mudler / mudler/vllm.cpp

BACKEND-ROCM: graph state checker rejects the quantized SiLU provider

Open
#3,056 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

The current private #3041 candidate c3c880d6abdfc6c698cee7e3bb1bbf4f73369062 completes the approved Qwen3.5-4B Q4_K_M gfx1100 smoke capture, but its state checker rejects that capture with missing required GetOp resolution.

scripts/check-rocm-graph-state.py:174–180 requires SiluAndMul for both admitted model profiles. The retained 4B capture reports MoeSiluMul: 96 and MatmulBTQuant: 387, and no SiluAndMul resolution. The 0.8B BF16 capture satisfied the existing assertion. Provider resolution is an observation of dispatch selection; it is not device invocation proof.

Reproduce with the exact candidate rocm_graph_state_capture binary SHA256 9dfcd1b3f37dde4cd0ffd2e5f5af6ff61f00c2472ad5ff87fba06990f9f7272c, approved image sha256:80aab4c182a1f3eeebe286173977e57fcaf10a049b41f475655b35d285de31dc, model SHA256 00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4, KV auto, scenario smoke, and ambient graph flags. The capture completed on physical gfx1100 device0 under the GPU mutex, with one capture and six replay counters. The independent state-reader gate remains FAIL; no comparison, default, oracle, or performance acceptance is claimed.

The repair must derive required providers from each validated loaded profile and its actual production chain. Preserve rejection of missing or wrong-profile providers. Add the retained quantized capture as a regression input, prove detector sensitivity, and use fresh review plus operator reruns. Do not change product providers or relax the state/logit/tensor comparison. This work belongs in the existing #3041 implementation flow and its owning spec.

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 scripts/check-rocm-graph-state.py:174–180 and read the existing #3041 implementation flow and owning spec. Trace how each validated loaded profile and its production chain determines required providers, while preserving rejection of missing or wrong-profile providers. Add the retained quantized capture as a regression input, prove detector sensitivity, and complete fresh review plus operator reruns.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp, python
Domain
backend, testing-qa
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
45/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.