Strata IQ3_S 速度優化前後與原始 Log
日期：2026-10-22（台北時間）
本文件把優化前、逐步調整、負對照和正式重跑的原始 Strata engine log 摘錄合併在同一份 Markdown。以下 log 行保留原文、來源檔名與原始行號，可直接複製到貼文，不依賴本機檔案連結。
結論
這次把速度推到 70 多 tok/s 的主要有效改動是 --pool-workers 從 15 增至 37。本機是 Xeon 38 核 / 76 執行緒；原先照用社群 16 核主機的 15 workers，沒有充分使用本機 CPU expert pool。
測試
15 workers
24 workers
37 workers
結果
短 prompt，1494 tokens、512 輸出，中位數
67.5 tok/s
75.5 tok/s
78.7 tok/s
37 比 15 快約 16.6%
同一組約 99K prompt、256 輸出，中位數
67.3 tok/s
未測長 prompt
75.2 tok/s
配對提升約 11.7%
正式重啟後同組 99K prompt，中位數
控制組 67.3 tok/s
—
74.6 tok/s
重跑仍快約 10.8%
正式服務短 prompt smoke
—
—
81.2 tok/s
單次 smoke，不當作中位數
測試條件
同一模型：GSQ-RCO abliterated IQ3_S，model ID gsq-rco-qwen3.8-flash-next-iq3_s。
同一引擎：Strata 0.1.45、HIP gfx1100、RX 7900 XTX。
15／24／37 workers 配對測試中，唯一設定差異是 --pool-workers；configuration.diff 的原文如下：
--- pool15-config.json
+++ pool37-config.json
@@ -14,7 +14,7 @@
"--pcie-frac",
"0.0",
"--pool-workers",
- "15",
+ "37",
"--prefill",
"auto:32768",
"--spec",
長 prompt 每次實際輸入 99,004–99,005 tokens，輸出上限 256 tokens，cache reuse=0。短 prompt 使用 1494-token fresh prompt，輸出上限 512 tokens。
benchmark 的 API sampling 固定為 temperature 0、top-k 1、top-p 1、min-p 0、seed 42、reasoning_effort=none；這是速度測試條件，不是 Hermes Agent 的日常 reasoning 設定。
Prefill 在各組約 1,712–1,735 tok/s，變動很小；提升主要出現在 decode。
原始 Strata Log：15 workers 優化前
來源 pool15-engine.log。啟動行顯示 15 workers；三筆相同 99K prompt 的 decode 為 57.6、67.3、70.1 tok/s，中位數 67.3。
pool15-engine.log:1835 [strata] 2026-10-22 00:53:00 engine started: /home/jaran/Strata-0.1.45/build-hip-rocm10/strata --serve --pack /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/pack --native /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf --ple-gguf /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00002-of-00002.gguf --expert-profile /home/jaran/Strata-0.1.45/data/expert-profile.bin --expert-cache 7700 --pcie-frac 0.0 --pool-workers 15 --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/strata/rt --max-context 262144 --kv k8v4 --kv-resident 32768 --vram-reserve-mib 700 --adapt-every 0 --prompt-cache 16 --prompt-cache-every 16384 --conversation-cache-mib 16384 --conversation-cache-slots 8 --conversation-cache-min-tokens 32768 --vision
pool15-engine.log:1853 strata generate: 15 pool workers on logical processors 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15, the host thread on 0 (draining too) (--host-core first)
pool15-engine.log:1862 strata generate: 15 expert-pool workers + the host thread
pool15-engine.log:1896 strata serve: prompt 99004 tokens = 0 reused + 99004 read in 57473 ms (1722.6 tok/s), 256 generated in 4445 ms (57.6 tok/s), drafts accepted 132 of 223, 4 checkpoints
pool15-engine.log:1897 strata serve: decode expert cache hit rate: 78.1% (130028 hits / 166560 lookups)
pool15-engine.log:1906 strata serve: prompt 99005 tokens = 0 reused + 99005 read in 57737 ms (1714.7 tok/s), 256 generated in 3806 ms (67.3 tok/s), drafts accepted 152 of 214, 4 checkpoints
pool15-engine.log:1907 strata serve: decode expert cache hit rate: 79.6% (121464 hits / 152640 lookups)
pool15-engine.log:1915 strata serve: prompt 99005 tokens = 0 reused + 99005 read in 57815 ms (1712.4 tok/s), 256 generated in 3652 ms (70.1 tok/s), drafts accepted 158 of 201, 4 checkpoints
pool15-engine.log:1916 strata serve: decode expert cache hit rate: 80.2% (115908 hits / 144480 lookups)
15-worker short prompt control 的三筆原始輸出：
pool15-engine.log:1780 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1925 ms (776.1 tok/s), 512 generated in 7587 ms (67.5 tok/s), drafts accepted 312 of 454, 1 checkpoints
pool15-engine.log:1795 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1936 ms (771.6 tok/s), 512 generated in 7450 ms (68.7 tok/s), drafts accepted 313 of 455, 1 checkpoints
pool15-engine.log:1800 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1934 ms (772.4 tok/s), 512 generated in 7702 ms (66.5 tok/s), drafts accepted 310 of 463, 1 checkpoints
中間值：24 workers
短 prompt 三筆為 75.4、75.5、75.6 tok/s；另有正式追加一筆 74.7。這是短 prompt 的中間檔位，沒有以 24 workers 跑同一組 99K 長 prompt。
