CHOI0126 commited on
Commit
47e2dc2
·
verified ·
1 Parent(s): e12e447

node D backup: eval (15 files)

Browse files
qwen3_4b/eval/lm_LCa3_v11/e0/sep.log ADDED
@@ -0,0 +1 @@
 
 
1
+
qwen3_4b/eval/lm_LCa3_v11/gpu_peak_mib CHANGED
@@ -1 +1 @@
1
- 40936
 
1
+ 42801
qwen3_4b/eval/lm_LCa3_v11/lm_torch_cand_int4tg.log ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ checkpoint /workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/ckpt-0005764/adapters.pt -> /workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_torch_cand_int4tg
2
+
3
+ Warning: please use at least NVCC 12.9 for the best DeepGEMM performance
4
+ AR token ids: 20 from cache, 0 decoded (/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache)
5
+ b=1: tau 2.880 [2.656, 3.149], miss rate 0.000, tokens/pass 2.864, rows/pass 30.9, exact 20/20; 62s
6
+ round ms by kind (median [p10, p90], capture rounds left out): start 17.7 [13.8, 18.1] n=19, fused 23.3 [19.1, 23.9] n=2319
7
+ wall clock vs harness AR on 20 prompts: 7.81 vs 17.04 ms/token, speedup 2.18x decode, 2.11x with prefill
8
+ no decodes differ from AR; wrote /workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_torch_cand_int4tg/q1_eval.json
qwen3_4b/eval/lm_LCa3_v11/lm_torch_cand_int4tg/q1_eval.json ADDED
@@ -0,0 +1,1324 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "tiny": false,
4
+ "checkpoint": [
5
+ "/workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/ckpt-0005764"
6
+ ],
7
+ "b": [
8
+ 1
9
+ ],
10
+ "benches": [
11
+ "gsm8k",
12
+ "math500",
13
+ "humaneval",
14
+ "mbpp",
15
+ "mt-bench"
16
+ ],
17
+ "limit": null,
18
+ "per_bench": 4,
19
+ "time_ar": true,
20
+ "skip_ar": false,
21
+ "graphs": true,
22
+ "attention": "triton",
23
+ "fused_ops": false,
24
+ "branch_memory": "copy",
25
+ "mask_attn_kernel": "branch_stream",
26
+ "static_max_len": 4096,
27
+ "max_new": 512,
28
+ "prompts_dir": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b",
29
+ "ar_cache": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache",
30
+ "reference": "/workspace/opt_runs/div2_v2_4b/code_lm/results/prompts/reference_numbers.json",
31
+ "contamination": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b/contamination.json",
32
+ "reps": 2000,
33
+ "batch_invariant": true,
34
+ "fp32_A_log": true,
35
+ "mask_lean": true,
36
+ "lora_fused_a": true,
37
+ "seed": 0,
38
+ "out": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_torch_cand_int4tg",
39
+ "threads": 1,
40
+ "fresh": false,
41
+ "policy": "fused",
42
+ "mtp": null,
43
+ "mtp_checkpoint": null,
44
+ "sglang_mtp": "/workspace/opt_runs/div2_v2_4b/code_lm/results/baselines/beta-off",
45
+ "drafter": null,
46
+ "mtp_vocab": null,
47
+ "dflash_vocab": null,
48
+ "draft_log": null,
49
+ "salvage": 0,
50
+ "calibration": null,
51
+ "mask_readout": "full",
52
+ "miss_policy": "recovery",
53
+ "dense_target": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c",
54
+ "late_commit": "on",
55
+ "lc_readout": "cand",
56
+ "late_merge": "on",
57
+ "late_merge_ops": "torch",
58
+ "late_weights": "int4",
59
+ "late_kernel": "tinygemm",
60
+ "dense_snapshot": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c"
61
+ },
62
+ "policy": "fused",
63
+ "checkpoint": "/workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/ckpt-0005764/adapters.pt",
64
+ "options": {
65
+ "attention": "triton",
66
+ "fused_ops": false,
67
+ "graphs": true,
68
+ "static_max_len": 4096,
69
+ "decode": {
70
+ "path": "static-graphed",
71
+ "attention": "triton",
72
+ "fused_ops": false,
73
+ "static_max_len": 4096
74
+ },
75
+ "mask_lean": true,
76
+ "lora_fused_a": true,
77
+ "mask_attn_kernel": "branch_stream",
78
+ "late_commit": "on",
79
+ "lc_readout": "cand"
80
+ },
81
+ "provenance": {
82
+ "git_head": null,
83
+ "row_git_heads": [
84
+ "None"
85
+ ]
86
+ },
87
+ "checkpoint_meta": {
88
+ "mask_from_layer": 31,
89
+ "max_block": 16,
90
+ "mask_attn_window": null,
91
+ "local_window": 0,
92
+ "mask_mode": "read",
93
+ "rank": 128,
94
+ "alpha": 90.50966799187809,
95
+ "targets": [
96
+ "q_proj",
97
+ "k_proj",
98
+ "v_proj",
99
+ "o_proj",
100
+ "gate_proj",
101
+ "up_proj",
102
+ "down_proj"
103
+ ],
104
+ "mask_block_causal": true,
105
+ "late_bonus_layers": 3,
106
+ "late_bonus_rank": 64
107
+ },
108
+ "checkpoint_extra": {
109
+ "step": 5764,
110
+ "config": {
111
+ "name": "k3q_qwen3_4b_LCa3_L31r128_lp16_causal_dsloss",
112
+ "description": "LCa3: SPL (ACc plus DSpine's backbone loss: 0.1 CE + 0.9 full-vocabulary L1 to the frozen target at row s+1+m, exp(-m/7) slot weights) plus late commit LC-a with k = 3 (docs/plans/late-commit-2026-11-14.md, sections 1-3 and 9): one MASK block per round. The last 3 MASK layers (33-35) run after verification: context <= the endpoint e, queries and the block's own keys at the actual positions e+1+m, and the bonus ids[e+1] injected right before layer 33 as in LB-3 (B(A e(b)), A [64, 2560], B [2560, 64] zero-initialised, B in the LoRA+ group). Layers 31-32 see only the confirmed prefix <= u = e - a (fork x_L0[u], positions u+1+m), a drawn per anchor from late_commit_accept_dist (its own stream; anchors and data order unchanged), u clamped at prompt_len - 1. The distribution is the per-row geometric mixture of SPL's 128-row walk (review F3): for each row of div2/v2_4b/results/SPL_L31r128_lp16_causal_dsloss_q_e0_128rows_perrow.jsonl a geometric capped at 15 whose mean equals that row's mean accept count (tokens/passes - 1), mixed with the rows' passes as weights (mean a 1.71766 = W128 2.71766 - 1; P(a=0) 0.412, P(a>=10) 0.022, P(a=15) 0.006). Labels ids[e+2+m], SPL's loss unchanged; with a = 0 the forward is LB-3's bit for bit. The eval adds walk_late_commit_full (W_LC: the 128 rows' rounds simulated, the next block from the lower layers <= u and the late layers <= e + bonus, a + 1 tokens per round, the full-vocabulary late readout) and walk_late_commit (cost variant: late slot m reads only the in-pass top-16 of slot m+a, the full vocabulary where m+a >= 15); the existing walk (block at the true endpoint: the perfect branch) stays beside them. Controls: SPL_L31r128_lp16_causal_dsloss_q and SPLLB3_L31r128_lp16_causal_dsloss_q at the same step (same seed, rows, anchors and eval rows). Verdict fixed before results (plan sections 4 and 9): the metric is walk_late_commit_full (the full-vocabulary late readout; plan section 9 replaced the top-16 readout, whose candidates were misaligned, before any result). With the better arm, W_LC_full / W128(SPL) >= 0.92 with the paired-bootstrap 95% lower bound >= 0.90 passes (the single-block structure is scaled up: 4B engine decode, Z1 cost model, then 27B); >= 1.00 is the strong pass; LC-a below 0.92 sends LC-b in, and if that also fails the branch/sibling structure stays. Needs late_commit.patch on top of native_training.patch. | SCREEN: quarter epoch (5764 steps), own cosine decay to 0",
113
+ "target": {
114
+ "kind": "hf_dense",
115
+ "dtype": "bfloat16",
116
+ "snapshot": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c"
117
+ },
118
+ "device": "cuda",
119
+ "train_data": "/workspace/opt_runs/div2_v2_4b/data/prepared/train",
120
+ "heldout_data": "/workspace/opt_runs/div2_v2_4b/data/prepared/heldout",
121
+ "mask_mode": "read",
122
+ "mask_from_layer": 31,
123
+ "mask_attn_window": null,
124
+ "local_window": 0,
125
+ "rank": 128,
126
+ "alpha": 90.50966799187809,
127
+ "targets": [
128
+ "q_proj",
129
+ "k_proj",
130
+ "v_proj",
131
+ "o_proj",
132
+ "gate_proj",
133
+ "up_proj",
134
+ "down_proj"
135
+ ],
136
+ "freeze_qkv_kv_rows": false,
137
+ "block": 15,
138
+ "max_block": 16,
139
+ "max_len": 3072,
140
+ "anchors_per_seq": 512,
141
+ "anchor_sampling": "uniform",
142
+ "hard_candidates": 4,
143
+ "seqs_per_step": 32,
144
+ "steps": 5764,
145
+ "arm_stop_after": null,
146
+ "lr": 0.0003,
147
+ "min_lr_ratio": 0.0,
148
+ "warmup_steps": 200,
149
+ "weight_decay": 0.0,
150
+ "betas": [
151
+ 0.9,
152
+ 0.999
153
+ ],
154
+ "adam_eps": 1e-08,
155
+ "grad_clip": 1.0,
156
+ "seed": 20260923,
157
+ "eval_every": 1441,
158
+ "eval_rows": 128,
159
+ "eval_anchors_per_seq": 32,
160
+ "eval_seed": 7,
161
+ "eval_walk_rows": 8,
162
+ "eval_walk_chunk": 128,
163
+ "ckpt_every": 1441,
164
+ "keep_last": 4,
165
+ "print_every": 20,
166
+ "context_scan": "segmented",
167
+ "mask_chunk": 512,
168
+ "checkpoint_steps": [
169
+ 5764
170
+ ],
171
+ "seq_assignment": "longest_first",
172
+ "dense_training_attention": "anchor_triton",
173
+ "lora_plus_ratio": 16.0,
174
+ "mask_block_causal": true,
175
+ "teacher_l1_weight": 0.9,
176
+ "slot_decay_gamma": 7.0,
177
+ "late_bonus_layers": 3,
178
+ "late_bonus_rank": 64,
179
+ "late_commit": "confirmed_prefix",
180
+ "late_commit_accept_dist": [
181
+ 0.411835,
182
+ 0.224416,
183
+ 0.129734,
184
+ 0.078494,
185
+ 0.049257,
186
+ 0.031871,
187
+ 0.021181,
188
+ 0.014418,
189
+ 0.01003,
190
+ 0.007117,
191
+ 0.005141,
192
+ 0.003775,
193
+ 0.002813,
194
+ 0.002123,
195
+ 0.001622,
196
+ 0.006173
197
+ ]
198
+ }
199
+ },
200
+ "ar_cache": {
201
+ "dir": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache",
202
+ "key": {
203
+ "target": "dense:models--Qwen--Qwen3-4B@1cfa9a7208912126459214e8b04321603b3df60c",
204
+ "dtype": "bfloat16",
205
+ "backend": "none",
206
+ "batch_invariant": true,
207
+ "fp32_A_log": false,
208
+ "max_new": 512,
209
+ "eos": [
210
+ 151645,
211
+ 151643
212
+ ],
213
+ "decode": {
214
+ "path": "static-graphed",
215
+ "attention": "triton",
216
+ "fused_ops": false,
217
+ "static_max_len": 4096
218
+ },
219
+ "harness": {
220
+ "source_sha1": "04e7703c37603bda2b76c602e94506442edd314a",
221
+ "source_files": [
222
+ "specdec/adapters/gated_lora.py",
223
+ "specdec/fused/branches.py",
224
+ "specdec/fused/hint.py",
225
+ "specdec/fused/late_bonus.py",
226
+ "specdec/fused/layout.py",
227
+ "specdec/fused/policy.py",
228
+ "specdec/fused/transfer.py",
229
+ "specdec/memory/gdn_state.py",
230
+ "specdec/paths/spec.py",
231
+ "specdec/run/decode.py",
232
+ "specdec/target/__init__.py",
233
+ "specdec/target/branch_memory.py",
234
+ "specdec/target/branch_triton.py",
235
+ "specdec/target/cache.py",
236
+ "specdec/target/kernels.py",
237
+ "specdec/target/lora_ops.py",
238
+ "specdec/target/qwen35.py",
239
+ "specdec/target/readout.py",
240
+ "specdec/target/static.py",
241
+ "specdec/target/triton_ops.py"
242
+ ],
243
+ "torch": "2.13.0+cu130",
244
+ "torch_cuda": "13.0",
245
+ "transformers": "5.12.1",
246
+ "sglang": "0.5.18",
247
+ "gpu": "NVIDIA H100 80GB HBM3"
248
+ },
249
+ "dense_sources_sha1": "03ba376b04f7686f84af14b96362f10f03e83351"
250
+ },
251
+ "tag": "5090a239b901",
252
+ "hits": 20,
253
+ "decoded": 0,
254
+ "used": true
255
+ },
256
+ "prompts": {
257
+ "gsm8k": [
258
+ "gsm8k:23",
259
+ "gsm8k:51",
260
+ "gsm8k:61",
261
+ "gsm8k:90"
262
+ ],
263
+ "math500": [
264
+ "math500:14",
265
+ "math500:34",
266
+ "math500:53",
267
+ "math500:79"
268
+ ],
269
+ "humaneval": [
270
+ "humaneval:4",
271
+ "humaneval:19",
272
+ "humaneval:58",
273
+ "humaneval:62"
274
+ ],
275
+ "mbpp": [
276
+ "mbpp:3",
277
+ "mbpp:77",
278
+ "mbpp:87",
279
+ "mbpp:92"
280
+ ],
281
+ "mt-bench": [
282
+ "mt-bench:21",
283
+ "mt-bench:33",
284
+ "mt-bench:46",
285
+ "mt-bench:59"
286
+ ]
287
+ },
288
+ "contamination": {
289
+ "file": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b/contamination.json",
290
+ "excluded_from_clean": {
291
+ "gsm8k": [],
292
+ "math500": [],
293
+ "humaneval": [],
294
+ "mbpp": [],
295
+ "mt-bench": []
296
+ }
297
+ },
298
+ "rows": {
299
+ "file": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_torch_cand_int4tg/q1_rows.jsonl",
300
+ "run_key": {
301
+ "ar": {
302
+ "target": "dense:models--Qwen--Qwen3-4B@1cfa9a7208912126459214e8b04321603b3df60c",
303
+ "dtype": "bfloat16",
304
+ "backend": "none",
305
+ "batch_invariant": true,
306
+ "fp32_A_log": false,
307
+ "max_new": 512,
308
+ "eos": [
309
+ 151645,
310
+ 151643
311
+ ],
312
+ "decode": {
313
+ "path": "static-graphed",
314
+ "attention": "triton",
315
+ "fused_ops": false,
316
+ "static_max_len": 4096
317
+ },
318
+ "harness": {
319
+ "source_sha1": "04e7703c37603bda2b76c602e94506442edd314a",
320
+ "source_files": [
321
+ "specdec/adapters/gated_lora.py",
322
+ "specdec/fused/branches.py",
323
+ "specdec/fused/hint.py",
324
+ "specdec/fused/late_bonus.py",
325
+ "specdec/fused/layout.py",
326
+ "specdec/fused/policy.py",
327
+ "specdec/fused/transfer.py",
328
+ "specdec/memory/gdn_state.py",
329
+ "specdec/paths/spec.py",
330
+ "specdec/run/decode.py",
331
+ "specdec/target/__init__.py",
332
+ "specdec/target/branch_memory.py",
333
+ "specdec/target/branch_triton.py",
334
+ "specdec/target/cache.py",
335
+ "specdec/target/kernels.py",
336
+ "specdec/target/lora_ops.py",
337
+ "specdec/target/qwen35.py",
338
+ "specdec/target/readout.py",
339
+ "specdec/target/static.py",
340
+ "specdec/target/triton_ops.py"
341
+ ],
342
+ "torch": "2.13.0+cu130",
343
+ "torch_cuda": "13.0",
344
+ "transformers": "5.12.1",
345
+ "sglang": "0.5.18",
346
+ "gpu": "NVIDIA H100 80GB HBM3"
347
+ },
348
+ "dense_sources_sha1": "03ba376b04f7686f84af14b96362f10f03e83351"
349
+ },
350
+ "checkpoint_sha1": "cc4c99cbfaaa0a869f5b219774db513c8fe917d2",
351
+ "mask_mode": "read",
352
+ "ar_checked": true,
353
+ "mask_options": {
354
+ "mask_lean": true,
355
+ "lora_fused_a": true
356
+ },
357
+ "mask_attn_kernel": "branch_stream",
358
+ "late_commit": {
359
+ "lc_readout": "cand",
360
+ "late_commit_sha1": "7766806b2bd5671e7e5038b76b03106b641aa6e4"
361
+ },
362
+ "late_weights": {
363
+ "format": "int4",
364
+ "late_quant_sha1": "edc27d2eece2380a26b5d19452fe9fa5e1130181",
365
+ "kernel": "tinygemm",
366
+ "group": 128
367
+ },
368
+ "late_merge": {
369
+ "variant": "merged",
370
+ "late_merge_sha1": "be3bb036283ed581d8cb0e33b64be1e926880dd0"
371
+ }
372
+ },
373
+ "resumed": 0
374
+ },
375
+ "definitions": {
376
+ "tau": "tokens emitted by non-recovery full passes / their count, misses included (specdec/fused/branches.py); no reference counterpart",
377
+ "miss_rate": "misses / non-recovery full passes",
378
+ "tokens_per_pass": "all output tokens / all passes (start and recovery included)",
379
+ "rows_per_pass": "mean over all passes of V + b x 15",
380
+ "tau_ci95": "95 % prompt-level bootstrap (percentile) of pooled tau",
381
+ "paired": "tau(b) - tau(b_max), both arms resampled with the same prompts",
382
+ "e2e": "wall clock against the harness's own AR on the same prompts (e2e_speedup): speedup_decode after the prompt, speedup_e2e with the shared prefill; engine: this harness, not SGLang",
383
+ "pooled_clean": "the same summary without the prompts found in the training split (contamination.excluded_from_clean); present only when some were excluded",
