x4rand: serving results + repo README
Browse files- README.md +1 -0
- dflash2_27b_traj_20260915/main-800k/x4rand/epoch4/serving/acceptance_summary.json +26 -0
- dflash2_27b_traj_20260915/main-800k/x4rand/epoch4/serving/acceptance_trace.jsonl +0 -0
- dflash2_27b_traj_20260915/main-800k/x4rand/epoch4/serving/client.log +46 -0
- dflash2_27b_traj_20260915/main-800k/x4rand/epoch4/serving/server.log +0 -0
README.md
CHANGED
|
@@ -34,6 +34,7 @@ greedy, concurrency 1, 500/500 completed. Same protocol as the dense and K4-jitt
|
|
| 34 |
| x4rand | 1 | 73,670 | 5.1485 | 5.2314 | +3.44% (dense 5.0575) |
|
| 35 |
| x4rand | 2 | 147,335 | 5.2056 | 5.2922 | +4.79% (dense 5.0502) |
|
| 36 |
| x4rand | 3 | 221,001 | 5.2202 | 5.3261 | +5.46% vs dense E2 5.0502 (dense stopped after E2) |
|
|
|
|
| 37 |
|
| 38 |
Reference arms (model-only, no optimizer) are in `ryan-0608/MoS-Aurora-Experiment-Archive` under
|
| 39 |
`dflash2_27b_traj_20260915/main-800k/`: `dense/epoch1`, `dense/epoch2`, `mos/epoch1` (K4 shared+1% jitter,
|
|
|
|
| 34 |
| x4rand | 1 | 73,670 | 5.1485 | 5.2314 | +3.44% (dense 5.0575) |
|
| 35 |
| x4rand | 2 | 147,335 | 5.2056 | 5.2922 | +4.79% (dense 5.0502) |
|
| 36 |
| x4rand | 3 | 221,001 | 5.2202 | 5.3261 | +5.46% vs dense E2 5.0502 (dense stopped after E2) |
|
| 37 |
+
| x4rand | 4 | 294,682 | 5.2201 | 5.3113 | +5.17% vs dense E2 5.0502; best epoch = 3 |
|
| 38 |
|
| 39 |
Reference arms (model-only, no optimizer) are in `ryan-0608/MoS-Aurora-Experiment-Archive` under
|
| 40 |
`dflash2_27b_traj_20260915/main-800k/`: `dense/epoch1`, `dense/epoch2`, `mos/epoch1` (K4 shared+1% jitter,
|
dflash2_27b_traj_20260915/main-800k/x4rand/epoch4/serving/acceptance_summary.json
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"concurrency": 1,
|
| 3 |
+
"draft_checkpoint": "/mnt/weka/mos/experiments/dflash2/qwen3.8-27b-traj/runs-b16-800k-traj-x4rand/mos/checkpoints/3/sglang_export",
|
| 4 |
+
"eval": "/workspace/weka-mirror/mos/data/aurora-100k/eval_balanced_500_seed20260719_thinknorm_v1.jsonl",
|
| 5 |
+
"finished_at_utc": "2026-09-25T09:36:39Z",
|
| 6 |
+
"latency_p50_seconds": 0.5017778598703444,
|
| 7 |
+
"latency_p95_seconds": 0.8580532639753073,
|
| 8 |
+
"max_tokens": 128,
|
| 9 |
+
"mean_of_per_request_ratios": 6.334351040787312,
|
| 10 |
+
"measured_wallclock_seconds": 259.594325958984,
|
| 11 |
+
"n_completed": 500,
|
| 12 |
+
"n_prompts": 500,
|
| 13 |
+
"n_successful": 500,
|
| 14 |
+
"output_token_throughput_per_second": 243.76110597270184,
|
| 15 |
+
"pooled_accept_length": 5.311314420010072,
|
| 16 |
+
"server_metric_accept_length": 4.125,
|
| 17 |
+
"speculative_algorithm": "DFLASH",
|
| 18 |
+
"speculative_eagle_topk": 0,
|
| 19 |
+
"speculative_num_draft_tokens": 16,
|
| 20 |
+
"speculative_num_steps": 0,
|
| 21 |
+
"successful_completion_tokens": 63279,
|
| 22 |
+
"target": "/mnt/weka/huggingface/hub/models--Qwen--Qwen3.8-27B/snapshots/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
|
| 23 |
+
"total_completion_tokens": 63279,
