| [2026-09-26 19:15:52 + 0.0m] === Testing HF upload to scaledown/vllm-inference-benchmarks === |
| [2026-09-26 19:15:52 + 0.0m] FAIL: $HF_TOKEN is not set in this shell |
| [2026-09-26 19:16:05 + 0.0m] === Testing HF upload to scaledown/vllm-inference-benchmarks === |
| [2026-09-26 19:16:05 + 0.0m] FAIL: $HF_TOKEN is not set in this shell |
| [2026-09-26 19:16:19 + 0.0m] === Testing HF upload to scaledown/vllm-inference-benchmarks === |
| [2026-09-26 19:16:20 + 0.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/_hf_upload_test.csv] |
| [2026-09-26 19:16:20 + 0.0m] If you see '[pushed to hf: ...]' above, the token/repo/write access all work. |
| [2026-09-26 19:16:20 + 0.0m] Check https://huggingface.co/datasets/scaledown/vllm-inference-benchmarks to confirm the file landed. |
| [2026-09-26 19:16:44 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 19:16:44 + 0.0m] [smoke 1/12] qwen3.6-35b-a3b-fp8::base -- starting server ... |
| [2026-09-26 19:17:39 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 19:17:39 + 0.0m] [smoke 1/12] qwen3.6-35b-a3b-fp8::base -- starting server ... |
| [2026-09-26 19:20:19 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 19:20:19 + 0.0m] [smoke 1/12] qwen3.6-35b-a3b-fp8::base -- starting server ... |
| [2026-09-26 19:23:33 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 19:23:33 + 0.0m] [smoke 1/12] qwen3.6-35b-a3b-fp8::base -- starting server ... |
| [2026-09-26 19:35:14 + 11.7m] [smoke 1/12] qwen3.6-35b-a3b-fp8::base -- server ready, running bench ... |
| [2026-09-26 19:35:47 + 12.2m] [smoke 1/12] qwen3.6-35b-a3b-fp8::base -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-35b-a3b-fp8::base/server_smoke.log) |
| [2026-09-26 19:35:47 + 12.2m] [smoke 2/12] qwen3.6-35b-a3b-fp8::mtp -- starting server ... |
| [2026-09-26 19:39:03 + 15.5m] [smoke 2/12] qwen3.6-35b-a3b-fp8::mtp -- server ready, running bench ... |
| [2026-09-26 19:39:38 + 16.1m] [smoke 2/12] qwen3.6-35b-a3b-fp8::mtp -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-35b-a3b-fp8::mtp/server_smoke.log) |
| [2026-09-26 19:39:38 + 16.1m] [smoke 3/12] qwen3.6-35b-a3b-fp8::dflash -- starting server ... |
| [2026-09-26 19:43:08 + 19.6m] [smoke 3/12] qwen3.6-35b-a3b-fp8::dflash -- server ready, running bench ... |
| [2026-09-26 19:43:47 + 20.2m] [smoke 3/12] qwen3.6-35b-a3b-fp8::dflash -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-35b-a3b-fp8::dflash/server_smoke.log) |
| [2026-09-26 19:43:47 + 20.2m] [smoke 4/12] qwen3.5-4b::base -- starting server ... |
| [2026-09-26 19:47:12 + 23.7m] [smoke 4/12] qwen3.5-4b::base -- server ready, running bench ... |
| [2026-09-26 19:47:42 + 24.2m] [smoke 4/12] qwen3.5-4b::base -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.5-4b::base/server_smoke.log) |
| [2026-09-26 19:47:42 + 24.2m] [smoke 5/12] qwen3.5-4b::mtp -- starting server ... |
| [2026-09-26 19:50:23 + 26.8m] [smoke 5/12] qwen3.5-4b::mtp -- server ready, running bench ... |
| [2026-09-26 19:50:57 + 27.4m] [smoke 5/12] qwen3.5-4b::mtp -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.5-4b::mtp/server_smoke.log) |
| [2026-09-26 19:50:57 + 27.4m] [smoke 6/12] qwen3.5-4b::dflash -- starting server ... |
| [2026-09-26 19:53:27 + 29.9m] [smoke 6/12] qwen3.5-4b::dflash -- server ready, running bench ... |
| [2026-09-26 19:54:01 + 30.5m] [smoke 6/12] qwen3.5-4b::dflash -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.5-4b::dflash/server_smoke.log) |
| [2026-09-26 19:54:01 + 30.5m] [smoke 7/12] muse-glimmer-30b::base -- starting server ... |
| [2026-09-26 19:57:26 + 33.9m] [smoke 7/12] muse-glimmer-30b::base -- FAIL (log: /workspace/benchmarks/logs/smoke/muse-glimmer-30b::base/server_smoke.log) |
| [2026-09-26 19:57:26 + 33.9m] [smoke 8/12] muse-glimmer-30b::mtp -- starting server ... |
| [2026-09-26 19:58:01 + 34.5m] [smoke 8/12] muse-glimmer-30b::mtp -- FAIL (log: /workspace/benchmarks/logs/smoke/muse-glimmer-30b::mtp/server_smoke.log) |
| [2026-09-26 19:58:01 + 34.5m] [smoke 9/12] muse-glimmer-30b::dflash -- starting server ... |
| [2026-09-26 20:00:11 + 36.6m] [smoke 9/12] muse-glimmer-30b::dflash -- FAIL (log: /workspace/benchmarks/logs/smoke/muse-glimmer-30b::dflash/server_smoke.log) |
| [2026-09-26 20:00:11 + 36.6m] [smoke 10/12] gemma-4-26b-a4b-it::base -- starting server ... |
| [2026-09-26 20:01:46 + 0.0m] download Qwen/Qwen3.5-4B failed (attempt 1/3): URL: https://huggingface.co/Qwen/Qwen3.5-4B/resolve/main/%2A.json |
| [2026-09-26 20:01:47 + 38.2m] [smoke 10/12] gemma-4-26b-a4b-it::base -- FAIL (log: /workspace/benchmarks/logs/smoke/gemma-4-26b-a4b-it::base/server_smoke.log) |
| [2026-09-26 20:01:47 + 38.2m] [smoke 11/12] gemma-4-26b-a4b-it::mtp -- starting server ... |
| [2026-09-26 20:02:17 + 0.5m] download Qwen/Qwen3.5-4B failed (attempt 2/3): URL: https://huggingface.co/Qwen/Qwen3.5-4B/resolve/main/%2A.json |
| [2026-09-26 20:02:22 + 38.8m] [smoke 11/12] gemma-4-26b-a4b-it::mtp -- FAIL (log: /workspace/benchmarks/logs/smoke/gemma-4-26b-a4b-it::mtp/server_smoke.log) |
| [2026-09-26 20:02:22 + 38.8m] [smoke 12/12] gemma-4-26b-a4b-it::dflash -- starting server ... |
| [2026-09-26 20:03:18 + 1.6m] download Qwen/Qwen3.5-4B failed (attempt 3/3): URL: https://huggingface.co/Qwen/Qwen3.5-4B/resolve/main/%2A.json |
| [2026-09-26 20:04:02 + 40.5m] [smoke 12/12] gemma-4-26b-a4b-it::dflash -- FAIL (log: /workspace/benchmarks/logs/smoke/gemma-4-26b-a4b-it::dflash/server_smoke.log) |
| [2026-09-26 20:04:02 + 40.5m] [hf upload failed, continuing: No module named 'huggingface_hub'] |
| [2026-09-26 20:05:42 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 20:05:42 + 0.0m] [smoke 1/4] muse-glimmer-30b::base -- downloading weights ... |
| [2026-09-26 20:06:06 + 0.4m] download meta-models/Muse-Glimmer-30B failed (attempt 1/3): RuntimeError: Task error: File reconstruction error: Internal Writer Error: Background writer channel closed |
| [2026-09-26 20:09:03 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 20:09:03 + 0.0m] [smoke 1/4] muse-glimmer-30b::base -- downloading weights ... |
| [2026-09-26 20:12:12 + 3.1m] [smoke 1/4] muse-glimmer-30b::base -- starting server ... |
| [2026-09-26 20:14:42 + 5.6m] [smoke 1/4] muse-glimmer-30b::base -- FAIL (log: /workspace/benchmarks/logs/smoke/muse-glimmer-30b::base/server_smoke.log) |
| [2026-09-26 20:14:44 + 5.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 20:14:44 + 5.7m] [smoke 2/4] muse-glimmer-30b::dflash -- downloading weights ... |
| [2026-09-26 20:15:04 + 6.0m] [smoke 2/4] muse-glimmer-30b::dflash -- starting server ... |
| [2026-09-26 20:16:33 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 20:16:33 + 0.0m] [smoke 1/4] muse-glimmer-30b::base -- downloading weights ... |
| [2026-09-26 20:16:34 + 0.0m] [smoke 1/4] muse-glimmer-30b::base -- starting server ... |
| [2026-09-26 20:18:44 + 2.2m] [smoke 1/4] muse-glimmer-30b::base -- server ready, running bench ... |
| [2026-09-26 20:19:15 + 2.7m] [smoke 1/4] muse-glimmer-30b::base -- PASS (log: /workspace/benchmarks/logs/smoke/muse-glimmer-30b::base/server_smoke.log) |
| [2026-09-26 20:19:16 + 2.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 20:19:16 + 2.7m] [smoke 2/4] muse-glimmer-30b::dflash -- downloading weights ... |
| [2026-09-26 20:19:19 + 2.8m] [smoke 2/4] muse-glimmer-30b::dflash -- starting server ... |
| [2026-09-26 20:24:19 + 7.8m] [smoke 2/4] muse-glimmer-30b::dflash -- server ready, running bench ... |
| [2026-09-26 20:24:50 + 8.3m] [smoke 2/4] muse-glimmer-30b::dflash -- PASS (log: /workspace/benchmarks/logs/smoke/muse-glimmer-30b::dflash/server_smoke.log) |
| [2026-09-26 20:24:51 + 8.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 20:24:51 + 8.3m] [smoke 3/4] gemma-4-26b-a4b-it::base -- downloading weights ... |
| [2026-09-26 20:27:53 + 11.3m] [smoke 3/4] gemma-4-26b-a4b-it::base -- starting server ... |
| [2026-09-26 20:32:19 + 15.8m] [smoke 3/4] gemma-4-26b-a4b-it::base -- server ready, running bench ... |
