Remove hybrid_linear_attn attention backend and refactor attention registry (#10816)
Co-authored-by: Yi Zhang <1109276519@qq.com>
This commit is contained in:
@@ -100,7 +100,6 @@ ATTENTION_BACKEND_CHOICES = [
|
||||
"trtllm_mla",
|
||||
"trtllm_mha",
|
||||
"dual_chunk_flash_attn",
|
||||
"hybrid_linear_attn",
|
||||
# AMD specific
|
||||
"aiter",
|
||||
"wave",
|
||||
@@ -801,7 +800,7 @@ class ServerArgs:
|
||||
self.speculative_algorithm is None
|
||||
), "Speculative decoding is currently not supported with Flex Attention backend"
|
||||
|
||||
if is_npu() and self.attention_backend in ["ascend", "hybrid_linear_attn"]:
|
||||
if is_npu() and self.attention_backend in ["ascend"]:
|
||||
logger.warning(
|
||||
"At this moment Ascend attention backend only supports a page_size of 128, change page_size to 128."
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user