Hugging Face's logo

Datasets:
quintic
/
rig-logs

Tags:
training-logs
scaling-laws
hyperparameter-transfer
tpu
jax
License:
Dataset card Files
xet
 Community
quintic commited on
Commit
e271fb0
·
verified ·
1 Parent(s): 72f164f

Add the 8k batch-size grid: 42 runs, dense and routed, batch 32 and 64 (part 2)

Browse files
Files changed (46) hide show
  1. .gitattributes +7 -0
  2. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1339/training.riglog +3 -0
  3. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1339/validation.csv +2 -0
  4. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/diagnostics.riglog +0 -0
  5. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/metrics.json +244 -0
  6. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/result.json +244 -0
  7. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/stderr.log +113 -0
  8. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/stdout.log +1 -0
  9. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/training.riglog +3 -0
  10. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/validation.csv +2 -0
  11. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/diagnostics.riglog +0 -0
  12. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/metrics.json +244 -0
  13. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/result.json +244 -0
  14. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/stderr.log +113 -0
  15. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/stdout.log +1 -0
  16. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/training.riglog +3 -0
  17. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/validation.csv +2 -0
  18. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/diagnostics.riglog +0 -0
  19. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/metrics.json +244 -0
  20. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/result.json +244 -0
  21. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/stderr.log +113 -0
  22. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/stdout.log +1 -0
  23. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/training.riglog +3 -0
  24. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/validation.csv +2 -0
  25. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/diagnostics.riglog +0 -0
  26. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/metrics.json +244 -0
  27. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/result.json +244 -0
  28. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/stderr.log +113 -0
  29. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/stdout.log +1 -0
  30. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/training.riglog +3 -0
  31. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/validation.csv +2 -0
  32. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/diagnostics.riglog +0 -0
  33. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/metrics.json +244 -0
  34. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/result.json +244 -0
  35. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/stderr.log +113 -0
  36. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/stdout.log +1 -0
  37. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/training.riglog +3 -0
  38. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/validation.csv +2 -0
  39. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/diagnostics.riglog +0 -0
  40. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/metrics.json +244 -0
  41. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/result.json +244 -0
  42. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/stderr.log +113 -0
  43. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/stdout.log +1 -0
  44. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/training.riglog +3 -0
  45. batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/validation.csv +2 -0
  46. batch-size-grid-8k/README.md +153 -0
.gitattributes CHANGED
@@ -360,3 +360,10 @@ batch-size-grid-8k/60m-moe-5tpp-bs32-lr2e-9-s1339/diagnostics.riglog filter=lfs
360
  batch-size-grid-8k/60m-moe-5tpp-bs32-lr2e-9-s1339/training.riglog filter=lfs diff=lfs merge=lfs -text
361
  batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1337/training.riglog filter=lfs diff=lfs merge=lfs -text
362
  batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1338/training.riglog filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
360
  batch-size-grid-8k/60m-moe-5tpp-bs32-lr2e-9-s1339/training.riglog filter=lfs diff=lfs merge=lfs -text
361
  batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1337/training.riglog filter=lfs diff=lfs merge=lfs -text
362
  batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1338/training.riglog filter=lfs diff=lfs merge=lfs -text
363
+ batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1339/training.riglog filter=lfs diff=lfs merge=lfs -text
364
+ batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/training.riglog filter=lfs diff=lfs merge=lfs -text
365
+ batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/training.riglog filter=lfs diff=lfs merge=lfs -text
366
+ batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/training.riglog filter=lfs diff=lfs merge=lfs -text
367
+ batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/training.riglog filter=lfs diff=lfs merge=lfs -text
368
+ batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/training.riglog filter=lfs diff=lfs merge=lfs -text
369
+ batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/training.riglog filter=lfs diff=lfs merge=lfs -text
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1339/training.riglog ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d864a8986d9f1923f5d5015c58db30c83c643a0362f29e5e249d99af8eeac91b
3
+ size 327744
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-7-s1339/validation.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ step,tokens_processed,kind,domain,validation_tokens,validation_loss,perplexity,validation_seconds,canonical
2
+ 571,299368448,fineweb,fineweb,4194304,5.097669720649719,163.64013542664168,1.8341180770657957,true
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/diagnostics.riglog ADDED
Binary file (75.1 kB). View file
 
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/metrics.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.216764271259308,
32
+ "perplexity": 184.33675429774732,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8335649539949372
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.0078125,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.00390625,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.8513281729704,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.00390625,
142
+ "diagnostic_compile_seconds": 91.04871499002911,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.12211701797787,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8335649539949372,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13655712003931145,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 814277.3583970269,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 175.4393858370604,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 61.26855382905342,
176
+ "train_loss": 5.185926914215088,
177
+ "train_seconds": 367.64923513203394,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.216764271259308,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1337,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/result.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.216764271259308,
32
+ "perplexity": 184.33675429774732,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8335649539949372
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.0078125,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.00390625,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.8513281729704,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.00390625,
142
+ "diagnostic_compile_seconds": 91.04871499002911,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.12211701797787,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8335649539949372,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13655712003931145,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 814277.3583970269,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 175.4393858370604,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 61.26855382905342,
176
+ "train_loss": 5.185926914215088,
177
+ "train_seconds": 367.64923513203394,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.216764271259308,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1337,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/stderr.log ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ E0819 11:04:18.835957 3436728 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
2
+ E0819 11:04:18.836072 3483082 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
3
+ E0819 11:04:18.836367 228211 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
4
+ E0819 11:04:18.836601 3383708 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
5
+
6
+ ◆ GPT TPU RIG reference / jax
7
+
8
+
9
+ ● Attention tile preflight resolving the shipped lookup or shape heuristic
10
+ ╭─ run configuration ────────────────────────────────────────────────────────╮
11
+ │ experiment config config.yaml · dev · sha256:a4bad1be29b5 │
12
+ │ devices 16 × TPU v4 │
13
+ │ JAX processes 4 (this rank 1) │
14
+ │ mesh data=16 (replicated model) │
15
+ │ dataset 79 train + 1 val shard(s) │
16
+ │ train / val tokens 7,900,000,000 / 100,000,000 │
17
+ │ downstream not requested │
18
+ │ model 60m · L12 D384 H6 RoPE RMSNorm GELU MLP×4 │
19
+ │ parameters 102.51M │
20
+ │ parameterization complete_d_p · mN=1 · mL=1 · mD=1 │
21
+ │ global batch 64 × 8192 tokens │
22
+ │ train sampling shuffled epochs · 7,899,979,776 unique targets/epoch │
23
+ │ compute bfloat16 │
24
+ │ attention tpu_flash │
25
+ │ attention tuning heuristic · key 6a5b398e13ed │
26
+ │ attention fwd q512 · kv512/256 │
27
+ │ attention dK/dV q512/256 · kv512/256 │
28
+ │ attention dQ q256 · kv512/256 │
29
+ │ output loss tiled CE (semantic 50,304, tile 2,048) │
30
+ │ diagnostics step 1 / every 10 / final │
31
+ │ duration 571 steps │
32
+ │ train tokens 299.37M │
33
+ │ traced FLOPs 220902.53T │
34
+ │ FLOP breakdown dot_general 1,638,200,770,560 (27.1%) · tpu_flash_ca… │
35
+ │ XProf disabled │
36
+ ╰────────────────────────────────────────────────────────────────────────────╯
37
+
38
+ ● Compiling train step compilation is outside train_seconds
39
+
40
+ ● Compiling sparse diagnostics separate executable; compilation is outside train_seconds
41
+
42
+ ● Compiling evaluation reused by probes and final validation
43
+
44
+ ● Training train compiled in 61.27s, eval in 23.12s; periodic validation disabled
45
+ 1/571 ────────────────── loss 10.9476 lr 1.37e-04 |g| 2.514 595.27K tok/s
46
+ 10/571 ────────────────── loss 8.5016 lr 1.37e-03 |g| 1.613 760.21K tok/s
47
+ 20/571 ━───────────────── loss 7.7458 lr 2.74e-03 |g| 0.747 787.30K tok/s
48
+ 30/571 ━───────────────── loss 7.7264 lr 4.11e-03 |g| 0.259 797.02K tok/s
49
+ 40/571 ━───────────────── loss 7.8550 lr 5.48e-03 |g| 0.567 802.00K tok/s
50
+ 50/571 ━━──────────────── loss 7.7473 lr 6.85e-03 |g| 0.309 805.11K tok/s
51
+ 60/571 ━━──────────────── loss 7.7232 lr 7.81e-03 |g| 0.865 807.05K tok/s
52
+ 70/571 ━━──────────────── loss 7.5656 lr 7.80e-03 |g| 0.332 808.41K tok/s
53
+ 80/571 ━━━─────────────── loss 7.5288 lr 7.78e-03 |g| 0.605 809.37K tok/s
54
+ 90/571 ━━━─────────────── loss 7.4382 lr 7.74e-03 |g| 0.378 810.11K tok/s
55
+ 100/571 ━━━─────────────── loss 7.3467 lr 7.69e-03 |g| 0.469 810.68K tok/s
56
+ 110/571 ━━━─────────────── loss 7.2200 lr 7.63e-03 |g| 0.387 811.13K tok/s
57
+ 120/571 ━━━━────────────── loss 7.1237 lr 7.56e-03 |g| 0.274 811.50K tok/s
58
+ 130/571 ━━━━────────────── loss 7.1441 lr 7.47e-03 |g| 0.282 811.80K tok/s
59
+ 140/571 ━━━━────────────── loss 6.9707 lr 7.37e-03 |g| 0.290 812.04K tok/s
60
+ 150/571 ━━━━━───────────── loss 6.9220 lr 7.26e-03 |g| 0.378 812.26K tok/s
61
+ 160/571 ━━━━━───────────── loss 6.8302 lr 7.14e-03 |g| 0.348 812.43K tok/s
62
+ 170/571 ━━━━━───────────── loss 6.7753 lr 7.01e-03 |g| 0.409 812.59K tok/s
63
+ 180/571 ━━━━━━──────────── loss 6.6405 lr 6.86e-03 |g| 0.360 812.73K tok/s
64
+ 190/571 ━━━━━━──────────── loss 6.5236 lr 6.71e-03 |g| 0.278 812.85K tok/s
65
+ 200/571 ━━━━━━──────────── loss 6.4873 lr 6.55e-03 |g| 0.145 812.96K tok/s
66
+ 210/571 ━━━━━━━─────────── loss 6.3776 lr 6.38e-03 |g| 0.223 813.06K tok/s
67
+ 220/571 ━━━━━━━─────────── loss 6.3280 lr 6.21e-03 |g| 0.205 813.15K tok/s
68
+ 230/571 ━━━━━━━─────────── loss 6.2979 lr 6.02e-03 |g| 0.402 813.24K tok/s
69
+ 240/571 ━━━━━━━━────────── loss 6.1941 lr 5.83e-03 |g| 0.175 813.31K tok/s
70
+ 250/571 ━━━━━━━━────────── loss 6.1359 lr 5.64e-03 |g| 0.174 813.38K tok/s
71
+ 260/571 ━━━━━━━━────────── loss 6.0785 lr 5.44e-03 |g| 0.223 813.44K tok/s
72
+ 270/571 ━━━━━━━━━───────── loss 6.0766 lr 5.23e-03 |g| 0.230 813.50K tok/s
73
+ 280/571 ━━━━━━━━━───────── loss 6.0063 lr 5.02e-03 |g| 0.199 813.55K tok/s
74
+ 290/571 ━━━━━━━━━───────── loss 6.0268 lr 4.81e-03 |g| 0.187 813.60K tok/s
75
+ 300/571 ━━━━━━━━━───────── loss 5.8269 lr 4.60e-03 |g| 0.237 813.65K tok/s
76
+ 310/571 ━━━━━━━━━━──────── loss 5.8347 lr 4.38e-03 |g| 0.183 813.69K tok/s
77
+ 320/571 ━━━━━━━━━━──────── loss 5.7255 lr 4.17e-03 |g| 0.157 813.74K tok/s
78
+ 330/571 ━━━━━━━━━━──────── loss 5.8136 lr 3.95e-03 |g| 0.225 813.77K tok/s
79
+ 340/571 ━━━━━━━━━━━─────── loss 5.7877 lr 3.74e-03 |g| 0.187 813.81K tok/s
80
+ 350/571 ━━━━━━━━━━━─────── loss 5.6581 lr 3.53e-03 |g| 0.131 813.85K tok/s
81
+ 360/571 ━━━━━━━━━━━─────── loss 5.6523 lr 3.32e-03 |g| 0.176 813.88K tok/s
82
+ 370/571 ━━━━━━━━━━━━────── loss 5.5914 lr 3.12e-03 |g| 0.147 813.91K tok/s
83
+ 380/571 ━━━━━━━━━━━━────── loss 5.5817 lr 2.92e-03 |g| 0.100 813.94K tok/s
84
+ 390/571 ━━━━━━━━━━━━────── loss 5.5905 lr 2.72e-03 |g| 0.128 813.97K tok/s
85
+ 400/571 ━━━━━━━━━━━━━───── loss 5.4892 lr 2.53e-03 |g| 0.115 813.99K tok/s
86
+ 410/571 ━━━━━━━━━━━━━───── loss 5.5201 lr 2.35e-03 |g| 0.128 814.01K tok/s
87
+ 420/571 ━━━━━━━━━━━━━───── loss 5.4634 lr 2.18e-03 |g| 0.139 814.04K tok/s
88
+ 430/571 ━━━━━━━━━━━━━━──── loss 5.3849 lr 2.01e-03 |g| 0.089 814.06K tok/s
89
+ 440/571 ━━━━━━━━━━━━━━──── loss 5.4176 lr 1.85e-03 |g| 0.083 814.09K tok/s
90
+ 450/571 ━━━━━━━━━━━━━━──── loss 5.4216 lr 1.70e-03 |g| 0.113 814.11K tok/s
91
+ 460/571 ━━━━━━━━━━━━━━━─── loss 5.3861 lr 1.56e-03 |g| 0.109 814.13K tok/s
92
+ 470/571 ━━━━━━━━━━━━━━━─── loss 5.3813 lr 1.43e-03 |g| 0.075 814.15K tok/s
93
+ 480/571 ━━━━━━━━━━━━━━━─── loss 5.3850 lr 1.31e-03 |g| 0.097 814.17K tok/s
94
+ 490/571 ━━━━━━━━━━━━━━━─── loss 5.3286 lr 1.20e-03 |g| 0.071 814.19K tok/s
95
+ 500/571 ━━━━━━━━━━━━━━━━── loss 5.2432 lr 1.11e-03 |g| 0.072 814.21K tok/s
96
+ 510/571 ━━━━━━━━━━━━━━━━── loss 5.2668 lr 1.02e-03 |g| 0.065 814.22K tok/s
97
+ 520/571 ━━━━━━━━━━━━━━━━── loss 5.3000 lr 9.51e-04 |g| 0.085 814.24K tok/s
98
+ 530/571 ━━━━━━━━━━━━━━━━━─ loss 5.2356 lr 8.91e-04 |g| 0.079 814.25K tok/s
99
+ 540/571 ━━━━━━━━━━━━━━━━━─ loss 5.2194 lr 8.44e-04 |g| 0.100 814.27K tok/s
100
+ 550/571 ━━━━━━━━━━━━━━━━━─ loss 5.2513 lr 8.10e-04 |g| 0.077 814.29K tok/s
101
+ 560/571 ━━━━━━━━━━━━━━━━━━ loss 5.2538 lr 7.89e-04 |g| 0.071 814.30K tok/s
102
+ 570/571 ━━━━━━━━━━━━━━━━━━ loss 5.2368 lr 7.81e-04 |g| 0.081 814.31K tok/s
103
+ 571/571 ━━━━━━━━━━━━━━━━━━ loss 5.1859 lr 7.81e-04 |g| 0.078 814.29K tok/s
104
+
105
+ ● Canonical validation 8 deterministic batches outside train_seconds
106
+
107
+ ● Fresh-domain validation skipped; no downstream data supplied
108
+
109
+ ● Artifacts training.riglog + validation.csv + diagnostics.riglog
110
+
111
+ ✓ synchronized training 367.649s (compilation excluded)
112
+ validation loss 5.2168 in 1.834s
113
+
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/stdout.log ADDED
@@ -0,0 +1 @@
 
