........................... None
246
+ kv_channels ..................................... 128
247
+ kv_lora_rank .................................... 512
248
+ lazy_mpu_init ................................... None
249
+ load ............................................ None
250
+ load_main_params_from_ckpt ...................... None
251
+ load_model_opt_format ........................... False
252
+ local_rank ...................................... 0
253
+ log_energy ...................................... False
254
+ log_interval .................................... 1
255
+ log_loss_scale_to_tensorboard ................... True
256
+ log_memory_to_tensorboard ....................... False
257
+ log_num_zeros_in_grad ........................... False
258
+ log_params_norm ................................. False
259
+ log_progress .................................... False
260
+ log_straggler ................................... False
261
+ log_throughput .................................. True
262
+ log_timers_to_tensorboard ....................... False
263
+ log_validation_ppl_to_tensorboard ............... False
264
+ log_world_size_to_tensorboard ................... False
265
+ logging_level ................................... None
266
+ loss_scale ...................................... None
267
+ loss_scale_window ............................... 1000
268
+ lr .............................................. 0.0001
269
+ lr_decay_iters .................................. 8
270
+ lr_decay_samples ................................ None
271
+ lr_decay_style .................................. cosine
272
+ lr_warmup_fraction .............................. None
273
+ lr_warmup_init .................................. 0.0
274
+ lr_warmup_iters ................................. 5
275
+ lr_warmup_samples ............................... 0
276
+ lr_wsd_decay_iters .............................. None
277
+ lr_wsd_decay_samples ............................ None
278
+ lr_wsd_decay_style .............................. exponential
279
+ main_grads_dtype ................................ torch.float32
280
+ main_params_dtype ............................... torch.float32
281
+ make_vocab_size_divisible_by .................... 128
282
+ mamba_head_dim .................................. 64
283
+ mamba_num_groups ................................ 8
284
+ mamba_num_heads ................................. None
285
+ mamba_state_dim ................................. 128
286
+ manual_gc ....................................... False
287
+ manual_gc_eval .................................. True
288
+ manual_gc_interval .............................. 0
289
+ mask_factor ..................................... 1.0
290
+ mask_prob ....................................... 0.15
291
+ mask_type ....................................... random
292
+ masked_softmax_fusion ........................... False
293
+ max_position_embeddings ......................... 4096
294
+ max_tokens_to_oom ............................... 12000
295
+ memory_snapshot_path ............................ snapshot.pickle
296
+ merge_file ...................................... /workspace/data/wikitext/gpt2-merges.txt
297
+ micro_batch_size ................................ 1
298
+ microbatch_group_size_per_vp_stage .............. None
299
+ mid_level_dataset_surplus ....................... 0.005
300
+ min_loss_scale .................................. 1.0
301
+ min_lr .......................................... 1e-05
302
+ mlp_chunks_for_prefill .......................... 1
303
+ mmap_bin_files .................................. True
304
+ mock_data ....................................... False
305
+ moe_apply_probs_on_input ........................ False
306
+ moe_aux_loss_coeff .............................. 0.0001
307
+ moe_deepep_num_sms .............................. 20
308
+ moe_echo_enable_random_offloading ............... False
309
+ moe_echo_expert_dispatch_overlap ................ False
310
+ moe_echo_expert_dispatcher_type ................. hybridep
311
+ moe_echo_recompute_expert_dispatch .............. False
312
+ moe_enable_deepep ............................... False
313
+ moe_enable_echo ................................. False
314
+ moe_enable_hybridep ............................. False
315
+ moe_enable_moonep ............................... True
316
+ moe_expert_capacity_factor ...................... None
317
+ moe_extended_tp ................................. False
318
+ moe_ffn_hidden_size ............................. 2048
319
+ moe_grouped_gemm ................................ True
320
+ moe_hybridep_num_sms ............................ 16
321
+ moe_input_jitter_eps ............................ None
322
+ moe_layer_freq .................................. [0, 0, 0, 1]
323
+ moe_layer_recompute ............................. False
324
+ moe_moonep_async_grad_reduce .................... True
325
+ moe_moonep_num_sms .............................. None
326
+ moe_num_echo_experts ............................ None
327
+ moe_pad_expert_input_to_capacity ................ False
328
+ moe_per_layer_logging ........................... False
329
+ moe_permute_fusion .............................. False
330
+ moe_received_token_capacity ..................... None
331
+ moe_router_bias_update_rate ..................... 0.001
332
+ moe_router_dtype ................................ fp32
333
+ moe_router_enable_expert_bias ................... True
334
+ moe_router_force_load_balancing ................. False
335
+ moe_router_group_topk ........................... 4
336
+ moe_router_load_balancing_type .................. seq_aux_loss
337
+ moe_router_num_groups ........................... 8
338
+ moe_router_padding_for_fp8 ...................... False
339
+ moe_router_pre_softmax .......................... False
340
+ moe_router_score_function ....................... sigmoid
341
+ moe_router_topk ................................. 8
342
+ moe_router_topk_scaling_factor .................. 2.5
343
+ moe_shared_expert_intermediate_size ............. 2048
344
+ moe_shared_expert_overlap ....................... False
345
+ moe_token_dispatcher_type ....................... flex
346
+ moe_token_drop_policy ........................... probs
347
+ moe_upcycling_granularity ....................... 1
348
+ moe_use_device_initiated_grouped_gemm ........... True
349
+ moe_use_legacy_grouped_gemm ..................... False
350
+ moe_use_upcycling ............................... False
351
+ moe_z_loss_coeff ................................ None
352
