EngineX-Mthreads/enginex-mthreads-vllm
4
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

init

Browse Source
This commit is contained in:
luopingyi
2026-01-09 13:34:11 +08:00
parent dfa6476b58
commit b2ef04d792
538 changed files with 105693 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/attention/__init__.py Normal file
View File

@@ -0,0 +1,13 @@
from vllm.attention.backends.abstract import (AttentionBackend,
AttentionMetadata,
AttentionMetadataPerStage)
from vllm.attention.layer import Attention
from vllm.attention.selector import get_attn_backend
__all__ = [
"AttentionBackend",
"AttentionMetadata",
"Attention",
"get_attn_backend",
"AttentionMetadataPerStage",
]

0
vllm/attention/backends/__init__.py Normal file
View File

127
vllm/attention/backends/abstract.py Normal file
View File

@@ -0,0 +1,127 @@
from abc import ABC, abstractmethod
from dataclasses import dataclass, fields
from typing import (Any, Dict, Generic, List, Optional, Set, Tuple, Type,
TypeVar)
import torch
class AttentionBackend(ABC):
"""Abstract class for attention backends."""
@staticmethod
@abstractmethod
def get_impl_cls() -> Type["AttentionImpl"]:
raise NotImplementedError
@staticmethod
@abstractmethod
def make_metadata(*args, **kwargs) -> "AttentionMetadataPerStage":
raise NotImplementedError
@staticmethod
@abstractmethod
def get_kv_cache_shape(
num_blocks: int,
block_size: int,
num_kv_heads: int,
head_size: int,
) -> Tuple[int, ...]:
raise NotImplementedError
@staticmethod
@abstractmethod
def swap_blocks(
src_kv_cache: torch.Tensor,
dst_kv_cache: torch.Tensor,
src_to_dst: Dict[int, int],
) -> None:
raise NotImplementedError
@staticmethod
@abstractmethod
def copy_blocks(
kv_caches: List[torch.Tensor],
src_to_dists: Dict[int, List[int]],
) -> None:
raise NotImplementedError
@dataclass
class AttentionMetadataPerStage:
"""Attention metadata for a specific stage. I.e., prefill or decode."""
def asdict_zerocopy(self,
skip_fields: Optional[Set[str]] = None
) -> Dict[str, Any]:
"""Similar to dataclasses.asdict, but avoids deepcopying."""
if skip_fields is None:
skip_fields = set()
# Note that if we add dataclasses as fields, they will need
# similar handling.
return {
field.name: getattr(self, field.name)
for field in fields(self) if field.name not in skip_fields
}
T = TypeVar("T", bound=AttentionMetadataPerStage)
@dataclass
class AttentionMetadata(Generic[T]):
"""Attention metadata for prefill and decode batched together."""
# Total number of prefill requests.
num_prefills: int
# Number of prefill tokens.
num_prefill_tokens: int
# Number of decode tokens. Note that it is equivalent to the number of
# decode requests.
num_decode_tokens: int
# The attention metadata for prefill requests in a batch.
# None if there's no prefill requests in a batch.
prefill_metadata: Optional[T]
# The attention metadata for decode requests in a batch.
# None if there's no decode requests in a batch.
decode_metadata: Optional[T]
# (num_tokens,). The indices of the token slots that input tokens will be
# stored into. E.g., if `slot_mapping` is [35, 2, 17] and the block size
# is 16, the three tokens are stored in the 3rd slot in block 2, 2nd slot
# in block 0, and 1st slot in block 1, respectively.
slot_mapping: torch.Tensor
# The kv cache's data type.
kv_cache_dtype: str
def __post_init__(self):
if self.num_prefill_tokens > 0:
assert self.num_prefills > 0
assert self.prefill_metadata is not None
if self.num_decode_tokens > 0:
assert self.decode_metadata is not None
class AttentionImpl(ABC):
@abstractmethod
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: Optional[int] = None,
alibi_slopes: Optional[List[float]] = None,
sliding_window: Optional[int] = None,
) -> None:
raise NotImplementedError
@abstractmethod
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
kv_cache: torch.Tensor,
attn_metadata: AttentionMetadata,
kv_scale: float,
) -> torch.Tensor:
raise NotImplementedError

283
vllm/attention/backends/flash_attn.py Normal file
View File

@@ -0,0 +1,283 @@
"""Attention layer with Flash and PagedAttention.
