[FEAT][ROCm] Integrate Paged Attention Kernel from AITER (#15001)

Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
vllmellm
2025-04-22 02:46:28 -07:00
committed by GitHub
co-authored by tjtanaa
parent 8f7bace7c3
commit 0e237f0035
5 changed files with 195 additions and 27 deletions
+2 -1
View File
@@ -118,7 +118,8 @@ def use_rocm_custom_paged_attention(qtype: torch.dtype, head_size: int,
and (head_size == 64 or head_size == 128)
and (block_size == 16 or block_size == 32)
and (gqa_ratio >= 1 and gqa_ratio <= 16) and max_seq_len <= 32768
and envs.VLLM_ROCM_CUSTOM_PAGED_ATTN)
and not (envs.VLLM_ROCM_USE_AITER_PAGED_ATTN
and envs.VLLM_ROCM_USE_AITER))
class RocmPlatform(Platform):