| [2026-09-26 20:32:52 + 16.3m] [smoke 3/4] gemma-4-26b-a4b-it::base -- PASS (log: /workspace/benchmarks/logs/smoke/gemma-4-26b-a4b-it::base/server_smoke.log) |
| [2026-09-26 20:32:53 + 16.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 20:32:53 + 16.3m] [smoke 4/4] gemma-4-26b-a4b-it::dflash -- downloading weights ... |
| [2026-09-26 20:33:01 + 16.5m] [smoke 4/4] gemma-4-26b-a4b-it::dflash -- starting server ... |
| [2026-09-26 20:37:21 + 20.8m] [smoke 4/4] gemma-4-26b-a4b-it::dflash -- server ready, running bench ... |
| [2026-09-26 20:37:52 + 21.3m] [smoke 4/4] gemma-4-26b-a4b-it::dflash -- PASS (log: /workspace/benchmarks/logs/smoke/gemma-4-26b-a4b-it::dflash/server_smoke.log) |
| [2026-09-26 20:37:53 + 21.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 20:37:55 + 21.4m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 21:24:14 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 21:24:14 + 0.0m] [smoke 1/32] qwen3.6-27b-fp8::base -- downloading weights ... |
| [2026-09-26 21:26:21 + 2.1m] [smoke 1/32] qwen3.6-27b-fp8::base -- starting server ... |
| [2026-09-26 21:34:22 + 10.1m] [smoke 1/32] qwen3.6-27b-fp8::base -- server ready, running bench ... |
| [2026-09-26 21:35:55 + 11.7m] [smoke 1/32] qwen3.6-27b-fp8::base -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-27b-fp8__base/server_smoke.log) |
| [2026-09-26 21:35:55 + 11.7m] [results CSV columns changed -- previous file moved to orchestrate_results.pre_20260926_213555.csv] |
| [2026-09-26 21:35:57 + 11.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.pre_20260926_213555.csv] |
| [2026-09-26 21:36:22 + 12.1m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 21:36:22 + 12.1m] [smoke 2/32] qwen3.6-27b-fp8::mtp -- downloading weights ... |
| [2026-09-26 21:36:24 + 12.2m] [smoke 2/32] qwen3.6-27b-fp8::mtp -- starting server ... |
| [2026-09-26 21:39:39 + 15.4m] [smoke 2/32] qwen3.6-27b-fp8::mtp -- server ready, running bench ... |
| [2026-09-26 21:41:16 + 17.0m] [smoke 2/32] qwen3.6-27b-fp8::mtp -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-27b-fp8__mtp/server_smoke.log) |
| [2026-09-26 21:41:17 + 17.1m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 21:41:17 + 17.1m] [smoke 3/32] qwen3.6-27b-fp8::dflash -- downloading weights ... |
| [2026-09-26 21:41:30 + 17.3m] [smoke 3/32] qwen3.6-27b-fp8::dflash -- starting server ... |
| [2026-09-26 21:44:26 + 20.2m] [smoke 3/32] qwen3.6-27b-fp8::dflash -- server ready, running bench ... |
| [2026-09-26 21:46:01 + 21.8m] [smoke 3/32] qwen3.6-27b-fp8::dflash -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-27b-fp8__dflash/server_smoke.log) |
| [2026-09-26 21:46:02 + 21.8m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 21:46:02 + 21.8m] [smoke 4/32] qwen3.6-27b::base -- downloading weights ... |
| [2026-09-26 21:49:21 + 25.1m] [smoke 4/32] qwen3.6-27b::base -- starting server ... |
| [2026-09-26 21:51:46 + 27.5m] [smoke 4/32] qwen3.6-27b::base -- server ready, running bench ... |
| [2026-09-26 21:53:22 + 29.1m] [smoke 4/32] qwen3.6-27b::base -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-27b__base/server_smoke.log) |
| [2026-09-26 21:53:23 + 29.2m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 21:53:23 + 29.2m] [smoke 5/32] qwen3.6-27b::mtp -- downloading weights ... |
| [2026-09-26 21:53:24 + 29.2m] [smoke 5/32] qwen3.6-27b::mtp -- starting server ... |
| [2026-09-26 21:53:51 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 21:53:51 + 0.0m] [smoke 1/32] qwen3.6-27b-fp8::base -- already passed, skipping |
| [2026-09-26 21:53:51 + 0.0m] [smoke 2/32] qwen3.6-27b-fp8::mtp -- already passed, skipping |
| [2026-09-26 21:53:51 + 0.0m] [smoke 3/32] qwen3.6-27b-fp8::dflash -- already passed, skipping |
| [2026-09-26 21:53:51 + 0.0m] [smoke 4/32] qwen3.6-27b::base -- already passed, skipping |
| [2026-09-26 21:53:51 + 0.0m] [smoke 5/32] qwen3.6-27b::mtp -- downloading weights ... |
| [2026-09-26 21:53:52 + 0.0m] [smoke 5/32] qwen3.6-27b::mtp -- starting server ... |
| [2026-09-26 21:56:17 + 2.4m] [smoke 5/32] qwen3.6-27b::mtp -- server ready, running bench ... |
| [2026-09-26 21:57:52 + 4.0m] [smoke 5/32] qwen3.6-27b::mtp -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-27b__mtp/server_smoke.log) |
| [2026-09-26 21:58:20 + 4.5m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 21:58:20 + 4.5m] [smoke 6/32] qwen3.6-27b::dflash -- downloading weights ... |
| [2026-09-26 21:58:22 + 4.5m] [smoke 6/32] qwen3.6-27b::dflash -- starting server ... |
| [2026-09-26 22:01:02 + 7.2m] [smoke 6/32] qwen3.6-27b::dflash -- server ready, running bench ... |
| [2026-09-26 22:02:41 + 8.8m] [smoke 6/32] qwen3.6-27b::dflash -- PASS (log: /workspace/benchmarks/logs/smoke/qwen3.6-27b__dflash/server_smoke.log) |
| [2026-09-26 22:02:42 + 8.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 22:02:42 + 8.9m] [smoke 7/32] qwen3.6-35b-a3b-fp8::base -- already passed, skipping |
| [2026-09-26 22:02:42 + 8.9m] [smoke 8/32] qwen3.6-35b-a3b-fp8::mtp -- already passed, skipping |
| [2026-09-26 22:02:42 + 8.9m] [smoke 9/32] qwen3.6-35b-a3b-fp8::dflash -- already passed, skipping |
| [2026-09-26 22:02:42 + 8.9m] [smoke 10/32] qwen3.6-35b-a3b::base -- downloading weights ... |
| [2026-09-26 22:04:29 + 10.6m] download Qwen/Qwen3.6-35B-A3B failed (attempt 1/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:06:34 + 12.7m] download Qwen/Qwen3.6-35B-A3B failed (attempt 2/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:09:10 + 15.3m] download Qwen/Qwen3.6-35B-A3B failed (attempt 3/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:09:10 + 15.3m] [smoke 10/32] qwen3.6-35b-a3b::base -- FAIL (download failed after retries) |
| [2026-09-26 22:09:12 + 15.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 22:09:12 + 15.3m] [smoke 11/32] qwen3.6-35b-a3b::mtp -- downloading weights ... |
| [2026-09-26 22:10:46 + 16.9m] download Qwen/Qwen3.6-35B-A3B failed (attempt 1/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:12:54 + 19.0m] download Qwen/Qwen3.6-35B-A3B failed (attempt 2/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:15:29 + 21.6m] download Qwen/Qwen3.6-35B-A3B failed (attempt 3/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:15:29 + 21.6m] [smoke 11/32] qwen3.6-35b-a3b::mtp -- FAIL (download failed after retries) |
| [2026-09-26 22:15:30 + 21.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 22:15:30 + 21.7m] [smoke 12/32] qwen3.6-35b-a3b::dflash -- downloading weights ... |
| [2026-09-26 22:17:05 + 23.2m] download Qwen/Qwen3.6-35B-A3B failed (attempt 1/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:19:10 + 25.3m] download Qwen/Qwen3.6-35B-A3B failed (attempt 2/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:21:44 + 27.9m] download Qwen/Qwen3.6-35B-A3B failed (attempt 3/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:21:44 + 27.9m] [smoke 12/32] qwen3.6-35b-a3b::dflash -- FAIL (download failed after retries) |
| [2026-09-26 22:21:46 + 27.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 22:21:46 + 27.9m] [smoke 13/32] qwen3.5-9b::base -- downloading weights ... |
| [2026-09-26 22:22:14 + 28.4m] download Qwen/Qwen3.5-9B failed (attempt 1/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:23:13 + 29.4m] download Qwen/Qwen3.5-9B failed (attempt 2/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:24:40 + 30.8m] download Qwen/Qwen3.5-9B failed (attempt 3/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:24:40 + 30.8m] [smoke 13/32] qwen3.5-9b::base -- FAIL (download failed after retries) |
| [2026-09-26 22:24:41 + 30.8m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-26 22:24:41 + 30.8m] [smoke 14/32] qwen3.5-9b::mtp -- downloading weights ... |
| [2026-09-26 22:25:11 + 31.3m] download Qwen/Qwen3.5-9B failed (attempt 1/3): RuntimeError: Task error: File reconstruction error: IO Error: Disk quota exceeded (os error 122) |
| [2026-09-26 22:35:19 + 0.0m] === Phase 0: smoke test === |