 
1
+ RIG_RESULT={"artifacts":{"diagnostics":"diagnostics.riglog","training_curve":"training.riglog","validation_curve":"validation.csv"},"checkpoint":null,"contract":{"dataset_id":"fineweb-8b-gpt2","model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"model_id":"reference-gpt-v3-family","sequence_length":8192,"tokenizer_id":"gpt2"},"evaluations":{"fineweb":{"canonical":true,"loss":5.216764271259308,"perplexity":184.33675429774732,"scored_tokens":4194304,"seconds":1.8335649539949372}},"implementation":{"attention_backend":"tpu_flash","attention_tuning":{"key_digest":"6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8","resolution_source":"heuristic","tiles":{"block_kv":512,"block_kv_compute":256,"block_kv_dkv":512,"block_kv_dkv_compute":256,"block_kv_dq":512,"block_kv_dq_compute":256,"block_q":512,"block_q_dkv":512,"block_q_dkv_compute":256,"block_q_dq":256},"tune_seconds":0.0},"configuration":{"overrides":{"study_batch_size":64,"tokens_per_parameter_micros":5000000},"path":"config.yaml","profile":"dev","resolved":{"evaluation":{"eval_batches":8,"val_every":0,"val_probe_batches":8},"kernels":{"attention_backend":"tpu_flash","loss_backend":"tiled","vocab_tile_size":2048},"logging":{"diagnostics_every":10,"log_every":10},"model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"optimizer":{"adam_epsilon":1e-08,"beta1":0.9,"beta2":0.95,"effective":{"adam_epsilon_horizon_multiplier":0.5,"beta1":0.6000000000000001,"beta2":0.7999999999999998,"global_peak_learning_rate":0.0078125,"weight_decay_horizon_multiplier":2.0},"grad_clip":0.0,"learning_rate":0.00390625,"min_lr_ratio":0.1,"warmup_steps":57,"weight_decay":0.1},"parameterization":{"attention_scale":"inverse_head_dim","base_depth":12,"base_width":384,"batch_multiplier":4.0,"data_multiplier":1.0,"depth_alpha":1.0,"depth_multiplier":1.0,"embeddings":"untied","init_std":0.02,"name":"complete_d_p","width_multiplier":1.0},"training":{"batch_size":64,"dtype":"bfloat16","sampling":"shuffled_epochs","seq_len":8192,"steps":571,"tokens_per_parameter":4.996285069139584,"train_tokens":299368448}},"schema_version":2,"sha256":"a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"},"loss_backend":"tiled","vocab_tile_size":2048},"metrics":{"achieved_tflops":600.8513281729704,"attention_tune_seconds":0.0,"base_learning_rate":0.00390625,"diagnostic_compile_seconds":91.04871499002911,"diagnostic_point_count":59,"diagnostics_every":10,"early_stopping_step":null,"estimated_total_flops":220902531230859264,"eval_compile_seconds":23.12211701797787,"expert_top_k":2,"experts":8,"final_validation_seconds":1.8335649539949372,"flop_accounting":{"by_site":{"dot_general":1638200770560,"pallas_call":695784701952,"tpu_flash_causal_attention_bwd_dkv":1236950581248,"tpu_flash_causal_attention_bwd_dq":1236950581248,"tpu_flash_causal_attention_fwd":1236950581248},"elementwise_per_sequence":9518381752,"matmul_per_sequence":6044837216256,"method":"traced-jaxpr","warnings":[]},"flops_per_token":737895168,"mfu_estimate":0.13655712003931145,"model_tier":"60m","parameter_count":59959680,"parameters":102509952,"schedule_steps":571,"tokens_per_parameter":4.996285069139584,"tokens_per_second":814277.3583970269,"tokens_processed":299368448,"total_compile_seconds":175.4393858370604,"total_parameter_count":102509952,"train_compile_seconds":61.26855382905342,"train_loss":5.185926914215088,"train_seconds":367.64923513203394,"training_data_epochs":0.037894837263947954,"training_data_sharding":"rank_disjoint_shuffled_windows","training_sampling":"shuffled_epochs","training_steps":571,"training_token_budget":299368448,"training_usable_tokens_per_epoch":7899979776,"validation_loss":5.216764271259308,"validation_probe_count":0,"validation_probe_seconds":0.0,"validation_tokens":4194304},"profile":"dev","schema_version":1,"seed":1337,"status":"ok","system":{"controller_process_index":1,"device_count":16,"device_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],"device_kinds":["TPU v4"],"jax_version":"0.11.0","jaxlib_version":"0.11.0","libtpu_version":"0.0.44.1","local_device_count":4,"platform":"tpu","process_count":4,"process_indices":[0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3],"python_version":"3.12.13"},"track":"open"}
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/training.riglog ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dabcf17d8aa423d3c0c379e04f2dfeefbe4339a1fbe0810c3d59985c063a8d12
3
+ size 327744
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1337/validation.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ step,tokens_processed,kind,domain,validation_tokens,validation_loss,perplexity,validation_seconds,canonical
2
+ 571,299368448,fineweb,fineweb,4194304,5.216764271259308,184.33675429774732,1.8335649539949372,true
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/diagnostics.riglog ADDED
Binary file (75.1 kB). View file
 
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/metrics.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.135362327098846,
32
+ "perplexity": 169.92587747222018,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8413573480211198
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.0078125,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.00390625,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 601.0030472181179,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.00390625,
142
+ "diagnostic_compile_seconds": 94.93301027396228,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 22.61157763702795,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8413573480211198,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13659160164048134,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 814482.9689657459,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 175.76108301396016,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 58.21649510296993,
176
+ "train_loss": 5.133830547332764,
177
+ "train_seconds": 367.55642463604454,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.135362327098846,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1338,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/result.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.135362327098846,
32
+ "perplexity": 169.92587747222018,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8413573480211198
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.0078125,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.00390625,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 601.0030472181179,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.00390625,
142
+ "diagnostic_compile_seconds": 94.93301027396228,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 22.61157763702795,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8413573480211198,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13659160164048134,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 814482.9689657459,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 175.76108301396016,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 58.21649510296993,
176
+ "train_loss": 5.133830547332764,
177
+ "train_seconds": 367.55642463604454,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.135362327098846,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1338,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/stderr.log ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ E0819 11:13:46.635471 3445595 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
2
+ E0819 11:13:46.635715 3491970 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
3
+ E0819 11:13:46.635780 238047 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
4
+ E0819 11:13:46.635789 3392591 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
5
+
6
+ ◆ GPT TPU RIG reference / jax
7
+
8
+
9
+ ● Attention tile preflight resolving the shipped lookup or shape heuristic
10
+ ╭─ run configuration ────────────────────────────────────────────────────────╮
11
+ │ experiment config config.yaml · dev · sha256:a4bad1be29b5 │
12
+ │ devices 16 × TPU v4 │
13
+ │ JAX processes 4 (this rank 1) │
14
+ │ mesh data=16 (replicated model) │
15
+ │ dataset 79 train + 1 val shard(s) │
16
+ │ train / val tokens 7,900,000,000 / 100,000,000 │
17
+ │ downstream not requested │
18
+ │ model 60m · L12 D384 H6 RoPE RMSNorm GELU MLP×4 │
19
+ │ parameters 102.51M │
20
+ │ parameterization complete_d_p · mN=1 · mL=1 · mD=1 │
21
+ │ global batch 64 × 8192 tokens │
22
+ │ train sampling shuffled epochs · 7,899,979,776 unique targets/epoch │
23
+ │ compute bfloat16 │
24
+ │ attention tpu_flash │
25
+ │ attention tuning heuristic · key 6a5b398e13ed │
26
+ │ attention fwd q512 · kv512/256 │
27
+ │ attention dK/dV q512/256 · kv512/256 │
28
+ │ attention dQ q256 · kv512/256 │
29
+ │ output loss tiled CE (semantic 50,304, tile 2,048) │
30
+ │ diagnostics step 1 / every 10 / final │
31
+ │ duration 571 steps │
32
+ │ train tokens 299.37M │
33
+ │ traced FLOPs 220902.53T │
34
+ │ FLOP breakdown dot_general 1,638,200,770,560 (27.1%) · tpu_flash_ca… │
35
+ │ XProf disabled │
36
+ ╰────────────────────────────────────────────────────────────────────────────╯
37
+
38
+ ● Compiling train step compilation is outside train_seconds
39
+
40
+ ● Compiling sparse diagnostics separate executable; compilation is outside train_seconds
41
+
42
+ ● Compiling evaluation reused by probes and final validation
43
+
44
+ ● Training train compiled in 58.22s, eval in 22.61s; periodic validation disabled
45
+ 1/571 ────────────────── loss 10.9063 lr 1.37e-04 |g| 2.266 703.58K tok/s
46
+ 10/571 ────────────────── loss 8.5151 lr 1.37e-03 |g| 1.842 776.12K tok/s
47
+ 20/571 ━───────────────── loss 7.7378 lr 2.74e-03 |g| 0.472 795.76K tok/s
48
+ 30/571 ━───────────────── loss 7.7823 lr 4.11e-03 |g| 0.639 802.48K tok/s
49
+ 40/571 ━───────────────── loss 7.6959 lr 5.48e-03 |g| 0.474 805.89K tok/s
50
+ 50/571 ━━──────────────── loss 7.7667 lr 6.85e-03 |g| 0.507 807.94K tok/s
51
+ 60/571 ━━──────────────── loss 7.6635 lr 7.81e-03 |g| 0.424 809.39K tok/s
52
+ 70/571 ━━──────────────── loss 7.6658 lr 7.80e-03 |g| 0.285 810.37K tok/s
53
+ 80/571 ━━━─────────────── loss 7.5694 lr 7.78e-03 |g| 0.345 811.01K tok/s
54
+ 90/571 ━━━─────────────── loss 7.5022 lr 7.74e-03 |g| 0.460 811.55K tok/s
55
+ 100/571 ━━━─────────────── loss 7.3004 lr 7.69e-03 |g| 0.304 811.95K tok/s
56
+ 110/571 ━━━─────────────── loss 7.3233 lr 7.63e-03 |g| 0.728 812.27K tok/s
57
+ 120/571 ━━━━────────────── loss 7.1510 lr 7.56e-03 |g| 0.643 812.52K tok/s
58
+ 130/571 ━━━━────────────── loss 6.9224 lr 7.47e-03 |g| 0.464 812.73K tok/s
59
+ 140/571 ━━━━────────────── loss 6.8478 lr 7.37e-03 |g| 0.430 812.89K tok/s
60
+ 150/571 ━━━━━───────────── loss 6.7447 lr 7.26e-03 |g| 0.494 813.04K tok/s
61
+ 160/571 ━━━━━───────────── loss 6.6706 lr 7.14e-03 |g| 0.461 813.17K tok/s
62
+ 170/571 ━━━━━───────────── loss 6.4182 lr 7.01e-03 |g| 0.296 813.28K tok/s
63
+ 180/571 ━━━━━━──────────── loss 6.4535 lr 6.86e-03 |g| 0.418 813.38K tok/s
64
+ 190/571 ━━━━━━──────────── loss 6.3154 lr 6.71e-03 |g| 0.379 813.47K tok/s
65
+ 200/571 ━━━━━━──────────── loss 6.2489 lr 6.55e-03 |g| 0.289 813.56K tok/s
66
+ 210/571 ━━━━━━━─────────── loss 6.1632 lr 6.38e-03 |g| 0.419 813.63K tok/s
67
+ 220/571 ━━━━━━━─────────── loss 6.1315 lr 6.21e-03 |g| 0.402 813.69K tok/s
68
+ 230/571 ━━━━━━━─────────── loss 6.1056 lr 6.02e-03 |g| 0.176 813.76K tok/s
69
+ 240/571 ━━━━━━━━────────── loss 6.0079 lr 5.83e-03 |g| 0.226 813.81K tok/s
70
+ 250/571 ━━━━━━━━────────── loss 5.9500 lr 5.64e-03 |g| 0.156 813.87K tok/s
71
+ 260/571 ━━━━━━━━────────── loss 5.8810 lr 5.44e-03 |g| 0.167 813.91K tok/s
72
+ 270/571 ━━━━━━━━━───────── loss 5.8430 lr 5.23e-03 |g| 0.238 813.96K tok/s
73
+ 280/571 ━━━━━━━━━───────── loss 5.8628 lr 5.02e-03 |g| 0.190 814.00K tok/s
74
+ 290/571 ━━━━━━━━━───────── loss 5.7581 lr 4.81e-03 |g| 0.207 814.04K tok/s
75
+ 300/571 ━━━━━━━━━───────── loss 5.7777 lr 4.60e-03 |g| 0.275 814.07K tok/s
76
+ 310/571 ━━━━━━━━━━──────── loss 5.7224 lr 4.38e-03 |g| 0.146 814.10K tok/s
77
+ 320/571 ━━━━━━━━━━──────── loss 5.6581 lr 4.17e-03 |g| 0.169 814.13K tok/s
78
+ 330/571 ━━━━━━━━━━──────── loss 5.6517 lr 3.95e-03 |g| 0.201 814.16K tok/s
79
+ 340/571 ━━━━━━━━━━━─────── loss 5.6744 lr 3.74e-03 |g| 0.118 814.19K tok/s
80
+ 350/571 ━━━━━━━━━━━─────── loss 5.5825 lr 3.53e-03 |g| 0.147 814.22K tok/s
81
+ 360/571 ━━━━━━━━━━━─────── loss 5.5180 lr 3.32e-03 |g| 0.112 814.24K tok/s
82
+ 370/571 ━━━━━━━━━━━━────── loss 5.5088 lr 3.12e-03 |g| 0.144 814.25K tok/s
83
+ 380/571 ━━━━━━━━━━━━────── loss 5.4606 lr 2.92e-03 |g| 0.149 814.27K tok/s
84
+ 390/571 ━━━━━━━━━━━━────── loss 5.4457 lr 2.72e-03 |g| 0.142 814.29K tok/s
85
+ 400/571 ━━━━━━━━━━━━━───── loss 5.4354 lr 2.53e-03 |g| 0.119 814.31K tok/s
86
+ 410/571 ━━━━━━━━━━━━━───── loss 5.3674 lr 2.35e-03 |g| 0.168 814.33K tok/s
87
+ 420/571 ━━━━━━━━━━━━━───── loss 5.4093 lr 2.18e-03 |g| 0.110 814.34K tok/s
88
+ 430/571 ━━━━━━━━━━━━━━──── loss 5.3597 lr 2.01e-03 |g| 0.118 814.36K tok/s
89
+ 440/571 ━━━━━━━━━━━━━━──── loss 5.3301 lr 1.85e-03 |g| 0.101 814.37K tok/s
90
+ 450/571 ━━━━━━━━━━━━━━──── loss 5.3176 lr 1.70e-03 |g| 0.119 814.39K tok/s
91
+ 460/571 ━━━━━━━━━━━━━━━─── loss 5.2994 lr 1.56e-03 |g| 0.110 814.40K tok/s
92
+ 470/571 ━━━━━━━━━━━━━━━─── loss 5.2402 lr 1.43e-03 |g| 0.080 814.41K tok/s
93
+ 480/571 ━━━━━━━━━━━━━━━─── loss 5.2008 lr 1.31e-03 |g| 0.119 814.43K tok/s
94
+ 490/571 ━━━━━━━━━━━━━━━─── loss 5.2437 lr 1.20e-03 |g| 0.109 814.44K tok/s
95
+ 500/571 ━━━━━━━━━━━━━━━━── loss 5.1855 lr 1.11e-03 |g| 0.091 814.45K tok/s
96
+ 510/571 ━━━━━━━━━━━━━━━━── loss 5.1762 lr 1.02e-03 |g| 0.073 814.46K tok/s
97
+ 520/571 ━━━━━━━━━━━━━━━━── loss 5.1596 lr 9.51e-04 |g| 0.084 814.47K tok/s
98
+ 530/571 ━━━━━━━━━━━━━━━━━─ loss 5.1847 lr 8.91e-04 |g| 0.073 814.48K tok/s
99
+ 540/571 ━━━━━━━━━━━━━━━━━─ loss 5.1405 lr 8.44e-04 |g| 0.086 814.49K tok/s
100
+ 550/571 ━━━━━━━━━━━━━━━━━─ loss 5.1518 lr 8.10e-04 |g| 0.083 814.50K tok/s
101
+ 560/571 ━━━━━━━━━━━━━━━━━━ loss 5.1538 lr 7.89e-04 |g| 0.091 814.51K tok/s
102
+ 570/571 ━━━━━━━━━━━━━━━━━━ loss 5.1279 lr 7.81e-04 |g| 0.079 814.52K tok/s
103
+ 571/571 ━━━━━━━━━━━━━━━━━━ loss 5.1338 lr 7.81e-04 |g| 0.072 814.50K tok/s
104
+
105
+ ● Canonical validation 8 deterministic batches outside train_seconds
106
+
107
+ ● Fresh-domain validation skipped; no downstream data supplied
108
+
109
+ ● Artifacts training.riglog + validation.csv + diagnostics.riglog
110
+
111
+ ✓ synchronized training 367.556s (compilation excluded)
112
+ validation loss 5.1354 in 1.841s
113
+
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/stdout.log ADDED
@@ -0,0 +1 @@
 