+ mrope_section ................................... None
353
+ mscale .......................................... 1.0
354
+ mscale_all_dim .................................. 1.0
355
+ mtp_loss_scaling_factor ......................... 0.1
356
+ mtp_num_layers .................................. None
357
+ multi_latent_attention .......................... True
358
+ nccl_all_reduce_for_prefill ..................... False
359
+ nccl_communicator_config_path ................... None
360
+ nccl_ub ......................................... False
361
+ no_load_optim ................................... None
362
+ no_load_rng ..................................... None
363
+ no_persist_layer_norm ........................... False
364
+ no_rope_freq .................................... None
365
+ no_save_optim ................................... None
366
+ no_save_rng ..................................... None
367
+ non_persistent_ckpt_type ........................ None
368
+ non_persistent_global_ckpt_dir .................. None
369
+ non_persistent_local_ckpt_algo .................. fully_parallel
370
+ non_persistent_local_ckpt_dir ................... None
371
+ non_persistent_save_interval .................... None
372
+ norm_epsilon .................................... 1e-06
373
+ normalization ................................... RMSNorm
374
+ num_attention_heads ............................. 128
375
+ num_channels .................................... 3
376
+ num_classes ..................................... 1000
377
+ num_dataset_builder_threads ..................... 1
378
+ num_distributed_optimizer_instances ............. 1
379
+ num_experts ..................................... 256
380
+ num_layers ...................................... 4
381
+ num_layers_at_end_in_bf16 ....................... 1
382
+ num_layers_at_start_in_bf16 ..................... 1
383
+ num_layers_per_virtual_pipeline_stage ........... None
384
+ num_query_groups ................................ 1
385
+ num_virtual_stages_per_pipeline_rank ............ None
386
+ num_workers ..................................... 2
387
+ object_storage_cache_path ....................... None
388
+ one_logger_async ................................ False
389
+ one_logger_project .............................. megatron-lm
390
+ one_logger_run_name ............................. None
391
+ onnx_safe ....................................... None
392
+ openai_gelu ..................................... False
393
+ optimizer ....................................... adam
394
+ optimizer_cpu_offload ........................... False
395
+ optimizer_offload_fraction ...................... 1.0
396
+ output_bert_embeddings .......................... False
397
+ overlap_cpu_optimizer_d2h_h2d ................... False
398
+ overlap_grad_reduce ............................. True
399
+ overlap_moe_expert_parallel_comm ................ False
400
+ overlap_p2p_comm ................................ False
401
+ overlap_p2p_comm_warmup_flush ................... False
402
+ overlap_param_gather ............................ True
403
+ overlap_param_gather_with_optimizer_step ........ False
404
+ override_opt_param_scheduler .................... False
405
+ params_dtype .................................... torch.bfloat16
406
+ patch_dim ....................................... 16
407
+ per_split_data_args_path ........................ None
408
+ perform_initialization .......................... True
409
+ pin_cpu_grads ................................... True
410
+ pin_cpu_params .................................. True
411
+ pipeline_model_parallel_comm_backend ............ None
412
+ pipeline_model_parallel_layout .................. None
413
+ pipeline_model_parallel_size .................... 1
414
+ position_embedding_type ......................... rope
415
+ pretrained_checkpoint ........................... None
416
+ profile ......................................... False
417
+ profile_ranks ................................... [0]
418
+ profile_step_end ................................ 12
419
+ profile_step_start .............................. 10
420
+ q_lora_rank ..................................... 1536
421
+ qk_head_dim ..................................... 128
422
+ qk_l2_norm ...................................... False
423
+ qk_layernorm .................................... True
424
+ qk_pos_emb_head_dim ............................. 64
425
+ query_in_block_prob ............................. 0.1
426
+ rampup_batch_size ............................... None
427
+ rank ............................................ 0
428
+ recompute_granularity ........................... None
429
+ recompute_method ................................ None
430
+ recompute_modules ............................... None
431
+ recompute_num_layers ............................ None
432
+ record_memory_history ........................... False
433
+ relative_attention_max_distance ................. 128
434
+ relative_attention_num_buckets .................. 32
435
+ replication ..................................... False
436
+ replication_factor .............................. 2
437
+ replication_jump ................................ None
438
+ rerun_mode ...................................... validate_results
439
+ reset_attention_mask ............................ False
440
+ reset_position_ids .............................. False
441
+ result_rejected_tracker_filename ................ None
442
+ retriever_report_topk_accuracies ................ []
443
+ retriever_score_scaling ......................... False
444
+ retriever_seq_length ............................ 256
445
+ retro_add_retriever ............................. False
446
+ retro_attention_gate ............................ 1
447
+ retro_cyclic_train_iters ........................ None
448
+ retro_encoder_attention_dropout ................. 0.1
449
+ retro_encoder_hidden_dropout .................... 0.1
450
+ retro_encoder_layers ............................ 2
451
+ retro_num_neighbors ............................. 2
452
+ retro_num_retrieved_chunks ...................... 2
453
+ retro_project_dir ............................... None
454
+ retro_verify_neighbor_count ..................... True
455
+ reuse_grad_buf_for_mxfp8_param_ag ............... False
456
+ rope_scaling_factor ............................. 8.0
457
+ rope_type ....................................... None
458
+ rotary_base ..................................... 10000
459
+ rotary_interleaved .............................. False
460
+ rotary_percent .................................. 1.0
