chore(cute-paged): retire warmup.py + closed-arc trace-capture scripts#16
Merged
Conversation
Pure deletion of dead/retired code. No behavior change. No new feature.
No perf claim. Not part of either active thread (SSM correctness or beta
kernel cost work) — this is housekeeping that reduces noise floor for
both.
Deletions:
- vllm/v1/attention/backends/cute_paged/warmup.py (165 lines) — module
was retired with the FULL+blessed AOT cache work; Dockerfile.gb10
explicitly removed the build-time invocation; no live importers in
vllm/ or scripts/ (only stale references in docs/superpowers/specs,
which are historical design records).
- scripts/phase_d_trace_capture.sh
- scripts/phase_d2_trace_capture.sh
- scripts/phase_d2e_trace_capture.sh
- scripts/phase_d3a_sweep.sh
- All four scripts belong to the closed D2.x cherry-pick bisection
arc; only referenced from their own trace-summary receipts under
benchmarks/nvllm/traces/cute_paged_mlp_fusion/ and from historical
docs/superpowers/specs entries.
- scripts/fusion_phaseb_diff.py — Phase B harness; the equivalent
diagnostic is now the CUTE_DEBUG_FUSION env-gated path in _backend.py.
Comment-only edit:
- _backend.py:71-73 — drop the stale file-path reference to warmup.py
(which no longer exists) from a comment about the disk-cache hook.
No code change.
NOT in this PR (per memory:feedback_comment_not_delete + post-tool-use
kernel-fusion-file hook): the 4 already-commented PHASE 3 ORIGINAL /
Phase 4 epsilon-locals blocks inside _backend.py. The audit identified
them as candidates but their comments self-justify retention for "Phase
B/C debug recovery" and re-enable scenarios. Leaving them in place.
Net: -936 lines / +3 lines across 7 files. Force-pushable rollback is
`git revert`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Swap serve.sh model row to ig1/Qwen3.5-27B-NVFP4 (default since 04/06)
- CuTe paged: "Prototype" -> "production decode path since v0.3.0"
- Document CUTE_WO_SPLIT={2,4,8} opt-in K-parallel W_O GEMV (PR #9)
- Drop --debug from serve-cute.sh launch line; PIECEWISE is default
- Veitner acknowledgment: "next direction" -> "applied in WO_SPLIT"
- Markdown lint fixups (blank lines after headings, table separators)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Pure deletion of dead/retired code. No behavior change. No new feature. No perf claim. Not part of either active thread (SSM correctness or β-kernel cost work) — housekeeping that reduces the noise floor for both.
Net: -936 lines / +3 lines across 7 files. Rollback is
git revert.Deletions
vllm/v1/attention/backends/cute_paged/warmup.pyDockerfile.gb10explicitly removed the build-time invocation; no live importers invllm/orscripts/(only stale references indocs/superpowers/specs/*, which are historical design records).scripts/phase_d_trace_capture.shbenchmarks/nvllm/traces/cute_paged_mlp_fusion/.scripts/phase_d2_trace_capture.shscripts/phase_d2e_trace_capture.shscripts/phase_d3a_sweep.shdocs/superpowers/plans/2026-04-19-phase-d3a-mlp-decode-retune.md.scripts/fusion_phaseb_diff.pyCUTE_DEBUG_FUSIONenv-gated path in_backend.py:1455-1470,1727-1760. Only self-reference in its own docstring.Comment-only edit
_backend.py:71-73— drop the stale file-path reference towarmup.py(which no longer exists) from a comment about the disk-cache hook. No code change.NOT in this PR
Per
memory:feedback_comment_not_delete+ the post-tool-use kernel/fusion-file hook reminder, the four already-commented PHASE 3 ORIGINAL / Phase 4 ε-locals blocks inside_backend.py(_backend.py:1384-1401,_backend.py:1499-1547,_backend.py:1554-1564) are deliberately left in place. The audit identified them as candidates, but their comments self-justify retention for "Phase B/C debug recovery" and re-enable scenarios. Touching them would conflict with the kernel-file convention; deferring those to a separate decision.Test plan
grep -rn "from .warmup\|cute_paged.warmup import"acrossvllm/andscripts/returned no live importers (only the doc-spec references).grep -rn "phase_d_trace_capture\|phase_d2_trace_capture\|phase_d2e_trace_capture\|phase_d3a_sweep"acrossscripts/andvllm/returned onlydocs/superpowers/{specs,plans}/mentions (historical, tolerated).grep -rn "fusion_phaseb_diff"returned only the script's own docstring self-reference._backend.py:71-73comment-only diff inspected; no code semantics changed.Audit provenance
/tmp/backend_housekeeping_audit.md(Tier A items A1, A2, A3 + Tier D D1 — D1 is the broken.contiguous()calls inside the deletedwarmup.py, subsumed by A1).🤖 Generated with Claude Code