384
+ "tau_pooled": "reference definition; ours = tau_pooled_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)",
385
+ "tau_raw_row": "reference definition; ours = tau_raw_row_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)",
386
+ "tau_prefill_corrected_row": "reference definition; ours = tau_prefill_corrected_row_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)"
387
+ },
388
+ "results": {
389
+ "b1": {
390
+ "pooled": {
391
+ "prompts": 20,
392
+ "tokens": 6702,
393
+ "passes": 2340,
394
+ "fused_passes": 2320,
395
+ "recovery_passes": 0,
396
+ "misses": 0,
397
+ "salvage_passes": 0,
398
+ "salvage_misses": 0,
399
+ "tokens_per_pass": 2.864102564102564,
400
+ "tau": 2.8801724137931033,
401
+ "miss_rate": 0.0,
402
+ "rows_per_pass": 30.871794871794872,
403
+ "rows_per_fused_pass": 31.0,
404
+ "tau_pooled_ref": 2.888793103448276,
405
+ "tau_raw_row_ref": 3.0297319604326525,
406
+ "tau_prefill_corrected_row_ref": 3.017755752034457,
407
+ "accepted_hist": {
408
+ "0": 762,
409
+ "1": 512,
410
+ "2": 371,
411
+ "3": 261,
412
+ "4": 146,
413
+ "5": 111,
414
+ "6": 55,
415
+ "7": 38,
416
+ "8": 30,
417
+ "9": 15,
418
+ "10": 6,
419
+ "11": 5,
420
+ "12": 3,
421
+ "13": 1,
422
+ "14": 3,
423
+ "15": 1
424
+ },
425
+ "committed_hist": {
426
+ "0": 758,
427
+ "1": 507,
428
+ "2": 367,
429
+ "3": 257,
430
+ "4": 146,
431
+ "5": 110,
432
+ "6": 54,
433
+ "7": 38,
434
+ "8": 30,
435
+ "9": 15,
436
+ "10": 6,
437
+ "11": 5,
438
+ "12": 3,
439
+ "13": 1,
440
+ "14": 2,
441
+ "15": 1
442
+ },
443
+ "overshoot": 6,
444
+ "decode_ms": 52344.67111492157,
445
+ "decode_ms_all": 53316.18374919891,
446
+ "capture_rounds": 2,
447
+ "capture_ms": 971.5126342773438,
448
+ "pass_ms_by_kind": {
449
+ "start": {
450
+ "n": 19,
451
+ "mean_ms": 16.90853214263916,
452
+ "median_ms": 17.692960739135742,
453
+ "p10_ms": 13.750227165222167,
454
+ "p90_ms": 18.090656280517578
455
+ },
456
+ "fused": {
457
+ "n": 2319,
458
+ "mean_ms": 22.43355282630937,
459
+ "median_ms": 23.347583770751953,
460
+ "p10_ms": 19.05339469909668,
461
+ "p90_ms": 23.852505111694338
462
+ }
463
+ },
464
+ "prefill_ms": 3422.352767944336,
465
+ "accounting_ok": true,
466
+ "exact": 20,
467
+ "tau_ci95": [
468
+ 2.655621301775148,
469
+ 3.1486169714017813
470
+ ],
471
+ "tau_pooled_ref_ci95": [
472
+ 2.663510848126233,
473
+ 3.1598062953995156
474
+ ]
475
+ },
476
+ "benches": {
477
+ "gsm8k": {
478
+ "stats": {
479
+ "prompts": 4,
480
+ "tokens": 987,
481
+ "passes": 308,
482
+ "fused_passes": 304,
483
+ "recovery_passes": 0,
484
+ "misses": 0,
485
+ "salvage_passes": 0,
486
+ "salvage_misses": 0,
487
+ "tokens_per_pass": 3.2045454545454546,
488
+ "tau": 3.2335526315789473,
489
+ "miss_rate": 0.0,
490
+ "rows_per_pass": 30.805194805194805,
491
+ "rows_per_fused_pass": 31.0,
492
+ "tau_pooled_ref": 3.2467105263157894,
493
+ "tau_raw_row_ref": 3.2688477511617076,
494
+ "tau_prefill_corrected_row_ref": 3.255419399031898,
495
+ "accepted_hist": {
496
+ "0": 98,
497
+ "1": 60,
498
+ "2": 47,
499
+ "3": 31,
500
+ "4": 23,
501
+ "5": 12,
502
+ "6": 7,
503
+ "7": 7,
504
+ "8": 7,
505
+ "9": 3,
506
+ "10": 3,
507
+ "11": 2,
508
+ "12": 1,
509
+ "13": 1,
510
+ "14": 1,
511
+ "15": 1
512
+ },
513
+ "committed_hist": {
514
+ "0": 98,
515
+ "1": 59,
516
+ "2": 45,
517
+ "3": 31,
518
+ "4": 23,
519
+ "5": 12,
520
+ "6": 6,
521
+ "7": 7,
522
+ "8": 7,
523
+ "9": 3,
524
+ "10": 3,
525
+ "11": 2,
526
+ "12": 1,
527
+ "13": 1,
528
+ "14": 1,
529
+ "15": 1
530
+ },
531
+ "overshoot": 0,
532
+ "decode_ms": 6861.245664596558,
533
+ "decode_ms_all": 7832.758298873901,
534
+ "capture_rounds": 2,
535
+ "capture_ms": 971.5126342773438,
536
+ "pass_ms_by_kind": {
537
+ "start": {
538
+ "n": 3,
539
+ "mean_ms": 18.383658091227215,
540
+ "median_ms": 17.654943466186523,
541
+ "p10_ms": 17.627346420288085,
542
+ "p90_ms": 19.431455612182617
543
+ },
544
+ "fused": {
545
+ "n": 303,
546
+ "mean_ms": 22.462358713936883,
547
+ "median_ms": 23.27680015563965,
548
+ "p10_ms": 18.804262161254883,
549
+ "p90_ms": 25.185849380493163
550
+ }
551
+ },
552
+ "prefill_ms": 1350.349479675293,
553
+ "accounting_ok": true,
554
+ "exact": 4,
555
+ "tau_ci95": [
556
+ 2.993485342019544,
557
+ 3.5213114754098362
558
+ ],
559
+ "tau_pooled_ref_ci95": [
560
+ 3.006514657980456,
561
+ 3.5344262295081967
562
+ ]
563
+ },
564
+ "prompts": [
565
+ {
566
+ "pid": "gsm8k:23",
567
+ "tokens": 203,
568
+ "passes": 62,
569
+ "identical": true,
570
+ "first_divergence": null,
571
+ "method_tally": [
572
+ 202,
573
+ 61
574
+ ],
575
+ "reference_tally": [
576
+ 203,
577
+ 61
578
+ ],
579
+ "misses": 0,
580
+ "finish_reason": "stop",
581
+ "decode_ms": 1335.9629745483398,
582
+ "capture_ms": 971.5126342773438
583
+ },
584
+ {
585
+ "pid": "gsm8k:51",
586
+ "tokens": 267,
587
+ "passes": 73,
588
+ "identical": true,
589
+ "first_divergence": null,
590
+ "method_tally": [
591
+ 266,
592
+ 72
593
+ ],
594
+ "reference_tally": [
595
+ 267,
596
+ 72
597
+ ],
598
+ "misses": 0,
599
+ "finish_reason": "stop",
600
+ "decode_ms": 1621.4309120178223,
601
+ "capture_ms": 0
602
+ },
603
+ {
604
+ "pid": "gsm8k:61",
605
+ "tokens": 240,
606
+ "passes": 83,
607
+ "identical": true,
608
+ "first_divergence": null,
609
+ "method_tally": [
610
+ 239,
611
+ 82
612
+ ],
613
+ "reference_tally": [
614
+ 240,
615
+ 82
616
+ ],
617
+ "misses": 0,
618
+ "finish_reason": "stop",
619
+ "decode_ms": 1889.212031364441,
620
+ "capture_ms": 0
621
+ },
622
+ {
623
+ "pid": "gsm8k:90",
624
+ "tokens": 277,
625
+ "passes": 90,
626
+ "identical": true,
627
+ "first_divergence": null,
628
+ "method_tally": [
629
+ 276,
630
+ 89
631
+ ],
632
+ "reference_tally": [
633
+ 277,
634
+ 89
635
+ ],
636
+ "misses": 0,
637
+ "finish_reason": "stop",
638
+ "decode_ms": 2014.6397466659546,
639
+ "capture_ms": 0
640
+ }
641
+ ],
642
+ "e2e": {
643
+ "prompts": 4,
644
+ "fused_ms_per_token": 6.951616681455479,
645
+ "ar_ms_per_token": 16.872936100945285,
646
+ "speedup_decode": 2.427195985353518,
647
+ "speedup_e2e": 2.1946660223276506
648
+ }
649
+ },
650
+ "math500": {
651
+ "stats": {
652
+ "prompts": 4,
653
+ "tokens": 1575,
654
+ "passes": 418,
655
+ "fused_passes": 414,
656
+ "recovery_passes": 0,
657
+ "misses": 0,
658
+ "salvage_passes": 0,
659
+ "salvage_misses": 0,
660
+ "tokens_per_pass": 3.7679425837320575,
661
+ "tau": 3.7946859903381642,
662
+ "miss_rate": 0.0,
663
+ "rows_per_pass": 30.85645933014354,
664
+ "rows_per_fused_pass": 31.0,
665
+ "tau_pooled_ref": 3.8043478260869565,
666
+ "tau_raw_row_ref": 3.923894514856766,
667
+ "tau_prefill_corrected_row_ref": 3.9118324280320573,
668
+ "accepted_hist": {
669
+ "0": 90,
670
+ "1": 82,
671
+ "2": 57,
672
+ "3": 53,
673
+ "4": 37,
674
+ "5": 26,
675
+ "6": 19,
676
+ "7": 19,
677
+ "8": 17,
678
+ "9": 7,
679
+ "10": 1,
680
+ "11": 3,
681
+ "12": 2,
682
+ "14": 1
683
+ },
684
+ "committed_hist": {
685
+ "0": 89,
686
+ "1": 81,
687
+ "2": 57,
688
+ "3": 52,
689
+ "4": 37,
690
+ "5": 26,
691
+ "6": 19,
692
+ "7": 19,
693
+ "8": 17,
694
+ "9": 7,
695
+ "10": 1,
696
+ "11": 3,
697
+ "12": 2
698
+ },
699
+ "overshoot": 3,
700
+ "decode_ms": 9319.681138038635,
701
+ "decode_ms_all": 9319.681138038635,
702
+ "capture_rounds": 0,
703
+ "capture_ms": 0,
704
+ "pass_ms_by_kind": {
705
+ "start": {
706
+ "n": 4,
707
+ "mean_ms": 16.523351907730103,
708
+ "median_ms": 17.246416091918945,
709
+ "p10_ms": 14.614627170562745,
710
+ "p90_ms": 17.853625297546387
711
+ },
712
+ "fused": {
713
+ "n": 414,
714
+ "mean_ms": 22.3516611845597,
715
+ "median_ms": 23.372400283813477,
716
+ "p10_ms": 18.913346672058104,
717
+ "p90_ms": 23.788560485839845
718
+ }
719
+ },
720
+ "prefill_ms": 495.2342300415039,
721
+ "accounting_ok": true,
722
+ "exact": 4,
723
+ "tau_ci95": [
724
+ 3.4499089253187614,
725
+ 4.1505791505791505
726
+ ],
727
+ "tau_pooled_ref_ci95": [
728
+ 3.4571948998178508,
729
+ 4.166023166023166
730
+ ]
731
+ },
732
+ "prompts": [
733
+ {
734
+ "pid": "math500:14",
735
+ "tokens": 512,
736
+ "passes": 154,
737
+ "identical": true,
738
+ "first_divergence": null,
739
+ "method_tally": [
740
+ 511,
741
+ 153
742
+ ],
743
+ "reference_tally": [
744
+ 512,
745
+ 153
746
+ ],
747
+ "misses": 0,
748
+ "finish_reason": "length",
749
+ "decode_ms": 3449.209053993225,
750
+ "capture_ms": 0
751
+ },
752
+ {
753
+ "pid": "math500:34",
754
+ "tokens": 512,
755
+ "passes": 128,
756
+ "identical": true,
757
+ "first_divergence": null,
758
+ "method_tally": [
759
+ 511,
760
+ 127
761
+ ],
762
+ "reference_tally": [
763
+ 512,
764
+ 127
765
+ ],
766
+ "misses": 0,
767
+ "finish_reason": "length",
768
+ "decode_ms": 2791.710760116577,
769
+ "capture_ms": 0
770
+ },
771
+ {
772
+ "pid": "math500:53",
773
+ "tokens": 362,
774
+ "passes": 91,
775
+ "identical": true,
776
+ "first_divergence": null,
777
+ "method_tally": [
778
+ 361,
779
+ 90
780
+ ],
781
+ "reference_tally": [
782
+ 362,
783
+ 90
784
+ ],
785
+ "misses": 0,
786
+ "finish_reason": "stop",
787
+ "decode_ms": 2062.390956878662,
788
+ "capture_ms": 0
789
+ },
790
+ {
791
+ "pid": "math500:79",
792
+ "tokens": 189,
793
+ "passes": 45,
794
+ "identical": true,
795
+ "first_divergence": null,
796
+ "method_tally": [
797
+ 188,
798
+ 44
799
+ ],
800
+ "reference_tally": [
801
+ 189,
802
+ 44
803
+ ],
804
+ "misses": 0,
805
+ "finish_reason": "stop",
806
+ "decode_ms": 1016.3703670501709,
807
+ "capture_ms": 0
808
+ }
809
+ ],
810
+ "e2e": {
811
+ "prompts": 4,
812
+ "fused_ms_per_token": 5.917257865421356,
813
+ "ar_ms_per_token": 16.996904987153552,
814
+ "speedup_decode": 2.87242932008731,
815
+ "speedup_e2e": 2.7950043464180805
816
+ }
817
+ },
818
+ "humaneval": {
819
+ "stats": {
820
+ "prompts": 4,
821
+ "tokens": 1615,
822
+ "passes": 613,
823
+ "fused_passes": 609,
824
+ "recovery_passes": 0,
825
+ "misses": 0,
826
+ "salvage_passes": 0,
827
+ "salvage_misses": 0,
828
+ "tokens_per_pass": 2.634584013050571,
829
+ "tau": 2.645320197044335,
830
+ "miss_rate": 0.0,
831
+ "rows_per_pass": 30.902120717781404,
832
+ "rows_per_fused_pass": 31.0,
833
+ "tau_pooled_ref": 2.651888341543514,
834
+ "tau_raw_row_ref": 2.6680471426250696,
835
+ "tau_prefill_corrected_row_ref": 2.6608926736285237,
836
+ "accepted_hist": {
837
+ "0": 199,
838
+ "1": 147,
839
+ "2": 103,
840
+ "3": 76,
841
+ "4": 34,
842
+ "5": 30,
843
+ "6": 12,
844
+ "7": 3,
845
+ "8": 2,
846
+ "9": 2,
847
+ "10": 1
848
+ },
849
+ "committed_hist": {
850
+ "0": 199,
851
+ "1": 145,
852
+ "2": 103,
853
+ "3": 74,
854
+ "4": 34,
855
+ "5": 30,
856
+ "6": 12,
857
+ "7": 3,
858
+ "8": 2,
859
+ "9": 2,
860
+ "10": 1
861
+ },
862
+ "overshoot": 2,
863
+ "decode_ms": 13829.814667701721,
864
+ "decode_ms_all": 13829.814667701721,
865
+ "capture_rounds": 0,
866
+ "capture_ms": 0,
867
+ "pass_ms_by_kind": {
868
+ "start": {
869
+ "n": 4,
870
+ "mean_ms": 16.362152099609375,
871
+ "median_ms": 16.53427219390869,
872
+ "p10_ms": 14.763119888305663,
873
+ "p90_ms": 17.823488235473633
874
+ },
875
+ "fused": {
876
+ "n": 609,
877
+ "mean_ms": 22.601586304274687,
878
+ "median_ms": 23.444480895996094,
879
+ "p10_ms": 19.177727890014648,
880
+ "p90_ms": 23.834905242919923
881
+ }
882
+ },
883
+ "prefill_ms": 506.28699493408203,
884
+ "accounting_ok": true,
885
+ "exact": 4,
886
+ "tau_ci95": [
887
+ 2.4624829467939975,
888
+ 2.8461538461538463
889
+ ],
890
+ "tau_pooled_ref_ci95": [
891
+ 2.4679399727148703,
892
+ 2.852307692307692
893
+ ]
894
+ },
895
+ "prompts": [
896
+ {
897
+ "pid": "humaneval:4",
898
+ "tokens": 318,
899
+ "passes": 123,
900
+ "identical": true,
901
+ "first_divergence": null,
902
+ "method_tally": [
903
+ 317,
904
+ 122
905
+ ],
906
+ "reference_tally": [
907
+ 318,
908
+ 122
909
+ ],
910
+ "misses": 0,
911
+ "finish_reason": "stop",
912
+ "decode_ms": 2739.100580215454,
913
+ "capture_ms": 0
914
+ },
915
+ {
916
+ "pid": "humaneval:19",
917
+ "tokens": 512,
918
+ "passes": 212,
919
+ "identical": true,
920
+ "first_divergence": null,
921
+ "method_tally": [
922
+ 511,
923
+ 211
924
+ ],
925
+ "reference_tally": [
926
+ 512,
927
+ 211
928
+ ],
929
+ "misses": 0,
930
+ "finish_reason": "length",
931
+ "decode_ms": 4778.178263664246,
932
+ "capture_ms": 0
933
+ },
934
+ {
935
+ "pid": "humaneval:58",
936
+ "tokens": 273,
937
+ "passes": 101,
938
+ "identical": true,
939
+ "first_divergence": null,
940
+ "method_tally": [
941
+ 272,
942
+ 100
943
+ ],
944
+ "reference_tally": [
945
+ 273,
946
+ 100
947
+ ],
948
+ "misses": 0,
949
+ "finish_reason": "stop",
950
+ "decode_ms": 2314.557309150696,
951
+ "capture_ms": 0
952
+ },
953
+ {
954
+ "pid": "humaneval:62",
955
+ "tokens": 512,
956
+ "passes": 177,
957
+ "identical": true,
958
+ "first_divergence": null,
959
+ "method_tally": [
960
+ 511,
961
+ 176
962
+ ],
963
+ "reference_tally": [
964
+ 512,
965
+ 176
966
+ ],
967
+ "misses": 0,
968
+ "finish_reason": "length",
969
+ "decode_ms": 3997.9785146713257,
970
+ "capture_ms": 0
971
+ }
972
+ ],
973
+ "e2e": {
974
+ "prompts": 4,
975
+ "fused_ms_per_token": 8.563352735419022,
976
+ "ar_ms_per_token": 17.069150008210457,
977
+ "speedup_decode": 1.993278863500562,
978
+ "speedup_e2e": 1.959658979859881
979
+ }
980
+ },
981
+ "mbpp": {
982
+ "stats": {
983
+ "prompts": 4,
984
+ "tokens": 1333,
985
+ "passes": 561,
986
+ "fused_passes": 557,
987
+ "recovery_passes": 0,
988
+ "misses": 0,
989
+ "salvage_passes": 0,
990
+ "salvage_misses": 0,
991
+ "tokens_per_pass": 2.376114081996435,
992
+ "tau": 2.385996409335727,
993
+ "miss_rate": 0.0,
994
+ "rows_per_pass": 30.893048128342247,
995
+ "rows_per_fused_pass": 31.0,
996
+ "tau_pooled_ref": 2.393177737881508,
997
+ "tau_raw_row_ref": 2.4702767220702,
998
+ "tau_prefill_corrected_row_ref": 2.461710837471707,
999
+ "accepted_hist": {
1000
+ "0": 222,
1001
+ "1": 137,
1002
+ "2": 84,
1003
+ "3": 55,
1004
+ "4": 23,
1005
+ "5": 19,
1006
+ "6": 8,
1007
+ "7": 5,
1008
+ "8": 3,
1009
+ "9": 1
1010
+ },
1011
+ "committed_hist": {
1012
+ "0": 220,
1013
+ "1": 137,
1014
+ "2": 83,
1015
+ "3": 54,
1016
+ "4": 23,
1017
+ "5": 19,
1018
+ "6": 8,
1019
+ "7": 5,
1020
+ "8": 3,
1021
+ "9": 1
1022
+ },
1023
+ "overshoot": 1,
1024
+ "decode_ms": 12362.202857017517,
1025
+ "decode_ms_all": 12362.202857017517,
1026
+ "capture_rounds": 0,
1027
+ "capture_ms": 0,
1028
+ "pass_ms_by_kind": {
1029
+ "start": {
1030
+ "n": 4,
1031
+ "mean_ms": 15.765079975128174,
1032
+ "median_ms": 15.722464084625244,
1033
+ "p10_ms": 13.693571090698242,
1034
+ "p90_ms": 17.87068157196045
1035
+ },
1036
+ "fused": {
1037
+ "n": 557,
1038
+ "mean_ms": 22.08104584760683,
1039
+ "median_ms": 23.281248092651367,
1040
+ "p10_ms": 18.42612419128418,
1041
+ "p90_ms": 23.909804153442384