461
+ rotary_scaling_factor ........................... 40.0
462
+ rotary_seq_len_interpolation_factor ............. None
463
+ run_workload_inspector_server ................... False
464
+ sample_rate ..................................... 1.0
465
+ save ............................................ None
466
+ save_interval ................................... None
467
+ save_retain_interval ............................ None
468
+ scatter_gather_tensors_in_pipeline .............. True
469
+ seed ............................................ 1234
470
+ seq_length ...................................... 4096
471
+ sequence_parallel ............................... False
472
+ sft ............................................. False
473
+ sft_tokenizer_prompt_format ..................... nemotron-h-aligned
474
+ sgd_momentum .................................... 0.9
475
+ short_seq_prob .................................. 0.1
476
+ skip_train ...................................... False
477
+ skipped_train_samples ........................... 0
478
+ spec ............................................ None
479
+ split ........................................... 100,0,0
480
+ squared_relu .................................... False
481
+ start_weight_decay .............................. 0.1
482
+ straggler_ctrlr_port ............................ 65535
483
+ straggler_minmax_count .......................... 1
484
+ suggested_communication_unit_size ............... None
485
+ swiglu .......................................... True
486
+ swin_backbone_type .............................. tiny
487
+ symmetric_ar_type ............................... None
488
+ te_rng_tracker .................................. True
489
+ tensor_model_parallel_size ...................... 1
490
+ tensorboard_dir ................................. None
491
+ tensorboard_log_interval ........................ 1
492
+ tensorboard_queue_size .......................... 1000
493
+ test_data_path .................................. None
494
+ test_mode ....................................... False
495
+ tiktoken_num_special_tokens ..................... 1000
496
+ tiktoken_pattern ................................ None
497
+ tiktoken_special_tokens ......................... None
498
+ timing_log_level ................................ 0
499
+ timing_log_option ............................... minmax
500
+ titles_data_path ................................ None
501
+ tokenizer_model ................................. None
502
+ tokenizer_type .................................. GPT2BPETokenizer
503
+ torch_fsdp2_reshard_after_forward ............... True
504
+ tp_comm_bootstrap_backend ....................... nccl
505
+ tp_comm_bulk_dgrad .............................. True
506
+ tp_comm_bulk_wgrad .............................. True
507
+ tp_comm_overlap ................................. False
508
+ tp_comm_overlap_ag .............................. True
509
+ tp_comm_overlap_cfg ............................. None
510
+ tp_comm_overlap_rs .............................. True
511
+ tp_comm_overlap_rs_dgrad ........................ False
512
+ tp_comm_split_ag ................................ True
513
+ tp_comm_split_rs ................................ True
514
+ train_data_path ................................. None
515
+ train_iters ..................................... 12
516
+ train_samples ................................... None
517
+ train_sync_interval ............................. None
518
+ transformer_impl ................................ transformer_engine
519
+ transformer_pipeline_model_parallel_size ........ 1
520
+ untie_embeddings_and_output_weights ............. True
521
+ use_checkpoint_args ............................. False
522
+ use_checkpoint_opt_param_scheduler .............. False
523
+ use_cpu_initialization .......................... None
524
+ use_custom_fsdp ................................. False
525
+ use_dist_ckpt ................................... True
526
+ use_dist_ckpt_deprecated ........................ False
527
+ use_distributed_optimizer ....................... True
528
+ use_flash_attn .................................. False
529
+ use_legacy_models ............................... False
530
+ use_mp_args_from_checkpoint_args ................ False
531
+ use_one_sent_docs ............................... False
532
+ use_persistent_ckpt_worker ...................... False
533
+ use_precision_aware_optimizer ................... False
534
+ use_pytorch_profiler ............................ False
535
+ use_ring_exchange_p2p ........................... False
536
+ use_rope_scaling ................................ False
537
+ use_rotary_position_embeddings .................. False
538
+ use_sharp ....................................... False
539
+ use_tokenizer_model_from_checkpoint_args ........ True
540
+ use_torch_fsdp2 ................................. False
541
+ use_torch_optimizer_for_cpu_offload ............. False
542
+ use_tp_pp_dp_mapping ............................ False
543
+ v_head_dim ...................................... 128
544
+ valid_data_path ................................. None
545
+ variable_seq_lengths ............................ False
546
+ virtual_pipeline_model_parallel_size ............ None
547
+ vision_backbone_type ............................ vit
548
+ vision_pretraining .............................. False
549
+ vision_pretraining_type ......................... classify
550
+ vocab_extra_ids ................................. 0
551
+ vocab_file ...................................... /workspace/data/wikitext/gpt2-vocab.json
552
+ vocab_size ...................................... None
553
+ wandb_exp_name ..................................
554
+ wandb_project ...................................
555
+ wandb_save_dir ..................................
556
+ weight_decay .................................... 0.1
557
+ weight_decay_incr_style ......................... constant
558
+ wgrad_deferral_limit ............................ 0
559
+ world_size ...................................... 8
560
+ yaml_cfg ........................................ None
561
+ -------------------- end of arguments ---------------------
562
+ INFO:megatron.core.num_microbatches_calculator:setting number of microbatches to constant 1
563
+ > building GPT2BPETokenizer tokenizer ...
564
+ > padded vocab (size: 50257) with 47 dummy tokens (new size: 50304)
565
+ WARNING:megatron.core.rerun_state_machine:RerunStateMachine initialized in mode RerunMode.VALIDATE_RESULTS
566
+ > initializing torch distributed ...
