Make FlashInfer cache dtype hardware-aware (fix bf16 on SM<80)#56
Open
os-gabe wants to merge 1 commit into
Open
Make FlashInfer cache dtype hardware-aware (fix bf16 on SM<80)#56os-gabe wants to merge 1 commit into
os-gabe wants to merge 1 commit into
Conversation
FlashInferKVCacheManager unconditionally mapped fp32 -> bf16, which fails at runtime on SM<80 (Turing/Volta, e.g. Titan RTX) where bf16 kernels aren't available. AggregatorStream also passed tokens.dtype, which autocast-exempt ops (LayerNorm) leak as fp32, so the bug fires even when demo.py selects fp16. - flashinfer_cache: hardware-aware fp32/None fallback (bf16 only on SM>=8). - aggregator/stream: prefer aggregator parameter dtype before falling through to the cache's default. SM>=80 behavior is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
FlashInferKVCacheManager unconditionally mapped fp32 -> bf16, which fails at runtime on SM<80 (Turing/Volta, e.g. Titan RTX) where bf16 kernels aren't available. AggregatorStream also passed tokens.dtype, which autocast-exempt ops (LayerNorm) leak as fp32, so the bug fires even when demo.py selects fp16.
SM>=80 behavior is unchanged.