1042
+ }
1043
+ },
1044
+ "prefill_ms": 551.9838714599609,
1045
+ "accounting_ok": true,
1046
+ "exact": 4,
1047
+ "tau_ci95": [
1048
+ 2.2472826086956523,
1049
+ 2.6560846560846563
1050
+ ],
1051
+ "tau_pooled_ref_ci95": [
1052
+ 2.2527173913043477,
1053
+ 2.6666666666666665
1054
+ ]
1055
+ },
1056
+ "prompts": [
1057
+ {
1058
+ "pid": "mbpp:3",
1059
+ "tokens": 317,
1060
+ "passes": 144,
1061
+ "identical": true,
1062
+ "first_divergence": null,
1063
+ "method_tally": [
1064
+ 316,
1065
+ 143
1066
+ ],
1067
+ "reference_tally": [
1068
+ 317,
1069
+ 143
1070
+ ],
1071
+ "misses": 0,
1072
+ "finish_reason": "stop",
1073
+ "decode_ms": 3270.4016313552856,
1074
+ "capture_ms": 0
1075
+ },
1076
+ {
1077
+ "pid": "mbpp:77",
1078
+ "tokens": 512,
1079
+ "passes": 226,
1080
+ "identical": true,
1081
+ "first_divergence": null,
1082
+ "method_tally": [
1083
+ 511,
1084
+ 225
1085
+ ],
1086
+ "reference_tally": [
1087
+ 512,
1088
+ 225
1089
+ ],
1090
+ "misses": 0,
1091
+ "finish_reason": "length",
1092
+ "decode_ms": 4758.26064491272,
1093
+ "capture_ms": 0
1094
+ },
1095
+ {
1096
+ "pid": "mbpp:87",
1097
+ "tokens": 311,
1098
+ "passes": 121,
1099
+ "identical": true,
1100
+ "first_divergence": null,
1101
+ "method_tally": [
1102
+ 310,
1103
+ 120
1104
+ ],
1105
+ "reference_tally": [
1106
+ 311,
1107
+ 120
1108
+ ],
1109
+ "misses": 0,
1110
+ "finish_reason": "stop",
1111
+ "decode_ms": 2739.6657638549805,
1112
+ "capture_ms": 0
1113
+ },
1114
+ {
1115
+ "pid": "mbpp:92",
1116
+ "tokens": 193,
1117
+ "passes": 70,
1118
+ "identical": true,
1119
+ "first_divergence": null,
1120
+ "method_tally": [
1121
+ 192,
1122
+ 69
1123
+ ],
1124
+ "reference_tally": [
1125
+ 193,
1126
+ 69
1127
+ ],
1128
+ "misses": 0,
1129
+ "finish_reason": "stop",
1130
+ "decode_ms": 1593.8748168945312,
1131
+ "capture_ms": 0
1132
+ }
1133
+ ],
1134
+ "e2e": {
1135
+ "prompts": 4,
1136
+ "fused_ms_per_token": 9.273970635421993,
1137
+ "ar_ms_per_token": 17.074735396085426,
1138
+ "speedup_decode": 1.8411461570590228,
1139
+ "speedup_e2e": 1.8079328675830206
1140
+ }
1141
+ },
1142
+ "mt-bench": {
1143
+ "stats": {
1144
+ "prompts": 4,
1145
+ "tokens": 1192,
1146
+ "passes": 440,
1147
+ "fused_passes": 436,
1148
+ "recovery_passes": 0,
1149
+ "misses": 0,
1150
+ "salvage_passes": 0,
1151
+ "salvage_misses": 0,
1152
+ "tokens_per_pass": 2.709090909090909,
1153
+ "tau": 2.7247706422018347,
1154
+ "miss_rate": 0.0,
1155
+ "rows_per_pass": 30.863636363636363,
1156
+ "rows_per_fused_pass": 31.0,
1157
+ "tau_pooled_ref": 2.7339449541284404,
1158
+ "tau_raw_row_ref": 2.8175936714495182,
1159
+ "tau_prefill_corrected_row_ref": 2.7989234220080994,
1160
+ "accepted_hist": {
1161
+ "0": 153,
1162
+ "1": 86,
1163
+ "2": 80,
1164
+ "3": 46,
1165
+ "4": 29,
1166
+ "5": 24,
1167
+ "6": 9,
1168
+ "7": 4,
1169
+ "8": 1,
1170
+ "9": 2,
1171
+ "10": 1,
1172
+ "14": 1
1173
+ },
1174
+ "committed_hist": {
1175
+ "0": 152,
1176
+ "1": 85,
1177
+ "2": 79,
1178
+ "3": 46,
1179
+ "4": 29,
1180
+ "5": 23,
1181
+ "6": 9,
1182
+ "7": 4,
1183
+ "8": 1,
1184
+ "9": 2,
1185
+ "10": 1,
1186
+ "14": 1
1187
+ },
1188
+ "overshoot": 0,
1189
+ "decode_ms": 9971.726787567139,
1190
+ "decode_ms_all": 9971.726787567139,
1191
+ "capture_rounds": 0,
1192
+ "capture_ms": 0,
1193
+ "pass_ms_by_kind": {
1194
+ "start": {
1195
+ "n": 4,
1196
+ "mean_ms": 17.87720012664795,
1197
+ "median_ms": 17.89081573486328,
1198
+ "p10_ms": 17.729085159301757,
1199
+ "p90_ms": 18.014422607421874
1200
+ },
1201
+ "fused": {
1202
+ "n": 436,
1203
+ "mean_ms": 22.706921988670977,
1204
+ "median_ms": 23.29379177093506,
1205
+ "p10_ms": 19.623552322387695,
1206
+ "p90_ms": 23.858479499816895
1207
+ }
1208
+ },
1209
+ "prefill_ms": 518.4981918334961,
1210
+ "accounting_ok": true,
1211
+ "exact": 4,
1212
+ "tau_ci95": [
1213
+ 2.3703703703703702,
1214
+ 2.989247311827957
1215
+ ],
1216
+ "tau_pooled_ref_ci95": [
1217
+ 2.380952380952381,
1218
+ 3.0
1219
+ ]
1220
+ },
1221
+ "prompts": [
1222
+ {
1223
+ "pid": "mt-bench:21",
1224
+ "tokens": 133,
1225
+ "passes": 65,
1226
+ "identical": true,
1227
+ "first_divergence": null,
1228
+ "method_tally": [
1229
+ 132,
1230
+ 64
1231
+ ],
1232
+ "reference_tally": [
1233
+ 133,
1234
+ 64
1235
+ ],
1236
+ "misses": 0,
1237
+ "finish_reason": "stop",
1238
+ "decode_ms": 1496.2943725585938,
1239
+ "capture_ms": 0
1240
+ },
1241
+ {
1242
+ "pid": "mt-bench:33",
1243
+ "tokens": 483,
1244
+ "passes": 166,
1245
+ "identical": true,
1246
+ "first_divergence": null,
1247
+ "method_tally": [
1248
+ 482,
1249
+ 165
1250
+ ],
1251
+ "reference_tally": [
1252
+ 483,
1253
+ 165
1254
+ ],
1255
+ "misses": 0,
1256
+ "finish_reason": "stop",
1257
+ "decode_ms": 3794.296356201172,
1258
+ "capture_ms": 0
1259
+ },
1260
+ {
1261
+ "pid": "mt-bench:46",
1262
+ "tokens": 501,
1263
+ "passes": 187,
1264
+ "identical": true,
1265
+ "first_divergence": null,
1266
+ "method_tally": [
1267
+ 500,
1268
+ 186
1269
+ ],
1270
+ "reference_tally": [
1271
+ 501,
1272
+ 186
1273
+ ],
1274
+ "misses": 0,
1275
+ "finish_reason": "stop",
1276
+ "decode_ms": 4183.002998352051,
1277
+ "capture_ms": 0
1278
+ },
1279
+ {
1280
+ "pid": "mt-bench:59",
1281
+ "tokens": 75,
1282
+ "passes": 22,
1283
+ "identical": true,
1284
+ "first_divergence": null,
1285
+ "method_tally": [
1286
+ 74,
1287
+ 21
1288
+ ],
1289
+ "reference_tally": [
1290
+ 75,
1291
+ 21
1292
+ ],
1293
+ "misses": 0,
1294
+ "finish_reason": "stop",
1295
+ "decode_ms": 498.13306045532227,
1296
+ "capture_ms": 0
1297
+ }
1298
+ ],
1299
+ "e2e": {
1300
+ "prompts": 4,
1301
+ "fused_ms_per_token": 8.365542607019412,
1302
+ "ar_ms_per_token": 17.135559822888982,
1303
+ "speedup_decode": 2.048350074567879,
1304
+ "speedup_e2e": 2.007012027306502
1305
+ }
1306
+ }
1307
+ },
1308
+ "e2e_pooled": {
1309
+ "prompts": 20,
1310
+ "fused_ms_per_token": 7.810306045198683,
1311
+ "ar_ms_per_token": 17.036198170922898,
1312
+ "speedup_decode": 2.1812459169120206,
1313
+ "speedup_e2e": 2.1144065251919493
1314
+ }
1315
+ }
1316
+ },
1317
+ "paired_vs_b_max": {},
1318
+ "mismatches_vs_ar": {},
1319
+ "reference_comparison": [
1320
+ {
1321
+ "note": "--dense-target: the reference numbers are SGLang runs of the 27B, not of this target"
1322
+ }
1323
+ ]
1324
+ }
qwen3_4b/eval/lm_LCa3_v11/lm_torch_cand_int4tg/q1_rows.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
qwen3_4b/eval/lm_LCa3_v11/lm_torch_cand_int4tg/q1_run.json ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ar": {
3
+ "target": "dense:models--Qwen--Qwen3-4B@1cfa9a7208912126459214e8b04321603b3df60c",
4
+ "dtype": "bfloat16",
5
+ "backend": "none",
6
+ "batch_invariant": true,
7
+ "fp32_A_log": false,
8
+ "max_new": 512,
9
+ "eos": [
10
+ 151645,
11
+ 151643
12
+ ],
13
+ "decode": {
14
+ "path": "static-graphed",
15
+ "attention": "triton",
16
+ "fused_ops": false,
17
+ "static_max_len": 4096
18
+ },
19
+ "harness": {
20
+ "source_sha1": "04e7703c37603bda2b76c602e94506442edd314a",
21
+ "source_files": [
22
+ "specdec/adapters/gated_lora.py",
23
+ "specdec/fused/branches.py",
24
+ "specdec/fused/hint.py",
25
+ "specdec/fused/late_bonus.py",
26
+ "specdec/fused/layout.py",
27
+ "specdec/fused/policy.py",
28
+ "specdec/fused/transfer.py",
29
+ "specdec/memory/gdn_state.py",
30
+ "specdec/paths/spec.py",
31
+ "specdec/run/decode.py",
32
+ "specdec/target/__init__.py",
33
+ "specdec/target/branch_memory.py",
34
+ "specdec/target/branch_triton.py",
35
+ "specdec/target/cache.py",
36
+ "specdec/target/kernels.py",
37
+ "specdec/target/lora_ops.py",
38
+ "specdec/target/qwen35.py",
39
+ "specdec/target/readout.py",
40
+ "specdec/target/static.py",
41
+ "specdec/target/triton_ops.py"
42
+ ],
43
+ "torch": "2.13.0+cu130",
44
+ "torch_cuda": "13.0",
45
+ "transformers": "5.12.1",
46
+ "sglang": "0.5.18",
47
+ "gpu": "NVIDIA H100 80GB HBM3"
48
+ },
49
+ "dense_sources_sha1": "03ba376b04f7686f84af14b96362f10f03e83351"
50
+ },
51
+ "checkpoint_sha1": "cc4c99cbfaaa0a869f5b219774db513c8fe917d2",
52
+ "mask_mode": "read",
53
+ "ar_checked": true,
54
+ "mask_options": {
55
+ "mask_lean": true,
56
+ "lora_fused_a": true
57
+ },
58
+ "mask_attn_kernel": "branch_stream",
59
+ "late_commit": {
60
+ "lc_readout": "cand",
61
+ "late_commit_sha1": "7766806b2bd5671e7e5038b76b03106b641aa6e4"
62
+ },
63
+ "late_weights": {
64
+ "format": "int4",
65
+ "late_quant_sha1": "edc27d2eece2380a26b5d19452fe9fa5e1130181",
66
+ "kernel": "tinygemm",
67
+ "group": 128
68
+ },
69
+ "late_merge": {
70
+ "variant": "merged",
71
+ "late_merge_sha1": "be3bb036283ed581d8cb0e33b64be1e926880dd0"
72
+ }
73
+ }
qwen3_4b/eval/lm_LCa3_v11/lm_triton_cand.log CHANGED
@@ -2,3 +2,7 @@ checkpoint /workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/
2
 
3
  Warning: please use at least NVCC 12.9 for the best DeepGEMM performance
4
  AR token ids: 20 from cache, 0 decoded (/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache)
 
 
 
 
 
2
 
3
  Warning: please use at least NVCC 12.9 for the best DeepGEMM performance
4
  AR token ids: 20 from cache, 0 decoded (/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache)
5
+ b=1: tau 2.896 [2.671, 3.164], miss rate 0.000, tokens/pass 2.880, rows/pass 30.9, exact 20/20; 60s
6
+ round ms by kind (median [p10, p90], capture rounds left out): start 17.5 [15.1, 19.7] n=19, fused 21.4 [17.2, 25.4] n=2306
7
+ wall clock vs harness AR on 20 prompts: 7.49 vs 17.04 ms/token, speedup 2.27x decode, 2.20x with prefill
8
+ no decodes differ from AR; wrote /workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_triton_cand/q1_eval.json
qwen3_4b/eval/lm_LCa3_v11/lm_triton_cand/q1_eval.json ADDED
@@ -0,0 +1,1313 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "tiny": false,
4
+ "checkpoint": [
5
+ "/workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/ckpt-0005764"
6
+ ],
7
+ "b": [
8
+ 1
9
+ ],
10
+ "benches": [
11
+ "gsm8k",
12
+ "math500",
13
+ "humaneval",
14
+ "mbpp",
15
+ "mt-bench"
16
+ ],
17
+ "limit": null,
18
+ "per_bench": 4,
19
+ "time_ar": true,
20
+ "skip_ar": false,
21
+ "graphs": true,
22
+ "attention": "triton",
23
+ "fused_ops": false,
24
+ "branch_memory": "copy",
25
+ "mask_attn_kernel": "branch_stream",
26
+ "static_max_len": 4096,
27
+ "max_new": 512,
28
+ "prompts_dir": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b",
29
+ "ar_cache": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache",
30
+ "reference": "/workspace/opt_runs/div2_v2_4b/code_lm/results/prompts/reference_numbers.json",
31
+ "contamination": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b/contamination.json",
32
+ "reps": 2000,
33
+ "batch_invariant": true,
34
+ "fp32_A_log": true,
35
+ "mask_lean": true,
36
+ "lora_fused_a": true,
37
+ "seed": 0,
38
+ "out": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_triton_cand",
39
+ "threads": 1,
40
+ "fresh": false,
41
+ "policy": "fused",
42
+ "mtp": null,
43
+ "mtp_checkpoint": null,
44
+ "sglang_mtp": "/workspace/opt_runs/div2_v2_4b/code_lm/results/baselines/beta-off",
45
+ "drafter": null,
46
+ "mtp_vocab": null,
47
+ "dflash_vocab": null,
48
+ "draft_log": null,
49
+ "salvage": 0,
50
+ "calibration": null,
51
+ "mask_readout": "full",
52
+ "miss_policy": "recovery",
53
+ "dense_target": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c",
54
+ "late_commit": "on",
55
+ "lc_readout": "cand",
56
+ "late_merge": "on",
57
+ "late_merge_ops": "triton",
58
+ "dense_snapshot": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c"
59
+ },
60
+ "policy": "fused",
61
+ "checkpoint": "/workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/ckpt-0005764/adapters.pt",
62
+ "options": {
63
+ "attention": "triton",
64
+ "fused_ops": false,
65
+ "graphs": true,
66
+ "static_max_len": 4096,
67
+ "decode": {
68
+ "path": "static-graphed",
69
+ "attention": "triton",
70
+ "fused_ops": false,
71
+ "static_max_len": 4096
72
+ },
73
+ "mask_lean": true,
74
+ "lora_fused_a": true,
75
+ "mask_attn_kernel": "branch_stream",
76
+ "late_commit": "on",
77
+ "lc_readout": "cand"
78
+ },
79
+ "provenance": {
80
+ "git_head": null,
81
+ "row_git_heads": [
82
+ "None"
83
+ ]
84
+ },
85
+ "checkpoint_meta": {
86
+ "mask_from_layer": 31,
87
+ "max_block": 16,
88
+ "mask_attn_window": null,
89
+ "local_window": 0,
90
+ "mask_mode": "read",
91
+ "rank": 128,
92
+ "alpha": 90.50966799187809,
93
+ "targets": [
94
+ "q_proj",
95
+ "k_proj",
96
+ "v_proj",
97
+ "o_proj",
98
+ "gate_proj",
99
+ "up_proj",
100
+ "down_proj"
101
+ ],
102
+ "mask_block_causal": true,
103
+ "late_bonus_layers": 3,
104
+ "late_bonus_rank": 64
105
+ },
106
+ "checkpoint_extra": {
107
+ "step": 5764,
108
+ "config": {
109
+ "name": "k3q_qwen3_4b_LCa3_L31r128_lp16_causal_dsloss",
110
+ "description": "LCa3: SPL (ACc plus DSpine's backbone loss: 0.1 CE + 0.9 full-vocabulary L1 to the frozen target at row s+1+m, exp(-m/7) slot weights) plus late commit LC-a with k = 3 (docs/plans/late-commit-2026-11-14.md, sections 1-3 and 9): one MASK block per round. The last 3 MASK layers (33-35) run after verification: context <= the endpoint e, queries and the block's own keys at the actual positions e+1+m, and the bonus ids[e+1] injected right before layer 33 as in LB-3 (B(A e(b)), A [64, 2560], B [2560, 64] zero-initialised, B in the LoRA+ group). Layers 31-32 see only the confirmed prefix <= u = e - a (fork x_L0[u], positions u+1+m), a drawn per anchor from late_commit_accept_dist (its own stream; anchors and data order unchanged), u clamped at prompt_len - 1. The distribution is the per-row geometric mixture of SPL's 128-row walk (review F3): for each row of div2/v2_4b/results/SPL_L31r128_lp16_causal_dsloss_q_e0_128rows_perrow.jsonl a geometric capped at 15 whose mean equals that row's mean accept count (tokens/passes - 1), mixed with the rows' passes as weights (mean a 1.71766 = W128 2.71766 - 1; P(a=0) 0.412, P(a>=10) 0.022, P(a=15) 0.006). Labels ids[e+2+m], SPL's loss unchanged; with a = 0 the forward is LB-3's bit for bit. The eval adds walk_late_commit_full (W_LC: the 128 rows' rounds simulated, the next block from the lower layers <= u and the late layers <= e + bonus, a + 1 tokens per round, the full-vocabulary late readout) and walk_late_commit (cost variant: late slot m reads only the in-pass top-16 of slot m+a, the full vocabulary where m+a >= 15); the existing walk (block at the true endpoint: the perfect branch) stays beside them. Controls: SPL_L31r128_lp16_causal_dsloss_q and SPLLB3_L31r128_lp16_causal_dsloss_q at the same step (same seed, rows, anchors and eval rows). Verdict fixed before results (plan sections 4 and 9): the metric is walk_late_commit_full (the full-vocabulary late readout; plan section 9 replaced the top-16 readout, whose candidates were misaligned, before any result). With the better arm, W_LC_full / W128(SPL) >= 0.92 with the paired-bootstrap 95% lower bound >= 0.90 passes (the single-block structure is scaled up: 4B engine decode, Z1 cost model, then 27B); >= 1.00 is the strong pass; LC-a below 0.92 sends LC-b in, and if that also fails the branch/sibling structure stays. Needs late_commit.patch on top of native_training.patch. | SCREEN: quarter epoch (5764 steps), own cosine decay to 0",