 
1
+ RIG_RESULT={"artifacts":{"diagnostics":"diagnostics.riglog","training_curve":"training.riglog","validation_curve":"validation.csv"},"checkpoint":null,"contract":{"dataset_id":"fineweb-8b-gpt2","model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"model_id":"reference-gpt-v3-family","sequence_length":8192,"tokenizer_id":"gpt2"},"evaluations":{"fineweb":{"canonical":true,"loss":5.135362327098846,"perplexity":169.92587747222018,"scored_tokens":4194304,"seconds":1.8413573480211198}},"implementation":{"attention_backend":"tpu_flash","attention_tuning":{"key_digest":"6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8","resolution_source":"heuristic","tiles":{"block_kv":512,"block_kv_compute":256,"block_kv_dkv":512,"block_kv_dkv_compute":256,"block_kv_dq":512,"block_kv_dq_compute":256,"block_q":512,"block_q_dkv":512,"block_q_dkv_compute":256,"block_q_dq":256},"tune_seconds":0.0},"configuration":{"overrides":{"study_batch_size":64,"tokens_per_parameter_micros":5000000},"path":"config.yaml","profile":"dev","resolved":{"evaluation":{"eval_batches":8,"val_every":0,"val_probe_batches":8},"kernels":{"attention_backend":"tpu_flash","loss_backend":"tiled","vocab_tile_size":2048},"logging":{"diagnostics_every":10,"log_every":10},"model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"optimizer":{"adam_epsilon":1e-08,"beta1":0.9,"beta2":0.95,"effective":{"adam_epsilon_horizon_multiplier":0.5,"beta1":0.6000000000000001,"beta2":0.7999999999999998,"global_peak_learning_rate":0.0078125,"weight_decay_horizon_multiplier":2.0},"grad_clip":0.0,"learning_rate":0.00390625,"min_lr_ratio":0.1,"warmup_steps":57,"weight_decay":0.1},"parameterization":{"attention_scale":"inverse_head_dim","base_depth":12,"base_width":384,"batch_multiplier":4.0,"data_multiplier":1.0,"depth_alpha":1.0,"depth_multiplier":1.0,"embeddings":"untied","init_std":0.02,"name":"complete_d_p","width_multiplier":1.0},"training":{"batch_size":64,"dtype":"bfloat16","sampling":"shuffled_epochs","seq_len":8192,"steps":571,"tokens_per_parameter":4.996285069139584,"train_tokens":299368448}},"schema_version":2,"sha256":"a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"},"loss_backend":"tiled","vocab_tile_size":2048},"metrics":{"achieved_tflops":601.0030472181179,"attention_tune_seconds":0.0,"base_learning_rate":0.00390625,"diagnostic_compile_seconds":94.93301027396228,"diagnostic_point_count":59,"diagnostics_every":10,"early_stopping_step":null,"estimated_total_flops":220902531230859264,"eval_compile_seconds":22.61157763702795,"expert_top_k":2,"experts":8,"final_validation_seconds":1.8413573480211198,"flop_accounting":{"by_site":{"dot_general":1638200770560,"pallas_call":695784701952,"tpu_flash_causal_attention_bwd_dkv":1236950581248,"tpu_flash_causal_attention_bwd_dq":1236950581248,"tpu_flash_causal_attention_fwd":1236950581248},"elementwise_per_sequence":9518381752,"matmul_per_sequence":6044837216256,"method":"traced-jaxpr","warnings":[]},"flops_per_token":737895168,"mfu_estimate":0.13659160164048134,"model_tier":"60m","parameter_count":59959680,"parameters":102509952,"schedule_steps":571,"tokens_per_parameter":4.996285069139584,"tokens_per_second":814482.9689657459,"tokens_processed":299368448,"total_compile_seconds":175.76108301396016,"total_parameter_count":102509952,"train_compile_seconds":58.21649510296993,"train_loss":5.133830547332764,"train_seconds":367.55642463604454,"training_data_epochs":0.037894837263947954,"training_data_sharding":"rank_disjoint_shuffled_windows","training_sampling":"shuffled_epochs","training_steps":571,"training_token_budget":299368448,"training_usable_tokens_per_epoch":7899979776,"validation_loss":5.135362327098846,"validation_probe_count":0,"validation_probe_seconds":0.0,"validation_tokens":4194304},"profile":"dev","schema_version":1,"seed":1338,"status":"ok","system":{"controller_process_index":1,"device_count":16,"device_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],"device_kinds":["TPU v4"],"jax_version":"0.11.0","jaxlib_version":"0.11.0","libtpu_version":"0.0.44.1","local_device_count":4,"platform":"tpu","process_count":4,"process_indices":[0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3],"python_version":"3.12.13"},"track":"open"}
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/training.riglog ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ee8e7328f7b1608486b23784dd2163d8b0992cfef49efb4f982efde8f6daa881
3
+ size 327744
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1338/validation.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ step,tokens_processed,kind,domain,validation_tokens,validation_loss,perplexity,validation_seconds,canonical
2
+ 571,299368448,fineweb,fineweb,4194304,5.135362327098846,169.92587747222018,1.8413573480211198,true
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/diagnostics.riglog ADDED
Binary file (75.1 kB). View file
 
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/metrics.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.078427672386169,
32
+ "perplexity": 160.52146507071186,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8400509939529002
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.0078125,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.00390625,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.4651041803471,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.00390625,
142
+ "diagnostic_compile_seconds": 96.30863135389518,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.83055985509418,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8400509939529002,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13646934185916978,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 813753.9453034434,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 179.4674630300142,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 59.32827182102483,
176
+ "train_loss": 5.105586528778076,
177
+ "train_seconds": 367.8857100820169,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.078427672386169,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1339,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/result.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.078427672386169,
32
+ "perplexity": 160.52146507071186,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8400509939529002
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.0078125,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.00390625,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.4651041803471,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.00390625,
142
+ "diagnostic_compile_seconds": 96.30863135389518,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.83055985509418,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8400509939529002,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13646934185916978,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 813753.9453034434,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 179.4674630300142,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 59.32827182102483,
176
+ "train_loss": 5.105586528778076,
177
+ "train_seconds": 367.8857100820169,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.078427672386169,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1339,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/stderr.log ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ E0819 11:23:14.645722 3401496 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
2
+ E0819 11:23:14.645970 3500850 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
3
+ E0819 11:23:14.647363 247898 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
4
+ E0819 11:23:14.647245 3454451 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
5
+
6
+ ◆ GPT TPU RIG reference / jax
7
+
8
+
9
+ ● Attention tile preflight resolving the shipped lookup or shape heuristic
10
+ ╭─ run configuration ────────────────────────────────────────────────────────╮
11
+ │ experiment config config.yaml · dev · sha256:a4bad1be29b5 │
12
+ │ devices 16 × TPU v4 │
13
+ │ JAX processes 4 (this rank 1) │
14
+ │ mesh data=16 (replicated model) │
15
+ │ dataset 79 train + 1 val shard(s) │
16
+ │ train / val tokens 7,900,000,000 / 100,000,000 │
17
+ │ downstream not requested │
18
+ │ model 60m · L12 D384 H6 RoPE RMSNorm GELU MLP×4 │
19
+ │ parameters 102.51M │
20
+ │ parameterization complete_d_p · mN=1 · mL=1 · mD=1 │
21
+ │ global batch 64 × 8192 tokens │
22
+ │ train sampling shuffled epochs · 7,899,979,776 unique targets/epoch │
23
+ │ compute bfloat16 │
24
+ │ attention tpu_flash │
25
+ │ attention tuning heuristic · key 6a5b398e13ed │
26
+ │ attention fwd q512 · kv512/256 │
27
+ │ attention dK/dV q512/256 · kv512/256 │
28
+ │ attention dQ q256 · kv512/256 │
29
+ │ output loss tiled CE (semantic 50,304, tile 2,048) │
30
+ │ diagnostics step 1 / every 10 / final │
31
+ │ duration 571 steps │
32
+ │ train tokens 299.37M │
33
+ │ traced FLOPs 220902.53T │
34
+ │ FLOP breakdown dot_general 1,638,200,770,560 (27.1%) · tpu_flash_ca… │
35
+ │ XProf disabled │
36
+ ╰────────────────────────────────────────────────────────────────────────────╯
37
+
38
+ ● Compiling train step compilation is outside train_seconds
39
+
40
+ ● Compiling sparse diagnostics separate executable; compilation is outside train_seconds
41
+
42
+ ● Compiling evaluation reused by probes and final validation
43
+
44
+ ● Training train compiled in 59.33s, eval in 23.83s; periodic validation disabled
45
+ 1/571 ────────────────── loss 10.8871 lr 1.37e-04 |g| 2.681 543.29K tok/s
46
+ 10/571 ────────────────── loss 8.3506 lr 1.37e-03 |g| 1.399 747.15K tok/s
47
+ 20/571 ━───────────────── loss 7.7726 lr 2.74e-03 |g| 0.202 779.93K tok/s
48
+ 30/571 ━───────────────── loss 7.7483 lr 4.11e-03 |g| 0.381 791.90K tok/s
49
+ 40/571 ━───────────────── loss 7.8917 lr 5.48e-03 |g| 1.216 798.06K tok/s
50
+ 50/571 ━━──────────────── loss 7.8108 lr 6.85e-03 |g| 0.770 801.77K tok/s
51
+ 60/571 ━━──────────────── loss 7.7198 lr 7.81e-03 |g| 0.611 804.22K tok/s
52
+ 70/571 ━━──────────────── loss 7.6627 lr 7.80e-03 |g| 0.427 805.96K tok/s
53
+ 80/571 ━━━─────────────── loss 7.5667 lr 7.78e-03 |g| 0.301 807.22K tok/s
54
+ 90/571 ━━━─────────────── loss 7.4644 lr 7.74e-03 |g| 0.331 808.15K tok/s
55
+ 100/571 ━━━─────────────── loss 7.3455 lr 7.69e-03 |g| 0.962 808.86K tok/s
56
+ 110/571 ━━━─────────────── loss 7.0786 lr 7.63e-03 |g| 0.355 809.40K tok/s
57
+ 120/571 ━━━━────────────── loss 6.9829 lr 7.56e-03 |g| 0.520 809.88K tok/s
58
+ 130/571 ━━━━────────────── loss 6.7121 lr 7.47e-03 |g| 0.312 810.26K tok/s
59
+ 140/571 ━━━━────────────── loss 6.6042 lr 7.37e-03 |g| 0.252 810.58K tok/s
60
+ 150/571 ━━━━━───────────── loss 6.5234 lr 7.26e-03 |g| 0.258 810.86K tok/s
61
+ 160/571 ━━━━━───────────── loss 6.4353 lr 7.14e-03 |g| 0.162 811.11K tok/s
62
+ 170/571 ━━━━━───────────── loss 6.4300 lr 7.01e-03 |g| 0.362 811.34K tok/s
63
+ 180/571 ━━━━━━──────────── loss 6.3033 lr 6.86e-03 |g| 0.250 811.54K tok/s
64
+ 190/571 ━━━━━━──────────── loss 6.2813 lr 6.71e-03 |g| 0.272 811.71K tok/s
65
+ 200/571 ━━━━━━──────────── loss 6.2195 lr 6.55e-03 |g| 0.357 811.86K tok/s
66
+ 210/571 ━━━━━━━─────────── loss 6.1542 lr 6.38e-03 |g| 0.301 812.00K tok/s
67
+ 220/571 ━━━━━━━─────────── loss 6.1013 lr 6.21e-03 |g| 0.288 812.13K tok/s
68
+ 230/571 ━━━━━━━─────────── loss 6.1131 lr 6.02e-03 |g| 0.316 812.25K tok/s
69
+ 240/571 ━━━━━━━━────────── loss 5.9241 lr 5.83e-03 |g| 0.174 812.35K tok/s
70
+ 250/571 ━━━━━━━━────────── loss 5.9043 lr 5.64e-03 |g| 0.261 812.45K tok/s
71
+ 260/571 ━━━━━━━━────────── loss 5.8669 lr 5.44e-03 |g| 0.234 812.55K tok/s
72
+ 270/571 ━━━━━━━━━───────── loss 5.8295 lr 5.23e-03 |g| 0.189 812.63K tok/s
73
+ 280/571 ━━━━━━━━━───────── loss 5.7798 lr 5.02e-03 |g| 0.223 812.71K tok/s
74
+ 290/571 ━━━━━━━━━───────── loss 5.7431 lr 4.81e-03 |g| 0.242 812.78K tok/s
75
+ 300/571 ━━━━━━━━━───────── loss 5.6641 lr 4.60e-03 |g| 0.238 812.86K tok/s
76
+ 310/571 ━━━━━━━━━━──────── loss 5.6333 lr 4.38e-03 |g| 0.151 812.92K tok/s
77
+ 320/571 ━━━━━━━━━━──────── loss 5.6100 lr 4.17e-03 |g| 0.143 812.98K tok/s
78
+ 330/571 ━━━━━━━━━━──────── loss 5.5401 lr 3.95e-03 |g| 0.114 813.04K tok/s
79
+ 340/571 ━━━━━━━━━━━─────── loss 5.5040 lr 3.74e-03 |g| 0.148 813.10K tok/s
80
+ 350/571 ━━━━━━━━━━━─────── loss 5.4756 lr 3.53e-03 |g| 0.151 813.15K tok/s
81
+ 360/571 ━━━━━━━━━━━─────── loss 5.5190 lr 3.32e-03 |g| 0.183 813.20K tok/s
82
+ 370/571 ━━━━━━━━━━━━────── loss 5.4473 lr 3.12e-03 |g| 0.141 813.24K tok/s
83
+ 380/571 ━━━━━━━━━━━━────── loss 5.4516 lr 2.92e-03 |g| 0.143 813.29K tok/s
84
+ 390/571 ━━━━━━━━━━━━────── loss 5.3922 lr 2.72e-03 |g| 0.111 813.33K tok/s
85
+ 400/571 ━━━━━━━━━━━━━───── loss 5.4323 lr 2.53e-03 |g| 0.132 813.37K tok/s
86
+ 410/571 ━━━━━━━━━━━━━───── loss 5.2905 lr 2.35e-03 |g| 0.111 813.41K tok/s
87
+ 420/571 ━━━━━━━━━━━━━───── loss 5.3259 lr 2.18e-03 |g| 0.119 813.44K tok/s
88
+ 430/571 ━━━━━━━━━━━━━━──── loss 5.2538 lr 2.01e-03 |g| 0.126 813.48K tok/s
89
+ 440/571 ━━━━━━━━━━━━━━──── loss 5.2614 lr 1.85e-03 |g| 0.117 813.51K tok/s
90
+ 450/571 ━━━━━━━━━━━━━━──── loss 5.2412 lr 1.70e-03 |g| 0.103 813.54K tok/s
91
+ 460/571 ━━━━━━━━━━━━━━━─── loss 5.2230 lr 1.56e-03 |g| 0.132 813.57K tok/s
92
+ 470/571 ━━━━━━━━━━━━━━━─── loss 5.2072 lr 1.43e-03 |g| 0.092 813.59K tok/s
93
+ 480/571 ━━━━━━━━━━━━━━━─── loss 5.1805 lr 1.31e-03 |g| 0.099 813.62K tok/s
94
+ 490/571 ━━━━━━━━━━━━━━━─── loss 5.1673 lr 1.20e-03 |g| 0.087 813.64K tok/s
95
+ 500/571 ━━━━━━━━━━━━━━━━── loss 5.2031 lr 1.11e-03 |g| 0.085 813.66K tok/s
96
+ 510/571 ━━━━━━━━━━━━━━━━── loss 5.1889 lr 1.02e-03 |g| 0.082 813.69K tok/s
97
+ 520/571 ━━━━━━━━━━━━━━━━── loss 5.1578 lr 9.51e-04 |g| 0.072 813.70K tok/s
98
+ 530/571 ━━━━━━━━━━━━━━━━━─ loss 5.0887 lr 8.91e-04 |g| 0.075 813.73K tok/s
99
+ 540/571 ━━━━━━━━━━━━━━━━━─ loss 5.1108 lr 8.44e-04 |g| 0.075 813.75K tok/s
100
+ 550/571 ━━━━━━━━━━━━━━━━━─ loss 5.1219 lr 8.10e-04 |g| 0.084 813.76K tok/s
101
+ 560/571 ━━━━━━━━━━━━━━━━━━ loss 5.0913 lr 7.89e-04 |g| 0.072 813.79K tok/s
102
+ 570/571 ━━━━━━━━━━━━━━━━━━ loss 5.0706 lr 7.81e-04 |g| 0.077 813.80K tok/s
103
+ 571/571 ━━━━━━━━━━━━━━━━━━ loss 5.1056 lr 7.81e-04 |g| 0.078 813.78K tok/s
104
+
105
+ ● Canonical validation 8 deterministic batches outside train_seconds
106
+
107
+ ● Fresh-domain validation skipped; no downstream data supplied
108
+
109
+ ● Artifacts training.riglog + validation.csv + diagnostics.riglog
110
+
111
+ ✓ synchronized training 367.886s (compilation excluded)
112
+ validation loss 5.0784 in 1.840s
113
+
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/stdout.log ADDED
@@ -0,0 +1 @@
 