| [2026-09-26 22:35:19 + 0.0m] [smoke 1/32] qwen3.6-27b-fp8::base -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 2/32] qwen3.6-27b-fp8::mtp -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 3/32] qwen3.6-27b-fp8::dflash -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 4/32] qwen3.6-27b::base -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 5/32] qwen3.6-27b::mtp -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 6/32] qwen3.6-27b::dflash -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 7/32] qwen3.6-35b-a3b-fp8::base -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 8/32] qwen3.6-35b-a3b-fp8::mtp -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 9/32] qwen3.6-35b-a3b-fp8::dflash -- already passed, skipping |
| [2026-09-26 22:35:19 + 0.0m] [smoke 10/32] qwen3.6-35b-a3b::base -- downloading weights ... |
| [2026-09-26 22:37:51 + 0.0m] === qwen3.6-27b-fp8: ensuring weights are on disk === |
| [2026-09-26 22:37:53 + 0.0m] === Phase 1: config sweep === |
| [2026-09-26 22:37:53 + 0.0m] [phase1 1/3] qwen3.6-27b-fp8::base -- sweeping 6 configs ... |
| [2026-09-26 22:37:53 + 0.0m] [phase1 1/3] [1/6] seqs=16 batched=4096 ... |
| [2026-09-26 22:45:36 + 7.8m] [phase1 1/3] [1/6] -> throughput=767.880612751248 (spread 0.37%) |
| [2026-09-26 22:45:36 + 7.8m] [phase1 1/3] [2/6] seqs=16 batched=6144 ... |
| [2026-09-26 22:52:57 + 15.1m] [phase1 1/3] [2/6] -> throughput=805.7353658286446 (spread 7.03%) |
| [2026-09-26 22:52:58 + 15.1m] [phase1 1/3] [3/6] seqs=32 batched=8192 ... |
| [2026-09-26 22:59:46 + 21.9m] [phase1 1/3] [3/6] -> throughput=970.419443213065 (spread 8.26%) |
| [2026-09-26 22:59:46 + 21.9m] [phase1 1/3] [4/6] seqs=32 batched=12288 ... |
| [2026-09-26 23:06:41 + 28.8m] [phase1 1/3] [4/6] -> throughput=977.4270140301653 (spread 8.73%) |
| [2026-09-26 23:06:41 + 28.8m] [phase1 1/3] [5/6] seqs=64 batched=16384 ... |
| [2026-09-26 23:13:27 + 35.6m] [phase1 1/3] [5/6] -> throughput=1022.5667448107928 (spread 9.84%) |
| [2026-09-26 23:13:27 + 35.6m] [phase1 1/3] [6/6] seqs=64 batched=24576 ... |
| [2026-09-26 23:20:15 + 42.4m] [phase1 1/3] [6/6] -> throughput=1057.963335175246 (spread 9.53%) |
| [2026-09-26 23:20:15 + 42.4m] [phase1 1/3] qwen3.6-27b-fp8::base -- done, 6/6 configs succeeded, kept seqs=16/batched=6144, seqs=32/batched=12288, seqs=64/batched=24576 |
| [2026-09-26 23:20:38 + 42.8m] [phase1 2/3] qwen3.6-27b-fp8::mtp -- sweeping 6 configs ... |
| [2026-09-26 23:20:38 + 42.8m] [phase1 2/3] [1/6] seqs=16 batched=4096 ... |
| [2026-09-26 23:27:17 + 49.4m] [phase1 2/3] [1/6] -> throughput=1078.2830607509804 (spread 7.39%) |
| [2026-09-26 23:27:17 + 49.4m] [phase1 2/3] [2/6] seqs=16 batched=6144 ... |
| [2026-09-26 23:34:00 + 56.2m] [phase1 2/3] [2/6] -> throughput=1074.7560006340664 (spread 5.69%) |
| [2026-09-26 23:34:00 + 56.2m] [phase1 2/3] [3/6] seqs=32 batched=8192 ... |
| [2026-09-26 23:40:18 + 62.5m] [phase1 2/3] [3/6] -> throughput=1354.4363547943522 (spread 12.06%) |
| [2026-09-26 23:40:18 + 62.5m] [phase1 2/3] [4/6] seqs=32 batched=12288 ... |
| [2026-09-26 23:46:34 + 68.7m] [phase1 2/3] [4/6] -> throughput=1337.170880390273 (spread 9.64%) |
| [2026-09-26 23:46:34 + 68.7m] [phase1 2/3] [5/6] seqs=64 batched=16384 ... |
| [2026-09-26 23:52:47 + 74.9m] [phase1 2/3] [5/6] -> throughput=1419.0670117520726 (spread 7.13%) |
| [2026-09-26 23:52:47 + 74.9m] [phase1 2/3] [6/6] seqs=64 batched=24576 ... |
| [2026-09-26 23:59:08 + 81.3m] [phase1 2/3] [6/6] -> throughput=1413.5953505538344 (spread 6.05%) |
| [2026-09-26 23:59:08 + 81.3m] [phase1 2/3] qwen3.6-27b-fp8::mtp -- done, 6/6 configs succeeded, kept seqs=16/batched=4096, seqs=32/batched=8192, seqs=64/batched=16384 |
| [2026-09-26 23:59:08 + 81.3m] [phase1 3/3] qwen3.6-27b-fp8::dflash -- sweeping 6 configs ... |
| [2026-09-26 23:59:08 + 81.3m] [phase1 3/3] [1/6] seqs=16 batched=4096 ... |
| [2026-09-27 00:05:09 + 0.0m] [registry skip: qwen3.6-27b, qwen3.6-35b-a3b, qwen3.5-9b, gemma-4-e4b] |
| [2026-09-27 00:05:09 + 0.0m] === qwen3.6-35b-a3b-fp8: ensuring weights are on disk === |
| [2026-09-27 00:05:11 + 0.0m] === Phase 1: config sweep === |
| [2026-09-27 00:05:11 + 0.0m] [phase1 1/3] qwen3.6-35b-a3b-fp8::base -- sweeping 6 configs ... |
| [2026-09-27 00:05:11 + 0.0m] [phase1 1/3] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 00:06:12 + 0.0m] [registry skip: qwen3.6-27b, qwen3.6-35b-a3b, qwen3.5-9b, gemma-4-e4b] |
| [2026-09-27 00:06:12 + 0.0m] === qwen3.6-35b-a3b-fp8: ensuring weights are on disk === |
| [2026-09-27 00:06:14 + 0.0m] === Phase 1: config sweep === |
| [2026-09-27 00:06:14 + 0.0m] [phase1 1/3] qwen3.6-35b-a3b-fp8::base -- sweeping 6 configs ... |
| [2026-09-27 00:06:14 + 0.0m] [phase1 1/3] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 00:18:29 + 12.3m] [phase1 1/3] [1/6] -> throughput=1201.9746027100593 (spread 0.92%) |
| [2026-09-27 00:18:29 + 12.3m] [phase1 1/3] [2/6] seqs=32 batched=12288 ... |
| [2026-09-27 00:24:46 + 18.6m] [phase1 1/3] [2/6] -> throughput=1236.7043351381249 (spread 5.35%) |
| [2026-09-27 00:24:46 + 18.6m] [phase1 1/3] [3/6] seqs=32 batched=16384 ... |
| [2026-09-27 00:31:00 + 24.8m] [phase1 1/3] [3/6] -> throughput=1233.5915367001248 (spread 4.36%) |
| [2026-09-27 00:31:00 + 24.8m] [phase1 1/3] [4/6] seqs=64 batched=16384 ... |
| [2026-09-27 00:37:02 + 30.8m] [phase1 1/3] [4/6] -> throughput=1328.1343183337658 (spread 11.95%) |
| [2026-09-27 00:37:02 + 30.8m] [phase1 1/3] [5/6] seqs=64 batched=24576 ... |
| [2026-09-27 00:43:07 + 36.9m] [phase1 1/3] [5/6] -> throughput=1333.0863062514204 (spread 9.64%) |
| [2026-09-27 00:43:07 + 36.9m] [phase1 1/3] [6/6] seqs=64 batched=32768 ... |
| [2026-09-27 00:49:14 + 43.0m] [phase1 1/3] [6/6] -> throughput=1332.1286699904697 (spread 9.25%) |
| [2026-09-27 00:49:14 + 43.0m] [phase1 1/3] qwen3.6-35b-a3b-fp8::base -- done, 6/6 configs succeeded, kept seqs=32/batched=12288, seqs=64/batched=24576 |
| [2026-09-27 00:49:39 + 43.4m] [phase1 2/3] qwen3.6-35b-a3b-fp8::mtp -- sweeping 6 configs ... |
| [2026-09-27 00:49:39 + 43.4m] [phase1 2/3] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 00:56:06 + 49.9m] [phase1 2/3] [1/6] -> throughput=1715.920836331525 (spread 16.36%) |
| [2026-09-27 00:56:06 + 49.9m] [phase1 2/3] [2/6] seqs=32 batched=12288 ... |
| [2026-09-27 01:01:53 + 55.7m] [phase1 2/3] [2/6] -> throughput=1658.4835563295978 (spread 9.04%) |
| [2026-09-27 01:01:53 + 55.7m] [phase1 2/3] [3/6] seqs=32 batched=16384 ... |
| [2026-09-27 01:07:38 + 61.4m] [phase1 2/3] [3/6] -> throughput=1779.4039891167693 (spread 12.38%) |
| [2026-09-27 01:07:38 + 61.4m] [phase1 2/3] [4/6] seqs=64 batched=16384 ... |
| [2026-09-27 01:13:14 + 67.0m] [phase1 2/3] [4/6] -> throughput=1941.8408703064983 (spread 5.7%) |
| [2026-09-27 01:13:14 + 67.0m] [phase1 2/3] [5/6] seqs=64 batched=24576 ... |
| [2026-09-27 01:18:54 + 72.7m] [phase1 2/3] [5/6] -> throughput=1933.6891029100284 (spread 10.99%) |
| [2026-09-27 01:18:54 + 72.7m] [phase1 2/3] [6/6] seqs=64 batched=32768 ... |
| [2026-09-27 01:24:55 + 78.7m] [phase1 2/3] [6/6] -> throughput=2018.4908673937903 (spread 7.12%) |
| [2026-09-27 01:24:55 + 78.7m] [phase1 2/3] qwen3.6-35b-a3b-fp8::mtp -- done, 6/6 configs succeeded, kept seqs=32/batched=16384, seqs=64/batched=32768 |
| [2026-09-27 01:24:55 + 78.7m] [phase1 3/3] qwen3.6-35b-a3b-fp8::dflash -- sweeping 6 configs ... |
| [2026-09-27 01:24:55 + 78.7m] [phase1 3/3] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 01:31:23 + 85.2m] [phase1 3/3] [1/6] -> throughput=1377.62372591635 (spread 12.58%) |
| [2026-09-27 01:31:23 + 85.2m] [phase1 3/3] [2/6] seqs=32 batched=12288 ... |
| [2026-09-27 01:37:21 + 91.1m] [phase1 3/3] [2/6] -> throughput=1563.5081125546621 (spread 6.5%) |
| [2026-09-27 01:37:21 + 91.1m] [phase1 3/3] [3/6] seqs=32 batched=16384 ... |
| [2026-09-27 01:43:37 + 97.4m] [phase1 3/3] [3/6] -> throughput=1362.4332237444023 (spread 11.58%) |
| [2026-09-27 01:43:37 + 97.4m] [phase1 3/3] [4/6] seqs=64 batched=16384 ... |
| [2026-09-27 01:49:28 + 103.3m] [phase1 3/3] [4/6] -> throughput=1687.4809337507463 (spread 0.75%) |
| [2026-09-27 01:49:28 + 103.3m] [phase1 3/3] [5/6] seqs=64 batched=24576 ... |