NOTE(woosuk): At the moment, this file includes a lot of duplicated code from
XFormers backend. The duplicated code will be removed once we use flash-attn or
flashinfer for all the attention operations.
"""
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple, Type
import torch
import torch_musa
from torch.nn.functional import scaled_dot_product_attention
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
AttentionMetadata,
AttentionMetadataPerStage)
from vllm.attention.ops.paged_attn import (PagedAttention,
PagedAttentionMetadata)
class FlashAttentionBackend(AttentionBackend):
@staticmethod
def get_impl_cls() -> Type["FlashAttentionImpl"]:
return FlashAttentionImpl
@staticmethod
def make_metadata(*args, **kwargs) -> "FlashAttentionMetadata":
return FlashAttentionMetadata(*args, **kwargs)
@staticmethod
def get_kv_cache_shape(
num_blocks: int,
block_size: int,
num_kv_heads: int,
head_size: int,
) -> Tuple[int, ...]:
return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
num_kv_heads, head_size)
@staticmethod
def swap_blocks(
src_kv_cache: torch.Tensor,
dst_kv_cache: torch.Tensor,
src_to_dst: Dict[int, int],
) -> None:
PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
@staticmethod
def copy_blocks(
kv_caches: List[torch.Tensor],
src_to_dists: Dict[int, List[int]],
) -> None:
PagedAttention.copy_blocks(kv_caches, src_to_dists)
@dataclass
class FlashAttentionMetadata(AttentionMetadataPerStage,
PagedAttentionMetadata):
"""Metadata for FlashAttentionBackend.
NOTE: Any python object stored here is not updated when it is
cuda-graph replayed. If you have values that need to be changed
dynamically, it should be stored in tensor. The tensor has to be
updated from `CUDAGraphRunner.forward` API.
"""
# Currently, input sequences can only contain all prompts
# or all decoding. True if all sequences are prompts.
is_prompt: bool
# (batch_size,). The sequence length per sequence. Sequence length means
# the computed tokens + new tokens None if it is a decoding.
seq_lens: Optional[List[int]]
# seq_lens stored as a tensor.
seq_lens_tensor: Optional[torch.Tensor]
# NOTE(sang): Definition of context_len, query_len, and seq_len.
# |---------- N-1 iteration --------|
# |---------------- N iteration ---------------------|
# |- tokenA -|......................|-- newTokens ---|
# |---------- context_len ----------|
# |-------------------- seq_len ----------------------|
# |-- query_len ---|
# Maximum query length in the batch.
max_query_len: Optional[int]
# Maximum sequence length in the batch.
max_seq_len: Optional[int]
# (batch_size + 1,). The cumulative subquery lengths of the sequences in
# the batch, used to index into subquery. E.g., if the subquery length
# is [4, 6], it is [0, 4, 10].
subquery_start_loc: Optional[torch.Tensor]
# (batch_size + 1,). The cumulative sequence lengths of the sequences in
# the batch, used to index into sequence. E.g., if the sequence length is
# [4, 6], it is [0, 4, 10].
seq_start_loc: Optional[torch.Tensor]
# (batch_size,) A tensor of context lengths (tokens that are computed
# so far).
context_lens_tensor: Optional[torch.Tensor]
# Whether or not if cuda graph is enabled.
# Cuda-graph is currently enabled for decoding only.
# TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
use_cuda_graph: bool
class FlashAttentionImpl(AttentionImpl):
"""
If the input tensors contain prompt tokens, the layout is as follows:
|<--------------- num_prefill_tokens ----------------->|
|<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
Otherwise, the layout is as follows:
|<----------------- num_decode_tokens ------------------>|
|<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
Generation tokens can contain padding when cuda-graph is used.
Currently, prompt tokens don't contain any padding.
The prompts might have different lengths, while the generation tokens
always have length 1.
If chunked prefill is enabled, prefill tokens and decode tokens can be
batched together in a flattened 1D query.
|<----- num_prefill_tokens ---->|<------- num_decode_tokens --------->|
|<-prefill_0->|...|<-prefill_N-1->|<--decode_0-->|...|<--decode_M-1-->|
Currently, cuda graph is disabled for chunked prefill, meaning there's no
padding between prefill and decode tokens.