 
1
+ RIG_RESULT={"artifacts":{"diagnostics":"diagnostics.riglog","training_curve":"training.riglog","validation_curve":"validation.csv"},"checkpoint":null,"contract":{"dataset_id":"fineweb-8b-gpt2","model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"model_id":"reference-gpt-v3-family","sequence_length":8192,"tokenizer_id":"gpt2"},"evaluations":{"fineweb":{"canonical":true,"loss":5.078427672386169,"perplexity":160.52146507071186,"scored_tokens":4194304,"seconds":1.8400509939529002}},"implementation":{"attention_backend":"tpu_flash","attention_tuning":{"key_digest":"6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8","resolution_source":"heuristic","tiles":{"block_kv":512,"block_kv_compute":256,"block_kv_dkv":512,"block_kv_dkv_compute":256,"block_kv_dq":512,"block_kv_dq_compute":256,"block_q":512,"block_q_dkv":512,"block_q_dkv_compute":256,"block_q_dq":256},"tune_seconds":0.0},"configuration":{"overrides":{"study_batch_size":64,"tokens_per_parameter_micros":5000000},"path":"config.yaml","profile":"dev","resolved":{"evaluation":{"eval_batches":8,"val_every":0,"val_probe_batches":8},"kernels":{"attention_backend":"tpu_flash","loss_backend":"tiled","vocab_tile_size":2048},"logging":{"diagnostics_every":10,"log_every":10},"model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"optimizer":{"adam_epsilon":1e-08,"beta1":0.9,"beta2":0.95,"effective":{"adam_epsilon_horizon_multiplier":0.5,"beta1":0.6000000000000001,"beta2":0.7999999999999998,"global_peak_learning_rate":0.0078125,"weight_decay_horizon_multiplier":2.0},"grad_clip":0.0,"learning_rate":0.00390625,"min_lr_ratio":0.1,"warmup_steps":57,"weight_decay":0.1},"parameterization":{"attention_scale":"inverse_head_dim","base_depth":12,"base_width":384,"batch_multiplier":4.0,"data_multiplier":1.0,"depth_alpha":1.0,"depth_multiplier":1.0,"embeddings":"untied","init_std":0.02,"name":"complete_d_p","width_multiplier":1.0},"training":{"batch_size":64,"dtype":"bfloat16","sampling":"shuffled_epochs","seq_len":8192,"steps":571,"tokens_per_parameter":4.996285069139584,"train_tokens":299368448}},"schema_version":2,"sha256":"a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"},"loss_backend":"tiled","vocab_tile_size":2048},"metrics":{"achieved_tflops":600.4651041803471,"attention_tune_seconds":0.0,"base_learning_rate":0.00390625,"diagnostic_compile_seconds":96.30863135389518,"diagnostic_point_count":59,"diagnostics_every":10,"early_stopping_step":null,"estimated_total_flops":220902531230859264,"eval_compile_seconds":23.83055985509418,"expert_top_k":2,"experts":8,"final_validation_seconds":1.8400509939529002,"flop_accounting":{"by_site":{"dot_general":1638200770560,"pallas_call":695784701952,"tpu_flash_causal_attention_bwd_dkv":1236950581248,"tpu_flash_causal_attention_bwd_dq":1236950581248,"tpu_flash_causal_attention_fwd":1236950581248},"elementwise_per_sequence":9518381752,"matmul_per_sequence":6044837216256,"method":"traced-jaxpr","warnings":[]},"flops_per_token":737895168,"mfu_estimate":0.13646934185916978,"model_tier":"60m","parameter_count":59959680,"parameters":102509952,"schedule_steps":571,"tokens_per_parameter":4.996285069139584,"tokens_per_second":813753.9453034434,"tokens_processed":299368448,"total_compile_seconds":179.4674630300142,"total_parameter_count":102509952,"train_compile_seconds":59.32827182102483,"train_loss":5.105586528778076,"train_seconds":367.8857100820169,"training_data_epochs":0.037894837263947954,"training_data_sharding":"rank_disjoint_shuffled_windows","training_sampling":"shuffled_epochs","training_steps":571,"training_token_budget":299368448,"training_usable_tokens_per_epoch":7899979776,"validation_loss":5.078427672386169,"validation_probe_count":0,"validation_probe_seconds":0.0,"validation_tokens":4194304},"profile":"dev","schema_version":1,"seed":1339,"status":"ok","system":{"controller_process_index":1,"device_count":16,"device_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],"device_kinds":["TPU v4"],"jax_version":"0.11.0","jaxlib_version":"0.11.0","libtpu_version":"0.0.44.1","local_device_count":4,"platform":"tpu","process_count":4,"process_indices":[0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3],"python_version":"3.12.13"},"track":"open"}
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/training.riglog ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:93d1d988729a5385e30dc692964835f5e21b0905ed9955793519f91746d3f9d2
3
+ size 327744
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-8-s1339/validation.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ step,tokens_processed,kind,domain,validation_tokens,validation_loss,perplexity,validation_seconds,canonical
2
+ 571,299368448,fineweb,fineweb,4194304,5.078427672386169,160.52146507071186,1.8400509939529002,true
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/diagnostics.riglog ADDED
Binary file (75.1 kB). View file
 
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/metrics.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.265833914279938,
32
+ "perplexity": 193.60769370796328,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.83323098893743
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.00390625,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.001953125,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.9627548715764,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.001953125,
142
+ "diagnostic_compile_seconds": 99.6910779599566,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.350994339911267,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.83323098893743,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13658244428899463,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 814428.3645337225,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 181.43496975686867,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 58.3928974570008,
176
+ "train_loss": 5.236739158630371,
177
+ "train_seconds": 367.58106794499326,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.265833914279938,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1337,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/result.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.265833914279938,
32
+ "perplexity": 193.60769370796328,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.83323098893743
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.00390625,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.001953125,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.9627548715764,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.001953125,
142
+ "diagnostic_compile_seconds": 99.6910779599566,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.350994339911267,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.83323098893743,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13658244428899463,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 814428.3645337225,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 181.43496975686867,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 58.3928974570008,
176
+ "train_loss": 5.236739158630371,
177
+ "train_seconds": 367.58106794499326,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.265833914279938,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1337,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/stderr.log ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ E0819 12:01:19.111668 287859 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
2
+ E0819 12:01:19.111737 3437166 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
3
+ E0819 12:01:19.112127 3536806 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
4
+ E0819 12:01:19.113869 3490220 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
5
+
6
+ ◆ GPT TPU RIG reference / jax
7
+
8
+
9
+ ● Attention tile preflight resolving the shipped lookup or shape heuristic
10
+ ╭─ run configuration ────────────────────────────────────────────────────────╮
11
+ │ experiment config config.yaml · dev · sha256:a4bad1be29b5 │
12
+ │ devices 16 × TPU v4 │
13
+ │ JAX processes 4 (this rank 1) │
14
+ │ mesh data=16 (replicated model) │
15
+ │ dataset 79 train + 1 val shard(s) │
16
+ │ train / val tokens 7,900,000,000 / 100,000,000 │
17
+ │ downstream not requested │
18
+ │ model 60m · L12 D384 H6 RoPE RMSNorm GELU MLP×4 │
19
+ │ parameters 102.51M │
20
+ │ parameterization complete_d_p · mN=1 · mL=1 · mD=1 │
21
+ │ global batch 64 × 8192 tokens │
22
+ │ train sampling shuffled epochs · 7,899,979,776 unique targets/epoch │
23
+ │ compute bfloat16 │
24
+ │ attention tpu_flash │
25
+ │ attention tuning heuristic · key 6a5b398e13ed │
26
+ │ attention fwd q512 · kv512/256 │
27
+ │ attention dK/dV q512/256 · kv512/256 │
28
+ │ attention dQ q256 · kv512/256 │
29
+ │ output loss tiled CE (semantic 50,304, tile 2,048) │
30
+ │ diagnostics step 1 / every 10 / final │
31
+ │ duration 571 steps │
32
+ │ train tokens 299.37M │
33
+ │ traced FLOPs 220902.53T │
34
+ │ FLOP breakdown dot_general 1,638,200,770,560 (27.1%) · tpu_flash_ca… │
35
+ │ XProf disabled │
36
+ ╰────────────────────────────────────────────────────────────────────────────╯
37
+
38
+ ● Compiling train step compilation is outside train_seconds
39
+
40
+ ● Compiling sparse diagnostics separate executable; compilation is outside train_seconds
41
+
42
+ ● Compiling evaluation reused by probes and final validation
43
+
44
+ ● Training train compiled in 58.39s, eval in 23.35s; periodic validation disabled
45
+ 1/571 ────────────────── loss 10.9476 lr 6.85e-05 |g| 2.514 643.21K tok/s
46
+ 10/571 ────────────────── loss 9.2624 lr 6.85e-04 |g| 1.833 767.34K tok/s
47
+ 20/571 ━───────────────── loss 7.7624 lr 1.37e-03 |g| 3.077 790.90K tok/s
48
+ 30/571 ━───────────────── loss 7.9681 lr 2.06e-03 |g| 7.728 799.10K tok/s
49
+ 40/571 ━───────────────── loss 7.6939 lr 2.74e-03 |g| 3.285 803.49K tok/s
50
+ 50/571 ━━──────────────── loss 7.6366 lr 3.43e-03 |g| 0.723 805.99K tok/s
51
+ 60/571 ━━──────────────── loss 7.5820 lr 3.91e-03 |g| 0.368 807.72K tok/s
52
+ 70/571 ━━──────────────── loss 7.6100 lr 3.90e-03 |g| 0.304 808.98K tok/s
53
+ 80/571 ━━━─────────────── loss 7.5723 lr 3.89e-03 |g| 0.306 809.87K tok/s
54
+ 90/571 ━━━─────────────── loss 7.6170 lr 3.87e-03 |g| 0.773 810.60K tok/s
55
+ 100/571 ━━━─────────────── loss 7.5296 lr 3.85e-03 |g| 0.214 811.17K tok/s
56
+ 110/571 ━━━─────────────── loss 7.4178 lr 3.81e-03 |g| 0.316 811.60K tok/s
57
+ 120/571 ━━━━────────────── loss 7.3656 lr 3.78e-03 |g| 0.426 811.96K tok/s
58
+ 130/571 ━━━━────────────── loss 7.3478 lr 3.73e-03 |g| 0.419 812.23K tok/s
59
+ 140/571 ━━━━────────────── loss 7.0625 lr 3.68e-03 |g| 0.282 812.46K tok/s
60
+ 150/571 ━━━━━───────────── loss 6.9500 lr 3.63e-03 |g| 0.403 812.64K tok/s
61
+ 160/571 ━━━━━───────────── loss 6.8693 lr 3.57e-03 |g| 0.782 812.79K tok/s
62
+ 170/571 ━━━━━───────────── loss 6.7857 lr 3.50e-03 |g| 0.309 812.93K tok/s
63
+ 180/571 ━━━━━━──────────── loss 6.7222 lr 3.43e-03 |g| 0.344 813.05K tok/s
64
+ 190/571 ━━━━━━──────────── loss 6.6071 lr 3.36e-03 |g| 0.596 813.16K tok/s
65
+ 200/571 ━━━━━━──────────── loss 6.5552 lr 3.28e-03 |g| 0.363 813.26K tok/s
66
+ 210/571 ━━━━━━━─────────── loss 6.4151 lr 3.19e-03 |g| 0.269 813.35K tok/s
67
+ 220/571 ━━━━━━━─────────── loss 6.3570 lr 3.10e-03 |g| 0.332 813.42K tok/s
68
+ 230/571 ━━━━━━━─────────── loss 6.3090 lr 3.01e-03 |g| 0.424 813.50K tok/s
69
+ 240/571 ━━━━━━━━────────── loss 6.2148 lr 2.92e-03 |g| 0.331 813.57K tok/s
70
+ 250/571 ━━━━━━━━────────── loss 6.1667 lr 2.82e-03 |g| 0.280 813.63K tok/s
71
+ 260/571 ━━━━━━━━────────── loss 6.0760 lr 2.72e-03 |g| 0.329 813.68K tok/s
72
+ 270/571 ━━━━━━━━━───────── loss 6.0496 lr 2.62e-03 |g| 0.322 813.74K tok/s
73
+ 280/571 ━━━━━━━━━───────── loss 5.9669 lr 2.51e-03 |g| 0.358 813.79K tok/s
74
+ 290/571 ━━━━━━━━━───────── loss 5.9765 lr 2.41e-03 |g| 0.227 813.83K tok/s
75
+ 300/571 ━━━━━━━━━───────── loss 5.8240 lr 2.30e-03 |g| 0.185 813.88K tok/s
76
+ 310/571 ━━━━━━━━━━──────── loss 5.8474 lr 2.19e-03 |g| 0.243 813.91K tok/s
77
+ 320/571 ━━━━━━━━━━──────── loss 5.7279 lr 2.08e-03 |g| 0.236 813.95K tok/s
78
+ 330/571 ━━━━━━━━━━──────── loss 5.8027 lr 1.98e-03 |g| 0.193 813.99K tok/s
79
+ 340/571 ━━━━━━━━━━━─────── loss 5.7890 lr 1.87e-03 |g| 0.227 814.02K tok/s
80
+ 350/571 ━━━━━━━━━━━─────── loss 5.6604 lr 1.76e-03 |g| 0.181 814.05K tok/s
81
+ 360/571 ━━━━━━━━━━━─────── loss 5.6547 lr 1.66e-03 |g| 0.189 814.08K tok/s
82
+ 370/571 ━━━━━━━━━━━━────── loss 5.6045 lr 1.56e-03 |g| 0.198 814.11K tok/s
83
+ 380/571 ━━━━━━━━━━━━────── loss 5.5984 lr 1.46e-03 |g| 0.239 814.14K tok/s
84
+ 390/571 ━━━━━━━━━━━━────── loss 5.5893 lr 1.36e-03 |g| 0.142 814.16K tok/s
85
+ 400/571 ━━━━━━━━━━━━━───── loss 5.5050 lr 1.27e-03 |g| 0.138 814.18K tok/s
86
+ 410/571 ━━━━━━━━━━━━━───── loss 5.5264 lr 1.18e-03 |g| 0.139 814.21K tok/s
87
+ 420/571 ━━━━━━━━━━━━━───── loss 5.4843 lr 1.09e-03 |g| 0.174 814.23K tok/s
88
+ 430/571 ━━━━━━━━━━━━━━──── loss 5.4149 lr 1.00e-03 |g| 0.160 814.25K tok/s
89
+ 440/571 ━━━━━━━━━━━━━━──── loss 5.4576 lr 9.25e-04 |g| 0.163 814.27K tok/s
90
+ 450/571 ━━━━━━━━━━━━━━──── loss 5.4543 lr 8.50e-04 |g| 0.134 814.29K tok/s
91
+ 460/571 ━━━━━━━━━━━━━━━─── loss 5.4189 lr 7.80e-04 |g| 0.141 814.31K tok/s
92
+ 470/571 ━━━━━━━━━━━━━━━─── loss 5.4211 lr 7.15e-04 |g| 0.127 814.33K tok/s
93
+ 480/571 ━━━━━━━━━━━━━━━─── loss 5.4214 lr 6.56e-04 |g| 0.144 814.34K tok/s
94
+ 490/571 ━━━━━━━━━━━━━━━─── loss 5.3714 lr 6.02e-04 |g| 0.118 814.36K tok/s
95
+ 500/571 ━━━━━━━━━━━━━━━━── loss 5.2899 lr 5.54e-04 |g| 0.130 814.37K tok/s
96
+ 510/571 ━━━━━━━━━━━━━━━━── loss 5.3148 lr 5.11e-04 |g| 0.105 814.39K tok/s
97
+ 520/571 ━━━━━━━━━━━━━━━━── loss 5.3428 lr 4.75e-04 |g| 0.128 814.40K tok/s
98
+ 530/571 ━━━━━━━━━━━━━━━━━─ loss 5.2800 lr 4.46e-04 |g| 0.097 814.42K tok/s
99
+ 540/571 ━━━━━━━━━━━━━━━━━─ loss 5.2661 lr 4.22e-04 |g| 0.116 814.43K tok/s
100
+ 550/571 ━━━━━━━━━━━━━━━━━─ loss 5.3037 lr 4.05e-04 |g| 0.119 814.44K tok/s
101
+ 560/571 ━━━━━━━━━━━━━━━━━━ loss 5.3040 lr 3.95e-04 |g| 0.120 814.46K tok/s
102
+ 570/571 ━━━━━━━━━━━━━━━━━━ loss 5.2874 lr 3.91e-04 |g| 0.131 814.47K tok/s
103
+ 571/571 ━━━━━━━━━━━━━━━━━━ loss 5.2367 lr 3.91e-04 |g| 0.116 814.45K tok/s
104
+
105
+ ● Canonical validation 8 deterministic batches outside train_seconds
106
+
107
+ ● Fresh-domain validation skipped; no downstream data supplied
108
+
109
+ ● Artifacts training.riglog + validation.csv + diagnostics.riglog
110
+
111
+ ✓ synchronized training 367.581s (compilation excluded)
112
+ validation loss 5.2658 in 1.833s
113
+
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/stdout.log ADDED
@@ -0,0 +1 @@
 