| [2026-09-27 01:55:29 + 109.3m] [phase1 3/3] [5/6] -> throughput=1588.0850599668638 (spread 12.48%) |
| [2026-09-27 01:55:29 + 109.3m] [phase1 3/3] [6/6] seqs=64 batched=32768 ... |
| [2026-09-27 02:01:32 + 115.3m] [phase1 3/3] [6/6] -> throughput=1516.2625922327302 (spread 21.24%) |
| [2026-09-27 02:01:32 + 115.3m] [phase1 3/3] qwen3.6-35b-a3b-fp8::dflash -- done, 6/6 configs succeeded, kept seqs=32/batched=12288, seqs=64/batched=16384 |
| [2026-09-27 02:01:33 + 115.4m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 02:01:35 + 115.4m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_state.json] |
| [2026-09-27 02:01:36 + 115.4m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_progress.log] |
| [2026-09-27 02:01:38 + 115.4m] [pushed to hf: scaledown/vllm-inference-benchmarks/logs] |
| [2026-09-27 02:01:38 + 115.4m] === Phase 2: concurrency curve === |
| [2026-09-27 02:01:38 + 115.4m] [phase2 1/3] qwen3.6-35b-a3b-fp8::base -- 2 configs, 10 (config, concurrency) points ... |
| [2026-09-27 02:01:38 + 115.4m] [phase2 1/3] [1/2] seqs=32 batched=12288 -- starting server ... |
| [2026-09-27 02:06:22 + 120.2m] [phase2 1/3] [1/2] conc=8 -> throughput=810.8021407062976 p99_e2el=7559.476108723322 (spread 6.31%) |
| [2026-09-27 02:08:47 + 122.6m] [phase2 1/3] [1/2] conc=16 -> throughput=1066.1632195210645 p99_e2el=11695.0721473433 (spread 9.87%) |
| [2026-09-27 02:11:52 + 125.7m] [phase2 1/3] [1/2] conc=24 -> throughput=1216.4418822808505 p99_e2el=15774.600832404107 (spread 2.08%) |
| [2026-09-27 02:15:12 + 129.0m] [phase2 1/3] [1/2] conc=32 -> throughput=1335.1366582368878 p99_e2el=19112.027153368108 (spread 0.56%) |
| [2026-09-27 02:15:13 + 129.0m] [phase2 1/3] [2/2] seqs=64 batched=24576 -- starting server ... |
| [2026-09-27 02:20:09 + 134.0m] [phase2 1/3] [2/2] conc=8 -> throughput=841.1100767189681 p99_e2el=7303.597593207378 (spread 5.23%) |
| [2026-09-27 02:22:34 + 136.4m] [phase2 1/3] [2/2] conc=16 -> throughput=1094.9688395110015 p99_e2el=11433.555923500087 (spread 3.07%) |
| [2026-09-27 02:25:29 + 139.3m] [phase2 1/3] [2/2] conc=24 -> throughput=1321.4123896035499 p99_e2el=14327.54705264467 (spread 3.3%) |
| [2026-09-27 02:28:39 + 142.4m] [phase2 1/3] [2/2] conc=32 -> throughput=1436.5283423222622 p99_e2el=17600.230673172045 (spread 2.81%) |
| [2026-09-27 02:32:36 + 146.4m] [phase2 1/3] [2/2] conc=48 -> throughput=1567.1151202636938 p99_e2el=24097.561099203296 (spread 0.83%) |
| [2026-09-27 02:37:09 + 150.9m] [phase2 1/3] [2/2] conc=64 -> throughput=1695.2860219749575 p99_e2el=30153.708891707927 (spread 3.18%) |
| [2026-09-27 02:37:10 + 151.0m] [phase2 1/3] qwen3.6-35b-a3b-fp8::base -- best: seqs=64 batched=24576 conc=24 throughput=1321.4123896035499 p99_e2el=14327.54705264467 |
| [2026-09-27 02:37:11 + 151.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 02:37:11 + 151.0m] [phase2 2/3] qwen3.6-35b-a3b-fp8::mtp -- 2 configs, 10 (config, concurrency) points ... |
| [2026-09-27 02:37:11 + 151.0m] [phase2 2/3] [1/2] seqs=32 batched=16384 -- starting server ... |
| [2026-09-27 02:41:59 + 155.8m] [phase2 2/3] [1/2] conc=8 -> throughput=990.2566305420269 p99_e2el=5974.3034427287075 (spread 6.45%) |
| [2026-09-27 02:44:03 + 157.8m] [phase2 2/3] [1/2] conc=16 -> throughput=1495.3549033485083 p99_e2el=7647.262660705015 (spread 4.39%) |
| [2026-09-27 02:46:27 + 160.2m] [phase2 2/3] [1/2] conc=24 -> throughput=1812.6201558458904 p99_e2el=9843.12744180905 (spread 2.1%) |
| [2026-09-27 02:49:13 + 163.0m] [phase2 2/3] [1/2] conc=32 -> throughput=1784.5227386637193 p99_e2el=13809.350763622198 (spread 4.27%) |
| [2026-09-27 02:49:14 + 163.0m] [phase2 2/3] [2/2] seqs=64 batched=32768 -- starting server ... |
| [2026-09-27 02:53:52 + 167.7m] [phase2 2/3] [2/2] conc=8 -> throughput=1002.3620200958702 p99_e2el=5935.395858804697 (spread 1.29%) |
| [2026-09-27 02:55:49 + 169.6m] [phase2 2/3] [2/2] conc=16 -> throughput=1519.4695937309907 p99_e2el=7720.933473242911 (spread 5.01%) |
| [2026-09-27 02:58:08 + 171.9m] [phase2 2/3] [2/2] conc=24 -> throughput=1930.7835908200382 p99_e2el=9417.822180467196 (spread 8.68%) |
| [2026-09-27 03:00:34 + 174.4m] [phase2 2/3] [2/2] conc=32 -> throughput=2228.204077920008 p99_e2el=11072.48854705831 (spread 1.77%) |
| [2026-09-27 03:03:30 + 177.3m] [phase2 2/3] [2/2] conc=48 -> throughput=2395.1147239483307 p99_e2el=15195.678483815846 (spread 8.45%) |
| [2026-09-27 03:06:53 + 180.7m] [phase2 2/3] [2/2] conc=64 -> throughput=2482.750070341842 p99_e2el=19385.364375307683 (spread 0.75%) |
| [2026-09-27 03:06:54 + 180.7m] [phase2 2/3] qwen3.6-35b-a3b-fp8::mtp -- best: seqs=64 batched=32768 conc=32 throughput=2228.204077920008 p99_e2el=11072.48854705831 |
| [2026-09-27 03:06:56 + 180.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 03:06:56 + 180.7m] [phase2 3/3] qwen3.6-35b-a3b-fp8::dflash -- 2 configs, 10 (config, concurrency) points ... |
| [2026-09-27 03:06:56 + 180.7m] [phase2 3/3] [1/2] seqs=32 batched=12288 -- starting server ... |
| [2026-09-27 03:11:37 + 185.4m] [phase2 3/3] [1/2] conc=8 -> throughput=1002.3153903301148 p99_e2el=6650.545087663924 (spread 10.9%) |
| [2026-09-27 03:13:52 + 187.7m] [phase2 3/3] [1/2] conc=16 -> throughput=1218.3479661620686 p99_e2el=9756.106221585065 (spread 2.58%) |
| [2026-09-27 03:16:40 + 190.5m] [phase2 3/3] [1/2] conc=24 -> throughput=1373.8661873070828 p99_e2el=14287.79016208836 (spread 4.96%) |
| [2026-09-27 03:19:46 + 193.6m] [phase2 3/3] [1/2] conc=32 -> throughput=1514.3817833843268 p99_e2el=17839.912263432045 (spread 11.29%) |
| [2026-09-27 03:19:48 + 193.6m] [phase2 3/3] [2/2] seqs=64 batched=16384 -- starting server ... |
| [2026-09-27 03:24:22 + 198.2m] [phase2 3/3] [2/2] conc=8 -> throughput=1117.800894178983 p99_e2el=5279.60593275726 (spread 4.76%) |
| [2026-09-27 03:26:31 + 200.3m] [phase2 3/3] [2/2] conc=16 -> throughput=1355.8367255822318 p99_e2el=8536.751333435526 (spread 2.06%) |
| [2026-09-27 03:29:04 + 202.9m] [phase2 3/3] [2/2] conc=24 -> throughput=1633.2585701231258 p99_e2el=12094.442657660698 (spread 0.37%) |
| [2026-09-27 03:31:46 + 205.6m] [phase2 3/3] [2/2] conc=32 -> throughput=1861.8811997402845 p99_e2el=13851.864057513865 (spread 4.98%) |
| [2026-09-27 03:34:58 + 208.8m] [phase2 3/3] [2/2] conc=48 -> throughput=2063.675634138652 p99_e2el=15885.832400307958 (spread 1.53%) |
| [2026-09-27 03:39:01 + 212.8m] [phase2 3/3] [2/2] conc=64 -> throughput=1982.8359011141167 p99_e2el=18790.001297390554 (spread 7.72%) |
| [2026-09-27 03:39:03 + 212.8m] [phase2 3/3] qwen3.6-35b-a3b-fp8::dflash -- best: seqs=64 batched=16384 conc=48 throughput=2063.675634138652 p99_e2el=15885.832400307958 |
| [2026-09-27 03:39:05 + 212.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 03:39:06 + 212.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 03:39:08 + 212.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_state.json] |
| [2026-09-27 03:39:09 + 212.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_progress.log] |
| [2026-09-27 03:39:12 + 213.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/logs] |
| [2026-09-27 03:39:12 + 213.0m] qwen3.6-35b-a3b-fp8 winner for phase 3: qwen3.6-35b-a3b-fp8::mtp (base=-0.68, mtp=1.00, dflash=-0.18) |
| [2026-09-27 03:39:12 + 213.0m] === Phase 3: full IO matrix for qwen3.6-35b-a3b-fp8::mtp === |
| [2026-09-27 03:39:12 + 213.0m] downloading weights for Qwen/Qwen3.6-35B-A3B-FP8 (deleted after its phase 2) ... |
| [2026-09-27 03:41:54 + 215.7m] 4 concurrencies [16, 32, 48, 64] x 12 IO cells = 48 points |
| [2026-09-27 03:41:54 + 215.7m] [phase3 1/48] conc=16 in=1024 out=1 ... |
| [2026-09-27 03:44:19 + 218.1m] [phase3 1/48] -> throughput=43.42289894916544 p99_e2el=412.0494277193211 (spread 1.44%) |
| [2026-09-27 03:44:19 + 218.1m] [phase3 2/48] conc=16 in=1024 out=10 ... |
| [2026-09-27 03:45:46 + 219.6m] [phase3 2/48] -> throughput=253.79191609376142 p99_e2el=1170.812459352892 (spread 9.35%) |
| [2026-09-27 03:45:46 + 219.6m] [phase3 3/48] conc=16 in=1024 out=50 ... |
