#6258 [Benchmark] AMD Radeon 880M (gfx1150, Linux) / Strata 0.2.12. swift-qwen38-flash-next-iq2xs-abliterated-IQ2_XS
closed · @kozhini · 1 コメント · GitHub で見る
BenchmarksSetup & installAMD / HIPNVIDIA / CUDAModels & quantsSecurityLinux
本文
# Community benchmark on AMD Radeon 880M Graphics Measured on 2026-11-19 with `tools/strata_bench.py` (suite 1, 3 runs per test, medians with ranges). Speeds are the engine's own timings; first-token times are measured by the client over HTTP. Speed only: no correctness check was run. Prefill preferred. ## Hardware and software - GPU and VRAM: AMD Radeon 880M Graphics (gfx1150, 12 CU), 48 GiB GTT carve-out (set by kernel boot parameters; rocminfo reports 0 GiB dedicated VRAM) - CPU: AMD Ryzen AI 9 365 w/ Radeon 880M (10 cores, 20 threads); installed RAM: 64 GiB - OS: Ubuntu 26.04.1 LTS, Linux 7.0.0-43-generic; GPU driver: ROCk 7.1.3.31500000; ROCm core 10.0 / HIP 7.15.26333; CUDA: not used - Strata: project 0.2.12 (git tag v0.2.12), commit 146df3c4; engine 0.1.58 - Background workloads: not recorded (GPU power limit/draw read above) ## Model and configuration - Model and quantization: swift-qwen38-flash-next-iq2xs-abliterated-IQ2_XS - Context: 131072 tokens; parallel slots: 1 - Settings: tuned configuration (hip backend; decode/MTP flags from issue #6146 combined with local prefill settings, expert cache 6474, 128K context); thinking off for the suite; temperature 0 ```text strata --pack pack --native Swift-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ2_XS-00001-of-00002.gguf --ple-gguf Swift-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ2_XS-00001-of-00002.gguf --expert-profile expert-profile.bin --mtp rt_full --mtp-draft-vocab draft_vocab.bin --expert-cache 6474 --prefill 32768 --kv int8 --spec 4 --spec-min-p 0.5 --mmap-experts --max-context 131072 ``` Environment of the config: `STRATA_HIPBLASLT_TUNING=gfx1150-hipblaslt-100401.txt`, `STRATA_HIP_WMMA=1`, `STRATA_PF_GEMM=1`, `STRATA_PA_FAST=1`, `STRATA_SELECT_WMMA=1`, `STRATA_HC_UPMIX=1`, `STRATA_GDN_CHUNKED=1`, `STRATA_CVEC_FUSE=1`, `STRATA_PF_PAD=1`, `STRATA_TOPK_REG_MIN_BLOCKS=4096`, `STRATA_QSA_FUSE_NR=1`, `STRATA_QSA_FUSE_GATE=1`, `STRATA_GDN_HEAD=1`, `STRATA_GDN_PP=2`, `STRATA_GDN_CONVL2=1`, `STRATA_GDN_NOY=1`, `STRATA_Q8_PACKED=1`, `STRATA_Q6_PACKED=1`, `STRATA_MMVF_ROWS=1`, `STRATA_ATTN_LANECELL=1`, `STRATA_EXPERT_V2=1`, `STRATA_TSUM=1`, `STRATA_LFUSE=1`, `STRATA_GDN_SPLIT=1`, `STRATA_QFUSE=1`, `STRATA_PLE_BATCH=1`, `STRATA_SH_STREAM=1` ## Method Run by `python tools/strata_bench.py --long` against the HTTP API of a Strata server (started by the benchmark from the install configuration); model loading is not timed. One warm-up chat and a 1K prompt are discarded first. Every prompt starts with a unique tag, so the conversation cache cannot help the fresh-prompt tests. - short chat: the three questions of `tools/ab_engine.py`, up to 100 new tokens, 3 runs - 4K / 16K prompts: generated word lists sized by a calibration request, 24 new tokens, 3 fresh prompts each - agent session: system prompt and four tools, then 5 turns that each add about 1500 words of a project file and continue from the model's own earlier answers (the way agent clients resend a conversation), 64 new tokens per turn, 3 sessions; the cache numbers come from the answer's usage (`cached_tokens`) and timings (`prompt_n`) - concurrency (when run): N requests sent at the same moment, total tokens over the wall time of the burst, latency over all requests ## Results | Configuration | Actual prompt tokens | Reused tokens | Generated tokens | Runs | Prompt tok/s median and range | Decode tok/s median and range | TTFT seconds median and range | | --- | ---: | ---: | ---: | ---: | --- | --- | --- | | short chat | 46 (43-51) | 0 | 100 (100-100) | 9 | not measured | 12.2 (11.1-17.5) | 1.84 (1.67-2.04) | | 4K prompt, fresh | 4019 (4019-4019) | 0 (0-0) | 24 cap | 3 | 246.5 (245.5-247.1) | not measured | 16.43 (16.38-16.49) | | 16K prompt, fresh | 16058 (16058-16058) | 0 (0-0) | 24 cap | 3 | 292.9 (292.3-294.5) | not measured | 54.96 (54.66-55.05) | | 32K prompt, fresh | 32049 (32049-32049) | 0 (0-0) | 24 cap | 3 | 288.8 (288.5-289.1) | not measured | 111.11 (110.99-111.22) | | agent session, follow-up turns | 13579 (13566-13593) at the last turn | see below | up to 64 | 3 | see below | 11.6 (10.3-12.9) | 14.03 (9.95-15.51) | Ranges are the lowest and highest of the runs; `Runs` counts requests that returned a token. ### Agent session (conversation cache) Cache hit over the follow-up turns: 69.8 % (69.8-69.8); earlier-turn tokens read again per session: 5 (5-5) (earlier turns that the cache did not hold; the 12725 (12725-12725) new tokens of the files are read once); first turn (nothing cached) first token: 6.90 s (6.80-7.05). | Turn (run 1) | Prompt tokens | Reused | Read | Prompt tok/s | First token (s) | Decode tok/s | | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | 0 | 609 | 0 | 609 | 91.1 | 6.801 | 13.0 | | 1 | 3825 | 620 | 3205 | 209.1 | 15.453 | 12.9 | | 2 | 5358 | 3888 | 1470 | 149.5 | 9.952 | 10.4 | | 3 | 8602 | 5421 | 3181 | 206.8 | 15.505 | 11.0 | | 4 | 10826 | 8662 | 2164 | 178.5 | 12.245 | 11.1 | | 5 | 13566 | 10856 | 2710 | 194.9 | 14.027 | 11.6 | Per-request data: `results.json`. Engine log lines from the start: `engine-log-excerpt.txt` (when this run started the server). ## Correctness and limitations Not measured: this is a speed suite. The conversation uses the model's own replies, so token counts differ a little between runs. Machine details come from the server, `nvidia-smi`/`rocm-smi` and the OS; background load is not recorded. Total time of the suite: 958 s. PS. 10.2 + 100500 or --expert-cache-auto is slower. and would be glad to see gfx1150 default support
関連リンク
インストール・モデル・リリースへのサイト内リンク。