111
+ "target": {
112
+ "kind": "hf_dense",
113
+ "dtype": "bfloat16",
114
+ "snapshot": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c"
115
+ },
116
+ "device": "cuda",
117
+ "train_data": "/workspace/opt_runs/div2_v2_4b/data/prepared/train",
118
+ "heldout_data": "/workspace/opt_runs/div2_v2_4b/data/prepared/heldout",
119
+ "mask_mode": "read",
120
+ "mask_from_layer": 31,
121
+ "mask_attn_window": null,
122
+ "local_window": 0,
123
+ "rank": 128,
124
+ "alpha": 90.50966799187809,
125
+ "targets": [
126
+ "q_proj",
127
+ "k_proj",
128
+ "v_proj",
129
+ "o_proj",
130
+ "gate_proj",
131
+ "up_proj",
132
+ "down_proj"
133
+ ],
134
+ "freeze_qkv_kv_rows": false,
135
+ "block": 15,
136
+ "max_block": 16,
137
+ "max_len": 3072,
138
+ "anchors_per_seq": 512,
139
+ "anchor_sampling": "uniform",
140
+ "hard_candidates": 4,
141
+ "seqs_per_step": 32,
142
+ "steps": 5764,
143
+ "arm_stop_after": null,
144
+ "lr": 0.0003,
145
+ "min_lr_ratio": 0.0,
146
+ "warmup_steps": 200,
147
+ "weight_decay": 0.0,
148
+ "betas": [
149
+ 0.9,
150
+ 0.999
151
+ ],
152
+ "adam_eps": 1e-08,
153
+ "grad_clip": 1.0,
154
+ "seed": 20260923,
155
+ "eval_every": 1441,
156
+ "eval_rows": 128,
157
+ "eval_anchors_per_seq": 32,
158
+ "eval_seed": 7,
159
+ "eval_walk_rows": 8,
160
+ "eval_walk_chunk": 128,
161
+ "ckpt_every": 1441,
162
+ "keep_last": 4,
163
+ "print_every": 20,
164
+ "context_scan": "segmented",
165
+ "mask_chunk": 512,
166
+ "checkpoint_steps": [
167
+ 5764
168
+ ],
169
+ "seq_assignment": "longest_first",
170
+ "dense_training_attention": "anchor_triton",
171
+ "lora_plus_ratio": 16.0,
172
+ "mask_block_causal": true,
173
+ "teacher_l1_weight": 0.9,
174
+ "slot_decay_gamma": 7.0,
175
+ "late_bonus_layers": 3,
176
+ "late_bonus_rank": 64,
177
+ "late_commit": "confirmed_prefix",
178
+ "late_commit_accept_dist": [
179
+ 0.411835,
180
+ 0.224416,
181
+ 0.129734,
182
+ 0.078494,
183
+ 0.049257,
184
+ 0.031871,
185
+ 0.021181,
186
+ 0.014418,
187
+ 0.01003,
188
+ 0.007117,
189
+ 0.005141,
190
+ 0.003775,
191
+ 0.002813,
192
+ 0.002123,
193
+ 0.001622,
194
+ 0.006173
195
+ ]
196
+ }
197
+ },
198
+ "ar_cache": {
199
+ "dir": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache",
200
+ "key": {
201
+ "target": "dense:models--Qwen--Qwen3-4B@1cfa9a7208912126459214e8b04321603b3df60c",
202
+ "dtype": "bfloat16",
203
+ "backend": "none",
204
+ "batch_invariant": true,
205
+ "fp32_A_log": false,
206
+ "max_new": 512,
207
+ "eos": [
208
+ 151645,
209
+ 151643
210
+ ],
211
+ "decode": {
212
+ "path": "static-graphed",
213
+ "attention": "triton",
214
+ "fused_ops": false,
215
+ "static_max_len": 4096
216
+ },
217
+ "harness": {
218
+ "source_sha1": "04e7703c37603bda2b76c602e94506442edd314a",
219
+ "source_files": [
220
+ "specdec/adapters/gated_lora.py",
221
+ "specdec/fused/branches.py",
222
+ "specdec/fused/hint.py",
223
+ "specdec/fused/late_bonus.py",
224
+ "specdec/fused/layout.py",
225
+ "specdec/fused/policy.py",
226
+ "specdec/fused/transfer.py",
227
+ "specdec/memory/gdn_state.py",
228
+ "specdec/paths/spec.py",
229
+ "specdec/run/decode.py",
230
+ "specdec/target/__init__.py",
231
+ "specdec/target/branch_memory.py",
232
+ "specdec/target/branch_triton.py",
233
+ "specdec/target/cache.py",
234
+ "specdec/target/kernels.py",
235
+ "specdec/target/lora_ops.py",
236
+ "specdec/target/qwen35.py",
237
+ "specdec/target/readout.py",
238
+ "specdec/target/static.py",
239
+ "specdec/target/triton_ops.py"
240
+ ],
241
+ "torch": "2.13.0+cu130",
242
+ "torch_cuda": "13.0",
243
+ "transformers": "5.12.1",
244
+ "sglang": "0.5.18",
245
+ "gpu": "NVIDIA H100 80GB HBM3"
246
+ },
247
+ "dense_sources_sha1": "03ba376b04f7686f84af14b96362f10f03e83351"
248
+ },
249
+ "tag": "5090a239b901",
250
+ "hits": 20,
251
+ "decoded": 0,
252
+ "used": true
253
+ },
254
+ "prompts": {
255
+ "gsm8k": [
256
+ "gsm8k:23",
257
+ "gsm8k:51",
258
+ "gsm8k:61",
259
+ "gsm8k:90"
260
+ ],
261
+ "math500": [
262
+ "math500:14",
263
+ "math500:34",
264
+ "math500:53",
265
+ "math500:79"
266
+ ],
267
+ "humaneval": [
268
+ "humaneval:4",
269
+ "humaneval:19",
270
+ "humaneval:58",
271
+ "humaneval:62"
272
+ ],
273
+ "mbpp": [
274
+ "mbpp:3",
275
+ "mbpp:77",
276
+ "mbpp:87",
277
+ "mbpp:92"
278
+ ],
279
+ "mt-bench": [
280
+ "mt-bench:21",
281
+ "mt-bench:33",
282
+ "mt-bench:46",
283
+ "mt-bench:59"
284
+ ]
285
+ },
286
+ "contamination": {
287
+ "file": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b/contamination.json",
288
+ "excluded_from_clean": {
289
+ "gsm8k": [],
290
+ "math500": [],
291
+ "humaneval": [],
292
+ "mbpp": [],
293
+ "mt-bench": []
294
+ }
295
+ },
296
+ "rows": {
297
+ "file": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_triton_cand/q1_rows.jsonl",
298
+ "run_key": {
299
+ "ar": {
300
+ "target": "dense:models--Qwen--Qwen3-4B@1cfa9a7208912126459214e8b04321603b3df60c",
301
+ "dtype": "bfloat16",
302
+ "backend": "none",
303
+ "batch_invariant": true,
304
+ "fp32_A_log": false,
305
+ "max_new": 512,
306
+ "eos": [
307
+ 151645,
308
+ 151643
309
+ ],
310
+ "decode": {
311
+ "path": "static-graphed",
312
+ "attention": "triton",
313
+ "fused_ops": false,
314
+ "static_max_len": 4096
315
+ },
316
+ "harness": {
317
+ "source_sha1": "04e7703c37603bda2b76c602e94506442edd314a",
318
+ "source_files": [
319
+ "specdec/adapters/gated_lora.py",
320
+ "specdec/fused/branches.py",
321
+ "specdec/fused/hint.py",
322
+ "specdec/fused/late_bonus.py",
323
+ "specdec/fused/layout.py",
324
+ "specdec/fused/policy.py",
325
+ "specdec/fused/transfer.py",
326
+ "specdec/memory/gdn_state.py",
327
+ "specdec/paths/spec.py",
328
+ "specdec/run/decode.py",
329
+ "specdec/target/__init__.py",
330
+ "specdec/target/branch_memory.py",
331
+ "specdec/target/branch_triton.py",
332
+ "specdec/target/cache.py",
333
+ "specdec/target/kernels.py",
334
+ "specdec/target/lora_ops.py",
335
+ "specdec/target/qwen35.py",
336
+ "specdec/target/readout.py",
337
+ "specdec/target/static.py",
338
+ "specdec/target/triton_ops.py"
339
+ ],
340
+ "torch": "2.13.0+cu130",
341
+ "torch_cuda": "13.0",
342
+ "transformers": "5.12.1",
343
+ "sglang": "0.5.18",
344
+ "gpu": "NVIDIA H100 80GB HBM3"
345
+ },
346
+ "dense_sources_sha1": "03ba376b04f7686f84af14b96362f10f03e83351"
347
+ },
348
+ "checkpoint_sha1": "cc4c99cbfaaa0a869f5b219774db513c8fe917d2",
349
+ "mask_mode": "read",
350
+ "ar_checked": true,
351
+ "mask_options": {
352
+ "mask_lean": true,
353
+ "lora_fused_a": true
354
+ },
355
+ "mask_attn_kernel": "branch_stream",
356
+ "late_commit": {
357
+ "lc_readout": "cand",
358
+ "late_commit_sha1": "7766806b2bd5671e7e5038b76b03106b641aa6e4"
359
+ },
360
+ "late_merge": {
361
+ "variant": "merged_triton",
362
+ "late_merge_sha1": "be3bb036283ed581d8cb0e33b64be1e926880dd0"
363
+ }
364
+ },
365
+ "resumed": 0
366
+ },
367
+ "definitions": {
368
+ "tau": "tokens emitted by non-recovery full passes / their count, misses included (specdec/fused/branches.py); no reference counterpart",
369
+ "miss_rate": "misses / non-recovery full passes",
370
+ "tokens_per_pass": "all output tokens / all passes (start and recovery included)",
371
+ "rows_per_pass": "mean over all passes of V + b x 15",
372
+ "tau_ci95": "95 % prompt-level bootstrap (percentile) of pooled tau",
373
+ "paired": "tau(b) - tau(b_max), both arms resampled with the same prompts",
374
+ "e2e": "wall clock against the harness's own AR on the same prompts (e2e_speedup): speedup_decode after the prompt, speedup_e2e with the shared prefill; engine: this harness, not SGLang",
375
+ "pooled_clean": "the same summary without the prompts found in the training split (contamination.excluded_from_clean); present only when some were excluded",
376
+ "tau_pooled": "reference definition; ours = tau_pooled_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)",
377
+ "tau_raw_row": "reference definition; ours = tau_raw_row_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)",
378
+ "tau_prefill_corrected_row": "reference definition; ours = tau_prefill_corrected_row_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)"
379
+ },
380
+ "results": {
381
+ "b1": {
382
+ "pooled": {
383
+ "prompts": 20,
384
+ "tokens": 6702,
385
+ "passes": 2327,
386
+ "fused_passes": 2307,
387
+ "recovery_passes": 0,
388
+ "misses": 0,
389
+ "salvage_passes": 0,
390
+ "salvage_misses": 0,
391
+ "tokens_per_pass": 2.8801031370863774,
392
+ "tau": 2.896402254009536,
393
+ "miss_rate": 0.0,
394
+ "rows_per_pass": 30.871078642028362,
395
+ "rows_per_fused_pass": 31.0,
396
+ "tau_pooled_ref": 2.905071521456437,
397
+ "tau_raw_row_ref": 3.042636491893952,
398
+ "tau_prefill_corrected_row_ref": 3.0306927650103224,
399
+ "accepted_hist": {
400
+ "0": 744,
401
+ "1": 516,
402
+ "2": 377,
403
+ "3": 252,
404
+ "4": 149,
405
+ "5": 102,
406
+ "6": 61,
407
+ "7": 44,
408
+ "8": 26,
409
+ "9": 17,
410
+ "10": 6,
411
+ "11": 5,
412
+ "12": 4,
413
+ "13": 1,
414
+ "14": 1,
415
+ "15": 2
416
+ },
417
+ "committed_hist": {
418
+ "0": 739,
419
+ "1": 512,
420
+ "2": 371,
421
+ "3": 250,
422
+ "4": 149,
423
+ "5": 102,
424
+ "6": 59,
425
+ "7": 44,
426
+ "8": 26,
427
+ "9": 17,
428
+ "10": 6,
429
+ "11": 5,
430
+ "12": 4,
431
+ "13": 1,
432
+ "15": 2
433
+ },
434
+ "overshoot": 12,
435
+ "decode_ms": 50210.76897716522,
436
+ "decode_ms_all": 51173.024958610535,
437
+ "capture_rounds": 2,
438
+ "capture_ms": 962.2559814453125,
439
+ "pass_ms_by_kind": {
440
+ "start": {
441
+ "n": 19,
442
+ "mean_ms": 18.249076341327868,
443
+ "median_ms": 17.521984100341797,
444
+ "p10_ms": 15.117145538330078,
445
+ "p90_ms": 19.71768379211425
446
+ },
447
+ "fused": {
448
+ "n": 2306,
449
+ "mean_ms": 21.623606472974846,
450
+ "median_ms": 21.415023803710938,
451
+ "p10_ms": 17.174799919128418,
452
+ "p90_ms": 25.37324810028076
453
+ }
454
+ },
455
+ "prefill_ms": 3542.4785537719727,
456
+ "accounting_ok": true,
457
+ "exact": 20,
458
+ "tau_ci95": [
459
+ 2.6706615532118887,
460
+ 3.163926499032882
461
+ ],
462
+ "tau_pooled_ref_ci95": [
463
+ 2.680249280920422,
464
+ 3.1744704570791527
465
+ ]
466
+ },
467
+ "benches": {
468
+ "gsm8k": {
469
+ "stats": {
470
+ "prompts": 4,
471
+ "tokens": 987,
472
+ "passes": 302,
473
+ "fused_passes": 298,
474
+ "recovery_passes": 0,
475
+ "misses": 0,
476
+ "salvage_passes": 0,
477
+ "salvage_misses": 0,
478
+ "tokens_per_pass": 3.2682119205298013,
479
+ "tau": 3.2986577181208054,
480
+ "miss_rate": 0.0,
481
+ "rows_per_pass": 30.801324503311257,
482
+ "rows_per_fused_pass": 31.0,
483
+ "tau_pooled_ref": 3.3120805369127515,
484
+ "tau_raw_row_ref": 3.3273406163247117,
485
+ "tau_prefill_corrected_row_ref": 3.3136891521642244,
486
+ "accepted_hist": {
487
+ "0": 96,
488
+ "1": 51,
489
+ "2": 48,
490
+ "3": 36,
491
+ "4": 21,
492
+ "5": 11,
493
+ "6": 8,
494
+ "7": 9,
495
+ "8": 4,
496
+ "9": 5,
497
+ "10": 3,
498
+ "11": 3,
499
+ "12": 1,
500
+ "13": 1,
501
+ "15": 1
502
+ },
503
+ "committed_hist": {
504
+ "0": 96,
505
+ "1": 50,
506
+ "2": 46,
507
+ "3": 36,
508
+ "4": 21,
509
+ "5": 11,
510
+ "6": 7,
511
+ "7": 9,
512
+ "8": 4,
513
+ "9": 5,
514
+ "10": 3,
515
+ "11": 3,
516
+ "12": 1,
517
+ "13": 1,
518
+ "15": 1
519
+ },
520
+ "overshoot": 0,
521
+ "decode_ms": 6451.08588218689,
522
+ "decode_ms_all": 7413.341863632202,
523
+ "capture_rounds": 2,
524
+ "capture_ms": 962.2559814453125,
525
+ "pass_ms_by_kind": {
526
+ "start": {
527
+ "n": 3,
528
+ "mean_ms": 16.64350954691569,
529
+ "median_ms": 17.390560150146484,
530
+ "p10_ms": 15.542931365966798,
531
+ "p90_ms": 17.445267486572266
532
+ },
533
+ "fused": {
534
+ "n": 297,
535
+ "mean_ms": 21.55271162810149,
536
+ "median_ms": 20.610368728637695,
537
+ "p10_ms": 16.941778945922852,
538
+ "p90_ms": 25.5650749206543
539
+ }
540
+ },
541
+ "prefill_ms": 1358.6704330444336,
542
+ "accounting_ok": true,
543
+ "exact": 4,
544
+ "tau_ci95": [
545
+ 3.053156146179402,
546
+ 3.591973244147157
547
+ ],
548
+ "tau_pooled_ref_ci95": [
549
+ 3.0664451827242525,
550
+ 3.6053511705685617
551
+ ]
552
+ },
553
+ "prompts": [
554
+ {
555
+ "pid": "gsm8k:23",
556
+ "tokens": 203,
557
+ "passes": 62,
558
+ "identical": true,
559
+ "first_divergence": null,
560
+ "method_tally": [
561
+ 202,
562
+ 61
563
+ ],
564
+ "reference_tally": [
565
+ 203,
566
+ 61
567
+ ],
568
+ "misses": 0,
569
+ "finish_reason": "stop",
570
+ "decode_ms": 1251.5390405654907,
571
+ "capture_ms": 962.2559814453125
572
+ },
573
+ {
574
+ "pid": "gsm8k:51",
575
+ "tokens": 267,
576
+ "passes": 72,
577
+ "identical": true,
578
+ "first_divergence": null,
579
+ "method_tally": [
580
+ 266,
581
+ 71
582
+ ],
583
+ "reference_tally": [
584
+ 267,
585
+ 71
586
+ ],
587
+ "misses": 0,
588
+ "finish_reason": "stop",
589
+ "decode_ms": 1532.4161596298218,
590
+ "capture_ms": 0
591
+ },
592
+ {
593
+ "pid": "gsm8k:61",
594
+ "tokens": 240,
595
+ "passes": 81,
596
+ "identical": true,
597
+ "first_divergence": null,
598
+ "method_tally": [
599
+ 239,
600
+ 80
601
+ ],
602
+ "reference_tally": [
603
+ 240,
604
+ 80
605
+ ],
606
+ "misses": 0,
607
+ "finish_reason": "stop",
608
+ "decode_ms": 1739.2376585006714,
609
+ "capture_ms": 0
610
+ },
611
+ {
612
+ "pid": "gsm8k:90",
613
+ "tokens": 277,
614
+ "passes": 87,
615
+ "identical": true,
616
+ "first_divergence": null,
617
+ "method_tally": [
618
+ 276,
619
+ 86
620
+ ],
621
+ "reference_tally": [
622
+ 277,
623
+ 86
624
+ ],
625
+ "misses": 0,
626
+ "finish_reason": "stop",
627
+ "decode_ms": 1927.8930234909058,
628
+ "capture_ms": 0
629
+ }
630
+ ],
631
+ "e2e": {
632
+ "prompts": 4,
633
+ "fused_ms_per_token": 6.5360545918813475,
634
+ "ar_ms_per_token": 16.872936100945285,
635
+ "speedup_decode": 2.5815170090384076,
636
+ "speedup_e2e": 2.3075179014065763
637
+ }
638
+ },
639
+ "math500": {
640
+ "stats": {
641
+ "prompts": 4,
642
+ "tokens": 1575,
643
+ "passes": 416,
644
+ "fused_passes": 412,
645
+ "recovery_passes": 0,
646
+ "misses": 0,
647
+ "salvage_passes": 0,
648
+ "salvage_misses": 0,
649
+ "tokens_per_pass": 3.7860576923076925,
650
+ "tau": 3.813106796116505,
651
+ "miss_rate": 0.0,
652
+ "rows_per_pass": 30.85576923076923,
653
+ "rows_per_fused_pass": 31.0,
654
+ "tau_pooled_ref": 3.8228155339805827,
655
+ "tau_raw_row_ref": 3.9567695994927803,
656
