mudler / mudler/vllm.cpp

Qualify HRX dense WMMA after Loom removes a live template provider

Open
#3,081 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: BACKEND-ROCM

Native HRX qualification for #3080 fails because Loom removes a live template provider during dense WMMA compilation. This issue owns the bounded compiler repair and its qualification evidence.

At ROCm/hrx-system 6bcd5a4ff111fa5bf160ab9f4592ca8e7cc810b1 and AMD-Ecosystem/llama.cpp 6319038132ed12f968ea68f37753f705da830ea8, test-hrx-ops on an RX 7900 XTX fails at the first Q4_K dense WMMA case: input_size=2048, output_size=128, token_count=2, output_accumulation=0, output_unary_op=23, weight_format=4. The native backend detects gfx1100 successfully.

Unmodified loom-link plus loom-compile reproduces the same failure without a GPU. The first template selection succeeds. inline-callables then removes template definitions while scalar template calls to ggml_unary_f32_apply remain. The next selection reports template.call callee has no template provider contract at loom/src/loom/transforms/symbol/template_selection.c:585.

The owning operator retains the exact linked recipe, pass snapshots, source pins, commands, and failed GPU run under the #3080 evidence. The repair must add an upstream-native regression that fails before the change, preserve live provider dependencies through inlining, compile the exact dense recipe, and pass independent review. Root reruns the native operation test. Do not weaken a verifier, tolerance, or operation contract, and do not treat this failure as a gfx1100 hardware limitation.

Use a committed scoped spec before implementation. The #3080 operator owns the work and the eventual disposition of the upstream patch. Native performance remains unaccepted until its declared correctness gate passes.

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

Start with the #3080 evidence and reproduce the failure using unmodified loom-link and loom-compile, then inspect loom/src/loom/transforms/symbol/template_selection.c:585 and the inline-callables behavior. Use test-hrx-ops and the exact dense WMMA recipe as qualification checks. Done means an upstream-native regression fails before the repair, live provider dependencies survive inlining, the exact recipe compiles, and the independent review and native operation test pass.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
compilers
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.