567
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
568
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
569
+ WARNING: one_logger package is required to enable e2e metrics tracking. please go to https://confluence.nvidia.com/display/MLWFO/Package+Repositories for details to install it
570
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
571
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
572
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
573
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
574
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
575
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
576
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
577
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
578
+ /workspace/Megatron-MoE-Echo/megatron/training/arguments.py:831: UserWarning: Disabling sequence parallelism because tensor model parallelism is disabled
579
+ warnings.warn("Disabling sequence parallelism because tensor model parallelism is disabled")
580
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
581
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
582
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
583
+ [Gloo] Rank 3 is connected to 7[Gloo] Rank 4 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
584
+ [Gloo] Rank 6 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
585
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
586
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
587
+ peer ranks. Expected number of connected peer ranks is : 7
588
+ [Gloo] Rank 1 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
589
+ [Gloo] Rank 0 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
590
+ [Gloo] Rank 2 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
591
+ [Gloo] Rank 3 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
592
+ [Gloo] Rank [Gloo] Rank 4 is connected to 67 is connected to peer ranks. 7Expected number of connected peer ranks is : peer ranks. 7Expected number of connected peer ranks is :
593
+ 7
594
+ [Gloo] Rank 5 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
595
+ [Gloo] Rank 7 is connected to 7 peer ranks. Expected number of connected peer ranks is : 7
596
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
597
+ > initialized tensor model parallel with size 1
598
+ > initialized pipeline model parallel with size 1
599
+ > setting random seeds to 1234 ...
600
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
601
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
602
+ [Gloo] Rank [Gloo] Rank 00 is connected to is connected to 00 peer ranks. peer ranks. Expected number of connected peer ranks is : Expected number of connected peer ranks is : 00
603
+
604
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
605
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
606
+ [Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
607
+ [rank2]:[W921 07:03:16.766656895 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 2] using GPU 2 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
608
+ [rank6]:[W921 07:03:16.766945196 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 6] using GPU 6 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
609
+ [rank5]:[W921 07:03:16.766991541 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 5] using GPU 5 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
610
+ [rank1]:[W921 07:03:16.767021800 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 1] using GPU 1 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
611
+ [rank7]:[W921 07:03:16.767049971 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 7] using GPU 7 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
612
+ [rank3]:[W921 07:03:16.767090367 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 3] using GPU 3 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
613
+ [rank4]:[W921 07:03:16.767686231 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 4] using GPU 4 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
614
+ > compiling dataset index builder ...
615
+ make: Entering directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
616
+ make: Nothing to be done for 'default'.
617
+ make: Leaving directory '/workspace/Megatron-MoE-Echo/megatron/core/datasets'
618
+ >>> done with dataset index builder. Compilation time: 0.038 seconds
619
+ WARNING: constraints for invoking optimized fused softmax kernel are not met. We default back to unfused kernel invocations.
620
+ > compiling and loading fused kernels ...
621
+ [rank0]:[W921 07:03:17.946978460 ProcessGroupNCCL.cpp:4789] [PG ID 0 PG GUID 0 Rank 0] using GPU 0 as device used by this process is currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect. You can specify device_id in init_process_group() to force use of a particular device.
622
+ >>> done with compiling and loading fused kernels. Compilation time: 3.972 seconds
623
+ time to initialize megatron (seconds): 7.546
624
+ [after megatron is initialized] datetime: 2026-09-21 07:03:23
625
+ building GPT model ...
626
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
627
+ warnings.warn(
628
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
629
+ warnings.warn(
630
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
631
+ warnings.warn(
632
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
633
+ warnings.warn(
634
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
635
+ warnings.warn(
636
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
637
+ warnings.warn(
638
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
639
+ warnings.warn(
640
+ /usr/local/lib/python3.12/dist-packages/transformer_engine/pytorch/cpu_offload.py:708: DeprecationWarning: Offloading weights is deprecated. Using offload_weights=True does not have any effect.
641
+ warnings.warn(
642
+ > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 4113898496
643
+ INFO:megatron.core.distributed.distributed_data_parallel:Setting up DistributedDataParallel with config DistributedDataParallelConfig(grad_reduce_in_fp32=True, overlap_grad_reduce=True, overlap_param_gather=True, align_param_gather=False, use_distributed_optimizer=True, num_distributed_optimizer_instances=1, check_for_nan_in_grad=False, check_for_large_grads=False, bucket_size=40000000, pad_buckets_for_high_nccl_busbw=False, average_in_collective=False, fp8_param_gather=False, reuse_grad_buf_for_mxfp8_param_ag=False, use_custom_fsdp=False, data_parallel_sharding_strategy='no_shard', gradient_reduce_div_fusion=True, suggested_communication_unit_size=None, preserve_fp32_weights=True, keep_fp8_transpose_cache_when_using_custom_fsdp=False, nccl_ub=False, fsdp_double_buffer=False)
644
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 17
645
+ Params for bucket 1 (360579072 elements, 360579072 padded size):
646
+ module.output_layer.weight
647
+ Params for bucket 2 (44047360 elements, 44047360 padded size):