pool24-engine.log:3 [strata] 2026-10-22 00:44:27 engine started: /home/jaran/Strata-0.1.45/build-hip-rocm10/strata --serve --pack /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/pack --native /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf --ple-gguf /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00002-of-00002.gguf --expert-profile /home/jaran/Strata-0.1.45/data/expert-profile.bin --expert-cache 7700 --pcie-frac 0.0 --pool-workers 24 --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/strata/rt --max-context 262144 --kv k8v4 --kv-resident 32768 --vram-reserve-mib 700 --adapt-every 0 --prompt-cache 16 --prompt-cache-every 16384 --conversation-cache-mib 16384 --conversation-cache-slots 8 --conversation-cache-min-tokens 32768 --vision
pool24-engine.log:21 strata generate: 24 pool workers on logical processors 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24, the host thread on 0 (draining too) (--host-core first)
pool24-engine.log:30 strata generate: 24 expert-pool workers + the host thread
pool24-engine.log:63 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1910 ms (782.4 tok/s), 512 generated in 6790 ms (75.4 tok/s), drafts accepted 314 of 461, 1 checkpoints
pool24-engine.log:94 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1917 ms (779.4 tok/s), 512 generated in 6784 ms (75.5 tok/s), drafts accepted 314 of 461, 1 checkpoints
pool24-engine.log:108 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1920 ms (778.0 tok/s), 512 generated in 6771 ms (75.6 tok/s), drafts accepted 309 of 466, 1 checkpoints
pool24-engine.log:112 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1921 ms (777.8 tok/s), 512 generated in 6853 ms (74.7 tok/s), drafts accepted 308 of 461, 1 checkpoints
原始 Strata Log：37 workers 優化後
短 prompt 三筆為 78.7、78.8、78.0 tok/s，中位數 78.7。99K 長 prompt 三筆為 65.3、75.2、77.9 tok/s，中位數 75.2。前三組中前兩組 MTP 接受數相同，但 decode 仍提高。
pool37-engine.log:3 [strata] 2026-10-22 00:47:14 engine started: /home/jaran/Strata-0.1.45/build-hip-rocm10/strata --serve --pack /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/pack --native /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf --ple-gguf /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00002-of-00002.gguf --expert-profile /home/jaran/Strata-0.1.45/data/expert-profile.bin --expert-cache 7700 --pcie-frac 0.0 --pool-workers 37 --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/strata/rt --max-context 262144 --kv k8v4 --kv-resident 32768 --vram-reserve-mib 700 --adapt-every 0 --prompt-cache 16 --prompt-cache-every 16384 --conversation-cache-mib 16384 --conversation-cache-slots 8 --conversation-cache-min-tokens 32768 --vision
pool37-engine.log:21 strata generate: 37 pool workers on logical processors 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37, the host thread on 0 (draining too) (--host-core first)
pool37-engine.log:30 strata generate: 37 expert-pool workers + the host thread
pool37-engine.log:63 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1906 ms (783.7 tok/s), 512 generated in 6505 ms (78.7 tok/s), drafts accepted 314 of 461, 1 checkpoints
pool37-engine.log:76 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1915 ms (780.3 tok/s), 512 generated in 6495 ms (78.8 tok/s), drafts accepted 309 of 466, 1 checkpoints
pool37-engine.log:89 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1909 ms (782.7 tok/s), 512 generated in 6565 ms (78.0 tok/s), drafts accepted 308 of 461, 1 checkpoints
pool37-engine.log:111 strata serve: prompt 99004 tokens = 0 reused + 99004 read in 57555 ms (1720.1 tok/s), 256 generated in 3923 ms (65.3 tok/s), drafts accepted 132 of 223, 4 checkpoints
pool37-engine.log:121 strata serve: prompt 99005 tokens = 0 reused + 99005 read in 57818 ms (1712.4 tok/s), 256 generated in 3405 ms (75.2 tok/s), drafts accepted 152 of 214, 4 checkpoints
pool37-engine.log:130 strata serve: prompt 99005 tokens = 0 reused + 99005 read in 58614 ms (1689.1 tok/s), 256 generated in 3288 ms (77.9 tok/s), drafts accepted 160 of 220, 4 checkpoints
負對照：37 workers 下測試 PCIe 0.10，但沒有保留
短 prompt 雖然到 80.6–82.9 tok/s，99K 長 prompt 波動很大：52.9、75.8、80.9，中位數 75.8，和 PCIe 0 的 75.2 差異不構成穩定收益，因此正式設定仍用 PCIe 0。
