[V1] Enable Triton(ROCm) Attention backend for Nvidia GPUs (#14071)

Signed-off-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Isotr0py
2025-03-21 03:14:19 +00:00
committed by GitHub
co-authored by Woosuk Kwon
parent b15fd2be2a
commit f8a08cb90d
5 changed files with 23 additions and 16 deletions
+8 -3
View File
@@ -213,9 +213,14 @@ class CudaPlatformBase(Platform):
return ("vllm.attention.backends."
"flashmla.FlashMLABackend")
if use_v1:
logger.info_once("Using Flash Attention backend on V1 engine.")
return ("vllm.v1.attention.backends.flash_attn."
"FlashAttentionBackend")
if selected_backend == _Backend.TRITON_ATTN_VLLM_V1:
logger.info_once("Using Triton backend on V1 engine.")
return ("vllm.v1.attention.backends."
"triton_attn.TritonAttentionBackend")
if cls.has_device_capability(80):
logger.info_once("Using Flash Attention backend on V1 engine.")
return ("vllm.v1.attention.backends."
"flash_attn.FlashAttentionBackend")
if selected_backend == _Backend.FLASHINFER:
logger.info("Using FlashInfer backend.")
return "vllm.attention.backends.flashinfer.FlashInferBackend"