"""
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: Optional[int] = None,
alibi_slopes: Optional[List[float]] = None,
sliding_window: Optional[int] = None,
) -> None:
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)
self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
self.sliding_window = -1
if alibi_slopes is not None:
alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
self.alibi_slopes = alibi_slopes
self.need_mask = (self.alibi_slopes is not None
or self.sliding_window is not None)
assert self.num_heads % self.num_kv_heads == 0
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
suppored_head_sizes = PagedAttention.get_supported_head_sizes()
if head_size not in suppored_head_sizes:
raise ValueError(
f"Head size {head_size} is not supported by PagedAttention. "
f"Supported head sizes are: {suppored_head_sizes}.")
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
kv_cache: torch.Tensor,
attn_metadata: AttentionMetadata[FlashAttentionMetadata],
kv_scale: float,
) -> torch.Tensor:
"""Forward pass with FlashAttention and PagedAttention.
Args:
query: shape = [num_tokens, num_heads * head_size]
key: shape = [num_tokens, num_kv_heads * head_size]
value: shape = [num_tokens, num_kv_heads * head_size]
kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
attn_metadata: Metadata for attention.
Returns:
shape = [num_tokens, num_heads * head_size]
"""
num_tokens, hidden_size = query.shape
# Reshape the query, key, and value tensors.
query = query.view(-1, self.num_heads, self.head_size)
key = key.view(-1, self.num_kv_heads, self.head_size)
value = value.view(-1, self.num_kv_heads, self.head_size)
# enable musa flash attention
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_math_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp(True)
if kv_cache is not None:
key_cache, value_cache = PagedAttention.split_kv_cache(
kv_cache, self.num_kv_heads, self.head_size)
# Reshape the input keys and values and store them in the cache.
# If kv_cache is not provided, the new key and value tensors are
# not cached. This happens during the initial memory profiling run.
PagedAttention.write_to_paged_cache(key, value, key_cache,
value_cache,
attn_metadata.slot_mapping,
attn_metadata.kv_cache_dtype,
kv_scale)
num_prefill_tokens = attn_metadata.num_prefill_tokens
num_decode_tokens = attn_metadata.num_decode_tokens
assert key.shape[0] == num_prefill_tokens + num_decode_tokens
assert value.shape[0] == num_prefill_tokens + num_decode_tokens
output = torch.empty_like(query)
# Query for decode. KV is not needed because it is already cached.
decode_query = query[num_prefill_tokens:]
# QKV for prefill.
query = query[:num_prefill_tokens]
key = key[:num_prefill_tokens]
value = value[:num_prefill_tokens]
query = query.movedim(0, query.dim() - 2).unsqueeze(0)
key = key.movedim(0, key.dim() - 2).unsqueeze(0)
value = value.movedim(0, value.dim() - 2).unsqueeze(0)
assert decode_query.shape[0] == num_decode_tokens
if prefill_meta := attn_metadata.prefill_metadata:
tensor = torch.full(
(1, 1, num_tokens, num_tokens),
dtype=torch.bool,
fill_value=1,
device=query.device)
att_mask = torch.tril(tensor, diagonal=0)
# Prompt run.
if kv_cache is None or prefill_meta.block_tables.numel() == 0:
# normal attention
# When block_tables are not filled, it means q and k are the
# prompt, and they have the same length.
attn_output = scaled_dot_product_attention(
query.contiguous(),
key.contiguous(),
value.contiguous(),
attn_mask=att_mask.contiguous(),
dropout_p=0.0,
is_causal=False,
)
attn_output = attn_output.squeeze(0).permute(1, 0, 2).contiguous()
assert output[:num_prefill_tokens].shape == attn_output.shape
output[:num_prefill_tokens] = attn_output
else:
# prefix-enabled attention
# TODO(Hai) this triton kernel has regression issue (broke) to
# deal with different data types between KV and FP8 KV cache,
# to be addressed separately.
output[:num_prefill_tokens] = PagedAttention.forward_prefix(
query,
key,
value,
key_cache,
value_cache,
prefill_meta.block_tables,
prefill_meta.subquery_start_loc,
prefill_meta.seq_lens_tensor,
prefill_meta.context_lens_tensor,
prefill_meta.max_query_len,
self.alibi_slopes,
self.sliding_window[0],
)
if decode_meta := attn_metadata.decode_metadata:
# Decoding run.
output[num_prefill_tokens:] = PagedAttention.forward_decode(
decode_query,
key_cache,
value_cache,
decode_meta.block_tables,
decode_meta.seq_lens_tensor,
decode_meta.max_seq_len,
attn_metadata.kv_cache_dtype,
self.num_kv_heads,
self.scale,
self.alibi_slopes,
kv_scale,
)
# Reshape the output tensor.
return output.view(num_tokens, hidden_size)

220
vllm/attention/backends/flashinfer.py Normal file
View File

@@ -0,0 +1,220 @@
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Set, Tuple, Type
try:
import flashinfer
from flash_attn import flash_attn_varlen_func
from flashinfer import BatchDecodeWithPagedKVCacheWrapper
except ImportError:
flashinfer = None
flash_attn_varlen_func = None
BatchDecodeWithPagedKVCacheWrapper = None
import torch
from vllm import _custom_ops as ops
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
AttentionMetadata,
AttentionMetadataPerStage)
class FlashInferBackend(AttentionBackend):
@staticmethod
def get_impl_cls() -> Type["FlashInferImpl"]:
return FlashInferImpl
@staticmethod
def make_metadata(*args, **kwargs) -> "FlashInferMetadata":
return FlashInferMetadata(*args, **kwargs)
@staticmethod
def get_kv_cache_shape(
num_blocks: int,
block_size: int,
num_kv_heads: int,
head_size: int,
) -> Tuple[int, ...]:
return (num_blocks, 2, block_size, num_kv_heads, head_size)
@staticmethod
def swap_blocks(
src_kv_cache: torch.Tensor,
dst_kv_cache: torch.Tensor,
src_to_dst: Dict[int, int],
) -> None:
raise NotImplementedError
@staticmethod
def copy_blocks(
kv_caches: List[torch.Tensor],
src_to_dists: Dict[int, List[int]],
) -> None:
raise NotImplementedError
@staticmethod
def get_supported_head_sizes() -> List[int]:
return [64, 128, 256]
@dataclass
class FlashInferMetadata(AttentionMetadataPerStage):
is_prompt: bool
use_cuda_graph: bool = False
decode_wrapper: Optional[BatchDecodeWithPagedKVCacheWrapper] = None
# Metadata for the prefill stage since we still
# use flash attention for prefill.
seq_start_loc: Optional[torch.Tensor] = None
max_seq_len: Optional[int] = None
block_tables: Optional[torch.Tensor] = None
# Metadata for the decode stage
# Workspace buffer required by the kernel, the buffer should not
# be allocated/deacollated by the FalshInfermetadata object.
workspace_buffer: Optional[torch.Tensor] = None
# An example for paged_kv_indices, paged_kv_indptr:
# request 1, page indices [0, 5, 8]
# request 2, page indices [1, 6, 7]
# request 3, page indices [3, 4]
# paged_kv_indices is a concatenation of page indices of all requests:
# [0, 5, 8, 1, 6, 7, 3, 4]
# paged_kv_indptr is used to index into paged_kv_indices:
# [0, 3, 6, 8]
# The indptr of the paged kv cache, shape: [batch_size + 1]
paged_kv_indptr: Optional[torch.Tensor] = None
# The page indices of the paged kv cache
paged_kv_indices: Optional[torch.Tensor] = None
# The number of entries in the last page of each request in
# the paged kv cache, shape: [batch_size]
paged_kv_last_page_len: Optional[torch.Tensor] = None
# The number of query/output heads
num_qo_heads: Optional[int] = None
# The number of key/value heads
num_kv_heads: Optional[int] = None
# The dimension of the attention heads
head_dim: Optional[int] = None
# Block size of vllm
page_size: Optional[int] = None
# The data type of the paged kv cache
data_type: torch.dtype = None
def __post_init__(self):
# Refer to
# https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157
supported_head_sizes = FlashInferBackend.get_supported_head_sizes()
if self.head_dim is not None and self.head_dim \
not in supported_head_sizes:
raise ValueError(
f"Only {supported_head_sizes} are supported for head_dim,",
f"received {self.head_dim}.")