| [2026-09-27 03:47:40 + 221.5m] [phase3 3/48] -> throughput=675.5771967724627 p99_e2el=1871.156342097092 (spread 14.42%) |
| [2026-09-27 03:47:40 + 221.5m] [phase3 4/48] conc=16 in=1024 out=100 ... |
| [2026-09-27 03:50:02 + 223.8m] [phase3 4/48] -> throughput=931.9166487337246 p99_e2el=2791.6441182070403 (spread 0.09%) |
| [2026-09-27 03:50:02 + 223.8m] [phase3 5/48] conc=16 in=1024 out=250 ... |
| [2026-09-27 03:53:20 + 227.1m] [phase3 5/48] -> throughput=1415.4105182827975 p99_e2el=4277.408063109033 (spread 1.09%) |
| [2026-09-27 03:53:20 + 227.1m] [phase3 6/48] conc=16 in=1024 out=500 ... |
| [2026-09-27 03:58:00 + 231.8m] [phase3 6/48] -> throughput=1808.9448862247937 p99_e2el=6452.111105176154 (spread 1.56%) |
| [2026-09-27 03:58:00 + 231.8m] [phase3 7/48] conc=16 in=2048 out=1 ... |
| [2026-09-27 03:59:30 + 233.3m] [phase3 7/48] -> throughput=26.42962934592587 p99_e2el=724.7903983690776 (spread 1.85%) |
| [2026-09-27 03:59:30 + 233.3m] [phase3 8/48] conc=16 in=2048 out=10 ... |
| [2026-09-27 04:00:31 + 234.3m] [phase3 8/48] -> FAILED |
| [2026-09-27 04:00:31 + 234.3m] [phase3 9/48] conc=16 in=2048 out=50 ... |
| [2026-09-27 04:01:33 + 235.3m] [phase3 9/48] -> FAILED |
| [2026-09-27 04:01:33 + 235.3m] [phase3 10/48] conc=16 in=2048 out=100 ... |
| [2026-09-27 04:02:33 + 236.3m] [phase3 10/48] -> FAILED |
| [2026-09-27 04:02:33 + 236.3m] [phase3 11/48] conc=16 in=2048 out=250 ... |
| [2026-09-27 04:03:34 + 237.4m] [phase3 11/48] -> FAILED |
| [2026-09-27 04:03:34 + 237.4m] [phase3 12/48] conc=16 in=2048 out=500 ... |
| [2026-09-27 04:04:35 + 238.4m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 04:04:35 + 238.4m] [phase3 12/48] -> FAILED |
| [2026-09-27 04:04:35 + 238.4m] [phase3 13/48] conc=32 in=1024 out=1 ... |
| [2026-09-27 04:05:32 + 239.3m] [phase3 13/48] -> FAILED |
| [2026-09-27 04:05:32 + 239.3m] [phase3 14/48] conc=32 in=1024 out=10 ... |
| [2026-09-27 04:06:28 + 240.3m] [phase3 14/48] -> FAILED |
| [2026-09-27 04:06:28 + 240.3m] [phase3 15/48] conc=32 in=1024 out=50 ... |
| [2026-09-27 04:07:27 + 241.3m] [phase3 15/48] -> FAILED |
| [2026-09-27 04:07:27 + 241.3m] [phase3 16/48] conc=32 in=1024 out=100 ... |
| [2026-09-27 04:08:27 + 242.2m] [phase3 16/48] -> FAILED |
| [2026-09-27 04:08:27 + 242.2m] [phase3 17/48] conc=32 in=1024 out=250 ... |
| [2026-09-27 04:09:26 + 243.2m] [phase3 17/48] -> FAILED |
| [2026-09-27 04:09:26 + 243.2m] [phase3 18/48] conc=32 in=1024 out=500 ... |
| [2026-09-27 04:10:25 + 244.2m] [phase3 18/48] -> FAILED |
| [2026-09-27 04:10:25 + 244.2m] [phase3 19/48] conc=32 in=2048 out=1 ... |
| [2026-09-27 04:11:26 + 245.2m] [phase3 19/48] -> FAILED |
| [2026-09-27 04:11:26 + 245.2m] [phase3 20/48] conc=32 in=2048 out=10 ... |
| [2026-09-27 04:13:07 + 246.9m] [phase3 20/48] -> FAILED |
| [2026-09-27 04:13:07 + 246.9m] [phase3 21/48] conc=32 in=2048 out=50 ... |
| [2026-09-27 04:15:22 + 249.2m] [phase3 21/48] -> FAILED |
| [2026-09-27 04:15:22 + 249.2m] [phase3 22/48] conc=32 in=2048 out=100 ... |
| [2026-09-27 04:17:10 + 251.0m] [phase3 22/48] -> FAILED |
| [2026-09-27 04:17:10 + 251.0m] [phase3 23/48] conc=32 in=2048 out=250 ... |
| [2026-09-27 04:18:52 + 252.7m] [phase3 23/48] -> FAILED |
| [2026-09-27 04:18:52 + 252.7m] [phase3 24/48] conc=32 in=2048 out=500 ... |
| [2026-09-27 04:20:29 + 254.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 04:20:29 + 254.3m] [phase3 24/48] -> FAILED |
| [2026-09-27 04:20:29 + 254.3m] [phase3 25/48] conc=48 in=1024 out=1 ... |
| [2026-09-27 04:22:04 + 255.9m] [phase3 25/48] -> FAILED |
| [2026-09-27 04:22:04 + 255.9m] [phase3 26/48] conc=48 in=1024 out=10 ... |
| [2026-09-27 04:23:45 + 257.5m] [phase3 26/48] -> FAILED |
| [2026-09-27 04:23:45 + 257.5m] [phase3 27/48] conc=48 in=1024 out=50 ... |
| [2026-09-27 04:25:42 + 259.5m] [phase3 27/48] -> FAILED |
| [2026-09-27 04:25:42 + 259.5m] [phase3 28/48] conc=48 in=1024 out=100 ... |
| [2026-09-27 04:27:33 + 261.3m] [phase3 28/48] -> FAILED |
| [2026-09-27 04:27:33 + 261.3m] [phase3 29/48] conc=48 in=1024 out=250 ... |
| [2026-09-27 04:29:28 + 263.3m] [phase3 29/48] -> FAILED |
| [2026-09-27 04:29:28 + 263.3m] [phase3 30/48] conc=48 in=1024 out=500 ... |
| [2026-09-27 04:31:18 + 265.1m] [phase3 30/48] -> FAILED |
| [2026-09-27 04:31:18 + 265.1m] [phase3 31/48] conc=48 in=2048 out=1 ... |
| [2026-09-27 04:32:59 + 266.8m] [phase3 31/48] -> FAILED |
| [2026-09-27 04:32:59 + 266.8m] [phase3 32/48] conc=48 in=2048 out=10 ... |
| [2026-09-27 04:34:08 + 267.9m] [phase3 32/48] -> FAILED |
| [2026-09-27 04:34:08 + 267.9m] [phase3 33/48] conc=48 in=2048 out=50 ... |
| [2026-09-27 04:35:15 + 269.0m] [phase3 33/48] -> FAILED |
| [2026-09-27 04:35:15 + 269.0m] [phase3 34/48] conc=48 in=2048 out=100 ... |
| [2026-09-27 04:36:23 + 270.2m] [phase3 34/48] -> FAILED |
| [2026-09-27 04:36:23 + 270.2m] [phase3 35/48] conc=48 in=2048 out=250 ... |
| [2026-09-27 04:37:30 + 271.3m] [phase3 35/48] -> FAILED |
| [2026-09-27 04:37:30 + 271.3m] [phase3 36/48] conc=48 in=2048 out=500 ... |
| [2026-09-27 04:38:39 + 272.5m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 04:38:39 + 272.5m] [phase3 36/48] -> FAILED |
| [2026-09-27 04:38:39 + 272.5m] [phase3 37/48] conc=64 in=1024 out=1 ... |
| [2026-09-27 04:39:42 + 273.5m] [phase3 37/48] -> FAILED |
| [2026-09-27 04:39:42 + 273.5m] [phase3 38/48] conc=64 in=1024 out=10 ... |
| [2026-09-27 04:40:48 + 274.6m] [phase3 38/48] -> FAILED |
| [2026-09-27 04:40:48 + 274.6m] [phase3 39/48] conc=64 in=1024 out=50 ... |
| [2026-09-27 04:41:47 + 275.6m] [phase3 39/48] -> FAILED |
| [2026-09-27 04:41:47 + 275.6m] [phase3 40/48] conc=64 in=1024 out=100 ... |
| [2026-09-27 04:42:47 + 276.6m] [phase3 40/48] -> FAILED |
| [2026-09-27 04:42:47 + 276.6m] [phase3 41/48] conc=64 in=1024 out=250 ... |
| [2026-09-27 04:43:47 + 277.6m] [phase3 41/48] -> FAILED |
| [2026-09-27 04:43:47 + 277.6m] [phase3 42/48] conc=64 in=1024 out=500 ... |
| [2026-09-27 04:44:46 + 278.6m] [phase3 42/48] -> FAILED |
| [2026-09-27 04:44:46 + 278.6m] [phase3 43/48] conc=64 in=2048 out=1 ... |
| [2026-09-27 04:45:52 + 279.7m] [phase3 43/48] -> FAILED |
| [2026-09-27 04:45:52 + 279.7m] [phase3 44/48] conc=64 in=2048 out=10 ... |
| [2026-09-27 04:46:57 + 280.7m] [phase3 44/48] -> FAILED |
| [2026-09-27 04:46:57 + 280.7m] [phase3 45/48] conc=64 in=2048 out=50 ... |
| [2026-09-27 04:48:03 + 281.8m] [phase3 45/48] -> FAILED |
| [2026-09-27 04:48:03 + 281.8m] [phase3 46/48] conc=64 in=2048 out=100 ... |
| [2026-09-27 04:49:13 + 283.0m] [phase3 46/48] -> FAILED |
| [2026-09-27 04:49:13 + 283.0m] [phase3 47/48] conc=64 in=2048 out=250 ... |
| [2026-09-27 04:50:18 + 284.1m] [phase3 47/48] -> FAILED |
| [2026-09-27 04:50:18 + 284.1m] [phase3 48/48] conc=64 in=2048 out=500 ... |
| [2026-09-27 04:57:23 + 291.2m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 04:57:23 + 291.2m] [phase3 48/48] -> FAILED |
| [2026-09-27 04:57:24 + 291.2m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 04:57:26 + 291.2m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_state.json] |
| [2026-09-27 04:57:29 + 291.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_progress.log] |
| [2026-09-27 04:57:32 + 291.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/logs] |
| [2026-09-27 04:57:36 + 291.4m] [deleted weights: Qwen/Qwen3.6-35B-A3B-FP8] |
| [2026-09-27 04:57:38 + 291.4m] [deleted weights: z-lab/Qwen3.6-35B-A3B-DFlash] |
| [2026-09-27 04:57:38 + 291.4m] === qwen3.5-4b: ensuring weights are on disk === |
| [2026-09-27 04:57:40 + 291.5m] === Phase 1: config sweep === |
| [2026-09-27 04:57:40 + 291.5m] [phase1 1/3] qwen3.5-4b::base -- sweeping 6 configs ... |
| [2026-09-27 04:57:40 + 291.5m] [phase1 1/3] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 05:02:48 + 296.6m] [phase1 1/3] [1/6] -> throughput=2802.038759229529 (spread 9.5%) |
| [2026-09-27 05:02:48 + 296.6m] [phase1 1/3] [2/6] seqs=32 batched=12288 ... |
| [2026-09-27 05:07:41 + 301.5m] [phase1 1/3] [2/6] -> throughput=2696.326493066303 (spread 8.53%) |