+ "tau_prefill_corrected_row_ref": 3.9445688876015206,
657
+ "accepted_hist": {
658
+ "0": 89,
659
+ "1": 80,
660
+ "2": 55,
661
+ "3": 52,
662
+ "4": 37,
663
+ "5": 28,
664
+ "6": 24,
665
+ "7": 19,
666
+ "8": 14,
667
+ "9": 7,
668
+ "10": 1,
669
+ "11": 2,
670
+ "12": 3,
671
+ "14": 1
672
+ },
673
+ "committed_hist": {
674
+ "0": 89,
675
+ "1": 79,
676
+ "2": 55,
677
+ "3": 51,
678
+ "4": 37,
679
+ "5": 28,
680
+ "6": 23,
681
+ "7": 19,
682
+ "8": 14,
683
+ "9": 7,
684
+ "10": 1,
685
+ "11": 2,
686
+ "12": 3
687
+ },
688
+ "overshoot": 9,
689
+ "decode_ms": 8918.562644958496,
690
+ "decode_ms_all": 8918.562644958496,
691
+ "capture_rounds": 0,
692
+ "capture_ms": 0,
693
+ "pass_ms_by_kind": {
694
+ "start": {
695
+ "n": 4,
696
+ "mean_ms": 16.89413619041443,
697
+ "median_ms": 17.443056106567383,
698
+ "p10_ms": 15.814899349212647,
699
+ "p90_ms": 17.534237098693847
700
+ },
701
+ "fused": {
702
+ "n": 412,
703
+ "mean_ms": 21.48297597135155,
704
+ "median_ms": 21.49619197845459,
705
+ "p10_ms": 16.927778434753417,
706
+ "p90_ms": 25.366207885742188
707
+ }
708
+ },
709
+ "prefill_ms": 508.09510803222656,
710
+ "accounting_ok": true,
711
+ "exact": 4,
712
+ "tau_ci95": [
713
+ 3.4249547920434,
714
+ 4.204545454545454
715
+ ],
716
+ "tau_pooled_ref_ci95": [
717
+ 3.4321880650994574,
718
+ 4.222727272727273
719
+ ]
720
+ },
721
+ "prompts": [
722
+ {
723
+ "pid": "math500:14",
724
+ "tokens": 512,
725
+ "passes": 155,
726
+ "identical": true,
727
+ "first_divergence": null,
728
+ "method_tally": [
729
+ 511,
730
+ 154
731
+ ],
732
+ "reference_tally": [
733
+ 512,
734
+ 154
735
+ ],
736
+ "misses": 0,
737
+ "finish_reason": "length",
738
+ "decode_ms": 3418.5534620285034,
739
+ "capture_ms": 0
740
+ },
741
+ {
742
+ "pid": "math500:34",
743
+ "tokens": 512,
744
+ "passes": 125,
745
+ "identical": true,
746
+ "first_divergence": null,
747
+ "method_tally": [
748
+ 511,
749
+ 124
750
+ ],
751
+ "reference_tally": [
752
+ 512,
753
+ 124
754
+ ],
755
+ "misses": 0,
756
+ "finish_reason": "length",
757
+ "decode_ms": 2651.048776626587,
758
+ "capture_ms": 0
759
+ },
760
+ {
761
+ "pid": "math500:53",
762
+ "tokens": 362,
763
+ "passes": 92,
764
+ "identical": true,
765
+ "first_divergence": null,
766
+ "method_tally": [
767
+ 361,
768
+ 91
769
+ ],
770
+ "reference_tally": [
771
+ 362,
772
+ 91
773
+ ],
774
+ "misses": 0,
775
+ "finish_reason": "stop",
776
+ "decode_ms": 1940.5727281570435,
777
+ "capture_ms": 0
778
+ },
779
+ {
780
+ "pid": "math500:79",
781
+ "tokens": 189,
782
+ "passes": 44,
783
+ "identical": true,
784
+ "first_divergence": null,
785
+ "method_tally": [
786
+ 188,
787
+ 43
788
+ ],
789
+ "reference_tally": [
790
+ 189,
791
+ 43
792
+ ],
793
+ "misses": 0,
794
+ "finish_reason": "stop",
795
+ "decode_ms": 908.3876781463623,
796
+ "capture_ms": 0
797
+ }
798
+ ],
799
+ "e2e": {
800
+ "prompts": 4,
801
+ "fused_ms_per_token": 5.662579457116506,
802
+ "ar_ms_per_token": 16.996904987153552,
803
+ "speedup_decode": 3.00161880568272,
804
+ "speedup_e2e": 2.910911740960628
805
+ }
806
+ },
807
+ "humaneval": {
808
+ "stats": {
809
+ "prompts": 4,
810
+ "tokens": 1615,
811
+ "passes": 613,
812
+ "fused_passes": 609,
813
+ "recovery_passes": 0,
814
+ "misses": 0,
815
+ "salvage_passes": 0,
816
+ "salvage_misses": 0,
817
+ "tokens_per_pass": 2.634584013050571,
818
+ "tau": 2.645320197044335,
819
+ "miss_rate": 0.0,
820
+ "rows_per_pass": 30.902120717781404,
821
+ "rows_per_fused_pass": 31.0,
822
+ "tau_pooled_ref": 2.651888341543514,
823
+ "tau_raw_row_ref": 2.6607403693316214,
824
+ "tau_prefill_corrected_row_ref": 2.653624249097236,
825
+ "accepted_hist": {
826
+ "0": 194,
827
+ "1": 145,
828
+ "2": 118,
829
+ "3": 68,
830
+ "4": 42,
831
+ "5": 21,
832
+ "6": 10,
833
+ "7": 6,
834
+ "8": 2,
835
+ "9": 2,
836
+ "10": 1
837
+ },
838
+ "committed_hist": {
839
+ "0": 193,
840
+ "1": 144,
841
+ "2": 116,
842
+ "3": 68,
843
+ "4": 42,
844
+ "5": 21,
845
+ "6": 10,
846
+ "7": 6,
847
+ "8": 2,
848
+ "9": 2,
849
+ "10": 1
850
+ },
851
+ "overshoot": 2,
852
+ "decode_ms": 13492.281768798828,
853
+ "decode_ms_all": 13492.281768798828,
854
+ "capture_rounds": 0,
855
+ "capture_ms": 0,
856
+ "pass_ms_by_kind": {
857
+ "start": {
858
+ "n": 4,
859
+ "mean_ms": 16.48423981666565,
860
+ "median_ms": 17.573375701904297,
861
+ "p10_ms": 14.457433605194092,
862
+ "p90_ms": 17.63973731994629
863
+ },
864
+ "fused": {
865
+ "n": 609,
866
+ "mean_ms": 22.046543201202244,
867
+ "median_ms": 23.11190414428711,
868
+ "p10_ms": 17.786343002319335,
869
+ "p90_ms": 25.10032730102539
870
+ }
871
+ },
872
+ "prefill_ms": 564.7402801513672,
873
+ "accounting_ok": true,
874
+ "exact": 4,
875
+ "tau_ci95": [
876
+ 2.4793956043956045,
877
+ 2.8505392912172574
878
+ ],
879
+ "tau_pooled_ref_ci95": [
880
+ 2.48489010989011,
881
+ 2.856702619414484
882
+ ]
883
+ },
884
+ "prompts": [
885
+ {
886
+ "pid": "humaneval:4",
887
+ "tokens": 318,
888
+ "passes": 125,
889
+ "identical": true,
890
+ "first_divergence": null,
891
+ "method_tally": [
892
+ 317,
893
+ 124
894
+ ],
895
+ "reference_tally": [
896
+ 318,
897
+ 124
898
+ ],
899
+ "misses": 0,
900
+ "finish_reason": "stop",
901
+ "decode_ms": 2711.967933654785,
902
+ "capture_ms": 0
903
+ },
904
+ {
905
+ "pid": "humaneval:19",
906
+ "tokens": 512,
907
+ "passes": 210,
908
+ "identical": true,
909
+ "first_divergence": null,
910
+ "method_tally": [
911
+ 511,
912
+ 209
913
+ ],
914
+ "reference_tally": [
915
+ 512,
916
+ 209
917
+ ],
918
+ "misses": 0,
919
+ "finish_reason": "length",
920
+ "decode_ms": 4634.605976104736,
921
+ "capture_ms": 0
922
+ },
923
+ {
924
+ "pid": "humaneval:58",
925
+ "tokens": 273,
926
+ "passes": 102,
927
+ "identical": true,
928
+ "first_divergence": null,
929
+ "method_tally": [
930
+ 272,
931
+ 101
932
+ ],
933
+ "reference_tally": [
934
+ 273,
935
+ 101
936
+ ],
937
+ "misses": 0,
938
+ "finish_reason": "stop",
939
+ "decode_ms": 2259.9564695358276,
940
+ "capture_ms": 0
941
+ },
942
+ {
943
+ "pid": "humaneval:62",
944
+ "tokens": 512,
945
+ "passes": 176,
946
+ "identical": true,
947
+ "first_divergence": null,
948
+ "method_tally": [
949
+ 511,
950
+ 175
951
+ ],
952
+ "reference_tally": [
953
+ 512,
954
+ 175
955
+ ],
956
+ "misses": 0,
957
+ "finish_reason": "length",
958
+ "decode_ms": 3885.751389503479,
959
+ "capture_ms": 0
960
+ }
961
+ ],
962
+ "e2e": {
963
+ "prompts": 4,
964
+ "fused_ms_per_token": 8.354354036407942,
965
+ "ar_ms_per_token": 17.069150008210457,
966
+ "speedup_decode": 2.0431442016730177,
967
+ "speedup_e2e": 2.0035443351966586
968
+ }
969
+ },
970
+ "mbpp": {
971
+ "stats": {
972
+ "prompts": 4,
973
+ "tokens": 1333,
974
+ "passes": 557,
975
+ "fused_passes": 553,
976
+ "recovery_passes": 0,
977
+ "misses": 0,
978
+ "salvage_passes": 0,
979
+ "salvage_misses": 0,
980
+ "tokens_per_pass": 2.393177737881508,
981
+ "tau": 2.403254972875226,
982
+ "miss_rate": 0.0,
983
+ "rows_per_pass": 30.892280071813286,
984
+ "rows_per_fused_pass": 31.0,
985
+ "tau_pooled_ref": 2.4104882459312837,
986
+ "tau_raw_row_ref": 2.4846920289855072,
987
+ "tau_prefill_corrected_row_ref": 2.476083721532091,
988
+ "accepted_hist": {
989
+ "0": 216,
990
+ "1": 144,
991
+ "2": 82,
992
+ "3": 49,
993
+ "4": 24,
994
+ "5": 18,
995
+ "6": 10,
996
+ "7": 5,
997
+ "8": 4,
998
+ "9": 1
999
+ },
1000
+ "committed_hist": {
1001
+ "0": 214,
1002
+ "1": 144,
1003
+ "2": 81,
1004
+ "3": 48,
1005
+ "4": 24,
1006
+ "5": 18,
1007
+ "6": 10,
1008
+ "7": 5,
1009
+ "8": 4,
1010
+ "9": 1
1011
+ },
1012
+ "overshoot": 1,
1013
+ "decode_ms": 11931.00333404541,
1014
+ "decode_ms_all": 11931.00333404541,
1015
+ "capture_rounds": 0,
1016
+ "capture_ms": 0,
1017
+ "pass_ms_by_kind": {
1018
+ "start": {
1019
+ "n": 4,
1020
+ "mean_ms": 17.607168197631836,
1021
+ "median_ms": 17.5380802154541,
1022
+ "p10_ms": 17.508051872253418,
1023
+ "p90_ms": 17.76155490875244
1024
+ },
1025
+ "fused": {
1026
+ "n": 553,
1027
+ "mean_ms": 21.4476937816544,
1028
+ "median_ms": 20.650175094604492,
1029
+ "p10_ms": 17.070034408569335,
1030
+ "p90_ms": 26.31941757202149
1031
+ }
1032
+ },
1033
+ "prefill_ms": 568.5144195556641,
1034
+ "accounting_ok": true,
1035
+ "exact": 4,
1036
+ "tau_ci95": [
1037
+ 2.271978021978022,
1038
+ 2.6560846560846563
1039
+ ],
1040
+ "tau_pooled_ref_ci95": [
1041
+ 2.2774725274725274,
1042
+ 2.6666666666666665
1043
+ ]
1044
+ },
1045
+ "prompts": [
1046
+ {
1047
+ "pid": "mbpp:3",
1048
+ "tokens": 317,
1049
+ "passes": 141,
1050
+ "identical": true,
1051
+ "first_divergence": null,
1052
+ "method_tally": [
1053
+ 316,
1054
+ 140
1055
+ ],
1056
+ "reference_tally": [
1057
+ 317,
1058
+ 140
1059
+ ],
1060
+ "misses": 0,
1061
+ "finish_reason": "stop",
1062
+ "decode_ms": 2991.5325422286987,
1063
+ "capture_ms": 0
1064
+ },
1065
+ {
1066
+ "pid": "mbpp:77",
1067
+ "tokens": 512,
1068
+ "passes": 225,
1069
+ "identical": true,
1070
+ "first_divergence": null,
1071
+ "method_tally": [
1072
+ 511,
1073
+ 224
1074
+ ],
1075
+ "reference_tally": [
1076
+ 512,
1077
+ 224
1078
+ ],
1079
+ "misses": 0,
1080
+ "finish_reason": "length",
1081
+ "decode_ms": 4822.021187782288,
1082
+ "capture_ms": 0
1083
+ },
1084
+ {
1085
+ "pid": "mbpp:87",
1086
+ "tokens": 311,
1087
+ "passes": 121,
1088
+ "identical": true,
1089
+ "first_divergence": null,
1090
+ "method_tally": [
1091
+ 310,
1092
+ 120
1093
+ ],
1094
+ "reference_tally": [
1095
+ 311,
1096
+ 120
1097
+ ],
1098
+ "misses": 0,
1099
+ "finish_reason": "stop",
1100
+ "decode_ms": 2660.395553588867,
1101
+ "capture_ms": 0
1102
+ },
1103
+ {
1104
+ "pid": "mbpp:92",
1105
+ "tokens": 193,
1106
+ "passes": 70,
1107
+ "identical": true,
1108
+ "first_divergence": null,
1109
+ "method_tally": [
1110
+ 192,
1111
+ 69
1112
+ ],
1113
+ "reference_tally": [
1114
+ 193,
1115
+ 69
1116
+ ],
1117
+ "misses": 0,
1118
+ "finish_reason": "stop",
1119
+ "decode_ms": 1457.0540504455566,
1120
+ "capture_ms": 0
1121
+ }
1122
+ ],
1123
+ "e2e": {
1124
+ "prompts": 4,
1125
+ "fused_ms_per_token": 8.950490123064824,
1126
+ "ar_ms_per_token": 17.074735396085426,
1127
+ "speedup_decode": 1.9076871949263376,
1128
+ "speedup_e2e": 1.870920221315493
1129
+ }
1130
+ },
1131
+ "mt-bench": {
1132
+ "stats": {
1133
+ "prompts": 4,
1134
+ "tokens": 1192,
1135
+ "passes": 439,
1136
+ "fused_passes": 435,
1137
+ "recovery_passes": 0,
1138
+ "misses": 0,
1139
+ "salvage_passes": 0,
1140
+ "salvage_misses": 0,
1141
+ "tokens_per_pass": 2.715261958997722,
1142
+ "tau": 2.731034482758621,
1143
+ "miss_rate": 0.0,
1144
+ "rows_per_pass": 30.863325740318906,
1145
+ "rows_per_fused_pass": 31.0,
1146
+ "tau_pooled_ref": 2.740229885057471,
1147
+ "tau_raw_row_ref": 2.783639845335141,
1148
+ "tau_prefill_corrected_row_ref": 2.76549781465654,
1149
+ "accepted_hist": {
1150
+ "0": 149,
1151
+ "1": 96,
1152
+ "2": 74,
1153
+ "3": 47,
1154
+ "4": 25,
1155
+ "5": 24,
1156
+ "6": 9,
1157
+ "7": 5,
1158
+ "8": 2,
1159
+ "9": 2,
1160
+ "10": 1,
1161
+ "15": 1
1162
+ },
1163
+ "committed_hist": {
1164
+ "0": 147,
1165
+ "1": 95,
1166
+ "2": 73,
1167
+ "3": 47,
1168
+ "4": 25,
1169
+ "5": 24,
1170
+ "6": 9,
1171
+ "7": 5,
1172
+ "8": 2,
1173
+ "9": 2,
1174
+ "10": 1,
1175
+ "15": 1
1176
+ },
1177
+ "overshoot": 0,
1178
+ "decode_ms": 9417.835347175598,
1179
+ "decode_ms_all": 9417.835347175598,
1180
+ "capture_rounds": 0,
1181
+ "capture_ms": 0,
1182
+ "pass_ms_by_kind": {
1183
+ "start": {
1184
+ "n": 4,
1185
+ "mean_ms": 23.21493625640869,
1186
+ "median_ms": 22.42342472076416,
1187
+ "p10_ms": 17.681142234802245,
1188
+ "p90_ms": 29.381939506530763
1189
+ },
1190
+ "fused": {
1191
+ "n": 435,
1192
+ "mean_ms": 21.436725522183824,
1193
+ "median_ms": 20.64240074157715,
1194
+ "p10_ms": 16.972576141357422,
1195
+ "p90_ms": 25.373004913330078
1196
+ }
1197
+ },
1198
+ "prefill_ms": 542.4583129882812,
1199
+ "accounting_ok": true,
1200
+ "exact": 4,
1201
+ "tau_ci95": [
1202
+ 2.3893333333333335,
1203
+ 2.9574468085106385
1204
+ ],
1205
+ "tau_pooled_ref_ci95": [
1206
+ 2.4,
1207
+ 2.9680851063829787
1208
+ ]
1209
+ },
1210
+ "prompts": [
1211
+ {
1212
+ "pid": "mt-bench:21",
1213
+ "tokens": 133,
1214
+ "passes": 65,
1215
+ "identical": true,
1216
+ "first_divergence": null,
1217
+ "method_tally": [
1218
+ 132,
1219
+ 64
1220
+ ],
1221
+ "reference_tally": [
1222
+ 133,
1223
+ 64
1224
+ ],
1225
+ "misses": 0,
1226
+ "finish_reason": "stop",
1227
+ "decode_ms": 1348.8949670791626,
1228
+ "capture_ms": 0
1229
+ },
1230
+ {
1231
+ "pid": "mt-bench:33",
1232
+ "tokens": 483,
1233
+ "passes": 167,
1234
+ "identical": true,
1235
+ "first_divergence": null,
1236
+ "method_tally": [
1237
+ 482,
1238
+ 166
1239
+ ],
1240
+ "reference_tally": [
1241
+ 483,
1242
+ 166
1243
+ ],
1244
+ "misses": 0,
1245
+ "finish_reason": "stop",
1246
+ "decode_ms": 3588.9172401428223,
1247
+ "capture_ms": 0
1248
+ },
1249
+ {
1250
+ "pid": "mt-bench:46",
1251
+ "tokens": 501,
1252
+ "passes": 184,
1253
+ "identical": true,
1254
+ "first_divergence": null,
1255
+ "method_tally": [
1256
+ 500,
1257
+ 183
1258
+ ],
1259
+ "reference_tally": [
1260
+ 501,
1261
+ 183
1262
+ ],
1263
+ "misses": 0,
1264
+ "finish_reason": "stop",
1265
+ "decode_ms": 3961.2701454162598,
1266
+ "capture_ms": 0
1267
+ },
1268
+ {
1269
+ "pid": "mt-bench:59",
1270
+ "tokens": 75,
1271
+ "passes": 23,
1272
+ "identical": true,
1273
+ "first_divergence": null,
1274
+ "method_tally": [
1275
+ 74,
1276
+ 22
1277
+ ],
1278
+ "reference_tally": [
1279
+ 75,
1280
+ 22
1281
+ ],
1282
+ "misses": 0,
1283
+ "finish_reason": "stop",
1284
+ "decode_ms": 518.7529945373535,
1285
+ "capture_ms": 0
1286
+ }
1287
+ ],
1288
+ "e2e": {
1289
+ "prompts": 4,
1290
+ "fused_ms_per_token": 7.900868579845301,
1291
+ "ar_ms_per_token": 17.135559822888982,
1292
+ "speedup_decode": 2.1688197506032303,
1293
+ "speedup_e2e": 2.1191738538315072
1294
+ }
1295
+ }
1296
+ },
1297
+ "e2e_pooled": {
1298
+ "prompts": 20,
1299
+ "fused_ms_per_token": 7.4919082329402,
1300
+ "ar_ms_per_token": 17.036198170922898,
1301
+ "speedup_decode": 2.2739464554595914,
1302
+ "speedup_e2e": 2.1969689818137583
1303
+ }
1304
+ }
1305
+ },
1306
+ "paired_vs_b_max": {},
1307
+ "mismatches_vs_ar": {},
1308
+ "reference_comparison": [
1309
+ {
1310
+ "note": "--dense-target: the reference numbers are SGLang runs of the 27B, not of this target"
1311
+ }
1312
+ ]
1313
+ }
qwen3_4b/eval/lm_LCa3_v11/lm_triton_cand/q1_rows.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
qwen3_4b/eval/lm_LCa3_v11/lm_triton_cand_eager.log ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ checkpoint /workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/ckpt-0005764/adapters.pt -> /workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_triton_cand_eager
2
+
