mudler / mudler/vllm.cpp

test(BACKEND-GATE-ROCM-SGLANG): calibrate a Qwen3-4B correctness proposal

Open
#3,077 3 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-GATE-ROCM-SGLANG

Owner: current Strix campaign operator. Parent: #3053. Spec: .agents/specs/strix-qwen3-4b-distributional-calibration.md (commit before implementation).

Pinned production vLLMe126687a9a828d513c01a07cd69f025f27d63280 showed repeated greedy token differences on Qwen3-4B BF16 at c1 and c4. Qualification802b32c9, result SHA1d518a475a79947d9921be92e9893ed74a4831359063284524b3c32a0db4f400, remains NOT_ACCEPTED. User approved scoping a distributional correctness proposal, not numerical tolerances or replacement of the strict gate.

Implement reviewed diagnostic score capture through public incremental generation, not whole-prefix prefill scoring labeled as decode. Ours6e3cbfb exposes a public logits callback; pinned vLLM exposes named importable LogitsProcessor classes, JSON-compatible extra_args and full-vocabulary sample logprobs. Verify these routes execute with the campaign production runner/graphs before calibration. No serialized callables or eager correctness/performance substitution. Compare original prompts plus identical consumed continuation prefixes; capture scores before any forcing. Record actual scheduler shapes and instrumentation equivalence, not requested concurrency alone.

Predeclare oracle-only calibration and held-out validation before viewing candidate scores. Separate c1/c4 and first-prefill/incremental phases. Proposed thresholds, sample counts and statistical rule require subsequent explicit ratification. Existing CUDA/27B/quantized tolerances are not transferable. Stable oracle behavior remains token-exact. Enforce hard failures for prefix/tokenizer/dtype/model changes, nonfinite/missing scores, wrong execution phase, swapped rows, post-forcing capture and skipped callbacks; include corruption controls. A forced output match is not a correctness pass.

Acceptance: source-bound diagnostic design, test-first implementation, independent mutation review, operator leased equivalence and oracle-only evidence, and a written proposed gate with held-out results. This issue does not itself authorize adoption of a tolerance or acceptance of speed. Stop if instrumentation changes the relevant numerical path or production capture cannot be established; name the exact unresolved interface.

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 .agents/specs/strix-qwen3-4b-distributional-calibration.md, then inspect the public logits callback in Ours6e3cbfb and the pinned vLLM named LogitsProcessor interfaces. Verify those routes with the campaign production runner and graphs before implementing test-first diagnostics. Done means reviewed source-bound capture, mutation controls, oracle-only calibration evidence, and a written held-out proposal without changing the strict gate.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
backend, machine-learning, testing-qa
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
28/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.