648
+ module.decoder.final_layernorm.weight
649
+ module.decoder.layers.3.mlp.shared_experts.linear_fc1.weight
650
+ module.decoder.layers.3.mlp.shared_experts.linear_fc2.weight
651
+ Params for bucket 3 (60497408 elements, 60497408 padded size):
652
+ module.decoder.layers.3.mlp.router.weight
653
+ module.decoder.layers.3.self_attention.linear_kv_down_proj.weight
654
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.layer_norm_weight
655
+ module.decoder.layers.3.self_attention.linear_q_up_proj.weight
656
+ module.decoder.layers.3.self_attention.linear_kv_up_proj.weight
657
+ module.decoder.layers.3.pre_mlp_layernorm.weight
658
+ Params for bucket 4 (128452096 elements, 128452096 padded size):
659
+ module.decoder.layers.3.self_attention.linear_q_up_proj.layer_norm_weight
660
+ module.decoder.layers.3.self_attention.linear_proj.weight
661
+ module.decoder.layers.3.self_attention.linear_q_down_proj.weight
662
+ Params for bucket 5 (132127744 elements, 132127744 padded size):
663
+ module.decoder.layers.3.input_layernorm.weight
664
+ module.decoder.layers.2.mlp.linear_fc2.weight
665
+ Params for bucket 6 (264241152 elements, 264241152 padded size):
666
+ module.decoder.layers.2.mlp.linear_fc1.weight
667
+ Params for bucket 7 (58662400 elements, 58662400 padded size):
668
+ module.decoder.layers.2.mlp.linear_fc1.layer_norm_weight
669
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.layer_norm_weight
670
+ module.decoder.layers.2.self_attention.linear_kv_up_proj.weight
671
+ module.decoder.layers.2.self_attention.linear_kv_down_proj.weight
672
+ module.decoder.layers.2.self_attention.linear_q_up_proj.weight
673
+ Params for bucket 8 (128452096 elements, 128452096 padded size):
674
+ module.decoder.layers.2.self_attention.linear_q_up_proj.layer_norm_weight
675
+ module.decoder.layers.2.self_attention.linear_proj.weight
676
+ module.decoder.layers.2.self_attention.linear_q_down_proj.weight
677
+ Params for bucket 9 (132127744 elements, 132127744 padded size):
678
+ module.decoder.layers.2.input_layernorm.weight
679
+ module.decoder.layers.1.mlp.linear_fc2.weight
680
+ Params for bucket 10 (264241152 elements, 264241152 padded size):
681
+ module.decoder.layers.1.mlp.linear_fc1.weight
682
+ Params for bucket 11 (58662400 elements, 58662400 padded size):
683
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.layer_norm_weight
684
+ module.decoder.layers.1.self_attention.linear_q_up_proj.weight
685
+ module.decoder.layers.1.self_attention.linear_kv_up_proj.weight
686
+ module.decoder.layers.1.mlp.linear_fc1.layer_norm_weight
687
+ module.decoder.layers.1.self_attention.linear_kv_down_proj.weight
688
+ Params for bucket 12 (128452096 elements, 128452096 padded size):
689
+ module.decoder.layers.1.self_attention.linear_q_up_proj.layer_norm_weight
690
+ module.decoder.layers.1.self_attention.linear_proj.weight
691
+ module.decoder.layers.1.self_attention.linear_q_down_proj.weight
692
+ Params for bucket 13 (132127744 elements, 132127744 padded size):
693
+ module.decoder.layers.1.input_layernorm.weight
694
+ module.decoder.layers.0.mlp.linear_fc2.weight
695
+ Params for bucket 14 (264241152 elements, 264241152 padded size):
696
+ module.decoder.layers.0.mlp.linear_fc1.weight
697
+ Params for bucket 15 (58662400 elements, 58662400 padded size):
698
+ module.decoder.layers.0.self_attention.linear_kv_down_proj.weight
699
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.layer_norm_weight
700
+ module.decoder.layers.0.mlp.linear_fc1.layer_norm_weight
701
+ module.decoder.layers.0.self_attention.linear_kv_up_proj.weight
702
+ module.decoder.layers.0.self_attention.linear_q_up_proj.weight
703
+ Params for bucket 16 (128452096 elements, 128452096 padded size):
704
+ module.decoder.layers.0.self_attention.linear_q_up_proj.layer_norm_weight
705
+ module.decoder.layers.0.self_attention.linear_q_down_proj.weight
706
+ module.decoder.layers.0.self_attention.linear_proj.weight
707
+ Params for bucket 17 (360586240 elements, 360586240 padded size):
708
+ module.decoder.layers.0.input_layernorm.weight
709
+ module.embedding.word_embeddings.weight
710
+ INFO:megatron.core.distributed.param_and_grad_buffer:Number of buckets for gradient all-reduce / reduce-scatter: 27
711
+ Params for bucket 1 (44040192 elements, 44040192 padded size):
712
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight0
713
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight2
714
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight1
715
+ Params for bucket 2 (44040192 elements, 44040192 padded size):
716
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight3
717
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight5
718
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight4
719
+ Params for bucket 3 (44040192 elements, 44040192 padded size):
720
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight6
721
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight8
722
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight7
723
+ Params for bucket 4 (44040192 elements, 44040192 padded size):
724
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight9
725
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight10
726
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight11
727
+ Params for bucket 5 (44040192 elements, 44040192 padded size):
728
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight12
729
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight13
730
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight14
731
+ Params for bucket 6 (44040192 elements, 44040192 padded size):
732
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight15
733
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight16
734
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight17
735
+ Params for bucket 7 (44040192 elements, 44040192 padded size):
736
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight18
737
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight19
738
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight20
739
+ Params for bucket 8 (44040192 elements, 44040192 padded size):
740
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight21
741
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight22
742
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight23
743
+ Params for bucket 9 (44040192 elements, 44040192 padded size):
744
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight24
745
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight25
746
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight26
747
+ Params for bucket 10 (44040192 elements, 44040192 padded size):
748
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight27
749
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight28
750
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight29
751
+ Params for bucket 11 (58720256 elements, 58720256 padded size):
