biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V0 Deprecation] Deprecate BlockSparse Attention & Phi3-Small (#21217)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Woosuk Kwon
2025-07-19 13:53:17 -07:00
committed by GitHub
parent 881e3cbe3b
commit 752c6ade2e
38 changed files with 65 additions and 2435 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/v1/attention/backends/cpu_attn.py
View File

@@ -1,7 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from typing import Any, Optional
from typing import Optional
import numpy as np
import torch
@@ -443,7 +443,6 @@ class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: str = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[str] = None,
@@ -451,9 +450,6 @@ class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
) -> None:
if kv_sharing_target_layer_name is not None:
raise NotImplementedError("KV sharing is not supported in V0.")
if blocksparse_params is not None:
raise ValueError(
"Torch SPDA does not support block-sparse attention.")
if logits_soft_cap is not None:
logger.warning_once("Torch SPDA does not support logits soft cap. "
"Outputs may be slightly off.")

6
vllm/v1/attention/backends/flash_attn.py
View File

@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Attention layer with FlashAttention."""
from dataclasses import dataclass
from typing import Any, ClassVar, Optional
from typing import ClassVar, Optional
import numpy as np
import torch
@@ -349,15 +349,11 @@ class FlashAttentionImpl(AttentionImpl):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: AttentionType = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[str] = None,
use_irope: bool = False,
) -> None:
if blocksparse_params is not None:
raise ValueError(
"FlashAttention does not support block-sparse attention.")
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)

3
vllm/v1/attention/backends/flashinfer.py
View File

@@ -4,7 +4,7 @@
from __future__ import annotations
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Optional
from typing import TYPE_CHECKING, Optional
import torch
from flashinfer import (BatchDecodeWithPagedKVCacheWrapper,
@@ -490,7 +490,6 @@ class FlashInferImpl(AttentionImpl):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: AttentionType = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[int] = None,

7
vllm/v1/attention/backends/flex_attention.py
View File

@@ -3,7 +3,7 @@
"""Attention layer with FlashAttention."""
from collections import defaultdict
from dataclasses import dataclass
from typing import Any, Optional
from typing import Optional
import torch
from torch.nn.attention.flex_attention import (BlockMask, _mask_mod_signature,
@@ -342,15 +342,10 @@ class FlexAttentionImpl(AttentionImpl):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: AttentionType = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[str] = None,
) -> None:
if blocksparse_params is not None:
# TODO we should support this :think
raise ValueError(
"FlashAttention does not support block-sparse attention.")
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)

3
vllm/v1/attention/backends/mla/common.py
View File

@@ -190,7 +190,7 @@ return curr_o @ W_O
import functools
from abc import abstractmethod
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Any, Generic, Optional, TypeVar, Union
from typing import TYPE_CHECKING, Generic, Optional, TypeVar, Union
import torch
@@ -754,7 +754,6 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]],
logits_soft_cap: Optional[float],
attn_type: str,
kv_sharing_target_layer_name: Optional[str],

12
vllm/v1/attention/backends/mla/cutlass_mla.py
View File

@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
from typing import Any, Optional
from typing import Optional
import torch
@@ -74,7 +74,6 @@ class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]],
logits_soft_cap: Optional[float],
attn_type: str,
kv_sharing_target_layer_name: Optional[str],
@@ -82,17 +81,14 @@ class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
**mla_args) -> None:
super().__init__(num_heads, head_size, scale, num_kv_heads,
alibi_slopes, sliding_window, kv_cache_dtype,
blocksparse_params, logits_soft_cap, attn_type,
logits_soft_cap, attn_type,
kv_sharing_target_layer_name, **mla_args)
unsupported_features = [
alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap
]
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
if any(unsupported_features):
raise NotImplementedError(
"CutlassMLAImpl does not support one of the following: "
"alibi_slopes, sliding_window, blocksparse_params, "
"logits_soft_cap")
"alibi_slopes, sliding_window, logits_soft_cap")
if attn_type != AttentionType.DECODER:
raise NotImplementedError("Encoder self-attention and "

12
vllm/v1/attention/backends/mla/flashmla.py
View File

@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from typing import Any, ClassVar, Optional
from typing import ClassVar, Optional
import torch
@@ -119,7 +119,6 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]],
logits_soft_cap: Optional[float],
attn_type: str,
kv_sharing_target_layer_name: Optional[str],
@@ -127,20 +126,17 @@ class FlashMLAImpl(MLACommonImpl[FlashMLAMetadata]):
**mla_args) -> None:
super().__init__(num_heads, head_size, scale, num_kv_heads,
alibi_slopes, sliding_window, kv_cache_dtype,
blocksparse_params, logits_soft_cap, attn_type,
logits_soft_cap, attn_type,
kv_sharing_target_layer_name, **mla_args)
assert is_flashmla_supported(), \
"FlashMLA is not supported on this device"
unsupported_features = [
alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap
]
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
if any(unsupported_features):
raise NotImplementedError(
"FlashMLAImpl does not support one of the following: "
"alibi_slopes, sliding_window, blocksparse_params, "
"logits_soft_cap")
"alibi_slopes, sliding_window, logits_soft_cap")
if attn_type != AttentionType.DECODER:
raise NotImplementedError("Encoder self-attention and "

