[autotuner][cute] demote tcgen05_tvm_ffi_launch to a seed-only knob - #3149
Draft
calebmkim wants to merge 1 commit into
Draft
[autotuner][cute] demote tcgen05_tvm_ffi_launch to a seed-only knob#3149calebmkim wants to merge 1 commit into
calebmkim wants to merge 1 commit into
Conversation
meta-cla
Bot
added
the
CLA Signed
calebmkim
force-pushed
the
calebmkim/stack/32
branch
from
July 24, 2026 17:09
dfd5530 to
917acec
Compare
calebmkim
force-pushed
the
calebmkim/stack/33
branch
from
July 24, 2026 17:09
7e05732 to
1598311
Compare
This was referenced Jul 24, 2026
Pinned to EnumFragment((True,)) in the search view, the FFI-launch flag caused
(1) population collapse -- every candidate forced ffi=True so the promotion
normalizer deduped the whole surface to one seed -- and (2) an encoder crash when a
partial seed back-filled default False into a (True,)-only slot ("Invalid enum
value False").
Fix: EnumFragment((True, False), search_choices=(False,)). Search draws only False
(FFI enters solely via the existing eligibility-gated CuteTcgen05ClusterM2FfiHeuristic
seed, #2681); choices carries True so the seed round-trips and default_config() still
promotes on eligible shapes.
Effect (single B200): distinct configs in a 400-draw search population on S3
2048x4096x4096 bf16, 2 -> 17, and a full S3 autotune now runs to completion with
heuristics on, with zero "Invalid enum value False" encoder crashes (the partial
formula seed that omits the key no longer aborts surrogate encoding at config #2).
SCOPE: this fixes one of the SIX clauses in _target1_tvm_ffi_promotion_requested,
which is what gates the wholesale projection onto the FFI seed. The other five
(tcgen05_flat_role_coordinates, tcgen05_layout_strategy='explicit_epi_tile', and the
three tcgen05_layout_overrides_* keys) are still independent search draws, so ~95% of
candidates continue to trip the gate and be overwritten (measured on this commit:
293/300 candidates -> 1 distinct config, with 0 of 30 keys varying). The 2 -> 17 gain
above is real but partial: it comes from restoring the small subset of draws that
escape all six clauses. The follow-up commit repairs those candidates away from the
FFI envelope instead of onto it, which is what actually restores a multi-config
search (299/300 distinct).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
stack-info: PR: #3149, branch: calebmkim/stack/33
calebmkim
force-pushed
the
calebmkim/stack/33
branch
from
July 27, 2026 22:13
1598311 to
cd3befd
Compare
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.
Stacked PRs:
[autotuner][cute] demote tcgen05_tvm_ffi_launch to a seed-only knob
Pinned to EnumFragment((True,)) in the search view, the FFI-launch flag caused
(1) population collapse -- every candidate forced ffi=True so the promotion
normalizer deduped the whole surface to one seed -- and (2) an encoder crash when a
partial seed back-filled default False into a (True,)-only slot ("Invalid enum
value False").
Fix: EnumFragment((True, False), search_choices=(False,)). Search draws only False
(FFI enters solely via the existing eligibility-gated CuteTcgen05ClusterM2FfiHeuristic
seed, #2681); choices carries True so the seed round-trips and default_config() still
promotes on eligible shapes.
Effect (single B200): distinct configs in a 400-draw search population on S3
2048x4096x4096 bf16, 2 -> 17, and a full S3 autotune now runs to completion with
heuristics on, with zero "Invalid enum value False" encoder crashes (the partial
formula seed that omits the key no longer aborts surrogate encoding at config #2).
SCOPE: this fixes one of the SIX clauses in target1_tvm_ffi_promotion_requested,
which is what gates the wholesale projection onto the FFI seed. The other five
(tcgen05_flat_role_coordinates, tcgen05_layout_strategy='explicit_epi_tile', and the
three tcgen05_layout_overrides* keys) are still independent search draws, so ~95% of
candidates continue to trip the gate and be overwritten (measured on this commit:
293/300 candidates -> 1 distinct config, with 0 of 30 keys varying). The 2 -> 17 gain
above is real but partial: it comes from restoring the small subset of draws that
escape all six clauses. The follow-up commit repairs those candidates away from the
FFI envelope instead of onto it, which is what actually restores a multi-config
search (299/300 distinct).
Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com