752
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight30
753
+ module.decoder.layers.3.mlp.experts.linear_fc2.weight31
754
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight0
755
+ Params for bucket 12 (58720256 elements, 58720256 padded size):
756
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight1
757
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight2
758
+ Params for bucket 13 (58720256 elements, 58720256 padded size):
759
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight3
760
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight4
761
+ Params for bucket 14 (58720256 elements, 58720256 padded size):
762
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight5
763
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight6
764
+ Params for bucket 15 (58720256 elements, 58720256 padded size):
765
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight7
766
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight8
767
+ Params for bucket 16 (58720256 elements, 58720256 padded size):
768
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight9
769
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight10
770
+ Params for bucket 17 (58720256 elements, 58720256 padded size):
771
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight11
772
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight12
773
+ Params for bucket 18 (58720256 elements, 58720256 padded size):
774
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight13
775
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight14
776
+ Params for bucket 19 (58720256 elements, 58720256 padded size):
777
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight15
778
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight16
779
+ Params for bucket 20 (58720256 elements, 58720256 padded size):
780
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight17
781
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight18
782
+ Params for bucket 21 (58720256 elements, 58720256 padded size):
783
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight19
784
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight20
785
+ Params for bucket 22 (58720256 elements, 58720256 padded size):
786
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight21
787
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight22
788
+ Params for bucket 23 (58720256 elements, 58720256 padded size):
789
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight23
790
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight24
791
+ Params for bucket 24 (58720256 elements, 58720256 padded size):
792
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight25
793
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight26
794
+ Params for bucket 25 (58720256 elements, 58720256 padded size):
795
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight27
796
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight28
797
+ Params for bucket 26 (58720256 elements, 58720256 padded size):
798
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight29
799
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight30
800
+ Params for bucket 27 (29360128 elements, 29360128 padded size):
801
+ module.decoder.layers.3.mlp.experts.linear_fc1.weight31
802
+ INFO:megatron.core.optimizer:Setting up optimizer with config OptimizerConfig(optimizer='adam', lr=0.0001, min_lr=1e-05, decoupled_lr=None, decoupled_min_lr=None, weight_decay=0.1, fp8_recipe='mxfp8', fp16=False, bf16=True, reuse_grad_buf_for_mxfp8_param_ag=False, params_dtype=torch.bfloat16, use_precision_aware_optimizer=False, store_param_remainders=True, main_grads_dtype=torch.float32, main_params_dtype=torch.float32, exp_avg_dtype=torch.float32, exp_avg_sq_dtype=torch.float32, loss_scale=None, initial_loss_scale=4294967296, min_loss_scale=1.0, loss_scale_window=1000, hysteresis=2, adam_beta1=0.9, adam_beta2=0.999, adam_eps=1e-08, sgd_momentum=0.9, use_distributed_optimizer=True, overlap_param_gather_with_optimizer_step=False, optimizer_cpu_offload=False, optimizer_offload_fraction=1.0, use_torch_optimizer_for_cpu_offload=False, overlap_cpu_optimizer_d2h_h2d=False, pin_cpu_grads=True, pin_cpu_params=True, clip_grad=1.0, log_num_zeros_in_grad=False, barrier_with_L1_time=True, timers=<megatron.core.timers.Timers object at 0x791ec3b76e70>, config_logger_dir='')
803
+ INFO:megatron.core.optimizer_param_scheduler:> learning rate decay style: cosine
804
+ [after model, optimizer, and learning rate scheduler are built] datetime: 2026-09-21 07:03:25
805
+ > building train, validation, and test datasets ...
806
+ > datasets target sizes (minimum size):
807
+ train: 96
808
+ validation: 0
809
+ test: 0
810
+ INFO:megatron.core.datasets.blended_megatron_dataset_config:Let split_matrix = [(0, 1.0), None, None]
811
+ > building train, validation, and test datasets for GPT ...
812
+ INFO:megatron.core.datasets.blended_megatron_dataset_builder:Building GPTDataset splits with sizes=(96, 0, 0) and config=GPTDatasetConfig(random_seed=1234, sequence_length=4096, blend=(['/workspace/data/wikitext/wikitext103_gpt2_text_document'], None), blend_per_split=None, split='100,0,0', split_matrix=[(0, 1.0), None, None], num_dataset_builder_threads=1, path_to_cache=None, mmap_bin_files=True, mock=False, tokenizer=<megatron.training.tokenizer.tokenizer._GPT2BPETokenizer object at 0x791ec3e02420>, mid_level_dataset_surplus=0.005, reset_position_ids=False, reset_attention_mask=False, eod_mask_loss=False, create_attention_mask=True, drop_last_partial_validation_sequence=True, add_extra_token_to_sequence=True, object_storage_cache_path=None)
813
+ INFO:megatron.core.datasets.indexed_dataset:Load the _IndexReader from /workspace/data/wikitext/wikitext103_gpt2_text_document.idx
814
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence lengths
815
+ INFO:megatron.core.datasets.indexed_dataset: Extract the sequence pointers
816
+ INFO:megatron.core.datasets.indexed_dataset: Extract the document indices
817
+ INFO:megatron.core.datasets.indexed_dataset:> total number of sequences: 1165029
818
+ INFO:megatron.core.datasets.indexed_dataset:> total number of documents: 1801350
819
+ INFO:megatron.core.datasets.gpt_dataset:Load the GPTDataset train indices
820
+ INFO:megatron.core.datasets.gpt_dataset: Load the document index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-document_index.npy
821
+ INFO:megatron.core.datasets.gpt_dataset: Load the sample index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-sample_index.npy
822
+ INFO:megatron.core.datasets.gpt_dataset: Load the shuffle index from 466b63c71b42dae59ebdc25db6336051-GPTDataset-train-shuffle_index.npy
823
+ INFO:megatron.core.datasets.gpt_dataset:> total number of samples: 29073
824
+ > finished creating GPT datasets ...
825
+ [after dataloaders are built] datetime: 2026-09-21 07:03:25
826
+ done with setup ...