pool37-pcie010-engine.log:3 [strata] 2026-10-22 00:57:30 engine started: /home/jaran/Strata-0.1.45/build-hip-rocm10/strata --serve --pack /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/pack --native /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf --ple-gguf /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00002-of-00002.gguf --expert-profile /home/jaran/Strata-0.1.45/data/expert-profile.bin --expert-cache 7700 --pcie-frac 0.10 --pool-workers 37 --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/strata/rt --max-context 262144 --kv k8v4 --kv-resident 32768 --vram-reserve-mib 700 --adapt-every 0 --prompt-cache 16 --prompt-cache-every 16384 --conversation-cache-mib 16384 --conversation-cache-slots 8 --conversation-cache-min-tokens 32768 --vision
pool37-pcie010-engine.log:21 strata generate: 37 pool workers on logical processors 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37, the host thread on 0 (draining too) (--host-core first)
pool37-pcie010-engine.log:64 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1907 ms (783.5 tok/s), 512 generated in 6352 ms (80.6 tok/s), drafts accepted 315 of 447, 1 checkpoints
pool37-pcie010-engine.log:77 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 5043 ms (296.3 tok/s), 512 generated in 6177 ms (82.9 tok/s), drafts accepted 328 of 434, 1 checkpoints
pool37-pcie010-engine.log:91 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1910 ms (782.1 tok/s), 512 generated in 6288 ms (81.4 tok/s), drafts accepted 319 of 440, 1 checkpoints
pool37-pcie010-engine.log:104 strata serve: prompt 99004 tokens = 0 reused + 99004 read in 57510 ms (1721.5 tok/s), 256 generated in 4840 ms (52.9 tok/s), drafts accepted 149 of 250, 4 checkpoints
pool37-pcie010-engine.log:114 strata serve: prompt 99005 tokens = 0 reused + 99005 read in 58265 ms (1699.2 tok/s), 256 generated in 3377 ms (75.8 tok/s), drafts accepted 156 of 227, 4 checkpoints
pool37-pcie010-engine.log:118 strata serve: prompt 99005 tokens = 0 reused + 99005 read in 58953 ms (1679.4 tok/s), 256 generated in 3165 ms (80.9 tok/s), drafts accepted 165 of 210, 4 checkpoints
正式設定重啟後重跑：37 workers
正式 launcher 重啟後，短 prompt smoke 是 81.2 tok/s；同組 99K prompt 重跑為 64.6、74.6、77.4，中位數 74.6。此組最低一筆低於 70，表示 70+ 是中位附近，不是每次固定速度。
active-final-engine.log:1936 [strata] 2026-10-22 01:03:00 engine started: /home/jaran/Strata-0.1.45/build-hip-rocm10/strata --serve --pack /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/pack --native /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00001-of-00002.gguf --ple-gguf /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-00002-of-00002.gguf --expert-profile /home/jaran/Strata-0.1.45/data/expert-profile.bin --expert-cache 7700 --pcie-frac 0.0 --pool-workers 37 --prefill auto:32768 --spec 4 --spec-min-p 0.5 --mtp /home/jaran/models/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ3_S-Strata-2938/model/strata/rt --max-context 262144 --kv k8v4 --kv-resident 32768 --vram-reserve-mib 700 --adapt-every 0 --prompt-cache 16 --prompt-cache-every 16384 --conversation-cache-mib 16384 --conversation-cache-slots 8 --conversation-cache-min-tokens 32768 --vision
active-final-engine.log:1954 strata generate: 37 pool workers on logical processors 1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37, the host thread on 0 (draining too) (--host-core first)
active-final-engine.log:1997 strata serve: prompt 1494 tokens = 0 reused + 1494 read in 1914 ms (780.4 tok/s), 512 generated in 6306 ms (81.2 tok/s), drafts accepted 316 of 453, 1 checkpoints
active-final-engine.log:2148 strata serve: prompt 99004 tokens = 0 reused + 99004 read in 57069 ms (1734.8 tok/s), 256 generated in 3964 ms (64.6 tok/s), drafts accepted 132 of 223, 4 checkpoints
active-final-engine.log:2161 strata serve: prompt 99005 tokens = 0 reused + 99005 read in 57501 ms (1721.8 tok/s), 256 generated in 3432 ms (74.6 tok/s), drafts accepted 152 of 214, 4 checkpoints
active-final-engine.log:2170 strata serve: prompt 99005 tokens = 0 reused + 99005 read in 57655 ms (1717.2 tok/s), 256 generated in 3307 ms (77.4 tok/s), drafts accepted 157 of 211, 4 checkpoints
結果界線
這是速度測試，不是 coding、Hermes 工具任務或 Telegram 任務測試；沒有測工具結果正確率、任務完成率或 soak。
短測 1494-token prompt 生成 512 tokens；99K prompt 生成 256 tokens。99K prompt 是本機固定合成英文段落，不是社群文章逐字 prompt。
速度改善最有力的證據是同模型／同 engine／同 prompt 下 worker-only 配對 A/B，再加正式重啟重跑；不能推論每一種 prompt 都會固定 75 tok/s。
15/24/37、PCIe 負對照和正式重跑的上述原始 engine log 行已直接嵌入本文，不需讀者存取本機檔案或點擊附件連結。
