[V0 deprecation] Remove _VLLM_V1 suffixes from attention backend names (#25489)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Signed-off-by: Matthew Bonanni <mbonanni001@gmail.com>
This commit is contained in:
+10
-11
@@ -241,9 +241,8 @@ class CudaPlatformBase(Platform):
|
||||
use_flashinfermla = selected_backend == _Backend.FLASHINFER_MLA or (
|
||||
selected_backend is None and cls.is_device_capability(100)
|
||||
and block_size in [32, 64])
|
||||
use_flashmla = selected_backend in [
|
||||
_Backend.FLASHMLA, _Backend.FLASHMLA_VLLM_V1
|
||||
] or (selected_backend is None and is_flashmla_supported()[0])
|
||||
use_flashmla = selected_backend == _Backend.FLASHMLA or (
|
||||
selected_backend is None and is_flashmla_supported()[0])
|
||||
use_flashattn = selected_backend == _Backend.FLASH_ATTN_MLA or (
|
||||
selected_backend is None and flash_attn_supports_mla())
|
||||
use_triton = selected_backend == _Backend.TRITON_MLA or (
|
||||
@@ -282,7 +281,7 @@ class CudaPlatformBase(Platform):
|
||||
if use_v1:
|
||||
FLASHINFER_V1 = "vllm.v1.attention.backends.flashinfer.FlashInferBackend" # noqa: E501
|
||||
FLEX_ATTENTION_V1 = "vllm.v1.attention.backends.flex_attention.FlexAttentionBackend" # noqa: E501
|
||||
TRITON_ATTN_VLLM_V1 = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend" # noqa: E501
|
||||
TRITON_ATTN = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend" # noqa: E501
|
||||
FLASH_ATTN_V1 = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend" # noqa: E501
|
||||
TREE_ATTN_V1 = "vllm.v1.attention.backends.tree_attn.TreeAttentionBackend" # noqa: E501
|
||||
XFORMERS_V1 = "vllm.v1.attention.backends.xformers.XFormersAttentionBackend" # noqa: E501
|
||||
@@ -300,16 +299,16 @@ class CudaPlatformBase(Platform):
|
||||
elif selected_backend == _Backend.FLEX_ATTENTION:
|
||||
logger.info_once("Using FlexAttention backend on V1 engine.")
|
||||
return FLEX_ATTENTION_V1
|
||||
elif selected_backend == _Backend.TRITON_ATTN_VLLM_V1:
|
||||
elif selected_backend == _Backend.TRITON_ATTN:
|
||||
logger.info_once("Using Triton backend on V1 engine.")
|
||||
return TRITON_ATTN_VLLM_V1
|
||||
return TRITON_ATTN
|
||||
elif selected_backend == _Backend.FLASH_ATTN:
|
||||
logger.info_once("Using Flash Attention backend on V1 engine.")
|
||||
return FLASH_ATTN_V1
|
||||
elif selected_backend == _Backend.TREE_ATTN:
|
||||
logger.info_once("Using Tree Attention backend on V1 engine.")
|
||||
return TREE_ATTN_V1
|
||||
elif selected_backend == _Backend.XFORMERS_VLLM_V1:
|
||||
elif selected_backend == _Backend.XFORMERS:
|
||||
logger.info_once("Using XFormers backend on V1 engine.")
|
||||
return XFORMERS_V1
|
||||
|
||||
@@ -341,7 +340,7 @@ class CudaPlatformBase(Platform):
|
||||
if (has_sink or
|
||||
use_fp8_kv_cache) and not cls.is_device_capability(90):
|
||||
logger.info_once("Using Triton backend on V1 engine.")
|
||||
return TRITON_ATTN_VLLM_V1
|
||||
return TRITON_ATTN
|
||||
elif is_default_backend_supported := is_attn_backend_supported(
|
||||
FLASH_ATTN_V1, head_size, dtype,
|
||||
allow_import_error=False):
|
||||
@@ -457,12 +456,12 @@ class CudaPlatformBase(Platform):
|
||||
else:
|
||||
# Default to FlashAttention
|
||||
if attention_backend is None:
|
||||
attention_backend = "FLASH_ATTN_VLLM_V1"
|
||||
attention_backend = "FLASH_ATTN"
|
||||
|
||||
# All Blackwell backends support fp8
|
||||
if cls.is_device_capability(100):
|
||||
supported = True
|
||||
elif attention_backend == "FLASH_ATTN_VLLM_V1":
|
||||
elif attention_backend == "FLASH_ATTN":
|
||||
if fp8_attention:
|
||||
from vllm.attention.utils.fa_utils import (
|
||||
flash_attn_supports_fp8)
|
||||
@@ -471,7 +470,7 @@ class CudaPlatformBase(Platform):
|
||||
supported = True
|
||||
elif attention_backend == "FLASHINFER":
|
||||
supported = True
|
||||
elif attention_backend == "TRITON_ATTN_VLLM_V1":
|
||||
elif attention_backend == "TRITON_ATTN":
|
||||
supported = cls.supports_fp8()
|
||||
return supported
|
||||
|
||||
|
||||
@@ -40,34 +40,26 @@ def in_wsl() -> bool:
|
||||
|
||||
class _Backend(enum.Enum):
|
||||
FLASH_ATTN = enum.auto()
|
||||
FLASH_ATTN_VLLM_V1 = enum.auto()
