Vulkan prefill ~10% slower since v0.4.9 on Qwen3.6-35B-A3B (RX 7900 XTX, Windows), same with a local build

未关闭
#203 2 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
45/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
cpp
领域
performance

调研方向

Start by running the reported llama-bench command on the RX 7900 XTX with the v0.4.8-dev and v0.5.0 builds, comparing prefill results at 512 and 4096 tokens. The issue names no source files or tests, so trace the version changes affecting Vulkan prefill after confirming the regression. Done means identifying and addressing the cause of the slower prefill while preserving token-generation performance.

由索引模型根据 Issue 内容生成。

描述

Disclosure: Claude Code ran these measurements on my machine and drafted this text. I reviewed it.

Setup: RX 7900 XTX 24 GB (gfx1100), i5-14600K, 48 GB DDR5, Windows 11 22631, Adrenalin 32.0.31041.1004. The display is on the iGPU, so the 7900 XTX runs only llama-server.

Builds:

  • v0.4.8-dev c88618c11: local build.
  • preview-v0.5.0 d51f2f5e0: local build with the same toolchain and flags (MSVC 14.44, Ninja, Release, GGML_VULKAN=ON, GGML_NATIVE=ON, GGML_OPENMP=ON, shared libs).
  • preview-v0.5.0: official bin-win-vulkan-x64 zip.
  • v0.4.9: official zip (one earlier run).

llama-bench, Qwen3.6-35B-A3B UD-IQ4_NL, -dev Vulkan0 -ngl 99 -fa 1 -ctk q6_0 -ctv q4_0 -ub 2048 -p 512,4096 -n 128 -r 5:

build pp512 pp4096 tg128
v0.4.8-dev (local) 4173 ± 72 4777 ± 20 155.2 ± 0.9
v0.5.0 (local) 3772 ± 180 4298 ± 8 156.9 ± 0.2
v0.5.0 (official) 3832 ± 34 4304 ± 11 157.3 ± 0.4
v0.4.9 (official, -p 512 -r 3, other session) 3631 ± 234 – 156.9 ± 0.9

The local and official v0.5.0 builds are the same, so the gap is in the code, not the build. Prefill is about 10% lower at 4096 tokens. Token generation is equal.

llama-server, same preset for each build, MTP on, ~7.5k-token prompt (2 runs), prefill t/s:

model v0.4.8-dev v0.5.0 local v0.5.0 official
Qwen3.6-35B-A3B UD-IQ4_NL (built-in MTP head) 3261 / 3539 2076 / 3313 3042 / 3327
Qwen3.8-27B Q4_K_M (built-in MTP head) 744 / 744 692 / 715 708 / 713
Qwen3.8-27B Q5_K_S (MTP sidecar) 728 / 729 721 / 746 740 / 746
  • The dense Q4_K_M model loses about 4%. The dense Q5_K_S model with an MTP sidecar does not lose, and its generation is 5-8% faster on v0.5.0 (57.1 → 62.2 t/s). Thanks for that.
  • At a full 262k context, the 35B prefill is 612 t/s on v0.4.8-dev and 583 t/s on v0.5.0 (-5%).

Full logs, presets and commands on request.

主要语言
C++
星标
1.2k
派生
80
平均合并
4 天 20 小时
30 天内合并 PR
2

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

Anbeeld/beellama.cpp 的其他 Issue

查看 Anbeeld/beellama.cpp 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。