827
+ (min, max) time across ranks (ms):
828
+ model-and-optimizer-setup ......................: (2252.22, 2252.42)
829
+ train/valid/test-data-iterators-setup ..........: (67.39, 70.03)
830
+ training ...
831
+ Overwriting rerun_state_machine.current_iteration from -1 to 0...
832
+ [before the start of training step] datetime: 2026-09-21 07:03:25
833
+ INFO:moonep.api:MoonEP comm buffer sizes:
834
+ config : R=8, S=4096, H=7168, E=256, K=8, NvS=40896, NvS_padded=40960
835
+ sm_count : num_sms=32, num_sms_dedup=148
836
+ hidden_buffer_size : 560.00 MiB
837
+ meta_buffer_size : 2.00 MiB
838
+ local_temp_buffer_size : 1.27 MiB
839
+ total_size : 563.27 MiB
840
+ [MoonEP selfcheck 1/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [768, 640, 640, 1024, 640, 640, 640, 512, 1280, 640, 896, 640, 1536, 2304, 1536, 896, 1408, 384, 1280, 1664, 1152, 640, 640, 1152, 1280, 1536, 768, 512, 1664, 384, 384, 896, 2432, 1792, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[215, 197, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [194, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [205, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [78, 91, 86, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [203, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [50, 32, 34, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
841
+ [2026-09-21 07:03:50] iteration 1/ 12 | consumed samples: 8 | elapsed time per iteration (ms): 24495.9 | throughput per GPU (TFLOP/s/GPU): 3.0 | learning rate: 2.000000E-05 | global batch size: 8 | lm loss: 1.226155E+01 | loss scale: 1.0 | grad norm: 15.011 | number of skipped iterations: 0 | number of nan iterations: 0 |
842
+ Number of parameters in transformer block in billions: 13.26
843
+ Number of parameters in embedding layers in billions: 0.72
844
+ Total number of parameters in billions: 13.98
845
+ Number of parameters in most loaded shard in billions: 13.9771
846
+ Theoretical memory footprints: weight and optimizer=99972.23 MB
847
+ [Rank 0] (after 1 iterations) memory (MB) | allocated: 46904.38818359375 | max allocated: 57624.96875 | reserved: 60652.0 | max reserved: 60652.0
848
+ [MoonEP selfcheck 2/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [768, 768, 768, 640, 512, 640, 768, 512, 1408, 640, 896, 768, 1408, 2560, 1408, 1024, 1280, 512, 1664, 1664, 1152, 640, 640, 1024, 1280, 1408, 640, 512, 1536, 384, 384, 896, 2176, 1664, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[205, 197, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [158, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [151, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, 78, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [50, 32, 34, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [91, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [215, 194, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
849
+ [2026-09-21 07:03:50] iteration 2/ 12 | consumed samples: 16 | elapsed time per iteration (ms): 200.4 | throughput per GPU (TFLOP/s/GPU): 371.5 | learning rate: 4.000000E-05 | global batch size: 8 | lm loss: 1.227770E+01 | loss scale: 1.0 | grad norm: 14.853 | number of skipped iterations: 0 | number of nan iterations: 0 |
850
+ [MoonEP selfcheck 3/4] dropped groups empty OK | covered rows [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 256, 257, 258, 259, 260, 261, 262, 263, 264, 265, 266, 267, 268, 269, 270, 271, 272, 273, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 284, 285, 286, 287] sizes [384, 640, 1152, 512, 384, 768, 896, 256, 896, 512, 1024, 768, 1664, 1920, 1024, 896, 1408, 256, 1152, 1792, 256, 768, 384, 896, 768, 1024, 384, 256, 640, 384, 384, 1024, 5248, 3200, 1280, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] | duplicated experts [[50, 34, 62, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [66, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [151, 158, 137, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [81, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [82, 78, 91, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [239, 252, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1], [52, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1]]
851
+ [2026-09-21 07:03:50] iteration 3/ 12 | consumed samples: 24 | elapsed time per iteration (ms): 199.0 | throughput per GPU (TFLOP/s/GPU): 374.1 | learning rate: 6.000000E-05 | global batch size: 8 | lm loss: 1.070648E+01 | loss scale: 1.0 | grad norm: 18.850 | number of skipped iterations: 0 | number of nan iterations: 0 |
852
+ Capture CUDA graph for training!!!
853
+ CUDA graph capture done!!!