# When using flashinfer, we are also creating the FlashInferMetadata,
# which will also call post_init by default, here we want to skip the
# post_init if it's the prefill phase.
if not self.is_prompt:
self.decode_wrapper = flashinfer.BatchDecodeWithPagedKVCacheWrapper(
self.workspace_buffer, "NHD")
self.decode_wrapper.begin_forward(
self.paged_kv_indptr,
self.paged_kv_indices,
self.paged_kv_last_page_len,
self.num_qo_heads,
self.num_kv_heads,
self.head_dim,
self.page_size,
# Disable flashinfer's pos encoding and use vllm's rope.
pos_encoding_mode="NONE",
data_type=self.data_type)
def asdict_zerocopy(self,
skip_fields: Optional[Set[str]] = None
) -> Dict[str, Any]:
if skip_fields is None:
skip_fields = set()
# We need to skip the decode_wrapper field since it cannot be
# broadcasted with nccl when TP is enabled.
skip_fields.add('decode_wrapper')
return super().asdict_zerocopy(skip_fields)
class FlashInferImpl(AttentionImpl):
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: Optional[int] = None,
alibi_slopes: Optional[List[float]] = None,
sliding_window: Optional[int] = None,
) -> None:
if sliding_window is not None:
raise ValueError("Sliding window is not supported in FlashInfer.")
self.sliding_window = (-1, -1)
self.alibi_slopes = alibi_slopes
self.scale = scale
self.num_heads = num_heads
self.head_size = head_size
self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
def forward(self, query: torch.Tensor, key: torch.Tensor,
value: torch.Tensor, kv_cache: Optional[torch.Tensor],
attn_metadata: AttentionMetadata[FlashInferMetadata],
kv_scale: float):
num_tokens, hidden_size = query.shape
query = query.view(-1, self.num_heads, self.head_size)
key = key.view(-1, self.num_kv_heads, self.head_size)
value = value.view(-1, self.num_kv_heads, self.head_size)
if attn_metadata.num_prefill_tokens > 0:
assert attn_metadata.num_decode_tokens == 0, (
"Chunked prefill is not supported with flashinfer yet.")
if attn_metadata.num_decode_tokens > 0:
assert attn_metadata.num_prefill_tokens == 0, (
"Chunked prefill is not supported with flashinfer yet.")
if kv_cache is not None:
# Use the same reshape and cache kernel as flash attention.
ops.reshape_and_cache_flash(
key,
value,
kv_cache[:, 0],
kv_cache[:, 1],
attn_metadata.slot_mapping.flatten(),
attn_metadata.kv_cache_dtype,
)
if prefill_meta := attn_metadata.prefill_metadata:
assert prefill_meta.block_tables is not None
if kv_cache is None or prefill_meta.block_tables.numel() == 0:
output = flash_attn_varlen_func(
q=query,
k=key,
v=value,
cu_seqlens_q=prefill_meta.seq_start_loc,
cu_seqlens_k=prefill_meta.seq_start_loc,
max_seqlen_q=prefill_meta.max_seq_len,
max_seqlen_k=prefill_meta.max_seq_len,
softmax_scale=self.scale,
causal=True,
window_size=self.sliding_window,
alibi_slopes=self.alibi_slopes,
)
else:
raise NotImplementedError(
"Prefix caching is not supported with flashinfer yet.")
else:
assert attn_metadata.decode_metadata is not None
assert attn_metadata.decode_metadata.decode_wrapper is not None
query = query.contiguous(
) # Flashinfer requires query to be contiguous
output = attn_metadata.decode_metadata.decode_wrapper.forward(
query,
kv_cache,
sm_scale=self.scale,
)
return output.view(num_tokens, hidden_size)

374
vllm/attention/backends/rocm_flash_attn.py Normal file
View File

@@ -0,0 +1,374 @@
"""Attention layer ROCm GPUs."""
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple, Type
import torch
import vllm.envs as envs
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
AttentionMetadata,
AttentionMetadataPerStage)
from vllm.attention.ops.paged_attn import (PagedAttention,
PagedAttentionMetadata)
from vllm.logger import init_logger
logger = init_logger(__name__)
class ROCmFlashAttentionBackend(AttentionBackend):
@staticmethod
def get_impl_cls() -> Type["ROCmFlashAttentionImpl"]:
return ROCmFlashAttentionImpl
@staticmethod
def make_metadata(*args, **kwargs) -> "ROCmFlashAttentionMetadata":
return ROCmFlashAttentionMetadata(*args, **kwargs)
@staticmethod
def get_kv_cache_shape(
num_blocks: int,
block_size: int,
num_kv_heads: int,
head_size: int,
) -> Tuple[int, ...]:
return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
num_kv_heads, head_size)
@staticmethod
def swap_blocks(
src_kv_cache: torch.Tensor,
dst_kv_cache: torch.Tensor,
src_to_dst: Dict[int, int],
) -> None:
PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
@staticmethod
def copy_blocks(
kv_caches: List[torch.Tensor],
src_to_dists: Dict[int, List[int]],
) -> None:
PagedAttention.copy_blocks(kv_caches, src_to_dists)
@dataclass
class ROCmFlashAttentionMetadata(AttentionMetadataPerStage,
PagedAttentionMetadata):
"""Metadata for FlashAttentionBackend.