3
+ Warning: please use at least NVCC 12.9 for the best DeepGEMM performance
4
+ AR token ids: 20 from cache, 0 decoded (/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache)
5
+ b=1: tau 2.896 [2.671, 3.164], miss rate 0.000, tokens/pass 2.880, rows/pass 30.9, exact 20/20; 162s
6
+ round ms by kind (median [p10, p90], capture rounds left out): start 62.4 [49.6, 84.8] n=20, fused 64.6 [49.9, 86.1] n=2307
7
+ wall clock vs harness AR on 20 prompts: 22.88 vs 51.05 ms/token, speedup 2.23x decode, 2.21x with prefill
8
+ no decodes differ from AR; wrote /workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_triton_cand_eager/q1_eval.json
qwen3_4b/eval/lm_LCa3_v11/lm_triton_cand_eager/q1_eval.json ADDED
@@ -0,0 +1,1313 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "tiny": false,
4
+ "checkpoint": [
5
+ "/workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/ckpt-0005764"
6
+ ],
7
+ "b": [
8
+ 1
9
+ ],
10
+ "benches": [
11
+ "gsm8k",
12
+ "math500",
13
+ "humaneval",
14
+ "mbpp",
15
+ "mt-bench"
16
+ ],
17
+ "limit": null,
18
+ "per_bench": 4,
19
+ "time_ar": true,
20
+ "skip_ar": false,
21
+ "graphs": false,
22
+ "attention": "triton",
23
+ "fused_ops": false,
24
+ "branch_memory": "copy",
25
+ "mask_attn_kernel": "branch_stream",
26
+ "static_max_len": 4096,
27
+ "max_new": 512,
28
+ "prompts_dir": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b",
29
+ "ar_cache": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache",
30
+ "reference": "/workspace/opt_runs/div2_v2_4b/code_lm/results/prompts/reference_numbers.json",
31
+ "contamination": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b/contamination.json",
32
+ "reps": 2000,
33
+ "batch_invariant": true,
34
+ "fp32_A_log": true,
35
+ "mask_lean": true,
36
+ "lora_fused_a": true,
37
+ "seed": 0,
38
+ "out": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_triton_cand_eager",
39
+ "threads": 1,
40
+ "fresh": false,
41
+ "policy": "fused",
42
+ "mtp": null,
43
+ "mtp_checkpoint": null,
44
+ "sglang_mtp": "/workspace/opt_runs/div2_v2_4b/code_lm/results/baselines/beta-off",
45
+ "drafter": null,
46
+ "mtp_vocab": null,
47
+ "dflash_vocab": null,
48
+ "draft_log": null,
49
+ "salvage": 0,
50
+ "calibration": null,
51
+ "mask_readout": "full",
52
+ "miss_policy": "recovery",
53
+ "dense_target": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c",
54
+ "late_commit": "on",
55
+ "lc_readout": "cand",
56
+ "late_merge": "on",
57
+ "late_merge_ops": "triton",
58
+ "dense_snapshot": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c"
59
+ },
60
+ "policy": "fused",
61
+ "checkpoint": "/workspace/opt_runs/div2_v2_4b/out/LCa3_L31r128_lp16_causal_dsloss_q/ckpt-0005764/adapters.pt",
62
+ "options": {
63
+ "attention": "triton",
64
+ "fused_ops": false,
65
+ "graphs": false,
66
+ "static_max_len": 4096,
67
+ "decode": {
68
+ "path": "dynamic-eager",
69
+ "attention": "triton",
70
+ "fused_ops": false,
71
+ "static_max_len": null
72
+ },
73
+ "mask_lean": true,
74
+ "lora_fused_a": true,
75
+ "mask_attn_kernel": "branch_stream",
76
+ "late_commit": "on",
77
+ "lc_readout": "cand"
78
+ },
79
+ "provenance": {
80
+ "git_head": null,
81
+ "row_git_heads": [
82
+ "None"
83
+ ]
84
+ },
85
+ "checkpoint_meta": {
86
+ "mask_from_layer": 31,
87
+ "max_block": 16,
88
+ "mask_attn_window": null,
89
+ "local_window": 0,
90
+ "mask_mode": "read",
91
+ "rank": 128,
92
+ "alpha": 90.50966799187809,
93
+ "targets": [
94
+ "q_proj",
95
+ "k_proj",
96
+ "v_proj",
97
+ "o_proj",
98
+ "gate_proj",
99
+ "up_proj",
100
+ "down_proj"
101
+ ],
102
+ "mask_block_causal": true,
103
+ "late_bonus_layers": 3,
104
+ "late_bonus_rank": 64
105
+ },
106
+ "checkpoint_extra": {
107
+ "step": 5764,
108
+ "config": {
109
+ "name": "k3q_qwen3_4b_LCa3_L31r128_lp16_causal_dsloss",
110
+ "description": "LCa3: SPL (ACc plus DSpine's backbone loss: 0.1 CE + 0.9 full-vocabulary L1 to the frozen target at row s+1+m, exp(-m/7) slot weights) plus late commit LC-a with k = 3 (docs/plans/late-commit-2026-11-14.md, sections 1-3 and 9): one MASK block per round. The last 3 MASK layers (33-35) run after verification: context <= the endpoint e, queries and the block's own keys at the actual positions e+1+m, and the bonus ids[e+1] injected right before layer 33 as in LB-3 (B(A e(b)), A [64, 2560], B [2560, 64] zero-initialised, B in the LoRA+ group). Layers 31-32 see only the confirmed prefix <= u = e - a (fork x_L0[u], positions u+1+m), a drawn per anchor from late_commit_accept_dist (its own stream; anchors and data order unchanged), u clamped at prompt_len - 1. The distribution is the per-row geometric mixture of SPL's 128-row walk (review F3): for each row of div2/v2_4b/results/SPL_L31r128_lp16_causal_dsloss_q_e0_128rows_perrow.jsonl a geometric capped at 15 whose mean equals that row's mean accept count (tokens/passes - 1), mixed with the rows' passes as weights (mean a 1.71766 = W128 2.71766 - 1; P(a=0) 0.412, P(a>=10) 0.022, P(a=15) 0.006). Labels ids[e+2+m], SPL's loss unchanged; with a = 0 the forward is LB-3's bit for bit. The eval adds walk_late_commit_full (W_LC: the 128 rows' rounds simulated, the next block from the lower layers <= u and the late layers <= e + bonus, a + 1 tokens per round, the full-vocabulary late readout) and walk_late_commit (cost variant: late slot m reads only the in-pass top-16 of slot m+a, the full vocabulary where m+a >= 15); the existing walk (block at the true endpoint: the perfect branch) stays beside them. Controls: SPL_L31r128_lp16_causal_dsloss_q and SPLLB3_L31r128_lp16_causal_dsloss_q at the same step (same seed, rows, anchors and eval rows). Verdict fixed before results (plan sections 4 and 9): the metric is walk_late_commit_full (the full-vocabulary late readout; plan section 9 replaced the top-16 readout, whose candidates were misaligned, before any result). With the better arm, W_LC_full / W128(SPL) >= 0.92 with the paired-bootstrap 95% lower bound >= 0.90 passes (the single-block structure is scaled up: 4B engine decode, Z1 cost model, then 27B); >= 1.00 is the strong pass; LC-a below 0.92 sends LC-b in, and if that also fails the branch/sibling structure stays. Needs late_commit.patch on top of native_training.patch. | SCREEN: quarter epoch (5764 steps), own cosine decay to 0",
111
+ "target": {
112
+ "kind": "hf_dense",
113
+ "dtype": "bfloat16",
114
+ "snapshot": "/workspace/opt_runs/div2_v2_4b/hf-home/hub/models--Qwen--Qwen3-4B/snapshots/1cfa9a7208912126459214e8b04321603b3df60c"
115
+ },
116
+ "device": "cuda",
117
+ "train_data": "/workspace/opt_runs/div2_v2_4b/data/prepared/train",
118
+ "heldout_data": "/workspace/opt_runs/div2_v2_4b/data/prepared/heldout",
119
+ "mask_mode": "read",
120
+ "mask_from_layer": 31,
121
+ "mask_attn_window": null,
122
+ "local_window": 0,
123
+ "rank": 128,
124
+ "alpha": 90.50966799187809,
125
+ "targets": [
126
+ "q_proj",
127
+ "k_proj",
128
+ "v_proj",
129
+ "o_proj",
130
+ "gate_proj",
131
+ "up_proj",
132
+ "down_proj"
133
+ ],
134
+ "freeze_qkv_kv_rows": false,
135
+ "block": 15,
136
+ "max_block": 16,
137
+ "max_len": 3072,
138
+ "anchors_per_seq": 512,
139
+ "anchor_sampling": "uniform",
140
+ "hard_candidates": 4,
141
+ "seqs_per_step": 32,
142
+ "steps": 5764,
143
+ "arm_stop_after": null,
144
+ "lr": 0.0003,
145
+ "min_lr_ratio": 0.0,
146
+ "warmup_steps": 200,
147
+ "weight_decay": 0.0,
148
+ "betas": [
149
+ 0.9,
150
+ 0.999
151
+ ],
152
+ "adam_eps": 1e-08,
153
+ "grad_clip": 1.0,
154
+ "seed": 20260923,
155
+ "eval_every": 1441,
156
+ "eval_rows": 128,
157
+ "eval_anchors_per_seq": 32,
158
+ "eval_seed": 7,
159
+ "eval_walk_rows": 8,
160
+ "eval_walk_chunk": 128,
161
+ "ckpt_every": 1441,
162
+ "keep_last": 4,
163
+ "print_every": 20,
164
+ "context_scan": "segmented",
165
+ "mask_chunk": 512,
166
+ "checkpoint_steps": [
167
+ 5764
168
+ ],
169
+ "seq_assignment": "longest_first",
170
+ "dense_training_attention": "anchor_triton",
171
+ "lora_plus_ratio": 16.0,
172
+ "mask_block_causal": true,
173
+ "teacher_l1_weight": 0.9,
174
+ "slot_decay_gamma": 7.0,
175
+ "late_bonus_layers": 3,
176
+ "late_bonus_rank": 64,
177
+ "late_commit": "confirmed_prefix",
178
+ "late_commit_accept_dist": [
179
+ 0.411835,
180
+ 0.224416,
181
+ 0.129734,
182
+ 0.078494,
183
+ 0.049257,
184
+ 0.031871,
185
+ 0.021181,
186
+ 0.014418,
187
+ 0.01003,
188
+ 0.007117,
189
+ 0.005141,
190
+ 0.003775,
191
+ 0.002813,
192
+ 0.002123,
193
+ 0.001622,
194
+ 0.006173
195
+ ]
196
+ }
197
+ },
198
+ "ar_cache": {
199
+ "dir": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/ar_cache",
200
+ "key": {
201
+ "target": "dense:models--Qwen--Qwen3-4B@1cfa9a7208912126459214e8b04321603b3df60c",
202
+ "dtype": "bfloat16",
203
+ "backend": "none",
204
+ "batch_invariant": true,
205
+ "fp32_A_log": false,
206
+ "max_new": 512,
207
+ "eos": [
208
+ 151645,
209
+ 151643
210
+ ],
211
+ "decode": {
212
+ "path": "dynamic-eager",
213
+ "attention": "triton",
214
+ "fused_ops": false,
215
+ "static_max_len": null
216
+ },
217
+ "harness": {
218
+ "source_sha1": "04e7703c37603bda2b76c602e94506442edd314a",
219
+ "source_files": [
220
+ "specdec/adapters/gated_lora.py",
221
+ "specdec/fused/branches.py",
222
+ "specdec/fused/hint.py",
223
+ "specdec/fused/late_bonus.py",
224
+ "specdec/fused/layout.py",
225
+ "specdec/fused/policy.py",
226
+ "specdec/fused/transfer.py",
227
+ "specdec/memory/gdn_state.py",
228
+ "specdec/paths/spec.py",
229
+ "specdec/run/decode.py",
230
+ "specdec/target/__init__.py",
231
+ "specdec/target/branch_memory.py",
232
+ "specdec/target/branch_triton.py",
233
+ "specdec/target/cache.py",
234
+ "specdec/target/kernels.py",
235
+ "specdec/target/lora_ops.py",
236
+ "specdec/target/qwen35.py",
237
+ "specdec/target/readout.py",
238
+ "specdec/target/static.py",
239
+ "specdec/target/triton_ops.py"
240
+ ],
241
+ "torch": "2.13.0+cu130",
242
+ "torch_cuda": "13.0",
243
+ "transformers": "5.12.1",
244
+ "sglang": "0.5.18",
245
+ "gpu": "NVIDIA H100 80GB HBM3"
246
+ },
247
+ "dense_sources_sha1": "03ba376b04f7686f84af14b96362f10f03e83351"
248
+ },
249
+ "tag": "925a25022406",
250
+ "hits": 20,
251
+ "decoded": 0,
252
+ "used": true
253
+ },
254
+ "prompts": {
255
+ "gsm8k": [
256
+ "gsm8k:23",
257
+ "gsm8k:51",
258
+ "gsm8k:61",
259
+ "gsm8k:90"
260
+ ],
261
+ "math500": [
262
+ "math500:14",
263
+ "math500:34",
264
+ "math500:53",
265
+ "math500:79"
266
+ ],
267
+ "humaneval": [
268
+ "humaneval:4",
269
+ "humaneval:19",
270
+ "humaneval:58",
271
+ "humaneval:62"
272
+ ],
273
+ "mbpp": [
274
+ "mbpp:3",
275
+ "mbpp:77",
276
+ "mbpp:87",
277
+ "mbpp:92"
278
+ ],
279
+ "mt-bench": [
280
+ "mt-bench:21",
281
+ "mt-bench:33",
282
+ "mt-bench:46",
283
+ "mt-bench:59"
284
+ ]
285
+ },
286
+ "contamination": {
287
+ "file": "/workspace/opt_runs/div2_v2_4b/code/results/prompts_dense/qwen3-4b/contamination.json",
288
+ "excluded_from_clean": {
289
+ "gsm8k": [],
290
+ "math500": [],
291
+ "humaneval": [],
292
+ "mbpp": [],
293
+ "mt-bench": []
294
+ }
295
+ },
296
+ "rows": {
297
+ "file": "/workspace/opt_runs/div2_v2_4b/out/eval/lm_LCa3_v11/lm_triton_cand_eager/q1_rows.jsonl",
298
+ "run_key": {
299
+ "ar": {
300
+ "target": "dense:models--Qwen--Qwen3-4B@1cfa9a7208912126459214e8b04321603b3df60c",
301
+ "dtype": "bfloat16",
302
+ "backend": "none",
303
+ "batch_invariant": true,
304
+ "fp32_A_log": false,
305
+ "max_new": 512,
306
+ "eos": [
307
+ 151645,
308
+ 151643
309
+ ],
310
+ "decode": {
311
+ "path": "dynamic-eager",
312
+ "attention": "triton",
313
+ "fused_ops": false,
314
+ "static_max_len": null
315
+ },
316
+ "harness": {
317
+ "source_sha1": "04e7703c37603bda2b76c602e94506442edd314a",
318
+ "source_files": [
319
+ "specdec/adapters/gated_lora.py",
320
+ "specdec/fused/branches.py",
321
+ "specdec/fused/hint.py",
322
+ "specdec/fused/late_bonus.py",
323
+ "specdec/fused/layout.py",
324
+ "specdec/fused/policy.py",
325
+ "specdec/fused/transfer.py",
326
+ "specdec/memory/gdn_state.py",
327
+ "specdec/paths/spec.py",
328
+ "specdec/run/decode.py",
329
+ "specdec/target/__init__.py",
330
+ "specdec/target/branch_memory.py",
331
+ "specdec/target/branch_triton.py",
332
+ "specdec/target/cache.py",
333
+ "specdec/target/kernels.py",
334
+ "specdec/target/lora_ops.py",
335
+ "specdec/target/qwen35.py",
336
+ "specdec/target/readout.py",
337
+ "specdec/target/static.py",
338
+ "specdec/target/triton_ops.py"
339
+ ],
340
+ "torch": "2.13.0+cu130",
341
+ "torch_cuda": "13.0",
342
+ "transformers": "5.12.1",
343
+ "sglang": "0.5.18",
344
+ "gpu": "NVIDIA H100 80GB HBM3"
345
+ },
346
+ "dense_sources_sha1": "03ba376b04f7686f84af14b96362f10f03e83351"
347
+ },
348
+ "checkpoint_sha1": "cc4c99cbfaaa0a869f5b219774db513c8fe917d2",
349
+ "mask_mode": "read",
350
+ "ar_checked": true,
351
+ "mask_options": {
352
+ "mask_lean": true,
353
+ "lora_fused_a": true
354
+ },
355
+ "mask_attn_kernel": "branch_stream",
356
+ "late_commit": {
357
+ "lc_readout": "cand",
358
+ "late_commit_sha1": "7766806b2bd5671e7e5038b76b03106b641aa6e4"
359
+ },
360
+ "late_merge": {
361
+ "variant": "merged_triton",
362
+ "late_merge_sha1": "be3bb036283ed581d8cb0e33b64be1e926880dd0"
363
+ }
364
+ },
365
+ "resumed": 0
366
+ },
367
+ "definitions": {
368
+ "tau": "tokens emitted by non-recovery full passes / their count, misses included (specdec/fused/branches.py); no reference counterpart",
369
+ "miss_rate": "misses / non-recovery full passes",
370
+ "tokens_per_pass": "all output tokens / all passes (start and recovery included)",
371
+ "rows_per_pass": "mean over all passes of V + b x 15",
372
+ "tau_ci95": "95 % prompt-level bootstrap (percentile) of pooled tau",
373
+ "paired": "tau(b) - tau(b_max), both arms resampled with the same prompts",
374
+ "e2e": "wall clock against the harness's own AR on the same prompts (e2e_speedup): speedup_decode after the prompt, speedup_e2e with the shared prefill; engine: this harness, not SGLang",
375
+ "pooled_clean": "the same summary without the prompts found in the training split (contamination.excluded_from_clean); present only when some were excluded",
376
+ "tau_pooled": "reference definition; ours = tau_pooled_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)",
377
+ "tau_raw_row": "reference definition; ours = tau_raw_row_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)",
378
+ "tau_prefill_corrected_row": "reference definition; ours = tau_prefill_corrected_row_ref (ours: output tokens = len(tokens); verify passes = all passes but the first (the V = 1 pass on the last prompt token emits the first token, as the reference's prefill does); recovery passes count as verify passes. Trajectories differ: harness greedy vs SGLang server greedy.)"
