fix(BACKEND-GATE-ROCM-SGLANG): isolate llama c4 input corruption
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-llama-c4-input-race.md.
The pinned llama.cpp 10bf611e533d81f739128304991c5e133c6aebd8 produces corrupt opening tokens for Qwen3-4B BF16 at concurrency four on gfx1151. The failure survives c4-first ordering and diagnostic graph disabling; c1 opening tokens remain sensible. Evidence: #3053 issuecomment-5586767582 and qualification-59a9c319-12/result.json, SHA256 1d518a475a79947d9921be92e9893ed74a4831359063284524b3c32a0db4f400.
Implement a public-API first-prefill probe: independent prompts, combined batch, and synchronized equivalent microbatch partition. Preserve token IDs, positions, sequence IDs and logits flags. For prompt lengths [6,5,6,7], pinned split_equal implies [20,1,2,1] microbatches. Establish the failure below sampling before selecting a repair. Candidate upstream ownership: ggml-org/llama.cpp#28056, #25992 and PR #27311; these are leads, not a reproduced diagnosis. Local #2557 concerns a different CPU/Q4_K case.
Acceptance: committed spec before code; test-first probe and mutation review; operator leased reproduction with source/binary/model hashes, complete logits evidence and failure status. A minimal tracked comparator patch requires reproduced buffer causality and a subsequent committed design; preserve the stock pin and label any patched comparator. No silent pin update, graph-disabled performance denominator, relaxed correctness, or speed claim. Keep the bug open until the fix lands; inconclusive diagnostics name the next hypothesis.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with .agents/specs/strix-llama-c4-input-race.md, #3053 issuecomment-5586767582, and qualification-59a9c319-12/result.json; verify the pinned source, binary, model, and logits evidence. Build the test-first public-API probe around the three prefill arrangements and review mutations before investigating repairs. Done requires reproduced buffer causality, a committed design and spec, and a tracked comparator or fix without changing the stock pin or correctness criteria.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- backend, performance, testing
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 28/100