NOTE: Any python object stored here is not updated when it is
cuda-graph replayed. If you have values that need to be changed
dynamically, it should be stored in tensor. The tensor has to be
updated from `CUDAGraphRunner.forward` API.
"""
# Currently, input sequences can only contain all prompts
# or all decoding. True if all sequences are prompts.
is_prompt: bool
# (batch_size,). The sequence length per sequence. Sequence length means
# the computed tokens + new tokens None if it is a decoding.
seq_lens: Optional[List[int]]
# seq_lens stored as a tensor.
seq_lens_tensor: Optional[torch.Tensor]
# NOTE(sang): Definition of context_len, query_len, and seq_len.
# |---------- N-1 iteration --------|
# |---------------- N iteration ---------------------|
# |- tokenA -|......................|-- newTokens ---|
# |---------- context_len ----------|
# |-------------------- seq_len ----------------------|
# |-- query_len ---|
# Maximum query length in the batch.
max_query_len: Optional[int]
# Maximum sequence length in the batch.
max_seq_len: Optional[int]
# (batch_size + 1,). The cumulative subquery lengths of the sequences in
# the batch, used to index into subquery. E.g., if the subquery length
# is [4, 6], it is [0, 4, 10].
subquery_start_loc: Optional[torch.Tensor]
# (batch_size + 1,). The cumulative sequence lengths of the sequences in
# the batch, used to index into sequence. E.g., if the sequence length is
# [4, 6], it is [0, 4, 10].
seq_start_loc: Optional[torch.Tensor]
# Whether or not if cuda graph is enabled.
# Cuda-graph is currently enabled for decoding only.
# TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
use_cuda_graph: bool
# (batch_size,) A tensor of context lengths (tokens that are computed
# so far).
context_lens_tensor: Optional[torch.Tensor]
class ROCmFlashAttentionImpl(AttentionImpl):
"""
If the input tensors contain prompt tokens, the layout is as follows:
|<--------------- num_prompt_tokens -------------->|
|<--prompt_0-->|<--prompt_1-->|...|<--prompt_N-1-->|
Otherwise, the layout is as follows:
|<------------------ num_generation_tokens (M) ----------------->|
|<--generation_0-->|..........|<--generation_M-1-->|<--padding-->|
Generation tokens can contain padding when cuda-graph is used.
Currently, prompt tokens don't contain any padding.
The prompts might have different lengths, while the generation tokens
always have length 1.
If chunked prefill is enabled, prefill tokens and decode tokens can be
batched together in a flattened 1D query.
|<----- num_prefill_tokens ---->|<------- num_decode_tokens ----------->|
|<-prompt_0->|...|<-prompt_N-1->|<-generation_0->|...|<-generation_M-1->|
Currently, cuda graph is disabled for chunked prefill, meaning there's no
padding between prefill and decode tokens.
"""
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: Optional[int] = None,
alibi_slopes: Optional[List[float]] = None,
sliding_window: Optional[int] = None,
) -> None:
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)
self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
self.sliding_window = ((sliding_window, sliding_window)
if sliding_window is not None else (-1, -1))
if alibi_slopes is not None:
alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
self.alibi_slopes = alibi_slopes
assert self.num_heads % self.num_kv_heads == 0
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
suppored_head_sizes = PagedAttention.get_supported_head_sizes()
if head_size not in suppored_head_sizes:
raise ValueError(
f"Head size {head_size} is not supported by PagedAttention. "
f"Supported head sizes are: {suppored_head_sizes}.")
