[Misc] Add multipstep chunked-prefill support for FlashInfer (#10467)
This commit is contained in:
@@ -32,7 +32,7 @@ logger = init_logger(__name__)
|
||||
MULTI_STEP_ATTENTION_BACKENDS = [
|
||||
"FLASH_ATTN", "ROCM_FLASH", "FLASHINFER", "NO_ATTENTION"
|
||||
]
|
||||
MULTI_STEP_CHUNKED_PREFILL_ATTENTION_BACKENDS = ["FLASH_ATTN"]
|
||||
MULTI_STEP_CHUNKED_PREFILL_ATTENTION_BACKENDS = ["FLASH_ATTN", "FLASHINFER"]
|
||||
|
||||
def _get_supported_attention_backends(chunked_prefill_enabled: bool) \
|
||||
-> List[str]:
|
||||
|
||||
Reference in New Issue
Block a user