|
| 24 |
+
"total_verify_ct": 11914,
|
| 25 |
+
"warmup_count": 10
|
| 26 |
+
}
|
dflash2_27b_traj_20260915/main-800k/x4rand/epoch4/serving/acceptance_trace.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
dflash2_27b_traj_20260915/main-800k/x4rand/epoch4/serving/client.log
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
warmup 1/10
|
| 2 |
+
warmup 2/10
|
| 3 |
+
warmup 3/10
|
| 4 |
+
warmup 4/10
|
| 5 |
+
warmup 5/10
|
| 6 |
+
warmup 6/10
|
| 7 |
+
warmup 7/10
|
| 8 |
+
warmup 8/10
|
| 9 |
+
warmup 9/10
|
| 10 |
+
warmup 10/10
|
| 11 |
+
request 50/500 id=q8eval_dc5eedcbc5aa725f0feb0366 al=5.333333333333333 error=None
|
| 12 |
+
request 100/500 id=q8eval_7a986284e914de87bbb616ab al=3.657142857142857 error=None
|
| 13 |
+
request 150/500 id=q8eval_960752a23f720cbafe8bb55e al=4.7407407407407405 error=None
|
| 14 |
+
request 200/500 id=q8eval_d8b172e9de37f6e592f86d87 al=4.7407407407407405 error=None
|
| 15 |
+
request 250/500 id=q8eval_e33d29487007e49b5656b7d8 al=4.923076923076923 error=None
|
| 16 |
+
request 300/500 id=q8eval_4868a37b6be51d452ffd7505 al=3.878787878787879 error=None
|
| 17 |
+
request 350/500 id=q8eval_95595366eb16b5393730d326 al=6.095238095238095 error=None
|
| 18 |
+
request 400/500 id=q8eval_9108cacb70a1386825c8be4f al=7.111111111111111 error=None
|
| 19 |
+
request 450/500 id=q8eval_263c3c4bfdf0771eb4f86851 al=9.846153846153847 error=None
|
| 20 |
+
request 500/500 id=q8eval_a0da1ae07ade386892f7acc0 al=4.129032258064516 error=None
|
| 21 |
+
{
|
| 22 |
+
"concurrency": 1,
|
| 23 |
+
"draft_checkpoint": "/mnt/weka/mos/experiments/dflash2/qwen3.8-27b-traj/runs-b16-800k-traj-x4rand/mos/checkpoints/3/sglang_export",
|
| 24 |
+
"eval": "/workspace/weka-mirror/mos/data/aurora-100k/eval_balanced_500_seed20260719_thinknorm_v1.jsonl",
|
| 25 |
+
"finished_at_utc": "2026-09-25T09:36:39Z",
|
| 26 |
+
"latency_p50_seconds": 0.5017778598703444,
|
| 27 |
+
"latency_p95_seconds": 0.8580532639753073,
|
| 28 |
+
"max_tokens": 128,
|
| 29 |
+
"mean_of_per_request_ratios": 6.334351040787312,
|
| 30 |
+
"measured_wallclock_seconds": 259.594325958984,
|
| 31 |
+
"n_completed": 500,
|
| 32 |
+
"n_prompts": 500,
|
| 33 |
+
"n_successful": 500,
|
| 34 |
+
"output_token_throughput_per_second": 243.76110597270184,
|
| 35 |
+
"pooled_accept_length": 5.311314420010072,
|
| 36 |
+
"server_metric_accept_length": 4.125,
|
| 37 |
+
"speculative_algorithm": "DFLASH",
|
| 38 |
+
"speculative_eagle_topk": 0,
|
| 39 |
+
"speculative_num_draft_tokens": 16,
|
| 40 |
+
"speculative_num_steps": 0,
|
| 41 |
+
"successful_completion_tokens": 63279,
|
| 42 |
+
"target": "/mnt/weka/huggingface/hub/models--Qwen--Qwen3.8-27B/snapshots/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
|
| 43 |
+
"total_completion_tokens": 63279,
|
| 44 |
+
"total_verify_ct": 11914,
|
| 45 |
+
"warmup_count": 10
|
| 46 |
+
}
|
dflash2_27b_traj_20260915/main-800k/x4rand/epoch4/serving/server.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|