| [2026-09-27 05:07:41 + 301.5m] [phase1 1/3] [3/6] seqs=32 batched=16384 ... |
| [2026-09-27 05:12:49 + 306.6m] [phase1 1/3] [3/6] -> throughput=2652.1234213651505 (spread 7.99%) |
| [2026-09-27 05:12:49 + 306.6m] [phase1 1/3] [4/6] seqs=64 batched=16384 ... |
| [2026-09-27 05:17:48 + 311.6m] [phase1 1/3] [4/6] -> throughput=2807.4734997890264 (spread 10.75%) |
| [2026-09-27 05:17:48 + 311.6m] [phase1 1/3] [5/6] seqs=64 batched=24576 ... |
| [2026-09-27 05:22:48 + 316.6m] [phase1 1/3] [5/6] -> throughput=2727.3317959276847 (spread 7.14%) |
| [2026-09-27 05:22:48 + 316.6m] [phase1 1/3] [6/6] seqs=64 batched=32768 ... |
| [2026-09-27 05:27:51 + 321.6m] [phase1 1/3] [6/6] -> throughput=2665.394936249848 (spread 9.35%) |
| [2026-09-27 05:27:51 + 321.6m] [phase1 1/3] qwen3.5-4b::base -- done, 6/6 configs succeeded, kept seqs=32/batched=8192, seqs=64/batched=16384 |
| [2026-09-27 05:27:51 + 321.6m] [phase1 2/3] qwen3.5-4b::mtp -- sweeping 6 configs ... |
| [2026-09-27 05:27:51 + 321.6m] [phase1 2/3] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 05:32:58 + 326.8m] [phase1 2/3] [1/6] -> throughput=2866.3087743698507 (spread 10.6%) |
| [2026-09-27 05:32:58 + 326.8m] [phase1 2/3] [2/6] seqs=32 batched=12288 ... |
| [2026-09-27 05:37:48 + 331.6m] [phase1 2/3] [2/6] -> throughput=2853.4876191614003 (spread 10.45%) |
| [2026-09-27 05:37:48 + 331.6m] [phase1 2/3] [3/6] seqs=32 batched=16384 ... |
| [2026-09-27 05:42:47 + 336.6m] [phase1 2/3] [3/6] -> throughput=2836.1270440806866 (spread 8.23%) |
| [2026-09-27 05:42:47 + 336.6m] [phase1 2/3] [4/6] seqs=64 batched=16384 ... |
| [2026-09-27 05:47:39 + 341.4m] [phase1 2/3] [4/6] -> throughput=2926.903603631148 (spread 14.96%) |
| [2026-09-27 05:47:39 + 341.4m] [phase1 2/3] [5/6] seqs=64 batched=24576 ... |
| [2026-09-27 05:52:43 + 346.5m] [phase1 2/3] [5/6] -> throughput=2912.5285099714283 (spread 11.21%) |
| [2026-09-27 05:52:43 + 346.5m] [phase1 2/3] [6/6] seqs=64 batched=32768 ... |
| [2026-09-27 05:57:47 + 351.6m] [phase1 2/3] [6/6] -> throughput=2883.667120273696 (spread 10.87%) |
| [2026-09-27 05:57:47 + 351.6m] [phase1 2/3] qwen3.5-4b::mtp -- done, 6/6 configs succeeded, kept seqs=32/batched=8192, seqs=64/batched=16384 |
| [2026-09-27 05:57:47 + 351.6m] [phase1 3/3] qwen3.5-4b::dflash -- sweeping 6 configs ... |
| [2026-09-27 05:57:47 + 351.6m] [phase1 3/3] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 06:03:26 + 357.2m] [phase1 3/3] [1/6] -> throughput=2131.5563884558796 (spread 20.24%) |
| [2026-09-27 06:03:26 + 357.2m] [phase1 3/3] [2/6] seqs=32 batched=12288 ... |
| [2026-09-27 06:08:59 + 362.8m] [phase1 3/3] [2/6] -> throughput=2273.2230507766585 (spread 7.24%) |
| [2026-09-27 06:08:59 + 362.8m] [phase1 3/3] [3/6] seqs=32 batched=16384 ... |
| [2026-09-27 06:14:15 + 368.0m] [phase1 3/3] [3/6] -> throughput=2227.9121242838564 (spread 7.05%) |
| [2026-09-27 06:14:15 + 368.0m] [phase1 3/3] [4/6] seqs=64 batched=16384 ... |
| [2026-09-27 06:19:34 + 373.4m] [phase1 3/3] [4/6] -> throughput=2404.485133835721 (spread 10.16%) |
| [2026-09-27 06:19:34 + 373.4m] [phase1 3/3] [5/6] seqs=64 batched=24576 ... |
| [2026-09-27 06:24:43 + 378.5m] [phase1 3/3] [5/6] -> throughput=2325.7857785065835 (spread 9.96%) |
| [2026-09-27 06:24:43 + 378.5m] [phase1 3/3] [6/6] seqs=64 batched=32768 ... |
| [2026-09-27 06:30:09 + 384.0m] [phase1 3/3] [6/6] -> throughput=2281.5057533911786 (spread 5.19%) |
| [2026-09-27 06:30:09 + 384.0m] [phase1 3/3] qwen3.5-4b::dflash -- done, 6/6 configs succeeded, kept seqs=32/batched=12288, seqs=64/batched=16384 |
| [2026-09-27 06:30:11 + 384.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 06:30:12 + 384.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_state.json] |
| [2026-09-27 06:30:14 + 384.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_progress.log] |
| [2026-09-27 06:30:16 + 384.1m] [pushed to hf: scaledown/vllm-inference-benchmarks/logs] |
| [2026-09-27 06:30:16 + 384.1m] === Phase 2: concurrency curve === |
| [2026-09-27 06:30:16 + 384.1m] [phase2 1/3] qwen3.5-4b::base -- 2 configs, 10 (config, concurrency) points ... |
| [2026-09-27 06:30:16 + 384.1m] [phase2 1/3] [1/2] seqs=32 batched=8192 -- starting server ... |
| [2026-09-27 06:34:41 + 388.5m] [phase2 1/3] [1/2] conc=8 -> throughput=1382.4704194886763 p99_e2el=4496.785503302235 (spread 3.24%) |
| [2026-09-27 06:36:30 + 390.3m] [phase2 1/3] [1/2] conc=16 -> throughput=1956.3292098617587 p99_e2el=6118.827587424306 (spread 1.7%) |
| [2026-09-27 06:38:29 + 392.3m] [phase2 1/3] [1/2] conc=24 -> throughput=2698.6565590981936 p99_e2el=6720.521317196546 (spread 2.17%) |
| [2026-09-27 06:40:43 + 394.5m] [phase2 1/3] [1/2] conc=32 -> throughput=2980.6650150136484 p99_e2el=8460.38014749996 (spread 2.74%) |
| [2026-09-27 06:40:44 + 394.5m] [phase2 1/3] [2/2] seqs=64 batched=16384 -- starting server ... |
| [2026-09-27 06:45:07 + 398.9m] [phase2 1/3] [2/2] conc=8 -> throughput=1361.7037027420433 p99_e2el=4516.908732482699 (spread 8.05%) |
| [2026-09-27 06:46:55 + 400.7m] [phase2 1/3] [2/2] conc=16 -> throughput=2081.2838150656535 p99_e2el=5750.532315776214 (spread 5.22%) |
| [2026-09-27 06:48:59 + 402.8m] [phase2 1/3] [2/2] conc=24 -> throughput=2737.1262606487708 p99_e2el=6689.1237444500475 (spread 4.3%) |
| [2026-09-27 06:51:07 + 404.9m] [phase2 1/3] [2/2] conc=32 -> throughput=3177.1727302083027 p99_e2el=7951.899707883127 (spread 0.65%) |
| [2026-09-27 06:53:25 + 407.2m] [phase2 1/3] [2/2] conc=48 -> throughput=3762.781997132622 p99_e2el=9786.471137348555 (spread 2.53%) |
| [2026-09-27 06:55:53 + 409.7m] [phase2 1/3] [2/2] conc=64 -> throughput=4417.555079336724 p99_e2el=11196.552431427406 (spread 1.66%) |
| [2026-09-27 06:55:54 + 409.7m] [phase2 1/3] qwen3.5-4b::base -- best: seqs=64 batched=16384 conc=24 throughput=2737.1262606487708 p99_e2el=6689.1237444500475 |
| [2026-09-27 06:55:56 + 409.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 06:55:56 + 409.7m] [phase2 2/3] qwen3.5-4b::mtp -- 2 configs, 10 (config, concurrency) points ... |
| [2026-09-27 06:55:56 + 409.7m] [phase2 2/3] [1/2] seqs=32 batched=8192 -- starting server ... |
| [2026-09-27 07:00:31 + 414.3m] [phase2 2/3] [1/2] conc=8 -> throughput=1287.5197979080717 p99_e2el=5207.8325141477435 (spread 1.73%) |
| [2026-09-27 07:02:28 + 416.3m] [phase2 2/3] [1/2] conc=16 -> throughput=1914.4650060984197 p99_e2el=6202.449175920335 (spread 14.48%) |
| [2026-09-27 07:04:35 + 418.4m] [phase2 2/3] [1/2] conc=24 -> throughput=2553.3307496177895 p99_e2el=7150.380029450172 (spread 8.86%) |
| [2026-09-27 07:06:48 + 420.6m] [phase2 2/3] [1/2] conc=32 -> throughput=2985.6027729677153 p99_e2el=8405.83817901788 (spread 9.52%) |
| [2026-09-27 07:06:49 + 420.6m] [phase2 2/3] [2/2] seqs=64 batched=16384 -- starting server ... |
| [2026-09-27 07:11:15 + 425.0m] [phase2 2/3] [2/2] conc=8 -> throughput=1159.3561012285613 p99_e2el=4805.276206142735 (spread 20.02%) |
| [2026-09-27 07:13:07 + 426.9m] [phase2 2/3] [2/2] conc=16 -> throughput=1884.2197868867875 p99_e2el=6157.578003644008 (spread 10.1%) |
| [2026-09-27 07:15:13 + 429.0m] [phase2 2/3] [2/2] conc=24 -> throughput=2600.791208975968 p99_e2el=7206.87676425325 (spread 16.4%) |
| [2026-09-27 07:17:20 + 431.1m] [phase2 2/3] [2/2] conc=32 -> throughput=3441.8126769358855 p99_e2el=6875.316595602781 (spread 3.53%) |
| [2026-09-27 07:19:33 + 433.3m] [phase2 2/3] [2/2] conc=48 -> throughput=4062.869998506481 p99_e2el=8902.087703694113 (spread 2.57%) |
| [2026-09-27 07:21:59 + 435.8m] [phase2 2/3] [2/2] conc=64 -> throughput=4543.185286484455 p99_e2el=10474.255268503903 (spread 4.05%) |
| [2026-09-27 07:22:00 + 435.8m] [phase2 2/3] qwen3.5-4b::mtp -- best: seqs=64 batched=16384 conc=32 throughput=3441.8126769358855 p99_e2el=6875.316595602781 |
| [2026-09-27 07:22:02 + 435.8m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 07:22:02 + 435.8m] [phase2 3/3] qwen3.5-4b::dflash -- 2 configs, 10 (config, concurrency) points ... |
| [2026-09-27 07:22:02 + 435.8m] [phase2 3/3] [1/2] seqs=32 batched=12288 -- starting server ... |