854
+ [2026-09-21 07:03:51] iteration 4/ 12 | consumed samples: 32 | elapsed time per iteration (ms): 751.7 | throughput per GPU (TFLOP/s/GPU): 99.1 | learning rate: 8.000000E-05 | global batch size: 8 | lm loss: 1.557410E+01 | loss scale: 1.0 | grad norm: 97.887 | number of skipped iterations: 0 | number of nan iterations: 0 |
855
+ [2026-09-21 07:03:51] iteration 5/ 12 | consumed samples: 40 | elapsed time per iteration (ms): 193.5 | throughput per GPU (TFLOP/s/GPU): 384.8 | learning rate: 1.000000E-04 | global batch size: 8 | lm loss: 1.447998E+01 | loss scale: 1.0 | grad norm: 97.902 | number of skipped iterations: 0 | number of nan iterations: 0 |
856
+ [2026-09-21 07:03:51] iteration 6/ 12 | consumed samples: 48 | elapsed time per iteration (ms): 192.6 | throughput per GPU (TFLOP/s/GPU): 386.7 | learning rate: 7.750000E-05 | global batch size: 8 | lm loss: 1.293705E+01 | loss scale: 1.0 | grad norm: 16.034 | number of skipped iterations: 0 | number of nan iterations: 0 |
857
+ [2026-09-21 07:03:51] iteration 7/ 12 | consumed samples: 56 | elapsed time per iteration (ms): 193.2 | throughput per GPU (TFLOP/s/GPU): 385.5 | learning rate: 3.250000E-05 | global batch size: 8 | lm loss: 1.689716E+01 | loss scale: 1.0 | grad norm: 100.878 | number of skipped iterations: 0 | number of nan iterations: 0 |
858
+ [2026-09-21 07:03:52] iteration 8/ 12 | consumed samples: 64 | elapsed time per iteration (ms): 192.4 | throughput per GPU (TFLOP/s/GPU): 387.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.098987E+01 | loss scale: 1.0 | grad norm: 18.714 | number of skipped iterations: 0 | number of nan iterations: 0 |
859
+ [2026-09-21 07:03:52] iteration 9/ 12 | consumed samples: 72 | elapsed time per iteration (ms): 193.0 | throughput per GPU (TFLOP/s/GPU): 385.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.048015E+01 | loss scale: 1.0 | grad norm: 16.061 | number of skipped iterations: 0 | number of nan iterations: 0 |
860
+ [2026-09-21 07:03:52] iteration 10/ 12 | consumed samples: 80 | elapsed time per iteration (ms): 192.3 | throughput per GPU (TFLOP/s/GPU): 387.1 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 1.017142E+01 | loss scale: 1.0 | grad norm: 11.421 | number of skipped iterations: 0 | number of nan iterations: 0 |
861
+ [2026-09-21 07:03:52] iteration 11/ 12 | consumed samples: 88 | elapsed time per iteration (ms): 192.5 | throughput per GPU (TFLOP/s/GPU): 386.9 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.971597E+00 | loss scale: 1.0 | grad norm: 11.756 | number of skipped iterations: 0 | number of nan iterations: 0 |
862
+ [2026-09-21 07:03:52] iteration 12/ 12 | consumed samples: 96 | elapsed time per iteration (ms): 193.3 | throughput per GPU (TFLOP/s/GPU): 385.2 | learning rate: 1.000000E-05 | global batch size: 8 | lm loss: 9.694415E+00 | loss scale: 1.0 | grad norm: 17.317 | number of skipped iterations: 0 | number of nan iterations: 0 |
863
+ [after training is done] datetime: 2026-09-21 07:03:52
864
+ [rank0]:[W921 07:03:54.044060769 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
865
+ [rank3]:[W921 07:03:55.131340753 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
866
+ [rank2]:[W921 07:03:55.134556809 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
867
+ [rank5]:[W921 07:03:55.135464656 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
868
+ [rank4]:[W921 07:03:55.136406419 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
869
+ [rank6]:[W921 07:03:55.177571237 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
870
+ [rank1]:[W921 07:03:55.541370244 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
871
+ [rank7]:[W921 07:03:55.663826523 ProcessGroupNCCL.cpp:1505] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())
872
+ W0921 07:04:00.293000 365649 torch/distributed/elastic/agent/server/api.py:719] Received 15 death signal, shutting down workers
873
+ W0921 07:04:00.294000 365649 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 365842 closing signal SIGTERM
874
+ W0921 07:04:00.294000 365649 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 365844 closing signal SIGTERM
875
+ W0921 07:04:00.295000 365649 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 365847 closing signal SIGTERM
876
+ W0921 07:04:00.295000 365649 torch/distributed/elastic/multiprocessing/api.py:900] Sending process 365848 closing signal SIGTERM
877
+ Traceback (most recent call last):
878
+ File "/usr/local/bin/torchrun", line 8, in <module>
879
+ sys.exit(main())
880
+ ^^^^^^
881
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper
882
+ return f(*args, **kwargs)
883
+ ^^^^^^^^^^^^^^^^^^
884
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 892, in main
885
+ run(args)
886
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 883, in run
887
+ elastic_launch(
888
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 139, in __call__
889
+ return launch_agent(self._config, self._entrypoint, list(args))
890
+ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
891
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 261, in launch_agent
892
+ result = agent.run()
893
+ ^^^^^^^^^^^
894
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper
895
+ result = f(*args, **kwargs)
896
+ ^^^^^^^^^^^^^^^^^^
897
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 711, in run
898
+ result = self._invoke_run(role)
899
+ ^^^^^^^^^^^^^^^^^^^^^^
900
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 870, in _invoke_run
901
+ time.sleep(monitor_interval)
902
+ File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 84, in _terminate_process_handler
903
+ raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval)
904
+ torch.distributed.elastic.multiprocessing.api.SignalException: Process 365649 got signal: 15
granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/results.tsv CHANGED
@@ -1,2 +1,4 @@
1
  config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
2
  DEEPSEEK_V3_L4 moonep matched 32 4096 8 none none 1 OK 192.67 192.30 97849 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1
 
 
 
1
  config backend policy slots seq gbs cap recompute rep status iter_ms_mean iter_ms_min peak_mib run_dir
2
  DEEPSEEK_V3_L4 moonep matched 32 4096 8 none none 1 OK 192.67 192.30 97849 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep1
3
+ DEEPSEEK_V3_L4 moonep matched 32 4096 8 none none 2 OK 192.97 192.10 97849 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep2
4
+ DEEPSEEK_V3_L4 moonep matched 32 4096 8 none none 3 OK 192.90 192.10 97849 /workspace/granularity_next_b200/tierb_models/attempts/DEEPSEEK_V3_L4/moonep/tierb/tierb_gbs8/DEEPSEEK_V3_L4__moonep__matched__seq4096__gbs8__capnone__async/rep3