379
+ },
380
+ "results": {
381
+ "b1": {
382
+ "pooled": {
383
+ "prompts": 20,
384
+ "tokens": 6702,
385
+ "passes": 2327,
386
+ "fused_passes": 2307,
387
+ "recovery_passes": 0,
388
+ "misses": 0,
389
+ "salvage_passes": 0,
390
+ "salvage_misses": 0,
391
+ "tokens_per_pass": 2.8801031370863774,
392
+ "tau": 2.896402254009536,
393
+ "miss_rate": 0.0,
394
+ "rows_per_pass": 30.871078642028362,
395
+ "rows_per_fused_pass": 31.0,
396
+ "tau_pooled_ref": 2.905071521456437,
397
+ "tau_raw_row_ref": 3.042636491893952,
398
+ "tau_prefill_corrected_row_ref": 3.0306927650103224,
399
+ "accepted_hist": {
400
+ "0": 744,
401
+ "1": 516,
402
+ "2": 377,
403
+ "3": 252,
404
+ "4": 149,
405
+ "5": 102,
406
+ "6": 61,
407
+ "7": 44,
408
+ "8": 26,
409
+ "9": 17,
410
+ "10": 6,
411
+ "11": 5,
412
+ "12": 4,
413
+ "13": 1,
414
+ "14": 1,
415
+ "15": 2
416
+ },
417
+ "committed_hist": {
418
+ "0": 739,
419
+ "1": 512,
420
+ "2": 371,
421
+ "3": 250,
422
+ "4": 149,
423
+ "5": 102,
424
+ "6": 59,
425
+ "7": 44,
426
+ "8": 26,
427
+ "9": 17,
428
+ "10": 6,
429
+ "11": 5,
430
+ "12": 4,
431
+ "13": 1,
432
+ "15": 2
433
+ },
434
+ "overshoot": 12,
435
+ "decode_ms": 153364.47620391846,
436
+ "decode_ms_all": 153364.47620391846,
437
+ "capture_rounds": 0,
438
+ "capture_ms": 0,
439
+ "pass_ms_by_kind": {
440
+ "start": {
441
+ "n": 20,
442
+ "mean_ms": 81.1545841217041,
443
+ "median_ms": 62.37579154968262,
444
+ "p10_ms": 49.59553413391113,
445
+ "p90_ms": 84.81053848266602
446
+ },
447
+ "fused": {
448
+ "n": 2307,
449
+ "mean_ms": 65.7743322589876,
450
+ "median_ms": 64.64998626708984,
451
+ "p10_ms": 49.90601577758789,
452
+ "p90_ms": 86.12802581787109
453
+ }
454
+ },
455
+ "prefill_ms": 3060.551239013672,
456
+ "accounting_ok": true,
457
+ "exact": 20,
458
+ "tau_ci95": [
459
+ 2.6706615532118887,
460
+ 3.163926499032882
461
+ ],
462
+ "tau_pooled_ref_ci95": [
463
+ 2.680249280920422,
464
+ 3.1744704570791527
465
+ ]
466
+ },
467
+ "benches": {
468
+ "gsm8k": {
469
+ "stats": {
470
+ "prompts": 4,
471
+ "tokens": 987,
472
+ "passes": 302,
473
+ "fused_passes": 298,
474
+ "recovery_passes": 0,
475
+ "misses": 0,
476
+ "salvage_passes": 0,
477
+ "salvage_misses": 0,
478
+ "tokens_per_pass": 3.2682119205298013,
479
+ "tau": 3.2986577181208054,
480
+ "miss_rate": 0.0,
481
+ "rows_per_pass": 30.801324503311257,
482
+ "rows_per_fused_pass": 31.0,
483
+ "tau_pooled_ref": 3.3120805369127515,
484
+ "tau_raw_row_ref": 3.3273406163247117,
485
+ "tau_prefill_corrected_row_ref": 3.3136891521642244,
486
+ "accepted_hist": {
487
+ "0": 96,
488
+ "1": 51,
489
+ "2": 48,
490
+ "3": 36,
491
+ "4": 21,
492
+ "5": 11,
493
+ "6": 8,
494
+ "7": 9,
495
+ "8": 4,
496
+ "9": 5,
497
+ "10": 3,
498
+ "11": 3,
499
+ "12": 1,
500
+ "13": 1,
501
+ "15": 1
502
+ },
503
+ "committed_hist": {
504
+ "0": 96,
505
+ "1": 50,
506
+ "2": 46,
507
+ "3": 36,
508
+ "4": 21,
509
+ "5": 11,
510
+ "6": 7,
511
+ "7": 9,
512
+ "8": 4,
513
+ "9": 5,
514
+ "10": 3,
515
+ "11": 3,
516
+ "12": 1,
517
+ "13": 1,
518
+ "15": 1
519
+ },
520
+ "overshoot": 0,
521
+ "decode_ms": 25879.782501220703,
522
+ "decode_ms_all": 25879.782501220703,
523
+ "capture_rounds": 0,
524
+ "capture_ms": 0,
525
+ "pass_ms_by_kind": {
526
+ "start": {
527
+ "n": 4,
528
+ "mean_ms": 168.88980674743652,
529
+ "median_ms": 85.48255920410156,
530
+ "p10_ms": 74.7245719909668,
531
+ "p90_ms": 329.7808395385743
532
+ },
533
+ "fused": {
534
+ "n": 298,
535
+ "mean_ms": 84.57793045043945,
536
+ "median_ms": 85.98331451416016,
537
+ "p10_ms": 73.8765480041504,
538
+ "p90_ms": 90.82000732421875
539
+ }
540
+ },
541
+ "prefill_ms": 1387.14697265625,
542
+ "accounting_ok": true,
543
+ "exact": 4,
544
+ "tau_ci95": [
545
+ 3.053156146179402,
546
+ 3.591973244147157
547
+ ],
548
+ "tau_pooled_ref_ci95": [
549
+ 3.0664451827242525,
550
+ 3.6053511705685617
551
+ ]
552
+ },
553
+ "prompts": [
554
+ {
555
+ "pid": "gsm8k:23",
556
+ "tokens": 203,
557
+ "passes": 62,
558
+ "identical": true,
559
+ "first_divergence": null,
560
+ "method_tally": [
561
+ 202,
562
+ 61
563
+ ],
564
+ "reference_tally": [
565
+ 203,
566
+ 61
567
+ ],
568
+ "misses": 0,
569
+ "finish_reason": "stop",
570
+ "decode_ms": 5695.499801635742,
571
+ "capture_ms": 0
572
+ },
573
+ {
574
+ "pid": "gsm8k:51",
575
+ "tokens": 267,
576
+ "passes": 72,
577
+ "identical": true,
578
+ "first_divergence": null,
579
+ "method_tally": [
580
+ 266,
581
+ 71
582
+ ],
583
+ "reference_tally": [
584
+ 267,
585
+ 71
586
+ ],
587
+ "misses": 0,
588
+ "finish_reason": "stop",
589
+ "decode_ms": 5881.466926574707,
590
+ "capture_ms": 0
591
+ },
592
+ {
593
+ "pid": "gsm8k:61",
594
+ "tokens": 240,
595
+ "passes": 81,
596
+ "identical": true,
597
+ "first_divergence": null,
598
+ "method_tally": [
599
+ 239,
600
+ 80
601
+ ],
602
+ "reference_tally": [
603
+ 240,
604
+ 80
605
+ ],
606
+ "misses": 0,
607
+ "finish_reason": "stop",
608
+ "decode_ms": 6895.553604125977,
609
+ "capture_ms": 0
610
+ },
611
+ {
612
+ "pid": "gsm8k:90",
613
+ "tokens": 277,
614
+ "passes": 87,
615
+ "identical": true,
616
+ "first_divergence": null,
617
+ "method_tally": [
618
+ 276,
619
+ 86
620
+ ],
621
+ "reference_tally": [
622
+ 277,
623
+ 86
624
+ ],
625
+ "misses": 0,
626
+ "finish_reason": "stop",
627
+ "decode_ms": 7407.262168884277,
628
+ "capture_ms": 0
629
+ }
630
+ ],
631
+ "e2e": {
632
+ "prompts": 4,
633
+ "fused_ms_per_token": 26.220650963749446,
634
+ "ar_ms_per_token": 51.62969007559702,
635
+ "speedup_decode": 1.969046845823014,
636
+ "speedup_e2e": 1.9252367087963405
637
+ }
638
+ },
639
+ "math500": {
640
+ "stats": {
641
+ "prompts": 4,
642
+ "tokens": 1575,
643
+ "passes": 416,
644
+ "fused_passes": 412,
645
+ "recovery_passes": 0,
646
+ "misses": 0,
647
+ "salvage_passes": 0,
648
+ "salvage_misses": 0,
649
+ "tokens_per_pass": 3.7860576923076925,
650
+ "tau": 3.813106796116505,
651
+ "miss_rate": 0.0,
652
+ "rows_per_pass": 30.85576923076923,
653
+ "rows_per_fused_pass": 31.0,
654
+ "tau_pooled_ref": 3.8228155339805827,
655
+ "tau_raw_row_ref": 3.9567695994927803,
656
+ "tau_prefill_corrected_row_ref": 3.9445688876015206,
657
+ "accepted_hist": {
658
+ "0": 89,
659
+ "1": 80,
660
+ "2": 55,
661
+ "3": 52,
662
+ "4": 37,
663
+ "5": 28,
664
+ "6": 24,
665
+ "7": 19,
666
+ "8": 14,
667
+ "9": 7,
668
+ "10": 1,
669
+ "11": 2,
670
+ "12": 3,
671
+ "14": 1
672
+ },
673
+ "committed_hist": {
674
+ "0": 89,
675
+ "1": 79,
676
+ "2": 55,
677
+ "3": 51,
678
+ "4": 37,
679
+ "5": 28,
680
+ "6": 23,
681
+ "7": 19,
682
+ "8": 14,
683
+ "9": 7,
684
+ "10": 1,
685
+ "11": 2,
686
+ "12": 3
687
+ },
688
+ "overshoot": 9,
689
+ "decode_ms": 28854.95544052124,
690
+ "decode_ms_all": 28854.95544052124,
691
+ "capture_rounds": 0,
692
+ "capture_ms": 0,
693
+ "pass_ms_by_kind": {
694
+ "start": {
695
+ "n": 4,
696
+ "mean_ms": 62.258169174194336,
697
+ "median_ms": 57.137441635131836,
698
+ "p10_ms": 50.812544631958005,
699
+ "p90_ms": 77.80037574768068
700
+ },
701
+ "fused": {
702
+ "n": 412,
703
+ "mean_ms": 69.43185136850597,
704
+ "median_ms": 66.85449600219727,
705
+ "p10_ms": 50.272729110717776,
706
+ "p90_ms": 87.90760345458985
707
+ }
708
+ },
709
+ "prefill_ms": 465.4977340698242,
710
+ "accounting_ok": true,
711
+ "exact": 4,
712
+ "tau_ci95": [
713
+ 3.4249547920434,
714
+ 4.204545454545454
715
+ ],
716
+ "tau_pooled_ref_ci95": [
717
+ 3.4321880650994574,
718
+ 4.222727272727273
719
+ ]
720
+ },
721
+ "prompts": [
722
+ {
723
+ "pid": "math500:14",
724
+ "tokens": 512,
725
+ "passes": 155,
726
+ "identical": true,
727
+ "first_divergence": null,
728
+ "method_tally": [
729
+ 511,
730
+ 154
731
+ ],
732
+ "reference_tally": [
733
+ 512,
734
+ 154
735
+ ],
736
+ "misses": 0,
737
+ "finish_reason": "length",
738
+ "decode_ms": 10941.528476715088,
739
+ "capture_ms": 0
740
+ },
741
+ {
742
+ "pid": "math500:34",
743
+ "tokens": 512,
744
+ "passes": 125,
745
+ "identical": true,
746
+ "first_divergence": null,
747
+ "method_tally": [
748
+ 511,
749
+ 124
750
+ ],
751
+ "reference_tally": [
752
+ 512,
753
+ 124
754
+ ],
755
+ "misses": 0,
756
+ "finish_reason": "length",
757
+ "decode_ms": 9684.818389892578,
758
+ "capture_ms": 0
759
+ },
760
+ {
761
+ "pid": "math500:53",
762
+ "tokens": 362,
763
+ "passes": 92,
764
+ "identical": true,
765
+ "first_divergence": null,
766
+ "method_tally": [
767
+ 361,
768
+ 91
769
+ ],
770
+ "reference_tally": [
771
+ 362,
772
+ 91
773
+ ],
774
+ "misses": 0,
775
+ "finish_reason": "stop",
776
+ "decode_ms": 5663.593776702881,
777
+ "capture_ms": 0
778
+ },
779
+ {
780
+ "pid": "math500:79",
781
+ "tokens": 189,
782
+ "passes": 44,
783
+ "identical": true,
784
+ "first_divergence": null,
785
+ "method_tally": [
786
+ 188,
787
+ 43
788
+ ],
789
+ "reference_tally": [
790
+ 189,
791
+ 43
792
+ ],
793
+ "misses": 0,
794
+ "finish_reason": "stop",
795
+ "decode_ms": 2565.0147972106934,
796
+ "capture_ms": 0
797
+ }
798
+ ],
799
+ "e2e": {
800
+ "prompts": 4,
801
+ "fused_ms_per_token": 18.320606628902375,
802
+ "ar_ms_per_token": 50.54916671026321,
803
+ "speedup_decode": 2.7591426274344775,
804
+ "speedup_e2e": 2.733625112564697
805
+ }
806
+ },
807
+ "humaneval": {
808
+ "stats": {
809
+ "prompts": 4,
810
+ "tokens": 1615,
811
+ "passes": 613,
812
+ "fused_passes": 609,
813
+ "recovery_passes": 0,
814
+ "misses": 0,
815
+ "salvage_passes": 0,
816
+ "salvage_misses": 0,
817
+ "tokens_per_pass": 2.634584013050571,
818
+ "tau": 2.645320197044335,
819
+ "miss_rate": 0.0,
820
+ "rows_per_pass": 30.902120717781404,
821
+ "rows_per_fused_pass": 31.0,
822
+ "tau_pooled_ref": 2.651888341543514,
823
+ "tau_raw_row_ref": 2.6607403693316214,
824
+ "tau_prefill_corrected_row_ref": 2.653624249097236,
825
+ "accepted_hist": {
826
+ "0": 194,
827
+ "1": 145,
828
+ "2": 118,
829
+ "3": 68,
830
+ "4": 42,
831
+ "5": 21,
832
+ "6": 10,
833
+ "7": 6,
834
+ "8": 2,
835
+ "9": 2,
836
+ "10": 1
837
+ },
838
+ "committed_hist": {
839
+ "0": 193,
840
+ "1": 144,
841
+ "2": 116,
842
+ "3": 68,
843
+ "4": 42,
844
+ "5": 21,
845
+ "6": 10,
846
+ "7": 6,
847
+ "8": 2,
848
+ "9": 2,
849
+ "10": 1
850
+ },
851
+ "overshoot": 2,
852
+ "decode_ms": 37648.55478668213,
853
+ "decode_ms_all": 37648.55478668213,
854
+ "capture_rounds": 0,
855
+ "capture_ms": 0,
856
+ "pass_ms_by_kind": {
857
+ "start": {
858
+ "n": 4,
859
+ "mean_ms": 59.87827205657959,
860
+ "median_ms": 62.37579154968262,
861
+ "p10_ms": 53.484596252441406,
862
+ "p90_ms": 64.27393226623535
863
+ },
864
+ "fused": {
865
+ "n": 609,
866
+ "mean_ms": 61.42699786281742,
867
+ "median_ms": 62.974910736083984,
868
+ "p10_ms": 49.818212890625,
869
+ "p90_ms": 71.24920806884765
870
+ }
871
+ },
872
+ "prefill_ms": 396.35987854003906,
873
+ "accounting_ok": true,
874
+ "exact": 4,
875
+ "tau_ci95": [
876
+ 2.4793956043956045,
877
+ 2.8505392912172574
878
+ ],
879
+ "tau_pooled_ref_ci95": [
880
+ 2.48489010989011,
881
+ 2.856702619414484
882
+ ]
883
+ },
884
+ "prompts": [
885
+ {
886
+ "pid": "humaneval:4",
887
+ "tokens": 318,
888
+ "passes": 125,
889
+ "identical": true,
890
+ "first_divergence": null,
891
+ "method_tally": [
892
+ 317,
893
+ 124
894
+ ],
895
+ "reference_tally": [
896
+ 318,
897
+ 124
898
+ ],
899
+ "misses": 0,
900
+ "finish_reason": "stop",
901
+ "decode_ms": 7438.806297302246,
902
+ "capture_ms": 0
