[FMHA] Support Vectorized KV Cache Layout and vLLM/SGLang block table in Batch Prefill kernel #1754
Conversation
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
…/8, block_size, 8], [num_blocks, num_kv_heads, block_size/8, head_size, 8]
…ayout Updated `mha_batch_prefill` API and tests to support vLLM-style block tables alongside SGLang-style page tables, while enforcing the new hardware-optimized 5D vectorized KV cache layout. **Key Changes:** * **API**: Added `block_table` and `seqlen_k` arguments to python/C++ interfaces. * **Layout Enforcement**: Added strict checks for 5D vectorized KV layout (swizzled x=8) in host bindings and python wrappers. * **CodeGen**: Automatically select `VLLM_BLOCK_TABLE_2D` or `SGLANG_PAGE_TABLE_1D` trait based on input arguments. * **Tests**: Added `test_batch_prefill_vllm` to verify block table correctness and updated existing tests to use the vectorized layout.
poyenc
reviewed
Dec 31, 2025
| if head_size_v_og % 8 != 0: | ||
| v = torch.nn.functional.pad(v, [0, 8 - head_size_v_og % 8]) | ||
| head_size_q_og = q.size(-1) | ||
| k_vector_size = 16 // k.element_size() |
Contributor
There was a problem hiding this comment.
Suggest adding a comment explaining that the magic number 16 corresponds to dwordx4
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Outdated suggestions cannot be applied.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.
Motivation
Introduces support for a vectorized KV cache memory layout (e.g., [num_blocks, num_kv_heads, head_size/8, block_size, 8]) to improve memory access efficiency and also support different type of block table such as vLLM and SGLang.
Technical Details
Key changes:
KV Cache Layout Optimization and Adjustment:
vLLM Block Table Integration:
Kernel Interface Updates:
Structure and Traits Updates:
Test Plan
Test Result
Submission Checklist