| [2026-09-27 07:26:30 + 440.3m] [phase2 3/3] [1/2] conc=8 -> throughput=1337.9705992920076 p99_e2el=4299.10745765781 (spread 8.29%) |
| [2026-09-27 07:28:27 + 442.2m] [phase2 3/3] [1/2] conc=16 -> throughput=1803.332090261689 p99_e2el=6448.375407834532 (spread 3.03%) |
| [2026-09-27 07:30:38 + 444.4m] [phase2 3/3] [1/2] conc=24 -> throughput=2221.254282111062 p99_e2el=8925.235627777878 (spread 0.48%) |
| [2026-09-27 07:33:14 + 447.0m] [phase2 3/3] [1/2] conc=32 -> throughput=2153.664440232534 p99_e2el=11516.361265724989 (spread 0.53%) |
| [2026-09-27 07:33:15 + 447.0m] [phase2 3/3] [2/2] seqs=64 batched=16384 -- starting server ... |
| [2026-09-27 07:37:45 + 451.6m] [phase2 3/3] [2/2] conc=8 -> throughput=1311.4151359675895 p99_e2el=4908.886131560432 (spread 1.92%) |
| [2026-09-27 07:39:41 + 453.5m] [phase2 3/3] [2/2] conc=16 -> throughput=1857.4243511383606 p99_e2el=6546.0376911750045 (spread 2.68%) |
| [2026-09-27 07:41:58 + 455.8m] [phase2 3/3] [2/2] conc=24 -> throughput=2204.6859608912982 p99_e2el=9266.638878998347 (spread 3.38%) |
| [2026-09-27 07:44:37 + 458.4m] [phase2 3/3] [2/2] conc=32 -> throughput=2249.8952043989093 p99_e2el=11251.29172846442 (spread 5.69%) |
| [2026-09-27 07:47:28 + 461.3m] [phase2 3/3] [2/2] conc=48 -> throughput=2575.2427177545055 p99_e2el=14091.132242332675 (spread 0.44%) |
| [2026-09-27 07:50:48 + 464.6m] [phase2 3/3] [2/2] conc=64 -> throughput=2908.4119969966796 p99_e2el=16387.99692353931 (spread 3.92%) |
| [2026-09-27 07:50:49 + 464.6m] [phase2 3/3] qwen3.5-4b::dflash -- best: seqs=32 batched=12288 conc=24 throughput=2221.254282111062 p99_e2el=8925.235627777878 |
| [2026-09-27 07:50:51 + 464.6m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 07:50:52 + 464.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 07:50:54 + 464.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_state.json] |
| [2026-09-27 07:50:56 + 464.7m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_progress.log] |
| [2026-09-27 07:50:59 + 464.8m] [pushed to hf: scaledown/vllm-inference-benchmarks/logs] |
| [2026-09-27 07:50:59 + 464.8m] qwen3.5-4b winner for phase 3: qwen3.5-4b::mtp (base=0.42, mtp=0.92, dflash=-1.00) |
| [2026-09-27 07:50:59 + 464.8m] === Phase 3: full IO matrix for qwen3.5-4b::mtp === |
| [2026-09-27 07:50:59 + 464.8m] downloading weights for Qwen/Qwen3.5-4B (deleted after its phase 2) ... |
| [2026-09-27 07:53:22 + 467.2m] 4 concurrencies [16, 32, 48, 64] x 12 IO cells = 48 points |
| [2026-09-27 07:53:22 + 467.2m] [phase3 1/48] conc=16 in=1024 out=1 ... |
| [2026-09-27 07:55:12 + 469.0m] [phase3 1/48] -> throughput=56.19317693629711 p99_e2el=362.9386209207587 (spread 2.95%) |
| [2026-09-27 07:55:12 + 469.0m] [phase3 2/48] conc=16 in=1024 out=10 ... |
| [2026-09-27 07:56:36 + 470.4m] [phase3 2/48] -> throughput=411.2857690836377 p99_e2el=526.52507163817 (spread 0.83%) |
| [2026-09-27 07:56:36 + 470.4m] [phase3 3/48] conc=16 in=1024 out=50 ... |
| [2026-09-27 07:58:14 + 472.0m] [phase3 3/48] -> throughput=1208.1587816808728 p99_e2el=997.1062562661242 (spread 3.85%) |
| [2026-09-27 07:58:14 + 472.0m] [phase3 4/48] conc=16 in=1024 out=100 ... |
| [2026-09-27 07:59:59 + 473.8m] [phase3 4/48] -> throughput=1746.9730614225598 p99_e2el=1307.9097158648071 (spread 1.3%) |
| [2026-09-27 07:59:59 + 473.8m] [phase3 5/48] conc=16 in=1024 out=250 ... |
| [2026-09-27 08:02:22 + 476.2m] [phase3 5/48] -> throughput=2560.797905724993 p99_e2el=2022.0175335137174 (spread 3.85%) |
| [2026-09-27 08:02:22 + 476.2m] [phase3 6/48] conc=16 in=1024 out=500 ... |
| [2026-09-27 08:05:43 + 479.5m] [phase3 6/48] -> throughput=2881.632581933668 p99_e2el=3841.247336377855 (spread 4.39%) |
| [2026-09-27 08:05:43 + 479.5m] [phase3 7/48] conc=16 in=2048 out=1 ... |
| [2026-09-27 08:07:13 + 481.0m] [phase3 7/48] -> throughput=29.82436710955609 p99_e2el=618.2033303449862 (spread 0.09%) |
| [2026-09-27 08:07:13 + 481.0m] [phase3 8/48] conc=16 in=2048 out=10 ... |
| [2026-09-27 08:08:46 + 482.6m] [phase3 8/48] -> throughput=253.59512476785017 p99_e2el=840.7425286434589 (spread 1.8%) |
| [2026-09-27 08:08:46 + 482.6m] [phase3 9/48] conc=16 in=2048 out=50 ... |
| [2026-09-27 08:10:33 + 484.3m] [phase3 9/48] -> throughput=914.9115084862553 p99_e2el=1253.9780498272735 (spread 4.53%) |
| [2026-09-27 08:10:33 + 484.3m] [phase3 10/48] conc=16 in=2048 out=100 ... |
| [2026-09-27 08:12:29 + 486.3m] [phase3 10/48] -> throughput=1463.7671238723447 p99_e2el=1545.8582455781288 (spread 3.09%) |
| [2026-09-27 08:12:29 + 486.3m] [phase3 11/48] conc=16 in=2048 out=250 ... |
| [2026-09-27 08:15:02 + 488.8m] [phase3 11/48] -> throughput=2235.303361602202 p99_e2el=2479.680108590983 (spread 2.12%) |
| [2026-09-27 08:15:02 + 488.8m] [phase3 12/48] conc=16 in=2048 out=500 ... |
| [2026-09-27 08:18:34 + 492.4m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 08:18:34 + 492.4m] [phase3 12/48] -> throughput=2641.296391183124 p99_e2el=4096.664342163131 (spread 0.32%) |
| [2026-09-27 08:18:34 + 492.4m] [phase3 13/48] conc=32 in=1024 out=1 ... |
| [2026-09-27 08:19:41 + 493.5m] [phase3 13/48] -> throughput=59.038116026519134 p99_e2el=626.7324358760379 (spread 0.1%) |
| [2026-09-27 08:19:41 + 493.5m] [phase3 14/48] conc=32 in=1024 out=10 ... |
| [2026-09-27 08:20:56 + 494.7m] [phase3 14/48] -> throughput=477.860316396208 p99_e2el=983.1316378968766 (spread 0.84%) |
| [2026-09-27 08:20:56 + 494.7m] [phase3 15/48] conc=32 in=1024 out=50 ... |
| [2026-09-27 08:22:24 + 496.2m] [phase3 15/48] -> throughput=1602.335023281478 p99_e2el=1487.8992720204396 (spread 2.23%) |
| [2026-09-27 08:22:24 + 496.2m] [phase3 16/48] conc=32 in=1024 out=100 ... |
| [2026-09-27 08:24:01 + 497.8m] [phase3 16/48] -> throughput=2428.3138660107443 p99_e2el=1898.5968019091508 (spread 5.46%) |
| [2026-09-27 08:24:01 + 497.8m] [phase3 17/48] conc=32 in=1024 out=250 ... |
| [2026-09-27 08:26:03 + 499.8m] [phase3 17/48] -> throughput=3683.8690346688863 p99_e2el=2911.276471288874 (spread 0.43%) |
| [2026-09-27 08:26:03 + 499.8m] [phase3 18/48] conc=32 in=1024 out=500 ... |
| [2026-09-27 08:28:49 + 502.6m] [phase3 18/48] -> throughput=4419.620238245054 p99_e2el=4684.405783864204 (spread 2.39%) |
| [2026-09-27 08:28:49 + 502.6m] [phase3 19/48] conc=32 in=2048 out=1 ... |
| [2026-09-27 08:30:21 + 504.1m] [phase3 19/48] -> throughput=29.74386881494563 p99_e2el=1274.6252269321121 (spread 0.11%) |
| [2026-09-27 08:30:21 + 504.1m] [phase3 20/48] conc=32 in=2048 out=10 ... |
| [2026-09-27 08:31:55 + 505.7m] [phase3 20/48] -> throughput=269.55012724037545 p99_e2el=1709.4645233266056 (spread 0.11%) |
| [2026-09-27 08:31:55 + 505.7m] [phase3 21/48] conc=32 in=2048 out=50 ... |
| [2026-09-27 08:33:35 + 507.4m] [phase3 21/48] -> throughput=1117.1475275267294 p99_e2el=2383.2553049852136 (spread 1.93%) |
| [2026-09-27 08:33:35 + 507.4m] [phase3 22/48] conc=32 in=2048 out=100 ... |
| [2026-09-27 08:35:26 + 509.2m] [phase3 22/48] -> throughput=1824.381432856197 p99_e2el=2705.0012357789074 (spread 2.63%) |
| [2026-09-27 08:35:26 + 509.2m] [phase3 23/48] conc=32 in=2048 out=250 ... |
| [2026-09-27 08:37:46 + 511.6m] [phase3 23/48] -> throughput=2953.657286840711 p99_e2el=3811.409963464829 (spread 0.17%) |
| [2026-09-27 08:37:46 + 511.6m] [phase3 24/48] conc=32 in=2048 out=500 ... |
| [2026-09-27 08:40:51 + 514.6m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 08:40:51 + 514.6m] [phase3 24/48] -> throughput=3812.831158779667 p99_e2el=5292.623742043505 (spread 0.64%) |
| [2026-09-27 08:40:51 + 514.6m] [phase3 25/48] conc=48 in=1024 out=1 ... |
| [2026-09-27 08:42:03 + 515.8m] [phase3 25/48] -> throughput=58.5635804166532 p99_e2el=1011.7745892424136 (spread 0.37%) |
| [2026-09-27 08:42:03 + 515.8m] [phase3 26/48] conc=48 in=1024 out=10 ... |
| [2026-09-27 08:43:19 + 517.1m] [phase3 26/48] -> throughput=505.0733949861659 p99_e2el=1498.7039315421134 (spread 1.4%) |
| [2026-09-27 08:43:19 + 517.1m] [phase3 27/48] conc=48 in=1024 out=50 ... |
| [2026-09-27 08:44:39 + 518.4m] [phase3 27/48] -> throughput=1876.0904060370149 p99_e2el=2025.2580325794406 (spread 2.54%) |