12
vllm/v1/attention/backends/mla/rocm_aiter_mla.py
View File

@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from typing import Any, ClassVar, Optional
from typing import ClassVar, Optional
import torch
@@ -167,7 +167,6 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]],
logits_soft_cap: Optional[float],
attn_type: str,
kv_sharing_target_layer_name: Optional[str],
@@ -175,20 +174,17 @@ class AiterMLAImpl(MLACommonImpl[AiterMLAMetadata]):
**mla_args) -> None:
super().__init__(num_heads, head_size, scale, num_kv_heads,
alibi_slopes, sliding_window, kv_cache_dtype,
blocksparse_params, logits_soft_cap, attn_type,
logits_soft_cap, attn_type,
kv_sharing_target_layer_name, **mla_args)
assert (num_heads == 16 or num_heads == 128), (
f"Aiter MLA only supports 16 or 128 number of heads.\n"
f"Provided {num_heads} number of heads.\n"
"Try adjusting tensor_parallel_size value.")
unsupported_features = [
alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap
]
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
if any(unsupported_features):
raise NotImplementedError(
"Aiter MLA does not support one of the following: "
"alibi_slopes, sliding_window, blocksparse_params, "
"logits_soft_cap")
"alibi_slopes, sliding_window, logits_soft_cap")
from aiter import flash_attn_varlen_func
self.flash_attn_varlen_func = flash_attn_varlen_func

12
vllm/v1/attention/backends/mla/triton_mla.py
View File

@@ -1,7 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any, Optional
from typing import Optional
import torch
@@ -42,7 +42,6 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]],
logits_soft_cap: Optional[float],
attn_type: str,
kv_sharing_target_layer_name: Optional[str],
@@ -50,17 +49,14 @@ class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
**mla_args) -> None:
super().__init__(num_heads, head_size, scale, num_kv_heads,
alibi_slopes, sliding_window, kv_cache_dtype,
blocksparse_params, logits_soft_cap, attn_type,
logits_soft_cap, attn_type,
kv_sharing_target_layer_name, **mla_args)
unsupported_features = [
alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap
]
unsupported_features = [alibi_slopes, sliding_window, logits_soft_cap]
if any(unsupported_features):
raise NotImplementedError(
"TritonMLAImpl does not support one of the following: "
"alibi_slopes, sliding_window, blocksparse_params, "
"logits_soft_cap")
"alibi_slopes, sliding_window, logits_soft_cap")
if attn_type != AttentionType.DECODER:
raise NotImplementedError("Encoder self-attention and "

8
vllm/v1/attention/backends/pallas.py
View File

@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from typing import Any, Optional
from typing import Optional
import torch
import torch_xla.core.xla_builder as xb
@@ -132,7 +132,6 @@ class PallasAttentionBackendImpl(AttentionImpl):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: str = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[int] = None,
@@ -142,9 +141,6 @@ class PallasAttentionBackendImpl(AttentionImpl):
logger.warning_once(
"Using irope in Pallas is not supported yet, it will fall back "
"to global attention for long context.")
if blocksparse_params is not None:
raise ValueError("Paged attention Pallas kernel does "
"not support block-sparse attention.")
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)
@@ -158,8 +154,6 @@ class PallasAttentionBackendImpl(AttentionImpl):
raise NotImplementedError("Alibi slopes is not supported.")
if kv_cache_dtype != "auto":
raise NotImplementedError("FP8 KV cache dtype is not supported.")
if blocksparse_params is not None:
raise NotImplementedError("Blocksparse is not supported.")
if attn_type != AttentionType.DECODER:
raise NotImplementedError("Encoder self-attention and "

6
vllm/v1/attention/backends/rocm_aiter_fa.py
View File

@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Attention layer with AiterFlashAttention."""
from dataclasses import dataclass
from typing import Any, Optional
from typing import Optional
import torch
@@ -334,15 +334,11 @@ class AiterFlashAttentionImpl(AttentionImpl):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: AttentionType = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[int] = None,
use_irope: bool = False,
) -> None:
if blocksparse_params is not None:
raise ValueError(
"AiterFlashAttention does not support block-sparse attention.")
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)

6
vllm/v1/attention/backends/triton_attn.py
View File

@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Attention layer with PagedAttention and Triton prefix prefill."""
from dataclasses import dataclass
from typing import Any, ClassVar, Optional
from typing import ClassVar, Optional
import torch
@@ -205,15 +205,11 @@ class TritonAttentionImpl(AttentionImpl):
alibi_slopes: Optional[list[float]],
sliding_window: Optional[int],
kv_cache_dtype: str,
blocksparse_params: Optional[dict[str, Any]] = None,
logits_soft_cap: Optional[float] = None,
attn_type: AttentionType = AttentionType.DECODER,
kv_sharing_target_layer_name: Optional[int] = None,
use_irope: bool = False,
) -> None:
if blocksparse_params is not None:
raise ValueError(
"TritonAttention does not support block-sparse attention.")
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)