mudler / mudler/vllm.cpp

feat(GFX1100-TG200): T25 keep ssm_out as Q5_K with runtime input permutation

Open
#2,806 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
C++
Stars
423
Forks
53
Avg merge
20h 26m
Merged PRs (30d)
310

Description

Row: `GFX1100-TG200`

T25 keeps the SSM output projection as Q5_K and adds a runtime V-head permutation kernel (`PermuteVHeadsKernelRocm`). The permutation reorders V-heads from grouped (krpk+r) to tiled (rnum_k+k) order before the K-quant GEMV, enabling the keep-quant path to avoid a dequantize-requantize round trip on the output projection.

Adds `kPermuteVHeads` op, `PermuteVHeads` dispatch function, and the ROCm kernel. Restores `kCastF16`/`CastF16`/`CastF16KernelRocm` that were lost during cherry-pick auto-merge resolution.

Depends on #2782 (F1), #2790 (T4a), #2792 (lever-C), #2796 (T27), #2800 (T24), and #2804 (T21).

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Read the dependent issues (#2782, #2790, #2792, #2796, #2800, and #2804), then trace kPermuteVHeads, PermuteVHeads, and PermuteVHeadsKernelRocm. Check the existing CastF16 symbols and the ROCm build path. Done means the runtime V-head permutation and keep-quant path are wired in, with the restored cast symbols present and the ROCm target building successfully.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
machine-learning, performance
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
42/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.