| [2026-09-27 08:44:39 + 518.4m] [phase3 28/48] conc=48 in=1024 out=100 ... |
| [2026-09-27 08:46:08 + 519.9m] [phase3 28/48] -> throughput=2851.128349931442 p99_e2el=2496.949089434929 (spread 0.57%) |
| [2026-09-27 08:46:08 + 519.9m] [phase3 29/48] conc=48 in=1024 out=250 ... |
| [2026-09-27 08:47:57 + 521.7m] [phase3 29/48] -> throughput=4255.9426113386 p99_e2el=3890.7127427821974 (spread 5.73%) |
| [2026-09-27 08:47:57 + 521.7m] [phase3 30/48] conc=48 in=1024 out=500 ... |
| [2026-09-27 08:50:23 + 524.2m] [phase3 30/48] -> throughput=5298.053294058312 p99_e2el=5592.800034261309 (spread 2.39%) |
| [2026-09-27 08:50:23 + 524.2m] [phase3 31/48] conc=48 in=2048 out=1 ... |
| [2026-09-27 08:51:57 + 525.8m] [phase3 31/48] -> throughput=29.71026926348332 p99_e2el=1779.5179906953126 (spread 0.19%) |
| [2026-09-27 08:51:57 + 525.8m] [phase3 32/48] conc=48 in=2048 out=10 ... |
| [2026-09-27 08:53:33 + 527.3m] [phase3 32/48] -> throughput=276.9398724138838 p99_e2el=2506.3197186775496 (spread 0.0%) |
| [2026-09-27 08:53:33 + 527.3m] [phase3 33/48] conc=48 in=2048 out=50 ... |
| [2026-09-27 08:55:13 + 529.0m] [phase3 33/48] -> throughput=1160.2646041167293 p99_e2el=3351.4941157307458 (spread 1.79%) |
| [2026-09-27 08:55:13 + 529.0m] [phase3 34/48] conc=48 in=2048 out=100 ... |
| [2026-09-27 08:57:03 + 530.9m] [phase3 34/48] -> throughput=1956.4202585012267 p99_e2el=3970.8697090158234 (spread 0.4%) |
| [2026-09-27 08:57:03 + 530.9m] [phase3 35/48] conc=48 in=2048 out=250 ... |
| [2026-09-27 08:59:16 + 533.1m] [phase3 35/48] -> throughput=3318.0936422976993 p99_e2el=5240.427681414876 (spread 0.57%) |
| [2026-09-27 08:59:16 + 533.1m] [phase3 36/48] conc=48 in=2048 out=500 ... |
| [2026-09-27 09:02:14 + 536.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 09:02:14 + 536.0m] [phase3 36/48] -> throughput=4268.1524849012785 p99_e2el=7408.222002417315 (spread 0.05%) |
| [2026-09-27 09:02:14 + 536.0m] [phase3 37/48] conc=64 in=1024 out=1 ... |
| [2026-09-27 09:03:26 + 537.2m] [phase3 37/48] -> throughput=58.47775618877699 p99_e2el=1270.767114290502 (spread 0.18%) |
| [2026-09-27 09:03:26 + 537.2m] [phase3 38/48] conc=64 in=1024 out=10 ... |
| [2026-09-27 09:04:41 + 538.5m] [phase3 38/48] -> throughput=519.3664373956368 p99_e2el=1550.3142174356617 (spread 0.99%) |
| [2026-09-27 09:04:41 + 538.5m] [phase3 39/48] conc=64 in=1024 out=50 ... |
| [2026-09-27 09:06:00 + 539.8m] [phase3 39/48] -> throughput=2034.105858688633 p99_e2el=2683.1140627502464 (spread 1.68%) |
| [2026-09-27 09:06:00 + 539.8m] [phase3 40/48] conc=64 in=1024 out=100 ... |
| [2026-09-27 09:07:29 + 541.3m] [phase3 40/48] -> throughput=3005.578507604845 p99_e2el=3106.965952832252 (spread 2.55%) |
| [2026-09-27 09:07:29 + 541.3m] [phase3 41/48] conc=64 in=1024 out=250 ... |
| [2026-09-27 09:09:17 + 543.1m] [phase3 41/48] -> throughput=4789.120068114762 p99_e2el=4520.574311020318 (spread 1.75%) |
| [2026-09-27 09:09:17 + 543.1m] [phase3 42/48] conc=64 in=1024 out=500 ... |
| [2026-09-27 09:11:37 + 545.4m] [phase3 42/48] -> throughput=5952.527027834622 p99_e2el=6920.018189554102 (spread 1.58%) |
| [2026-09-27 09:11:37 + 545.4m] [phase3 43/48] conc=64 in=2048 out=1 ... |
| [2026-09-27 09:13:08 + 546.9m] [phase3 43/48] -> throughput=29.627278284252903 p99_e2el=2291.6336202877574 (spread 0.03%) |
| [2026-09-27 09:13:08 + 546.9m] [phase3 44/48] conc=64 in=2048 out=10 ... |
| [2026-09-27 09:14:41 + 548.5m] [phase3 44/48] -> throughput=277.0154197718353 p99_e2el=3674.3496711319312 (spread 0.13%) |
| [2026-09-27 09:14:41 + 548.5m] [phase3 45/48] conc=64 in=2048 out=50 ... |
| [2026-09-27 09:16:22 + 550.2m] [phase3 45/48] -> throughput=1174.6968759441024 p99_e2el=4504.880332809406 (spread 0.18%) |
| [2026-09-27 09:16:22 + 550.2m] [phase3 46/48] conc=64 in=2048 out=100 ... |
| [2026-09-27 09:18:06 + 551.9m] [phase3 46/48] -> throughput=1997.9354260363116 p99_e2el=5228.168547146487 (spread 0.9%) |
| [2026-09-27 09:18:06 + 551.9m] [phase3 47/48] conc=64 in=2048 out=250 ... |
| [2026-09-27 09:20:17 + 554.1m] [phase3 47/48] -> throughput=3454.171634156587 p99_e2el=6883.646148391998 (spread 0.26%) |
| [2026-09-27 09:20:17 + 554.1m] [phase3 48/48] conc=64 in=2048 out=500 ... |
| [2026-09-27 09:23:05 + 556.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 09:23:05 + 556.9m] [phase3 48/48] -> throughput=4518.175883717349 p99_e2el=9453.931180511136 (spread 0.04%) |
| [2026-09-27 09:23:07 + 556.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 09:23:09 + 556.9m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_state.json] |
| [2026-09-27 09:23:11 + 557.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_progress.log] |
| [2026-09-27 09:23:14 + 557.0m] [pushed to hf: scaledown/vllm-inference-benchmarks/logs] |
| [2026-09-27 09:23:16 + 557.1m] [deleted weights: Qwen/Qwen3.5-4B] |
| [2026-09-27 09:23:17 + 557.1m] [deleted weights: z-lab/Qwen3.5-4B-DFlash] |
| [2026-09-27 09:23:17 + 557.1m] === muse-glimmer-30b: ensuring weights are on disk === |
| [2026-09-27 09:23:19 + 557.1m] === Phase 1: config sweep === |
| [2026-09-27 09:23:19 + 557.1m] [phase1 1/2] muse-glimmer-30b::base -- sweeping 6 configs ... |
| [2026-09-27 09:23:19 + 557.1m] [phase1 1/2] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 09:30:46 + 564.6m] [phase1 1/2] [1/6] -> throughput=1148.9368950026046 (spread 4.88%) |
| [2026-09-27 09:30:46 + 564.6m] [phase1 1/2] [2/6] seqs=32 batched=12288 ... |
| [2026-09-27 09:36:47 + 570.6m] [phase1 1/2] [2/6] -> throughput=1140.7683323099263 (spread 6.09%) |
| [2026-09-27 09:36:47 + 570.6m] [phase1 1/2] [3/6] seqs=32 batched=16384 ... |
| [2026-09-27 09:42:47 + 576.6m] [phase1 1/2] [3/6] -> throughput=1142.5916224322068 (spread 6.34%) |
| [2026-09-27 09:42:47 + 576.6m] [phase1 1/2] [4/6] seqs=64 batched=16384 ... |
| [2026-09-27 09:48:47 + 582.6m] [phase1 1/2] [4/6] -> throughput=1163.4905245848242 (spread 6.22%) |
| [2026-09-27 09:48:47 + 582.6m] [phase1 1/2] [5/6] seqs=64 batched=24576 ... |
| [2026-09-27 09:54:48 + 588.6m] [phase1 1/2] [5/6] -> throughput=1142.8214292797443 (spread 6.33%) |
| [2026-09-27 09:54:48 + 588.6m] [phase1 1/2] [6/6] seqs=64 batched=32768 ... |
| [2026-09-27 10:00:55 + 594.7m] [phase1 1/2] [6/6] -> throughput=1150.0936071506 (spread 5.12%) |
| [2026-09-27 10:00:55 + 594.7m] [phase1 1/2] muse-glimmer-30b::base -- done, 6/6 configs succeeded, kept seqs=32/batched=8192, seqs=64/batched=16384 |
| [2026-09-27 10:00:55 + 594.7m] [phase1 2/2] muse-glimmer-30b::dflash -- sweeping 6 configs ... |
| [2026-09-27 10:00:55 + 594.7m] [phase1 2/2] [1/6] seqs=32 batched=8192 ... |
| [2026-09-27 10:06:56 + 600.7m] [phase1 2/2] [1/6] -> throughput=1436.6649132855614 (spread 4.7%) |
| [2026-09-27 10:06:56 + 600.7m] [phase1 2/2] [2/6] seqs=32 batched=12288 ... |
| [2026-09-27 10:12:33 + 606.3m] [phase1 2/2] [2/6] -> throughput=1425.1457226322657 (spread 5.09%) |
| [2026-09-27 10:12:33 + 606.3m] [phase1 2/2] [3/6] seqs=32 batched=16384 ... |
| [2026-09-27 10:18:13 + 612.0m] [phase1 2/2] [3/6] -> throughput=1439.6366836090324 (spread 5.73%) |
| [2026-09-27 10:18:13 + 612.0m] [phase1 2/2] [4/6] seqs=64 batched=16384 ... |
| [2026-09-27 10:23:54 + 617.7m] [phase1 2/2] [4/6] -> throughput=1409.263152936588 (spread 6.02%) |
| [2026-09-27 10:23:54 + 617.7m] [phase1 2/2] [5/6] seqs=64 batched=24576 ... |
| [2026-09-27 10:29:43 + 623.5m] [phase1 2/2] [5/6] -> throughput=1402.1112710819825 (spread 6.55%) |
| [2026-09-27 10:29:43 + 623.5m] [phase1 2/2] [6/6] seqs=64 batched=32768 ... |
| [2026-09-27 10:35:30 + 629.3m] [phase1 2/2] [6/6] -> throughput=1404.1261577435369 (spread 6.26%) |
| [2026-09-27 10:35:30 + 629.3m] [phase1 2/2] muse-glimmer-30b::dflash -- done, 6/6 configs succeeded, kept seqs=32/batched=16384, seqs=64/batched=16384 |
| [2026-09-27 10:35:31 + 629.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_results.csv] |
| [2026-09-27 10:35:33 + 629.3m] [pushed to hf: scaledown/vllm-inference-benchmarks/orchestrate_state.json] |
|
|