 
1
+ RIG_RESULT={"artifacts":{"diagnostics":"diagnostics.riglog","training_curve":"training.riglog","validation_curve":"validation.csv"},"checkpoint":null,"contract":{"dataset_id":"fineweb-8b-gpt2","model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"model_id":"reference-gpt-v3-family","sequence_length":8192,"tokenizer_id":"gpt2"},"evaluations":{"fineweb":{"canonical":true,"loss":5.265833914279938,"perplexity":193.60769370796328,"scored_tokens":4194304,"seconds":1.83323098893743}},"implementation":{"attention_backend":"tpu_flash","attention_tuning":{"key_digest":"6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8","resolution_source":"heuristic","tiles":{"block_kv":512,"block_kv_compute":256,"block_kv_dkv":512,"block_kv_dkv_compute":256,"block_kv_dq":512,"block_kv_dq_compute":256,"block_q":512,"block_q_dkv":512,"block_q_dkv_compute":256,"block_q_dq":256},"tune_seconds":0.0},"configuration":{"overrides":{"study_batch_size":64,"tokens_per_parameter_micros":5000000},"path":"config.yaml","profile":"dev","resolved":{"evaluation":{"eval_batches":8,"val_every":0,"val_probe_batches":8},"kernels":{"attention_backend":"tpu_flash","loss_backend":"tiled","vocab_tile_size":2048},"logging":{"diagnostics_every":10,"log_every":10},"model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"optimizer":{"adam_epsilon":1e-08,"beta1":0.9,"beta2":0.95,"effective":{"adam_epsilon_horizon_multiplier":0.5,"beta1":0.6000000000000001,"beta2":0.7999999999999998,"global_peak_learning_rate":0.00390625,"weight_decay_horizon_multiplier":2.0},"grad_clip":0.0,"learning_rate":0.001953125,"min_lr_ratio":0.1,"warmup_steps":57,"weight_decay":0.1},"parameterization":{"attention_scale":"inverse_head_dim","base_depth":12,"base_width":384,"batch_multiplier":4.0,"data_multiplier":1.0,"depth_alpha":1.0,"depth_multiplier":1.0,"embeddings":"untied","init_std":0.02,"name":"complete_d_p","width_multiplier":1.0},"training":{"batch_size":64,"dtype":"bfloat16","sampling":"shuffled_epochs","seq_len":8192,"steps":571,"tokens_per_parameter":4.996285069139584,"train_tokens":299368448}},"schema_version":2,"sha256":"a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"},"loss_backend":"tiled","vocab_tile_size":2048},"metrics":{"achieved_tflops":600.9627548715764,"attention_tune_seconds":0.0,"base_learning_rate":0.001953125,"diagnostic_compile_seconds":99.6910779599566,"diagnostic_point_count":59,"diagnostics_every":10,"early_stopping_step":null,"estimated_total_flops":220902531230859264,"eval_compile_seconds":23.350994339911267,"expert_top_k":2,"experts":8,"final_validation_seconds":1.83323098893743,"flop_accounting":{"by_site":{"dot_general":1638200770560,"pallas_call":695784701952,"tpu_flash_causal_attention_bwd_dkv":1236950581248,"tpu_flash_causal_attention_bwd_dq":1236950581248,"tpu_flash_causal_attention_fwd":1236950581248},"elementwise_per_sequence":9518381752,"matmul_per_sequence":6044837216256,"method":"traced-jaxpr","warnings":[]},"flops_per_token":737895168,"mfu_estimate":0.13658244428899463,"model_tier":"60m","parameter_count":59959680,"parameters":102509952,"schedule_steps":571,"tokens_per_parameter":4.996285069139584,"tokens_per_second":814428.3645337225,"tokens_processed":299368448,"total_compile_seconds":181.43496975686867,"total_parameter_count":102509952,"train_compile_seconds":58.3928974570008,"train_loss":5.236739158630371,"train_seconds":367.58106794499326,"training_data_epochs":0.037894837263947954,"training_data_sharding":"rank_disjoint_shuffled_windows","training_sampling":"shuffled_epochs","training_steps":571,"training_token_budget":299368448,"training_usable_tokens_per_epoch":7899979776,"validation_loss":5.265833914279938,"validation_probe_count":0,"validation_probe_seconds":0.0,"validation_tokens":4194304},"profile":"dev","schema_version":1,"seed":1337,"status":"ok","system":{"controller_process_index":1,"device_count":16,"device_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],"device_kinds":["TPU v4"],"jax_version":"0.11.0","jaxlib_version":"0.11.0","libtpu_version":"0.0.44.1","local_device_count":4,"platform":"tpu","process_count":4,"process_indices":[0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3],"python_version":"3.12.13"},"track":"open"}
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/training.riglog ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6f9dab5be886947d12f9d0ea70a85a90a25c943195fe8124b7188e6e6cde0b90
3
+ size 327744
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1337/validation.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ step,tokens_processed,kind,domain,validation_tokens,validation_loss,perplexity,validation_seconds,canonical
2
+ 571,299368448,fineweb,fineweb,4194304,5.265833914279938,193.60769370796328,1.83323098893743,true
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/diagnostics.riglog ADDED
Binary file (75.1 kB). View file
 
