[Attention] MLA decode optimizations (#12528)
Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com> Signed-off-by: simon-mo <xmo@berkeley.edu> Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu> Co-authored-by: simon-mo <simon.mo@hey.com> Co-authored-by: Michael Goin <mgoin64@gmail.com> Co-authored-by: Zhuohan Li <zhuohan123@gmail.com> Co-authored-by: Tyler Michael Smith <tysmith@redhat.com> Co-authored-by: Alexander Matveev <59768536+alexm-neuralmagic@users.noreply.github.com> Co-authored-by: simon-mo <xmo@berkeley.edu>
This commit is contained in:
co-authored by
Woosuk Kwon
simon-mo
Michael Goin
Zhuohan Li
Tyler Michael Smith
Alexander Matveev
simon-mo
parent
a1fc18c030
commit
cabaf4eff3
@@ -31,7 +31,8 @@ class CpuPlatform(Platform):
|
||||
@classmethod
|
||||
def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,
|
||||
dtype: torch.dtype, kv_cache_dtype: Optional[str],
|
||||
block_size: int, use_v1: bool) -> str:
|
||||
block_size: int, use_v1: bool,
|
||||
use_mla: bool) -> str:
|
||||
if selected_backend != _Backend.TORCH_SDPA:
|
||||
logger.info("Cannot use %s backend on CPU.", selected_backend)
|
||||
logger.info("Using Torch SDPA backend.")
|
||||
|
||||
@@ -157,10 +157,14 @@ class CudaPlatformBase(Platform):
|
||||
|
||||
@classmethod
|
||||
def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
|
||||
kv_cache_dtype, block_size, use_v1) -> str:
|
||||
kv_cache_dtype, block_size, use_v1,
|
||||
use_mla) -> str:
|
||||
if use_v1:
|
||||
logger.info("Using Flash Attention backend on V1 engine.")
|
||||
return "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"
|
||||
if use_mla:
|
||||
logger.info("Using Triton MLA backend.")
|
||||
return "vllm.attention.backends.triton_mla.TritonMLABackend"
|
||||
if selected_backend == _Backend.FLASHINFER:
|
||||
logger.info("Using FlashInfer backend.")
|
||||
return "vllm.attention.backends.flashinfer.FlashInferBackend"
|
||||
@@ -171,7 +175,8 @@ class CudaPlatformBase(Platform):
|
||||
pass
|
||||
elif selected_backend:
|
||||
raise ValueError(
|
||||
f"Invalid attention backend for {cls.device_name}")
|
||||
f"Invalid attention backend for {cls.device_name}, "
|
||||
f"with use_v1: {use_v1} use_mla: {use_mla}")
|
||||
|
||||
target_backend = _Backend.FLASH_ATTN
|
||||
if not cls.has_device_capability(80):
|
||||
|
||||
@@ -27,7 +27,8 @@ class HpuPlatform(Platform):
|
||||
@classmethod
|
||||
def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,
|
||||
dtype: torch.dtype, kv_cache_dtype: Optional[str],
|
||||
block_size: int, use_v1: bool) -> str:
|
||||
block_size: int, use_v1: bool,
|
||||
use_mla: bool) -> str:
|
||||
logger.info("Using HPUAttention backend.")
|
||||
return "vllm.attention.backends.hpu_attn.HPUAttentionBackend"
|
||||
|
||||
|
||||
@@ -30,6 +30,7 @@ class _Backend(enum.Enum):
|
||||
TORCH_SDPA = enum.auto()
|
||||
OPENVINO = enum.auto()
|
||||
FLASHINFER = enum.auto()
|
||||
TRITON_MLA = enum.auto()
|
||||
HPU_ATTN = enum.auto()
|
||||
PALLAS = enum.auto()
|
||||
IPEX = enum.auto()
|
||||
@@ -139,7 +140,8 @@ class Platform:
|
||||
@classmethod
|
||||
def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,
|
||||
dtype: torch.dtype, kv_cache_dtype: Optional[str],
|
||||
block_size: int, use_v1: bool) -> str:
|
||||
block_size: int, use_v1: bool,
|
||||
use_mla: bool) -> str:
|
||||
"""Get the attention backend class of a device."""
|
||||
return ""
|
||||
|
||||
|
||||
@@ -30,7 +30,8 @@ class OpenVinoPlatform(Platform):
|
||||
@classmethod
|
||||
def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,
|
||||
dtype: torch.dtype, kv_cache_dtype: Optional[str],
|
||||
block_size: int, use_v1: bool) -> str:
|
||||
block_size: int, use_v1: bool,
|
||||
use_mla: bool) -> str:
|
||||
if selected_backend != _Backend.OPENVINO:
|
||||
logger.info("Cannot use %s backend on OpenVINO.", selected_backend)
|
||||
logger.info("Using OpenVINO Attention backend.")
|
||||
|
||||
@@ -75,7 +75,8 @@ class RocmPlatform(Platform):
|
||||
|
||||
@classmethod
|
||||
def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
|
||||
kv_cache_dtype, block_size, use_v1) -> str:
|
||||
kv_cache_dtype, block_size, use_v1,
|
||||
use_mla) -> str:
|
||||
selected_backend = (_Backend.ROCM_FLASH if selected_backend
|
||||
== _Backend.FLASH_ATTN else selected_backend)
|
||||
if selected_backend == _Backend.ROCM_FLASH:
|
||||
|
||||
@@ -29,7 +29,8 @@ class TpuPlatform(Platform):
|
||||
@classmethod
|
||||
def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,
|
||||
dtype: torch.dtype, kv_cache_dtype: Optional[str],
|
||||
block_size: int, use_v1: bool) -> str:
|
||||
block_size: int, use_v1: bool,
|
||||
use_mla: bool) -> str:
|
||||
if selected_backend != _Backend.PALLAS:
|
||||
logger.info("Cannot use %s backend on TPU.", selected_backend)
|
||||
logger.info("Using Pallas backend.")
|
||||
|
||||
@@ -27,7 +27,8 @@ class XPUPlatform(Platform):
|
||||
@classmethod
|
||||
def get_attn_backend_cls(cls, selected_backend: _Backend, head_size: int,
|
||||
dtype: torch.dtype, kv_cache_dtype: Optional[str],
|
||||
block_size: int, use_v1: bool) -> str:
|
||||
block_size: int, use_v1: bool,
|
||||
use_mla: bool) -> str:
|
||||
if selected_backend != _Backend.IPEX:
|
||||
logger.info("Cannot use %s backend on XPU.", selected_backend)
|
||||
logger.info("Using IPEX attention backend.")
|
||||
|
||||
Reference in New Issue
Block a user