903
+ },
904
+ {
905
+ "pid": "humaneval:19",
906
+ "tokens": 512,
907
+ "passes": 210,
908
+ "identical": true,
909
+ "first_divergence": null,
910
+ "method_tally": [
911
+ 511,
912
+ 209
913
+ ],
914
+ "reference_tally": [
915
+ 512,
916
+ 209
917
+ ],
918
+ "misses": 0,
919
+ "finish_reason": "length",
920
+ "decode_ms": 13521.155044555664,
921
+ "capture_ms": 0
922
+ },
923
+ {
924
+ "pid": "humaneval:58",
925
+ "tokens": 273,
926
+ "passes": 102,
927
+ "identical": true,
928
+ "first_divergence": null,
929
+ "method_tally": [
930
+ 272,
931
+ 101
932
+ ],
933
+ "reference_tally": [
934
+ 273,
935
+ 101
936
+ ],
937
+ "misses": 0,
938
+ "finish_reason": "stop",
939
+ "decode_ms": 6169.830215454102,
940
+ "capture_ms": 0
941
+ },
942
+ {
943
+ "pid": "humaneval:62",
944
+ "tokens": 512,
945
+ "passes": 176,
946
+ "identical": true,
947
+ "first_divergence": null,
948
+ "method_tally": [
949
+ 511,
950
+ 175
951
+ ],
952
+ "reference_tally": [
953
+ 512,
954
+ 175
955
+ ],
956
+ "misses": 0,
957
+ "finish_reason": "length",
958
+ "decode_ms": 10518.763229370117,
959
+ "capture_ms": 0
960
+ }
961
+ ],
962
+ "e2e": {
963
+ "prompts": 4,
964
+ "fused_ms_per_token": 23.31179862952454,
965
+ "ar_ms_per_token": 51.77458472000925,
966
+ "speedup_decode": 2.220960533454352,
967
+ "speedup_e2e": 2.208343774801641
968
+ }
969
+ },
970
+ "mbpp": {
971
+ "stats": {
972
+ "prompts": 4,
973
+ "tokens": 1333,
974
+ "passes": 557,
975
+ "fused_passes": 553,
976
+ "recovery_passes": 0,
977
+ "misses": 0,
978
+ "salvage_passes": 0,
979
+ "salvage_misses": 0,
980
+ "tokens_per_pass": 2.393177737881508,
981
+ "tau": 2.403254972875226,
982
+ "miss_rate": 0.0,
983
+ "rows_per_pass": 30.892280071813286,
984
+ "rows_per_fused_pass": 31.0,
985
+ "tau_pooled_ref": 2.4104882459312837,
986
+ "tau_raw_row_ref": 2.4846920289855072,
987
+ "tau_prefill_corrected_row_ref": 2.476083721532091,
988
+ "accepted_hist": {
989
+ "0": 216,
990
+ "1": 144,
991
+ "2": 82,
992
+ "3": 49,
993
+ "4": 24,
994
+ "5": 18,
995
+ "6": 10,
996
+ "7": 5,
997
+ "8": 4,
998
+ "9": 1
999
+ },
1000
+ "committed_hist": {
1001
+ "0": 214,
1002
+ "1": 144,
1003
+ "2": 81,
1004
+ "3": 48,
1005
+ "4": 24,
1006
+ "5": 18,
1007
+ "6": 10,
1008
+ "7": 5,
1009
+ "8": 4,
1010
+ "9": 1
1011
+ },
1012
+ "overshoot": 1,
1013
+ "decode_ms": 33849.65217590332,
1014
+ "decode_ms_all": 33849.65217590332,
1015
+ "capture_rounds": 0,
1016
+ "capture_ms": 0,
1017
+ "pass_ms_by_kind": {
1018
+ "start": {
1019
+ "n": 4,
1020
+ "mean_ms": 54.99039173126221,
1021
+ "median_ms": 52.03236770629883,
1022
+ "p10_ms": 49.51867561340332,
1023
+ "p90_ms": 62.8285270690918
1024
+ },
1025
+ "fused": {
1026
+ "n": 553,
1027
+ "mean_ms": 60.8131837413712,
1028
+ "median_ms": 62.1063346862793,
1029
+ "p10_ms": 49.83096923828125,
1030
+ "p90_ms": 69.77151184082031
1031
+ }
1032
+ },
1033
+ "prefill_ms": 393.0157165527344,
1034
+ "accounting_ok": true,
1035
+ "exact": 4,
1036
+ "tau_ci95": [
1037
+ 2.271978021978022,
1038
+ 2.6560846560846563
1039
+ ],
1040
+ "tau_pooled_ref_ci95": [
1041
+ 2.2774725274725274,
1042
+ 2.6666666666666665
1043
+ ]
1044
+ },
1045
+ "prompts": [
1046
+ {
1047
+ "pid": "mbpp:3",
1048
+ "tokens": 317,
1049
+ "passes": 141,
1050
+ "identical": true,
1051
+ "first_divergence": null,
1052
+ "method_tally": [
1053
+ 316,
1054
+ 140
1055
+ ],
1056
+ "reference_tally": [
1057
+ 317,
1058
+ 140
1059
+ ],
1060
+ "misses": 0,
1061
+ "finish_reason": "stop",
1062
+ "decode_ms": 8291.162788391113,
1063
+ "capture_ms": 0
1064
+ },
1065
+ {
1066
+ "pid": "mbpp:77",
1067
+ "tokens": 512,
1068
+ "passes": 225,
1069
+ "identical": true,
1070
+ "first_divergence": null,
1071
+ "method_tally": [
1072
+ 511,
1073
+ 224
1074
+ ],
1075
+ "reference_tally": [
1076
+ 512,
1077
+ 224
1078
+ ],
1079
+ "misses": 0,
1080
+ "finish_reason": "length",
1081
+ "decode_ms": 13552.30948638916,
1082
+ "capture_ms": 0
1083
+ },
1084
+ {
1085
+ "pid": "mbpp:87",
1086
+ "tokens": 311,
1087
+ "passes": 121,
1088
+ "identical": true,
1089
+ "first_divergence": null,
1090
+ "method_tally": [
1091
+ 310,
1092
+ 120
1093
+ ],
1094
+ "reference_tally": [
1095
+ 311,
1096
+ 120
1097
+ ],
1098
+ "misses": 0,
1099
+ "finish_reason": "stop",
1100
+ "decode_ms": 7684.127578735352,
1101
+ "capture_ms": 0
1102
+ },
1103
+ {
1104
+ "pid": "mbpp:92",
1105
+ "tokens": 193,
1106
+ "passes": 70,
1107
+ "identical": true,
1108
+ "first_divergence": null,
1109
+ "method_tally": [
1110
+ 192,
1111
+ 69
1112
+ ],
1113
+ "reference_tally": [
1114
+ 193,
1115
+ 69
1116
+ ],
1117
+ "misses": 0,
1118
+ "finish_reason": "stop",
1119
+ "decode_ms": 4322.052322387695,
1120
+ "capture_ms": 0
1121
+ }
1122
+ ],
1123
+ "e2e": {
1124
+ "prompts": 4,
1125
+ "fused_ms_per_token": 25.393587528809693,
1126
+ "ar_ms_per_token": 50.21653410159638,
1127
+ "speedup_decode": 1.9775281473964401,
1128
+ "speedup_e2e": 1.9663086966659638
1129
+ }
1130
+ },
1131
+ "mt-bench": {
1132
+ "stats": {
1133
+ "prompts": 4,
1134
+ "tokens": 1192,
1135
+ "passes": 439,
1136
+ "fused_passes": 435,
1137
+ "recovery_passes": 0,
1138
+ "misses": 0,
1139
+ "salvage_passes": 0,
1140
+ "salvage_misses": 0,
1141
+ "tokens_per_pass": 2.715261958997722,
1142
+ "tau": 2.731034482758621,
1143
+ "miss_rate": 0.0,
1144
+ "rows_per_pass": 30.863325740318906,
1145
+ "rows_per_fused_pass": 31.0,
1146
+ "tau_pooled_ref": 2.740229885057471,
1147
+ "tau_raw_row_ref": 2.783639845335141,
1148
+ "tau_prefill_corrected_row_ref": 2.76549781465654,
1149
+ "accepted_hist": {
1150
+ "0": 149,
1151
+ "1": 96,
1152
+ "2": 74,
1153
+ "3": 47,
1154
+ "4": 25,
1155
+ "5": 24,
1156
+ "6": 9,
1157
+ "7": 5,
1158
+ "8": 2,
1159
+ "9": 2,
1160
+ "10": 1,
1161
+ "15": 1
1162
+ },
1163
+ "committed_hist": {
1164
+ "0": 147,
1165
+ "1": 95,
1166
+ "2": 73,
1167
+ "3": 47,
1168
+ "4": 25,
1169
+ "5": 24,
1170
+ "6": 9,
1171
+ "7": 5,
1172
+ "8": 2,
1173
+ "9": 2,
1174
+ "10": 1,
1175
+ "15": 1
1176
+ },
1177
+ "overshoot": 0,
1178
+ "decode_ms": 27131.531299591064,
1179
+ "decode_ms_all": 27131.531299591064,
1180
+ "capture_rounds": 0,
1181
+ "capture_ms": 0,
1182
+ "pass_ms_by_kind": {
1183
+ "start": {
1184
+ "n": 4,
1185
+ "mean_ms": 59.75628089904785,
1186
+ "median_ms": 57.05011177062988,
1187
+ "p10_ms": 47.77764472961426,
1188
+ "p90_ms": 73.89985237121583
1189
+ },
1190
+ "fused": {
1191
+ "n": 435,
1192
+ "mean_ms": 61.821853278149135,
1193
+ "median_ms": 62.877471923828125,
1194
+ "p10_ms": 49.637918853759764,
1195
+ "p90_ms": 75.07242736816406
1196
+ }
1197
+ },
1198
+ "prefill_ms": 418.5309371948242,
1199
+ "accounting_ok": true,
1200
+ "exact": 4,
1201
+ "tau_ci95": [
1202
+ 2.3893333333333335,
1203
+ 2.9574468085106385
1204
+ ],
1205
+ "tau_pooled_ref_ci95": [
1206
+ 2.4,
1207
+ 2.9680851063829787
1208
+ ]
1209
+ },
1210
+ "prompts": [
1211
+ {
1212
+ "pid": "mt-bench:21",
1213
+ "tokens": 133,
1214
+ "passes": 65,
1215
+ "identical": true,
1216
+ "first_divergence": null,
1217
+ "method_tally": [
1218
+ 132,
1219
+ 64
1220
+ ],
1221
+ "reference_tally": [
1222
+ 133,
1223
+ 64
1224
+ ],
1225
+ "misses": 0,
1226
+ "finish_reason": "stop",
1227
+ "decode_ms": 4415.579372406006,
1228
+ "capture_ms": 0
1229
+ },
1230
+ {
1231
+ "pid": "mt-bench:33",
1232
+ "tokens": 483,
1233
+ "passes": 167,
1234
+ "identical": true,
1235
+ "first_divergence": null,
1236
+ "method_tally": [
1237
+ 482,
1238
+ 166
1239
+ ],
1240
+ "reference_tally": [
1241
+ 483,
1242
+ 166
1243
+ ],
1244
+ "misses": 0,
1245
+ "finish_reason": "stop",
1246
+ "decode_ms": 10584.269901275635,
1247
+ "capture_ms": 0
1248
+ },
1249
+ {
1250
+ "pid": "mt-bench:46",
1251
+ "tokens": 501,
1252
+ "passes": 184,
1253
+ "identical": true,
1254
+ "first_divergence": null,
1255
+ "method_tally": [
1256
+ 500,
1257
+ 183
1258
+ ],
1259
+ "reference_tally": [
1260
+ 501,
1261
+ 183
1262
+ ],
1263
+ "misses": 0,
1264
+ "finish_reason": "stop",
1265
+ "decode_ms": 10808.679847717285,
1266
+ "capture_ms": 0
1267
+ },
1268
+ {
1269
+ "pid": "mt-bench:59",
1270
+ "tokens": 75,
1271
+ "passes": 23,
1272
+ "identical": true,
1273
+ "first_divergence": null,
1274
+ "method_tally": [
1275
+ 74,
1276
+ 22
1277
+ ],
1278
+ "reference_tally": [
1279
+ 75,
1280
+ 22
1281
+ ],
1282
+ "misses": 0,
1283
+ "finish_reason": "stop",
1284
+ "decode_ms": 1323.0021781921387,
1285
+ "capture_ms": 0
1286
+ }
1287
+ ],
1288
+ "e2e": {
1289
+ "prompts": 4,
1290
+ "fused_ms_per_token": 22.7613517614019,
1291
+ "ar_ms_per_token": 51.20760074397862,
1292
+ "speedup_decode": 2.2497609667812046,
1293
+ "speedup_e2e": 2.2307750340380825
1294
+ }
1295
+ }
1296
+ },
1297
+ "e2e_pooled": {
1298
+ "prompts": 20,
1299
+ "fused_ms_per_token": 22.883389466415764,
1300
+ "ar_ms_per_token": 51.05453536859807,
1301
+ "speedup_decode": 2.2310740042914965,
1302
+ "speedup_e2e": 2.208577107358479
1303
+ }
1304
+ }
1305
+ },
1306
+ "paired_vs_b_max": {},
1307
+ "mismatches_vs_ar": {},
1308
+ "reference_comparison": [
1309
+ {
1310
+ "note": "--dense-target: the reference numbers are SGLang runs of the 27B, not of this target"
1311
+ }
1312
+ ]
1313
+ }
qwen3_4b/eval/lm_LCa3_v11/lm_triton_cand_eager/q1_rows.jsonl ADDED
The diff for this file is too large to render. See raw diff
 
qwen3_4b/eval/lm_LCa3_v11/lm_triton_cand_eager/q1_run.json ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "ar": {
3
+ "target": "dense:models--Qwen--Qwen3-4B@1cfa9a7208912126459214e8b04321603b3df60c",
4
+ "dtype": "bfloat16",
5
+ "backend": "none",
6
+ "batch_invariant": true,
7
+ "fp32_A_log": false,
8
+ "max_new": 512,
9
+ "eos": [
10
+ 151645,
11
+ 151643
12
+ ],
13
+ "decode": {
14
+ "path": "dynamic-eager",
15
+ "attention": "triton",
16
+ "fused_ops": false,
17
+ "static_max_len": null
18
+ },
19
+ "harness": {
20
+ "source_sha1": "04e7703c37603bda2b76c602e94506442edd314a",
21
+ "source_files": [
22
+ "specdec/adapters/gated_lora.py",
23
+ "specdec/fused/branches.py",
24
+ "specdec/fused/hint.py",
25
+ "specdec/fused/late_bonus.py",
26
+ "specdec/fused/layout.py",
27
+ "specdec/fused/policy.py",
28
+ "specdec/fused/transfer.py",
29
+ "specdec/memory/gdn_state.py",
30
+ "specdec/paths/spec.py",
31
+ "specdec/run/decode.py",
32
+ "specdec/target/__init__.py",
33
+ "specdec/target/branch_memory.py",
34
+ "specdec/target/branch_triton.py",
35
+ "specdec/target/cache.py",
36
+ "specdec/target/kernels.py",
37
+ "specdec/target/lora_ops.py",
38
+ "specdec/target/qwen35.py",
39
+ "specdec/target/readout.py",
40
+ "specdec/target/static.py",
41
+ "specdec/target/triton_ops.py"
42
+ ],
43
+ "torch": "2.13.0+cu130",
44
+ "torch_cuda": "13.0",
45
+ "transformers": "5.12.1",
46
+ "sglang": "0.5.18",
47
+ "gpu": "NVIDIA H100 80GB HBM3"
48
+ },
49
+ "dense_sources_sha1": "03ba376b04f7686f84af14b96362f10f03e83351"
50
+ },
51
+ "checkpoint_sha1": "cc4c99cbfaaa0a869f5b219774db513c8fe917d2",
52
+ "mask_mode": "read",
53
+ "ar_checked": true,
54
+ "mask_options": {
55
+ "mask_lean": true,
56
+ "lora_fused_a": true
57
+ },
58
+ "mask_attn_kernel": "branch_stream",
59
+ "late_commit": {
60
+ "lc_readout": "cand",
61
+ "late_commit_sha1": "7766806b2bd5671e7e5038b76b03106b641aa6e4"
62
+ },
63
+ "late_merge": {
64
+ "variant": "merged_triton",
65
+ "late_merge_sha1": "be3bb036283ed581d8cb0e33b64be1e926880dd0"
66
+ }
67
+ }