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/metrics.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.343035042285919,
32
+ "perplexity": 209.1465165097913,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8331210270989686
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.00390625,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.001953125,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.5862729377122,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.001953125,
142
+ "diagnostic_compile_seconds": 91.81675400596578,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.191538011073135,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8331210270989686,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13649688021311643,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 813918.1539371624,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 175.10236649704166,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 60.094074480002746,
176
+ "train_loss": 5.334192276000977,
177
+ "train_seconds": 367.81148884794675,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.343035042285919,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1338,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/result.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.343035042285919,
32
+ "perplexity": 209.1465165097913,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8331210270989686
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.00390625,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.001953125,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.5862729377122,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.001953125,
142
+ "diagnostic_compile_seconds": 91.81675400596578,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.191538011073135,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8331210270989686,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13649688021311643,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 813918.1539371624,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 175.10236649704166,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 60.094074480002746,
176
+ "train_loss": 5.334192276000977,
177
+ "train_seconds": 367.81148884794675,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.343035042285919,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1338,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/stderr.log ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ E0819 12:10:53.939621 3499222 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
2
+ E0819 12:10:53.940003 3545831 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
3
+ E0819 12:10:53.940986 297851 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
4
+ E0819 12:10:53.941081 3446297 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
5
+
6
+ ◆ GPT TPU RIG reference / jax
7
+
8
+
9
+ ● Attention tile preflight resolving the shipped lookup or shape heuristic
10
+ ╭─ run configuration ────────────────────────────────────────────────────────╮
11
+ │ experiment config config.yaml · dev · sha256:a4bad1be29b5 │
12
+ │ devices 16 × TPU v4 │
13
+ │ JAX processes 4 (this rank 1) │
14
+ │ mesh data=16 (replicated model) │
15
+ │ dataset 79 train + 1 val shard(s) │
16
+ │ train / val tokens 7,900,000,000 / 100,000,000 │
17
+ │ downstream not requested │
18
+ │ model 60m · L12 D384 H6 RoPE RMSNorm GELU MLP×4 │
19
+ │ parameters 102.51M │
20
+ │ parameterization complete_d_p · mN=1 · mL=1 · mD=1 │
21
+ │ global batch 64 × 8192 tokens │
22
+ │ train sampling shuffled epochs · 7,899,979,776 unique targets/epoch │
23
+ │ compute bfloat16 │
24
+ │ attention tpu_flash │
25
+ │ attention tuning heuristic · key 6a5b398e13ed │
26
+ │ attention fwd q512 · kv512/256 │
27
+ │ attention dK/dV q512/256 · kv512/256 │
28
+ │ attention dQ q256 · kv512/256 │
29
+ │ output loss tiled CE (semantic 50,304, tile 2,048) │
30
+ │ diagnostics step 1 / every 10 / final │
31
+ │ duration 571 steps │
32
+ │ train tokens 299.37M │
33
+ │ traced FLOPs 220902.53T │
34
+ │ FLOP breakdown dot_general 1,638,200,770,560 (27.1%) · tpu_flash_ca… │
35
+ │ XProf disabled │
36
+ ╰────────────────────────────────────────────────────────────────────────────╯
37
+
38
+ ● Compiling train step compilation is outside train_seconds
39
+
40
+ ● Compiling sparse diagnostics separate executable; compilation is outside train_seconds
41
+
42
+ ● Compiling evaluation reused by probes and final validation
43
+
44
+ ● Training train compiled in 60.09s, eval in 23.19s; periodic validation disabled
45
+ 1/571 ────────────────── loss 10.9063 lr 6.85e-05 |g| 2.266 697.52K tok/s
46
+ 10/571 ────────────────── loss 9.2615 lr 6.85e-04 |g| 2.044 747.74K tok/s
47
+ 20/571 ━───────────────── loss 7.6818 lr 1.37e-03 |g| 0.948 780.49K tok/s
48
+ 30/571 ━───────────────── loss 7.6517 lr 2.06e-03 |g| 1.644 792.00K tok/s
49
+ 40/571 ━───────────────── loss 7.7416 lr 2.74e-03 |g| 0.270 798.13K tok/s
50
+ 50/571 ━━──────────────── loss 7.7910 lr 3.43e-03 |g| 0.619 801.97K tok/s
51
+ 60/571 ━━──────────────── loss 8.1727 lr 3.91e-03 |g| 2.074 804.47K tok/s
52
+ 70/571 ━━──────────────── loss 7.6729 lr 3.90e-03 |g| 0.464 806.12K tok/s
53
+ 80/571 ━━━─────────────── loss 7.7205 lr 3.89e-03 |g| 1.189 807.30K tok/s
54
+ 90/571 ━━━─────────────── loss 7.4029 lr 3.87e-03 |g| 0.340 808.22K tok/s
55
+ 100/571 ━━━─────────────── loss 7.4355 lr 3.85e-03 |g| 0.453 808.95K tok/s
56
+ 110/571 ━━━─────────────── loss 7.2449 lr 3.81e-03 |g| 0.287 809.55K tok/s
57
+ 120/571 ━━━━────────────── loss 7.1421 lr 3.78e-03 |g| 0.436 810.02K tok/s
58
+ 130/571 ━━━━────────────── loss 6.9155 lr 3.73e-03 |g| 0.454 810.41K tok/s
59
+ 140/571 ━━━━────────────── loss 6.8567 lr 3.68e-03 |g| 0.295 810.74K tok/s
60
+ 150/571 ━━━━━───────────── loss 6.7766 lr 3.63e-03 |g| 0.588 811.03K tok/s
61
+ 160/571 ━━━━━───────────── loss 6.6766 lr 3.57e-03 |g| 0.337 811.27K tok/s
62
+ 170/571 ━━━━━───────────── loss 6.5355 lr 3.50e-03 |g| 0.258 811.50K tok/s
63
+ 180/571 ━━━━━━──────────── loss 6.4862 lr 3.43e-03 |g| 0.276 811.69K tok/s
64
+ 190/571 ━━━━━━──────────── loss 6.4010 lr 3.36e-03 |g| 0.305 811.87K tok/s
65
+ 200/571 ━━━━━━──────────── loss 6.4117 lr 3.28e-03 |g| 0.317 812.02K tok/s
66
+ 210/571 ━━━━━━━─────────── loss 6.3494 lr 3.19e-03 |g| 0.320 812.17K tok/s
67
+ 220/571 ━━━━━━━─────────── loss 6.3065 lr 3.10e-03 |g| 0.389 812.30K tok/s
68
+ 230/571 ━━━━━━━─────────── loss 6.3033 lr 3.01e-03 |g| 0.368 812.41K tok/s
69
+ 240/571 ━━━━━━━━────────── loss 6.1950 lr 2.92e-03 |g| 0.293 812.52K tok/s
70
+ 250/571 ━━━━━━━━────────── loss 6.1670 lr 2.82e-03 |g| 0.274 812.62K tok/s
71
+ 260/571 ━━━━━━━━────────── loss 6.1492 lr 2.72e-03 |g| 0.439 812.71K tok/s
72
+ 270/571 ━━━━━━━━━───────── loss 5.9822 lr 2.62e-03 |g| 0.202 812.79K tok/s
73
+ 280/571 ━━━━━━━━━───────── loss 6.0481 lr 2.51e-03 |g| 0.275 812.87K tok/s
74
+ 290/571 ━━━━━━━━━───────── loss 5.9827 lr 2.41e-03 |g| 0.310 812.94K tok/s
75
+ 300/571 ━━━━━━━━━───────── loss 5.9447 lr 2.30e-03 |g| 0.192 813.01K tok/s
76
+ 310/571 ━━━━━━━━━━──────── loss 5.9078 lr 2.19e-03 |g| 0.268 813.08K tok/s
77
+ 320/571 ━━━━━━━━━━──────── loss 5.8559 lr 2.08e-03 |g| 0.296 813.14K tok/s
78
+ 330/571 ━━━━━━━━━━──────── loss 5.8325 lr 1.98e-03 |g| 0.226 813.19K tok/s
79
+ 340/571 ━━━━━━━━━━━─────── loss 5.8810 lr 1.87e-03 |g| 0.211 813.25K tok/s
80
+ 350/571 ━━━━━━━━━━━─────── loss 5.7645 lr 1.76e-03 |g| 0.182 813.30K tok/s
81
+ 360/571 ━━━━━━━━━━━─────── loss 5.7002 lr 1.66e-03 |g| 0.152 813.34K tok/s
82
+ 370/571 ━━━━━━━━━━━━────── loss 5.6963 lr 1.56e-03 |g| 0.179 813.39K tok/s
83
+ 380/571 ━━━━━━━━━━━━────── loss 5.6530 lr 1.46e-03 |g| 0.201 813.43K tok/s
84
+ 390/571 ━━━━━━━━━━━━────── loss 5.6420 lr 1.36e-03 |g| 0.259 813.47K tok/s
85
+ 400/571 ━━━━━━━━━━━━━───── loss 5.6232 lr 1.27e-03 |g| 0.147 813.51K tok/s
86
+ 410/571 ━━━━━━━━━━━━━───── loss 5.5477 lr 1.18e-03 |g| 0.152 813.55K tok/s
87
+ 420/571 ━━━━━━━━━━━━━───── loss 5.5924 lr 1.09e-03 |g| 0.164 813.58K tok/s
88
+ 430/571 ━━━━━━━━━━━━━━──── loss 5.5538 lr 1.00e-03 |g| 0.131 813.61K tok/s
89
+ 440/571 ━━━━━━━━━━━━━━──── loss 5.5317 lr 9.25e-04 |g| 0.171 813.64K tok/s
90
+ 450/571 ━━━━━━━━━━━━━━──── loss 5.5062 lr 8.50e-04 |g| 0.108 813.67K tok/s
91
+ 460/571 ━━━━━━━━━━━━━━━─── loss 5.4919 lr 7.80e-04 |g| 0.112 813.70K tok/s
92
+ 470/571 ━━━━━━━━━━━━━━━─── loss 5.4406 lr 7.15e-04 |g| 0.114 813.73K tok/s
93
+ 480/571 ━━━━━━━━━━━━━━━─── loss 5.4009 lr 6.56e-04 |g| 0.142 813.76K tok/s
94
+ 490/571 ━━━━━━━━━━━━━━━─── loss 5.4421 lr 6.02e-04 |g| 0.127 813.79K tok/s
95
+ 500/571 ━━━━━━━━━━━━━━━━── loss 5.3921 lr 5.54e-04 |g| 0.140 813.81K tok/s
96
+ 510/571 ━━━━━━━━━━━━━━━━── loss 5.3781 lr 5.11e-04 |g| 0.091 813.84K tok/s
97
+ 520/571 ━━━━━━━━━━━━━━━━── loss 5.3662 lr 4.75e-04 |g| 0.127 813.86K tok/s
98
+ 530/571 ━━━━━━━━━━━━━━━━━─ loss 5.3878 lr 4.46e-04 |g| 0.108 813.88K tok/s
99
+ 540/571 ━━━━━━━━━━━━━━━━━─ loss 5.3410 lr 4.22e-04 |g| 0.111 813.90K tok/s
100
+ 550/571 ━━━━━━━━━━━━━━━━━─ loss 5.3519 lr 4.05e-04 |g| 0.091 813.92K tok/s
101
+ 560/571 ━━━━━━━━━━━━━━━━━━ loss 5.3538 lr 3.95e-04 |g| 0.085 813.94K tok/s
102
+ 570/571 ━━━━━━━━━━━━━━━━━━ loss 5.3301 lr 3.91e-04 |g| 0.097 813.96K tok/s
103
+ 571/571 ━━━━━━━━━━━━━━━━━━ loss 5.3342 lr 3.91e-04 |g| 0.098 813.94K tok/s
104
+
105
+ ● Canonical validation 8 deterministic batches outside train_seconds
106
+
107
+ ● Fresh-domain validation skipped; no downstream data supplied
108
+
109
+ ● Artifacts training.riglog + validation.csv + diagnostics.riglog
110
+
111
+ ✓ synchronized training 367.811s (compilation excluded)
112
+ validation loss 5.3430 in 1.833s
113
+
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/stdout.log ADDED
@@ -0,0 +1 @@
 
 
1
+ RIG_RESULT={"artifacts":{"diagnostics":"diagnostics.riglog","training_curve":"training.riglog","validation_curve":"validation.csv"},"checkpoint":null,"contract":{"dataset_id":"fineweb-8b-gpt2","model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"model_id":"reference-gpt-v3-family","sequence_length":8192,"tokenizer_id":"gpt2"},"evaluations":{"fineweb":{"canonical":true,"loss":5.343035042285919,"perplexity":209.1465165097913,"scored_tokens":4194304,"seconds":1.8331210270989686}},"implementation":{"attention_backend":"tpu_flash","attention_tuning":{"key_digest":"6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8","resolution_source":"heuristic","tiles":{"block_kv":512,"block_kv_compute":256,"block_kv_dkv":512,"block_kv_dkv_compute":256,"block_kv_dq":512,"block_kv_dq_compute":256,"block_q":512,"block_q_dkv":512,"block_q_dkv_compute":256,"block_q_dq":256},"tune_seconds":0.0},"configuration":{"overrides":{"study_batch_size":64,"tokens_per_parameter_micros":5000000},"path":"config.yaml","profile":"dev","resolved":{"evaluation":{"eval_batches":8,"val_every":0,"val_probe_batches":8},"kernels":{"attention_backend":"tpu_flash","loss_backend":"tiled","vocab_tile_size":2048},"logging":{"diagnostics_every":10,"log_every":10},"model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"optimizer":{"adam_epsilon":1e-08,"beta1":0.9,"beta2":0.95,"effective":{"adam_epsilon_horizon_multiplier":0.5,"beta1":0.6000000000000001,"beta2":0.7999999999999998,"global_peak_learning_rate":0.00390625,"weight_decay_horizon_multiplier":2.0},"grad_clip":0.0,"learning_rate":0.001953125,"min_lr_ratio":0.1,"warmup_steps":57,"weight_decay":0.1},"parameterization":{"attention_scale":"inverse_head_dim","base_depth":12,"base_width":384,"batch_multiplier":4.0,"data_multiplier":1.0,"depth_alpha":1.0,"depth_multiplier":1.0,"embeddings":"untied","init_std":0.02,"name":"complete_d_p","width_multiplier":1.0},"training":{"batch_size":64,"dtype":"bfloat16","sampling":"shuffled_epochs","seq_len":8192,"steps":571,"tokens_per_parameter":4.996285069139584,"train_tokens":299368448}},"schema_version":2,"sha256":"a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"},"loss_backend":"tiled","vocab_tile_size":2048},"metrics":{"achieved_tflops":600.5862729377122,"attention_tune_seconds":0.0,"base_learning_rate":0.001953125,"diagnostic_compile_seconds":91.81675400596578,"diagnostic_point_count":59,"diagnostics_every":10,"early_stopping_step":null,"estimated_total_flops":220902531230859264,"eval_compile_seconds":23.191538011073135,"expert_top_k":2,"experts":8,"final_validation_seconds":1.8331210270989686,"flop_accounting":{"by_site":{"dot_general":1638200770560,"pallas_call":695784701952,"tpu_flash_causal_attention_bwd_dkv":1236950581248,"tpu_flash_causal_attention_bwd_dq":1236950581248,"tpu_flash_causal_attention_fwd":1236950581248},"elementwise_per_sequence":9518381752,"matmul_per_sequence":6044837216256,"method":"traced-jaxpr","warnings":[]},"flops_per_token":737895168,"mfu_estimate":0.13649688021311643,"model_tier":"60m","parameter_count":59959680,"parameters":102509952,"schedule_steps":571,"tokens_per_parameter":4.996285069139584,"tokens_per_second":813918.1539371624,"tokens_processed":299368448,"total_compile_seconds":175.10236649704166,"total_parameter_count":102509952,"train_compile_seconds":60.094074480002746,"train_loss":5.334192276000977,"train_seconds":367.81148884794675,"training_data_epochs":0.037894837263947954,"training_data_sharding":"rank_disjoint_shuffled_windows","training_sampling":"shuffled_epochs","training_steps":571,"training_token_budget":299368448,"training_usable_tokens_per_epoch":7899979776,"validation_loss":5.343035042285919,"validation_probe_count":0,"validation_probe_seconds":0.0,"validation_tokens":4194304},"profile":"dev","schema_version":1,"seed":1338,"status":"ok","system":{"controller_process_index":1,"device_count":16,"device_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],"device_kinds":["TPU v4"],"jax_version":"0.11.0","jaxlib_version":"0.11.0","libtpu_version":"0.0.44.1","local_device_count":4,"platform":"tpu","process_count":4,"process_indices":[0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3],"python_version":"3.12.13"},"track":"open"}
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/training.riglog ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2198aa232a0e104059ac2ceb020feedabcd54ba78d3702d6b0b32b95866775d
3
+ size 327744
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1338/validation.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ step,tokens_processed,kind,domain,validation_tokens,validation_loss,perplexity,validation_seconds,canonical
2
+ 571,299368448,fineweb,fineweb,4194304,5.343035042285919,209.1465165097913,1.8331210270989686,true
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/diagnostics.riglog ADDED
Binary file (75.1 kB). View file
 