|
||||
TRITON_ATTN_VLLM_V1 = enum.auto()
|
||||
TRITON_ATTN = enum.auto()
|
||||
XFORMERS = enum.auto()
|
||||
ROCM_FLASH = enum.auto()
|
||||
ROCM_AITER_MLA = enum.auto() # Supported by V1
|
||||
ROCM_AITER_MLA_VLLM_V1 = enum.auto()
|
||||
ROCM_AITER_FA = enum.auto() # used for ViT attn backend
|
||||
TORCH_SDPA = enum.auto()
|
||||
TORCH_SDPA_VLLM_V1 = enum.auto()
|
||||
FLASHINFER = enum.auto()
|
||||
FLASHINFER_VLLM_V1 = enum.auto()
|
||||
FLASHINFER_MLA = enum.auto()
|
||||
TRITON_MLA = enum.auto() # Supported by V1
|
||||
TRITON_MLA_VLLM_V1 = enum.auto()
|
||||
CUTLASS_MLA = enum.auto()
|
||||
FLASHMLA = enum.auto() # Supported by V1
|
||||
FLASHMLA_VLLM_V1 = enum.auto()
|
||||
FLASH_ATTN_MLA = enum.auto() # Supported by V1
|
||||
PALLAS = enum.auto()
|
||||
PALLAS_VLLM_V1 = enum.auto()
|
||||
IPEX = enum.auto()
|
||||
DUAL_CHUNK_FLASH_ATTN = enum.auto()
|
||||
DIFFERENTIAL_FLASH_ATTN = enum.auto()
|
||||
NO_ATTENTION = enum.auto()
|
||||
FLEX_ATTENTION = enum.auto()
|
||||
TREE_ATTN = enum.auto()
|
||||
XFORMERS_VLLM_V1 = enum.auto()
|
||||
ROCM_ATTN_VLLM_V1 = enum.auto()
|
||||
ROCM_ATTN = enum.auto()
|
||||
|
||||
|
||||
class PlatformEnum(enum.Enum):
|
||||
|
||||
@@ -218,8 +218,7 @@ class RocmPlatform(Platform):
|
||||
raise ValueError(
|
||||
f" The selected backend, {selected_backend.name},"
|
||||
f"does not support block size {block_size}.")
|
||||
if selected_backend in (_Backend.ROCM_AITER_MLA,
|
||||
_Backend.ROCM_AITER_MLA_VLLM_V1):
|
||||
if selected_backend == _Backend.ROCM_AITER_MLA:
|
||||
if block_size == 1:
|
||||
logger.info("Using AITER MLA backend on V1 engine.")
|
||||
return "vllm.v1.attention.backends.mla.rocm_aiter_mla.AiterMLABackend" # noqa: E501
|
||||
@@ -240,7 +239,7 @@ class RocmPlatform(Platform):
|
||||
elif (envs.VLLM_ROCM_USE_AITER and
|
||||
envs.VLLM_USE_AITER_UNIFIED_ATTENTION) or \
|
||||
envs.VLLM_V1_USE_PREFILL_DECODE_ATTENTION or \
|
||||
selected_backend == _Backend.ROCM_ATTN_VLLM_V1:
|
||||
selected_backend == _Backend.ROCM_ATTN:
|
||||
# rocm specific backend, with aiter and/or
|
||||
# triton prefix-prefill
|
||||
logger.info("Using Rocm/Aiter Attention backend on V1 engine.")
|
||||
|
||||
@@ -50,8 +50,7 @@ class TpuPlatform(Platform):
|
||||
dtype: torch.dtype, kv_cache_dtype: Optional[str],
|
||||
block_size: int, use_v1: bool, use_mla: bool,
|
||||
has_sink) -> str:
|
||||
if (selected_backend != _Backend.PALLAS
|
||||
and selected_backend != _Backend.PALLAS_VLLM_V1):
|
||||
if selected_backend != _Backend.PALLAS:
|
||||
logger.info("Cannot use %s backend on TPU.", selected_backend)
|
||||
|
||||
if not use_v1:
|
||||
|
||||
@@ -40,14 +40,14 @@ class XPUPlatform(Platform):
|
||||
use_v1 = envs.VLLM_USE_V1
|
||||
if not use_v1:
|
||||
raise ValueError("XPU backend only supports V1.")
|
||||
TRITON_ATTN_VLLM_V1 = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend" # noqa: E501
|
||||
FLASH_ATTN_V1 = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend" # noqa: E501
|
||||
if selected_backend == _Backend.TRITON_ATTN_VLLM_V1:
|
||||
TRITON_ATTN = "vllm.v1.attention.backends.triton_attn.TritonAttentionBackend" # noqa: E501
|
||||
FLASH_ATTN = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend" # noqa: E501
|
||||
if selected_backend == _Backend.TRITON_ATTN:
|
||||
logger.info_once("Using Triton backend on V1 engine.")
|
||||
return TRITON_ATTN_VLLM_V1
|
||||
return TRITON_ATTN
|
||||
elif selected_backend == _Backend.FLASH_ATTN:
|
||||
logger.info_once("Using Flash Attention backend on V1 engine.")
|
||||
return FLASH_ATTN_V1
|
||||
return FLASH_ATTN
|
||||
elif selected_backend:
|
||||
raise ValueError(
|
||||
f"Invalid attention backend for {cls.device_name}, "
|
||||
@@ -64,7 +64,7 @@ class XPUPlatform(Platform):
|
||||
XPU only support fp8 kv cache with triton backend.
|
||||
"""
|
||||
if envs.is_set("VLLM_ATTENTION_BACKEND") and \
|
||||
envs.VLLM_ATTENTION_BACKEND == "TRITON_ATTN_VLLM_V1":
|
||||
envs.VLLM_ATTENTION_BACKEND == "TRITON_ATTN":
|
||||
return kv_cache_dtype in ["fp8_e4m3", "fp8_e5m2", "fp8"]
|
||||
|
||||
return False
|
||||
|
||||
Reference in New Issue
Block a user