KV-WARMUP-PROFILE: reconcile actual startup and generation config observations
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
- 难度
- 5/5
- 预计耗时
- 一周以上
- 新手友好度
- 25/100
- Issue 类型
- 缺陷
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
- 领域
- performance, testing
调研方向
阅读 warmup-memory-profiling.md Gate 0，然后针对固定的定义运行 red-before CPU reproducer。跟踪 upstream_state_observer.py:86–92、vllm/config/compilation.py:37–48、vllm/platforms/rocm.py:894 和 vllm/config/vllm.py:1427 之间的转换。完成的标准是 instrumentation 保持严格拒绝和原始有序证据，并通过指定的 mutation、CPU/image、review 和 physical-control gates。
由索引模型根据 Issue 内容生成。
描述
Row: KV-WARMUP-PROFILE
The private startup observer for #3046 refuses the first fresh physical control after the actual graph-manager repair #3059. The unchanged pinned vLLM model run succeeds, but the separate finalizer exits 1 with PROFILE3046_REFUSED startup/generation configuration differs: compilation.custom_ops. This is an instrument join failure, not an accepted memory measurement or a native model failure.
The operator ran source seal 39072f980e4028863c417eb78e83a9a11205f42cbeac14681fc55db05be8c1da after fresh review and independent CPU gates: 50 permanent tests, 8 upstream cases, 83 mutations, 31 paired controls, and 20 approved-image CPU methods. The native auto-KV control ran under /home/vikash/gpu.lock from 2026-09-08 05:50:50 to 05:58:00 UTC and exited 0. The finalizer exited 1. Source, model, and state-instrument bytes remain unchanged. The other eight prepared controls were not executed.
A read-only enumeration using the unchanged finalizer helpers finds two differences among all 30 overlapping fields:
compilation.custom_ops: startup records["+sparse_attn_indexer", "none"]; generation records["+sparse_attn_indexer", "none", "+sparse_attn_indexer"].compilation.mode: startup comparison produces"VLLM_COMPILE"; generation records integer3.
The active pin is e126687a9a828d513c01a07cd69f025f27d63280. vllm/config/compilation.py:37–48 defines CompilationMode as an IntEnum with VLLM_COMPILE=3. The accepted state observer's plain function takes integer subclasses before Enum at upstream_state_observer.py:86–92; the startup finalizer's comparable returns the enum name. The permanent fixtures did not expose this actual representation difference.
For the custom-op change, vllm/platforms/rocm.py:894 unconditionally appends +sparse_attn_indexer, called from VllmConfig.__post_init__ at vllm/config/vllm.py:1427. These are source leads. The actual intervening call and its object identity have not yet been established. Do not deduplicate, ignore this field, infer harmlessness, or accept changed dispatch without tracing the complete transition.
Owner: the gfx1100 campaign operator and a fresh private observer implementer, row KV-WARMUP-PROFILE. The committed warmup-memory-profiling.md Gate 0 requires exact source-bound startup/generation identity. Repair the instrument against that contract, with real pinned definitions and both actual encodings in the CPU reproducer. Preserve strict refusal for changed modes, custom-op selections, phases, workers, source identities, and incomplete observations. Keep raw ordered custom-op lists and all original failed evidence. No upstream/runtime/model/config/default changes or state-observer replacement are authorized by this issue.
Acceptance: red-before reproduction, smallest complete repair, effective mutations with paired detector-removal/restoration controls, complete CPU/image gates, fresh immutable-head review, operator rerun, and new physical controls. The original physical run remains refused. The owning integration closes the issue when the repaired instrumentation and evidence are accepted.
Evidence: profile-gap-ownership/observer-implementation/runtime-graph-manager-repair-3059/operator/startup-generation-config-differences.json, SHA256 854bc04bc4fd1c8272342b76f75062e30cc6b6f93cf840931fa41ad16d54672c. The corresponding gpu-runs/auto-enabled-1 contains the exact Docker vectors, actual launcher PID witness, native log, finalizer refusal, and all raw file hashes. Approved image: sha256:80aab4c182a1f3eeebe286173977e57fcaf10a049b41f475655b35d285de31dc. Model: pinned Qwen3.5-0.8B BF16, native production graphs, utilization 0.92, unchanged scheduling, eight generated tokens.
- 主要语言
- C++
- 星标
- 423
- 派生
- 53
- 平均合并
- 1 天 7 小时
- 30 天内合并 PR
- 380
环境准备
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
mudler/vllm.cpp 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 84/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 82/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 70/100
维护者通常 1 天内回复
-
难度 1/5 1 小时以内 新手友好度 88/100
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 84/100
维护者通常 1 天内回复
相似的 Issue
-
Unconfirmed bug
难度 1/5 1 小时以内 新手友好度 88/100
luanti-org/luanti#17605 · 1 条评论 ·
维护者通常 2 天内回复
-
area: config area: firmware priority: P2 - medium size: S type: bug
难度 2/5 1-3 小时 新手友好度 76/100
Mizithra/ActiveTerrain#16 ·
-
难度 2/5 1-3 小时 新手友好度 84/100
grumpycoders/pcsx-redux#2171 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 70/100
维护者通常 2 天内回复
-
难度 2/5 1-3 小时 新手友好度 88/100
bytedance/trae-agent#524 · 1 条评论 ·
维护者通常 1 天内回复