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/metrics.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.254374384880066,
32
+ "perplexity": 191.4017045894264,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8720559159992263
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.00390625,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.001953125,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.7166387097175,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.001953125,
142
+ "diagnostic_compile_seconds": 102.93548001605086,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.249745324952528,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8720559159992263,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13652650879766307,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 814094.8264208141,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 184.8709388379939,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 58.6857134969905,
176
+ "train_loss": 5.280473232269287,
177
+ "train_seconds": 367.7316674719332,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.254374384880066,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1339,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/result.json ADDED
@@ -0,0 +1,244 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "artifacts": {
3
+ "diagnostics": "diagnostics.riglog",
4
+ "training_curve": "training.riglog",
5
+ "validation_curve": "validation.csv"
6
+ },
7
+ "checkpoint": null,
8
+ "contract": {
9
+ "dataset_id": "fineweb-8b-gpt2",
10
+ "model": {
11
+ "d_model": 384,
12
+ "heads": 6,
13
+ "layers": 12,
14
+ "mlp_activation": "gelu",
15
+ "mlp_mult": 4,
16
+ "normalization": "rms_norm",
17
+ "parameterization": "complete_d_p",
18
+ "position_encoding": "rope_base_10000",
19
+ "semantic_vocab_size": 50304,
20
+ "tied_embeddings": false,
21
+ "tier": "60m",
22
+ "vocab_size": 50304
23
+ },
24
+ "model_id": "reference-gpt-v3-family",
25
+ "sequence_length": 8192,
26
+ "tokenizer_id": "gpt2"
27
+ },
28
+ "evaluations": {
29
+ "fineweb": {
30
+ "canonical": true,
31
+ "loss": 5.254374384880066,
32
+ "perplexity": 191.4017045894264,
33
+ "scored_tokens": 4194304,
34
+ "seconds": 1.8720559159992263
35
+ }
36
+ },
37
+ "implementation": {
38
+ "attention_backend": "tpu_flash",
39
+ "attention_tuning": {
40
+ "key_digest": "6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8",
41
+ "resolution_source": "heuristic",
42
+ "tiles": {
43
+ "block_kv": 512,
44
+ "block_kv_compute": 256,
45
+ "block_kv_dkv": 512,
46
+ "block_kv_dkv_compute": 256,
47
+ "block_kv_dq": 512,
48
+ "block_kv_dq_compute": 256,
49
+ "block_q": 512,
50
+ "block_q_dkv": 512,
51
+ "block_q_dkv_compute": 256,
52
+ "block_q_dq": 256
53
+ },
54
+ "tune_seconds": 0.0
55
+ },
56
+ "configuration": {
57
+ "overrides": {
58
+ "study_batch_size": 64,
59
+ "tokens_per_parameter_micros": 5000000
60
+ },
61
+ "path": "config.yaml",
62
+ "profile": "dev",
63
+ "resolved": {
64
+ "evaluation": {
65
+ "eval_batches": 8,
66
+ "val_every": 0,
67
+ "val_probe_batches": 8
68
+ },
69
+ "kernels": {
70
+ "attention_backend": "tpu_flash",
71
+ "loss_backend": "tiled",
72
+ "vocab_tile_size": 2048
73
+ },
74
+ "logging": {
75
+ "diagnostics_every": 10,
76
+ "log_every": 10
77
+ },
78
+ "model": {
79
+ "d_model": 384,
80
+ "heads": 6,
81
+ "layers": 12,
82
+ "mlp_activation": "gelu",
83
+ "mlp_mult": 4,
84
+ "normalization": "rms_norm",
85
+ "parameterization": "complete_d_p",
86
+ "position_encoding": "rope_base_10000",
87
+ "semantic_vocab_size": 50304,
88
+ "tied_embeddings": false,
89
+ "tier": "60m",
90
+ "vocab_size": 50304
91
+ },
92
+ "optimizer": {
93
+ "adam_epsilon": 1e-08,
94
+ "beta1": 0.9,
95
+ "beta2": 0.95,
96
+ "effective": {
97
+ "adam_epsilon_horizon_multiplier": 0.5,
98
+ "beta1": 0.6000000000000001,
99
+ "beta2": 0.7999999999999998,
100
+ "global_peak_learning_rate": 0.00390625,
101
+ "weight_decay_horizon_multiplier": 2.0
102
+ },
103
+ "grad_clip": 0.0,
104
+ "learning_rate": 0.001953125,
105
+ "min_lr_ratio": 0.1,
106
+ "warmup_steps": 57,
107
+ "weight_decay": 0.1
108
+ },
109
+ "parameterization": {
110
+ "attention_scale": "inverse_head_dim",
111
+ "base_depth": 12,
112
+ "base_width": 384,
113
+ "batch_multiplier": 4.0,
114
+ "data_multiplier": 1.0,
115
+ "depth_alpha": 1.0,
116
+ "depth_multiplier": 1.0,
117
+ "embeddings": "untied",
118
+ "init_std": 0.02,
119
+ "name": "complete_d_p",
120
+ "width_multiplier": 1.0
121
+ },
122
+ "training": {
123
+ "batch_size": 64,
124
+ "dtype": "bfloat16",
125
+ "sampling": "shuffled_epochs",
126
+ "seq_len": 8192,
127
+ "steps": 571,
128
+ "tokens_per_parameter": 4.996285069139584,
129
+ "train_tokens": 299368448
130
+ }
131
+ },
132
+ "schema_version": 2,
133
+ "sha256": "a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"
134
+ },
135
+ "loss_backend": "tiled",
136
+ "vocab_tile_size": 2048
137
+ },
138
+ "metrics": {
139
+ "achieved_tflops": 600.7166387097175,
140
+ "attention_tune_seconds": 0.0,
141
+ "base_learning_rate": 0.001953125,
142
+ "diagnostic_compile_seconds": 102.93548001605086,
143
+ "diagnostic_point_count": 59,
144
+ "diagnostics_every": 10,
145
+ "early_stopping_step": null,
146
+ "estimated_total_flops": 220902531230859264,
147
+ "eval_compile_seconds": 23.249745324952528,
148
+ "expert_top_k": 2,
149
+ "experts": 8,
150
+ "final_validation_seconds": 1.8720559159992263,
151
+ "flop_accounting": {
152
+ "by_site": {
153
+ "dot_general": 1638200770560,
154
+ "pallas_call": 695784701952,
155
+ "tpu_flash_causal_attention_bwd_dkv": 1236950581248,
156
+ "tpu_flash_causal_attention_bwd_dq": 1236950581248,
157
+ "tpu_flash_causal_attention_fwd": 1236950581248
158
+ },
159
+ "elementwise_per_sequence": 9518381752,
160
+ "matmul_per_sequence": 6044837216256,
161
+ "method": "traced-jaxpr",
162
+ "warnings": []
163
+ },
164
+ "flops_per_token": 737895168,
165
+ "mfu_estimate": 0.13652650879766307,
166
+ "model_tier": "60m",
167
+ "parameter_count": 59959680,
168
+ "parameters": 102509952,
169
+ "schedule_steps": 571,
170
+ "tokens_per_parameter": 4.996285069139584,
171
+ "tokens_per_second": 814094.8264208141,
172
+ "tokens_processed": 299368448,
173
+ "total_compile_seconds": 184.8709388379939,
174
+ "total_parameter_count": 102509952,
175
+ "train_compile_seconds": 58.6857134969905,
176
+ "train_loss": 5.280473232269287,
177
+ "train_seconds": 367.7316674719332,
178
+ "training_data_epochs": 0.037894837263947954,
179
+ "training_data_sharding": "rank_disjoint_shuffled_windows",
180
+ "training_sampling": "shuffled_epochs",
181
+ "training_steps": 571,
182
+ "training_token_budget": 299368448,
183
+ "training_usable_tokens_per_epoch": 7899979776,
184
+ "validation_loss": 5.254374384880066,
185
+ "validation_probe_count": 0,
186
+ "validation_probe_seconds": 0.0,
187
+ "validation_tokens": 4194304
188
+ },
189
+ "profile": "dev",
190
+ "schema_version": 1,
191
+ "seed": 1339,
192
+ "status": "ok",
193
+ "system": {
194
+ "controller_process_index": 1,
195
+ "device_count": 16,
196
+ "device_ids": [
197
+ 0,
198
+ 1,
199
+ 2,
200
+ 3,
201
+ 4,
202
+ 5,
203
+ 6,
204
+ 7,
205
+ 8,
206
+ 9,
207
+ 10,
208
+ 11,
209
+ 12,
210
+ 13,
211
+ 14,
212
+ 15
213
+ ],
214
+ "device_kinds": [
215
+ "TPU v4"
216
+ ],
217
+ "jax_version": "0.11.0",
218
+ "jaxlib_version": "0.11.0",
219
+ "libtpu_version": "0.0.44.1",
220
+ "local_device_count": 4,
221
+ "platform": "tpu",
222
+ "process_count": 4,
223
+ "process_indices": [
224
+ 0,
225
+ 0,
226
+ 0,
227
+ 0,
228
+ 1,
229
+ 1,
230
+ 1,
231
+ 1,
232
+ 2,
233
+ 2,
234
+ 2,
235
+ 2,
236
+ 3,
237
+ 3,
238
+ 3,
239
+ 3
240
+ ],
241
+ "python_version": "3.12.13"
242
+ },
243
+ "track": "open"
244
+ }
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/stderr.log ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ E0819 12:20:22.789940 3554784 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
2
+ E0819 12:20:22.789954 3508142 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
3
+ E0819 12:20:22.790718 3455195 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
4
+ E0819 12:20:22.791455 307817 hugepage_text.cc:340] RAW: File offset incorrectly aligned for file-backed THP: (600000 & ~ffffffffffe00000) = 0 != 1000 = (201000 & ~ffffffffffe00000)
5
+
6
+ ◆ GPT TPU RIG reference / jax
7
+
8
+
9
+ ● Attention tile preflight resolving the shipped lookup or shape heuristic
10
+ ╭─ run configuration ────────────────────────────────────────────────────────╮
11
+ │ experiment config config.yaml · dev · sha256:a4bad1be29b5 │
12
+ │ devices 16 × TPU v4 │
13
+ │ JAX processes 4 (this rank 1) │
14
+ │ mesh data=16 (replicated model) │
15
+ │ dataset 79 train + 1 val shard(s) │
16
+ │ train / val tokens 7,900,000,000 / 100,000,000 │
17
+ │ downstream not requested │
18
+ │ model 60m · L12 D384 H6 RoPE RMSNorm GELU MLP×4 │
19
+ │ parameters 102.51M │
20
+ │ parameterization complete_d_p · mN=1 · mL=1 · mD=1 │
21
+ │ global batch 64 × 8192 tokens │
22
+ │ train sampling shuffled epochs · 7,899,979,776 unique targets/epoch │
23
+ │ compute bfloat16 │
24
+ │ attention tpu_flash │
25
+ │ attention tuning heuristic · key 6a5b398e13ed │
26
+ │ attention fwd q512 · kv512/256 │
27
+ │ attention dK/dV q512/256 · kv512/256 │
28
+ │ attention dQ q256 · kv512/256 │
29
+ │ output loss tiled CE (semantic 50,304, tile 2,048) │
30
+ │ diagnostics step 1 / every 10 / final │
31
+ │ duration 571 steps │
32
+ │ train tokens 299.37M │
33
+ │ traced FLOPs 220902.53T │
34
+ │ FLOP breakdown dot_general 1,638,200,770,560 (27.1%) · tpu_flash_ca… │
35
+ │ XProf disabled │
36
+ ╰────────────────────────────────────────────────────────────────────────────╯
37
+
38
+ ● Compiling train step compilation is outside train_seconds
39
+
40
+ ● Compiling sparse diagnostics separate executable; compilation is outside train_seconds
41
+
42
+ ● Compiling evaluation reused by probes and final validation
43
+
44
+ ● Training train compiled in 58.69s, eval in 23.25s; periodic validation disabled
45
+ 1/571 ────────────────── loss 10.8871 lr 6.85e-05 |g| 2.681 587.06K tok/s
46
+ 10/571 ────────────────── loss 9.1372 lr 6.85e-04 |g| 1.577 760.09K tok/s
47
+ 20/571 ━───────────────── loss 7.6899 lr 1.37e-03 |g| 0.671 786.93K tok/s
48
+ 30/571 ━───────────────── loss 7.5302 lr 2.06e-03 |g| 0.407 796.25K tok/s
49
+ 40/571 ━───────────────── loss 7.6072 lr 2.74e-03 |g| 0.370 801.22K tok/s
50
+ 50/571 ━━──────────────── loss 7.5658 lr 3.43e-03 |g| 0.916 804.18K tok/s
51
+ 60/571 ━━──────────────── loss 7.5213 lr 3.91e-03 |g| 0.570 806.09K tok/s
52
+ 70/571 ━━──────────────── loss 7.4393 lr 3.90e-03 |g| 0.413 807.45K tok/s
53
+ 80/571 ━━━─────────────── loss 7.3816 lr 3.89e-03 |g| 0.725 808.45K tok/s
54
+ 90/571 ━━━─────────────── loss 7.3361 lr 3.87e-03 |g| 0.869 809.22K tok/s
55
+ 100/571 ━━━─────────────── loss 7.0859 lr 3.85e-03 |g| 0.425 809.82K tok/s
56
+ 110/571 ━━━─────────────── loss 6.9719 lr 3.81e-03 |g| 0.247 810.31K tok/s
57
+ 120/571 ━━━━────────────── loss 6.8888 lr 3.78e-03 |g| 0.276 810.70K tok/s
58
+ 130/571 ━━━━────────────── loss 6.8179 lr 3.73e-03 |g| 0.371 811.04K tok/s
59
+ 140/571 ━━━━────────────── loss 6.6586 lr 3.68e-03 |g| 0.315 811.32K tok/s
60
+ 150/571 ━━━━━───────────── loss 6.6408 lr 3.63e-03 |g| 0.547 811.57K tok/s
61
+ 160/571 ━━━━━───────────── loss 6.5504 lr 3.57e-03 |g| 0.547 811.79K tok/s
62
+ 170/571 ━━━━━───────────── loss 6.5749 lr 3.50e-03 |g| 0.570 811.98K tok/s
63
+ 180/571 ━━━━━━──────────── loss 6.4157 lr 3.43e-03 |g| 0.271 812.12K tok/s
64
+ 190/571 ━━━━━━──────────── loss 6.3916 lr 3.36e-03 |g| 0.284 812.28K tok/s
65
+ 200/571 ━━━━━━──────────── loss 6.3463 lr 3.28e-03 |g| 0.481 812.41K tok/s
66
+ 210/571 ━━━━━━━─────────── loss 6.3026 lr 3.19e-03 |g| 0.284 812.54K tok/s
67
+ 220/571 ━━━━━━━─────────── loss 6.2790 lr 3.10e-03 |g| 0.323 812.65K tok/s
68
+ 230/571 ━━━━━━━─────────── loss 6.1845 lr 3.01e-03 |g| 0.189 812.75K tok/s
69
+ 240/571 ━━━━━━━━────────── loss 6.1513 lr 2.92e-03 |g| 0.502 812.85K tok/s
70
+ 250/571 ━━━━━━━━────────── loss 6.1203 lr 2.82e-03 |g| 0.311 812.94K tok/s
71
+ 260/571 ━━━━━━━━────────── loss 6.0623 lr 2.72e-03 |g| 0.287 813.02K tok/s
72
+ 270/571 ━━━━━━━━━───────── loss 6.0136 lr 2.62e-03 |g| 0.306 813.10K tok/s
73
+ 280/571 ━━━━━━━━━───────── loss 5.9230 lr 2.51e-03 |g| 0.220 813.17K tok/s
74
+ 290/571 ━━━━━━━━━───────── loss 5.8959 lr 2.41e-03 |g| 0.210 813.23K tok/s
75
+ 300/571 ━━━━━━━━━───────── loss 5.8038 lr 2.30e-03 |g| 0.244 813.29K tok/s
76
+ 310/571 ━━━━━━━━━━──────── loss 5.8204 lr 2.19e-03 |g| 0.238 813.35K tok/s
77
+ 320/571 ━━━━━━━━━━──────── loss 5.7668 lr 2.08e-03 |g| 0.209 813.40K tok/s
78
+ 330/571 ━━━━━━━━━━──────── loss 5.7297 lr 1.98e-03 |g| 0.254 813.45K tok/s
79
+ 340/571 ━━━━━━━━━━━─────── loss 5.6694 lr 1.87e-03 |g| 0.226 813.50K tok/s
80
+ 350/571 ━━━━━━━━━━━─────── loss 5.6426 lr 1.76e-03 |g| 0.153 813.53K tok/s
81
+ 360/571 ━━━━━━━━━━━─────── loss 5.6963 lr 1.66e-03 |g| 0.198 813.57K tok/s
82
+ 370/571 ━━━━━━━━━━━━────── loss 5.6138 lr 1.56e-03 |g| 0.206 813.61K tok/s
83
+ 380/571 ━━━━━━━━━━━━────── loss 5.6172 lr 1.46e-03 |g| 0.189 813.65K tok/s
84
+ 390/571 ━━━━━━━━━━━━────── loss 5.5650 lr 1.36e-03 |g| 0.168 813.69K tok/s
85
+ 400/571 ━━━━━━━━━━━━━───── loss 5.5994 lr 1.27e-03 |g| 0.195 813.72K tok/s
86
+ 410/571 ━━━━━━━━━━━━━───── loss 5.4641 lr 1.18e-03 |g| 0.170 813.76K tok/s
87
+ 420/571 ━━━━━━━━━━━━━───── loss 5.5093 lr 1.09e-03 |g| 0.195 813.79K tok/s
88
+ 430/571 ━━━━━━━━━━━━━━──── loss 5.4275 lr 1.00e-03 |g| 0.163 813.82K tok/s
89
+ 440/571 ━━━━━━━━━━━━━━──── loss 5.4346 lr 9.25e-04 |g| 0.108 813.85K tok/s
90
+ 450/571 ━━━━━━━━━━━━━━──── loss 5.4257 lr 8.50e-04 |g| 0.185 813.88K tok/s
91
+ 460/571 ━━━━━━━━━━━━━━━─── loss 5.3950 lr 7.80e-04 |g| 0.170 813.91K tok/s
92
+ 470/571 ━━━━━━━━━━━━━━━─── loss 5.3806 lr 7.15e-04 |g| 0.146 813.93K tok/s
93
+ 480/571 ━━━━━━━━━━━━━━━─── loss 5.3560 lr 6.56e-04 |g| 0.114 813.95K tok/s
94
+ 490/571 ━━━━━━━━━━━━━━━─── loss 5.3431 lr 6.02e-04 |g| 0.096 813.98K tok/s
95
+ 500/571 ━━━━━━━━━━━━━━━━── loss 5.3796 lr 5.54e-04 |g| 0.140 814.00K tok/s
96
+ 510/571 ━━━━━━━━━━━━━━━━── loss 5.3610 lr 5.11e-04 |g| 0.100 814.02K tok/s
97
+ 520/571 ━━━━━━━━━━━━━━━━── loss 5.3288 lr 4.75e-04 |g| 0.089 814.05K tok/s
98
+ 530/571 ━━━━━━━━━━━━━━━━━─ loss 5.2656 lr 4.46e-04 |g| 0.109 814.06K tok/s
99
+ 540/571 ━━━━━━━━━━━━━━━━━─ loss 5.2860 lr 4.22e-04 |g| 0.088 814.08K tok/s
100
+ 550/571 ━━━━━━━━━━━━━━━━━─ loss 5.2900 lr 4.05e-04 |g| 0.094 814.10K tok/s
101
+ 560/571 ━━━━━━━━━━━━━━━━━━ loss 5.2739 lr 3.95e-04 |g| 0.113 814.12K tok/s
102
+ 570/571 ━━━━━━━━━━━━━━━━━━ loss 5.2502 lr 3.91e-04 |g| 0.087 814.13K tok/s
103
+ 571/571 ━━━━━━━━━━━━━━━━━━ loss 5.2805 lr 3.91e-04 |g| 0.086 814.11K tok/s
104
+
105
+ ● Canonical validation 8 deterministic batches outside train_seconds
106
+
107
+ ● Fresh-domain validation skipped; no downstream data supplied
108
+
109
+ ● Artifacts training.riglog + validation.csv + diagnostics.riglog
110
+
111
+ ✓ synchronized training 367.732s (compilation excluded)
112
+ validation loss 5.2544 in 1.872s
113
+
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/stdout.log ADDED
@@ -0,0 +1 @@
 
