Repository navigation
[backlog] perf(rdna): retune the 6 TunableOp rows slower than the heuristic - #74
Draft
BlivionIaG wants to merge 1 commit into
Draft
BlivionIaG wants to merge 1 commit into
BlivionIaG wants to merge 1 commit into
Conversation
Tracking branch for a backlog item; the plan is in the draft PR.
wsantos
added a commit
to wsantos/vllm-rdna
that referenced
this pull request
Oct 12, 2026
The row set currently in service on this box resolves through the *fallback* path: the loaded /opt/rocm/core-10.0/lib/librocblas.so.5.6 hashes to sha256[:12] = d05939967b5f, and no profile registered in profiles.json matches that hash, so configure_tunableop falls back to the directory named after it. Until now that directory was untracked, so the tables actually serving traffic were not in version control. 296 GemmTunableOp_Half_TN rows across the four TP ranks (74/rank), captured for librocblas.so.5.6 @ d05939967b5f, torch 2.13.0+gitc1d415f, hip 7.15.26333, float16, seed 620, batch_tokens [800..8192]. QUALIFIED records that all 70 FP16 TN shapes matched independent FP32 calculations on all four V620s. Deliberately NOT done here, because it needs a human owner: - no profiles.json registry entry and no profile README.md, so this is still the per-user fallback rather than a first-class named profile - no lookup-hit proof (the README requires it per profile); the hit rate for these rows has not been measured - no capture/tune/curate campaign notes; these rows are a raw capture, not curated against the heuristic the way rocm7.14-rocblas5.5 was in opengfx1030#81 Refs: opengfx1030/vllm-rdna tunableop/README.md (profile requirements), PR opengfx1030#81 (the curate-and-measure protocol), opengfx1030#74 (open retune backlog). Co-authored-by: Pi <noreply@pi.dev>
wsantos
added a commit
to wsantos/vllm-rdna
that referenced
this pull request
Oct 12, 2026
The row set currently in service on this box resolves through the *fallback* path: the loaded /opt/rocm/core-10.0/lib/librocblas.so.5.6 hashes to sha256[:12] = d05939967b5f, and no profile registered in profiles.json matches that hash, so configure_tunableop falls back to the directory named after it. Until now that directory was untracked, so the tables actually serving traffic were not in version control. 296 GemmTunableOp_Half_TN rows across the four TP ranks (74/rank), captured for librocblas.so.5.6 @ d05939967b5f, torch 2.13.0+gitc1d415f, hip 7.15.26333, float16, seed 620, batch_tokens [800..8192]. QUALIFIED records that all 70 FP16 TN shapes matched independent FP32 calculations on all four V620s. Deliberately NOT done here, because it needs a human owner: - no profiles.json registry entry and no profile README.md, so this is still the per-user fallback rather than a first-class named profile - no lookup-hit proof (the README requires it per profile); the hit rate for these rows has not been measured - no capture/tune/curate campaign notes; these rows are a raw capture, not curated against the heuristic the way rocm7.14-rocblas5.5 was in opengfx1030#81 Refs: opengfx1030/vllm-rdna tunableop/README.md (profile requirements), PR opengfx1030#81 (the curate-and-measure protocol), opengfx1030#74 (open retune backlog).
|
Measured, and partly refuted: the whole 168-shape grid was re-tuned and compared heuristic / current / new in one session on build
|
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
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.
Backlog item, draft until the work lands on this branch.
Why
The coverage campaign kept 6 Flash-Next rows measured 3-10 % slower than the rocBLAS heuristic (listed in
tunableop/rocm7.14-rocblas5.5/provenance.json).Plan
Re-tune or drop those rows; regenerate the profile.
Validation
lookup hit check + affected cells.