 
1
+ RIG_RESULT={"artifacts":{"diagnostics":"diagnostics.riglog","training_curve":"training.riglog","validation_curve":"validation.csv"},"checkpoint":null,"contract":{"dataset_id":"fineweb-8b-gpt2","model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"model_id":"reference-gpt-v3-family","sequence_length":8192,"tokenizer_id":"gpt2"},"evaluations":{"fineweb":{"canonical":true,"loss":5.254374384880066,"perplexity":191.4017045894264,"scored_tokens":4194304,"seconds":1.8720559159992263}},"implementation":{"attention_backend":"tpu_flash","attention_tuning":{"key_digest":"6a5b398e13ed702d07dce43defd3e6c9511f03f0501b7aa889512dafaa291aa8","resolution_source":"heuristic","tiles":{"block_kv":512,"block_kv_compute":256,"block_kv_dkv":512,"block_kv_dkv_compute":256,"block_kv_dq":512,"block_kv_dq_compute":256,"block_q":512,"block_q_dkv":512,"block_q_dkv_compute":256,"block_q_dq":256},"tune_seconds":0.0},"configuration":{"overrides":{"study_batch_size":64,"tokens_per_parameter_micros":5000000},"path":"config.yaml","profile":"dev","resolved":{"evaluation":{"eval_batches":8,"val_every":0,"val_probe_batches":8},"kernels":{"attention_backend":"tpu_flash","loss_backend":"tiled","vocab_tile_size":2048},"logging":{"diagnostics_every":10,"log_every":10},"model":{"d_model":384,"heads":6,"layers":12,"mlp_activation":"gelu","mlp_mult":4,"normalization":"rms_norm","parameterization":"complete_d_p","position_encoding":"rope_base_10000","semantic_vocab_size":50304,"tied_embeddings":false,"tier":"60m","vocab_size":50304},"optimizer":{"adam_epsilon":1e-08,"beta1":0.9,"beta2":0.95,"effective":{"adam_epsilon_horizon_multiplier":0.5,"beta1":0.6000000000000001,"beta2":0.7999999999999998,"global_peak_learning_rate":0.00390625,"weight_decay_horizon_multiplier":2.0},"grad_clip":0.0,"learning_rate":0.001953125,"min_lr_ratio":0.1,"warmup_steps":57,"weight_decay":0.1},"parameterization":{"attention_scale":"inverse_head_dim","base_depth":12,"base_width":384,"batch_multiplier":4.0,"data_multiplier":1.0,"depth_alpha":1.0,"depth_multiplier":1.0,"embeddings":"untied","init_std":0.02,"name":"complete_d_p","width_multiplier":1.0},"training":{"batch_size":64,"dtype":"bfloat16","sampling":"shuffled_epochs","seq_len":8192,"steps":571,"tokens_per_parameter":4.996285069139584,"train_tokens":299368448}},"schema_version":2,"sha256":"a4bad1be29b57d20ec7ae7b0d539f2e777d11f5539607655550b7afca185affe"},"loss_backend":"tiled","vocab_tile_size":2048},"metrics":{"achieved_tflops":600.7166387097175,"attention_tune_seconds":0.0,"base_learning_rate":0.001953125,"diagnostic_compile_seconds":102.93548001605086,"diagnostic_point_count":59,"diagnostics_every":10,"early_stopping_step":null,"estimated_total_flops":220902531230859264,"eval_compile_seconds":23.249745324952528,"expert_top_k":2,"experts":8,"final_validation_seconds":1.8720559159992263,"flop_accounting":{"by_site":{"dot_general":1638200770560,"pallas_call":695784701952,"tpu_flash_causal_attention_bwd_dkv":1236950581248,"tpu_flash_causal_attention_bwd_dq":1236950581248,"tpu_flash_causal_attention_fwd":1236950581248},"elementwise_per_sequence":9518381752,"matmul_per_sequence":6044837216256,"method":"traced-jaxpr","warnings":[]},"flops_per_token":737895168,"mfu_estimate":0.13652650879766307,"model_tier":"60m","parameter_count":59959680,"parameters":102509952,"schedule_steps":571,"tokens_per_parameter":4.996285069139584,"tokens_per_second":814094.8264208141,"tokens_processed":299368448,"total_compile_seconds":184.8709388379939,"total_parameter_count":102509952,"train_compile_seconds":58.6857134969905,"train_loss":5.280473232269287,"train_seconds":367.7316674719332,"training_data_epochs":0.037894837263947954,"training_data_sharding":"rank_disjoint_shuffled_windows","training_sampling":"shuffled_epochs","training_steps":571,"training_token_budget":299368448,"training_usable_tokens_per_epoch":7899979776,"validation_loss":5.254374384880066,"validation_probe_count":0,"validation_probe_seconds":0.0,"validation_tokens":4194304},"profile":"dev","schema_version":1,"seed":1339,"status":"ok","system":{"controller_process_index":1,"device_count":16,"device_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],"device_kinds":["TPU v4"],"jax_version":"0.11.0","jaxlib_version":"0.11.0","libtpu_version":"0.0.44.1","local_device_count":4,"platform":"tpu","process_count":4,"process_indices":[0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3],"python_version":"3.12.13"},"track":"open"}
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/training.riglog ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:38b9cd6efe0d58487151a6f0c88506ded687453786bd0ab811c287c35b06c553
3
+ size 327744
batch-size-grid-8k/60m-moe-5tpp-bs64-lr2e-9-s1339/validation.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ step,tokens_processed,kind,domain,validation_tokens,validation_loss,perplexity,validation_seconds,canonical
2
+ 571,299368448,fineweb,fineweb,4194304,5.254374384880066,191.4017045894264,1.8720559159992263,true
batch-size-grid-8k/README.md ADDED
@@ -0,0 +1,153 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Batch size at 8k context, dense and routed
2
+
3
+ The 1,024-context ladder swept batch size at a fixed token budget and settled
4
+ on 128 sequences. This repeats that sweep at 8,192 context, on both the dense
5
+ 8k model and the routed one, and gets the opposite answer: **batch 16 wins, and
6
+ every doubling above it costs.** The learning rate is swept at each batch, so
7
+ the result is not an artifact of a rate tuned for one of them.
8
+
9
+ **42 runs · 22 MB · full resolution**
10
+
11
+ ## Setup
12
+
13
+ - **models:** `reference_8k` (dense) and `reference_moe` (top-2-of-8 routed), 8,192 context, document masking on
14
+ - **tiers:** 60M at both families, 125M routed only
15
+ - **batch sizes:** 32 and 64 sequences — 262,144 and 524,288 tokens per step
16
+ - **learning rates:** 2^-7, 2^-8, 2^-9 at every batch
17
+ - **seeds:** 3 per cell at 60M (1337, 1338, 1339), 1 per cell at 125M
18
+ - **budget:** 5 tokens per parameter, sized by *active* parameters — **held fixed across batch sizes**, so a bigger batch buys fewer optimizer steps
19
+ - **parameterization:** Complete(d)P, where `m_B` counts sequences against the recipe default of 16, so the effective peak rate already rises with `sqrt(m_B)` — batch 64 runs at twice the effective rate of batch 16 at the same base rate
20
+ - **hardware:** TPU v4 — 4 processes, 16 chips
21
+ - **dataset:** FineWeb-Edu, GPT-2 tokenizer (`fineweb-8b-gpt2`)
22
+ - **optimizer:** AdamW, beta1 0.9, beta2 0.95, weight decay 0.1, no gradient clipping
23
+ - **schedule:** 10% linear warmup, cosine decay to 10% of peak
24
+ - **precision:** bfloat16 compute, float32 diagnostics
25
+
26
+ **The batch-16 arm of this comparison is not in this folder.** It is the
27
+ existing ladder it extends: dense 8k lives in
28
+ [`lr-sweep-8k-60M`](../lr-sweep-8k-60M), routed lives in
29
+ [`moe-lr-sweep-8k`](../moe-lr-sweep-8k). Those were run first, as learning-rate
30
+ sweeps at the recipe's default batch of 16, and every comparison below reads
31
+ its batch-16 column from them.
32
+
33
+ Batch 16 is the floor on this slice: 16 chips cannot hold half a sequence each,
34
+ so the ladder point is the low edge and cannot be bracketed from below. Batch 32
35
+ and 64 are local batch 2 and 4 — the cases the segment block-spec fix (`b100d0e`)
36
+ unblocked. Neither could compile before it.
37
+
38
+ ## Reproducing
39
+
40
+ ```bash
41
+ for recipe in reference_8k reference_moe; do
42
+ tag=$([ "$recipe" = reference_8k ] && echo 8k || echo moe)
43
+ for batch in 32 64; do
44
+ for lr in 0.0078125 0.00390625 0.001953125; do
45
+ for seed in 1337 1338 1339; do
46
+ rig run "$recipe" --cluster v4-32 --profile dev --track open \
47
+ --tier 60m --tokens-per-parameter 5 --study-batch-size "$batch" \
48
+ --base-learning-rate "$lr" --seed "$seed" --checkpoint-policy none \
49
+ --name "60m-${tag}-bs${batch}-lr${lr}-s${seed}"
50
+ done
51
+ done
52
+ done
53
+ done
54
+
55
+ # 125M routed, one seed per cell
56
+ for batch in 32 64; do
57
+ for lr in 0.0078125 0.00390625 0.001953125; do
58
+ rig run reference_moe --cluster v4-32 --profile dev --track open \
59
+ --tier 125m --tokens-per-parameter 5 --study-batch-size "$batch" \
60
+ --base-learning-rate "$lr" --seed 1337 --checkpoint-policy none \
61
+ --name "125m-moe-bs${batch}-lr${lr}-s1337"
62
+ done
63
+ done
64
+ ```
65
+
66
+ ## What it showed
67
+
68
+ **Raising the batch only costs.** Each column is the best of the three learning
69
+ rates tried at that batch, so this is the best each batch could do:
70
+
71
+ | model | batch 16 | batch 32 | batch 64 |
72
+ |---|--:|--:|--:|
73
+ | 60M dense 8k | **4.0030** | 4.3963 (+0.393) | 5.3026 (+1.300) |
74
+ | 60M routed 8k | **3.9278** | 4.1956 (+0.268) | 5.0789 (+1.151) |
75
+ | 125M routed 8k | **3.5723** | 3.6260 (+0.054) | 3.8838 (+0.312) |
76
+
77
+ The penalty shrinks with scale — 1.30 nats at 60M against 0.31 at 125M — but it
78
+ never turns into a gain.
79
+
80
+ **The mechanism is steps, not learning rate.** The token budget is held fixed,
81
+ so batch 64 spends the whole budget in 571 steps where batch 16 gets 2,286:
82
+
83
+ | batch | tokens per step | steps at 60M | steps at 125M |
84
+ |---|--:|--:|--:|
85
+ | 16 | 131,072 | 2,286 | 4,709 |
86
+ | 32 | 262,144 | 1,143 | 2,355 |
87
+ | 64 | 524,288 | 571 | 1,177 |
88
+
89
+ Under Complete(d)P the effective peak rate already scales as `sqrt(m_B)`, so
90
+ the large-batch runs were not starved of learning rate. They were starved of
91
+ updates.
92
+
93
+ **And it buys no speed.** Throughput is flat across the whole grid, so there is
94
+ nothing on the other side of the ledger:
95
+
96
+ | model | batch 16 | batch 32 | batch 64 |
97
+ |---|--:|--:|--:|
98
+ | 60M dense 8k | 1,041 | 1,100 | 1,093 |
99
+ | 60M routed 8k | 597 | 594 | 601 |
100
+ | 125M routed 8k | 544 | 563 | 571 |
101
+
102
+ TFLOP/s achieved. At 1,024 context the larger batches finished sooner on the
103
+ same budget, which is what made batch 128 worth its loss penalty there. At 8k
104
+ a sequence is already eight times longer, so batch 16 saturates the chips on
105
+ its own and the usual reason to grow the batch is gone.
106
+
107
+ **Does the optimal learning rate drift?** Not resolvably. The best rate per
108
+ cell jumps around — 2^-7, then 2^-8, then 2^-9, with no consistent direction —
109
+ but the seed spread grows faster than the gaps do:
110
+
111
+ | batch | median seed spread within a cell | largest |
112
+ |---|--:|--:|
113
+ | 16 | 0.0106 | 0.0432 |
114
+ | 32 | 0.0458 | 0.1816 |
115
+ | 64 | 0.0678 | 0.1050 |
116
+
117
+ Seed spread is the standard deviation of the three seeds in one cell — how far
118
+ apart identical configurations land. At batch 16 it is small enough that a
119
+ 0.04-nat gap between learning rates is real; at batch 32 and 64 the gaps
120
+ between rates are the same size as the spread, so the apparent optimum in those
121
+ rows is noise. Separating them would need more seeds, and given that every
122
+ large-batch cell is far worse than batch 16 regardless of rate, it was not
123
+ worth the machine time. What the grid does establish is that no learning rate
124
+ in this range rescues the larger batches.
125
+
126
+ The growing spread is itself the same story told twice: fewer optimizer steps
127
+ means the run ends further from convergence, and a run that has not converged
128
+ is more sensitive to where it started.
129
+
130
+ ## Files
131
+
132
+ Each run holds `training.riglog` (loss, learning rate, gradient norm at every
133
+ step, plus the full routing series for the routed runs — balance loss, busiest
134
+ and idlest expert share, entropy, top-1 gate, logit RMS, model-wide and per
135
+ layer, plus per-expert load), `diagnostics.riglog`, `result.json`,
136
+ `metrics.json`, and `validation.csv`.
137
+
138
+ ```python
139
+ from huggingface_hub import hf_hub_download
140
+ from rig import logpack
141
+
142
+ path = hf_hub_download("quintic/rig-logs",
143
+ "batch-size-grid-8k/60m-8k-5tpp-bs32-lr2e-8-s1337/training.riglog",
144
+ repo_type="dataset")
145
+ log = logpack.read_log(path)
146
+ log.series("train_loss")
147
+ log.axis("tokens_processed") # same total for every batch in this study
148
+ ```
149
+
150
+ Run names are `<tier>-<family>-5tpp-bs<batch>-lr2e-<exponent>-s<seed>`. The
151
+ `family` field is `8k` for dense and `moe` for routed; it exists only in this
152
+ study, because it is the only one holding both families at the same tier, batch,
153
+ rate, and seed, which would otherwise collide.