self.use_naive_attn = False
# NOTE: Allow for switching between Triton and CK. Defaulting to triton.
self.use_triton_flash_attn = envs.VLLM_USE_TRITON_FLASH_ATTN
if self.use_triton_flash_attn:
from vllm.attention.ops.triton_flash_attention import ( # noqa: F401
triton_attention)
self.attn_func = triton_attention
logger.debug("Using Triton FA in ROCmBackend")
else:
# if not using triton, navi3x not use flash-attn either
if torch.cuda.get_device_capability()[0] == 11:
self.use_naive_attn = True
else:
try:
from flash_attn import flash_attn_varlen_func # noqa: F401
self.attn_func = flash_attn_varlen_func
logger.debug("Using CK FA in ROCmBackend")
except ModuleNotFoundError:
self.use_naive_attn = True
if self.use_naive_attn:
self.attn_func = _naive_attention
logger.debug("Using naive attention in ROCmBackend")
def repeat_kv(self, x: torch.Tensor, n_rep: int) -> torch.Tensor:
"""torch.repeat_interleave(x, dim=1, repeats=n_rep)"""
tokens, n_kv_heads, head_dim = x.shape
return (x[:, :,
None, :].expand(tokens, n_kv_heads, n_rep,
head_dim).reshape(tokens, n_kv_heads * n_rep,
head_dim))
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
kv_cache: torch.Tensor,
attn_metadata: AttentionMetadata[ROCmFlashAttentionMetadata],
kv_scale: float = 1.0,
) -> torch.Tensor:
"""Forward pass with FlashAttention and PagedAttention.
Args:
query: shape = [num_tokens, num_heads * head_size]
key: shape = [num_tokens, num_kv_heads * head_size]
value: shape = [num_tokens, num_kv_heads * head_size]
kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
attn_metadata: Metadata for attention.
Returns:
shape = [num_tokens, num_heads * head_size]
"""
num_tokens, hidden_size = query.shape
# Reshape the query, key, and value tensors.
query = query.view(-1, self.num_heads, self.head_size)
key = key.view(-1, self.num_kv_heads, self.head_size)
value = value.view(-1, self.num_kv_heads, self.head_size)
if kv_cache is not None:
key_cache, value_cache = PagedAttention.split_kv_cache(
kv_cache, self.num_kv_heads, self.head_size)
# Reshape the input keys and values and store them in the cache.
# If kv_cache is not provided, the new key and value tensors are
# not cached. This happens during the initial memory profiling run.
PagedAttention.write_to_paged_cache(
key,
value,
key_cache,
value_cache,
attn_metadata.slot_mapping,
attn_metadata.kv_cache_dtype,
kv_scale,
)
num_prefill_tokens = attn_metadata.num_prefill_tokens
num_decode_tokens = attn_metadata.num_decode_tokens
assert key.shape[0] == num_prefill_tokens + num_decode_tokens
assert value.shape[0] == num_prefill_tokens + num_decode_tokens
output = torch.empty_like(query)
# Query for decode. KV is not needed because it is already cached.
decode_query = query[num_prefill_tokens:]
# QKV for prefill.
query = query[:num_prefill_tokens]
key = key[:num_prefill_tokens]
value = value[:num_prefill_tokens]
assert query.shape[0] == num_prefill_tokens
assert decode_query.shape[0] == num_decode_tokens
if prefill_meta := attn_metadata.prefill_metadata:
# Prompt run.
assert prefill_meta.seq_lens is not None
if kv_cache is None or prefill_meta.block_tables.numel() == 0:
# triton attention
# When block_tables are not filled, it means q and k are the
# prompt, and they have the same length.
if self.use_triton_flash_attn:
out, _ = self.attn_func(
query,
key,
value,
None,
prefill_meta.seq_start_loc,
prefill_meta.seq_start_loc,
prefill_meta.max_seq_len,
prefill_meta.max_seq_len,
True,
self.scale,
)
elif self.use_naive_attn:
if self.num_kv_heads != self.num_heads:
# Interleave for MQA workaround.
key = self.repeat_kv(key, self.num_queries_per_kv)
value = self.repeat_kv(value, self.num_queries_per_kv)
out = self.attn_func(
query,
key,
value,
prefill_meta.seq_lens,
self.scale,
)
else:
out = self.attn_func(
q=query,
k=key,
v=value,
cu_seqlens_q=prefill_meta.seq_start_loc,
cu_seqlens_k=prefill_meta.seq_start_loc,
max_seqlen_q=prefill_meta.max_seq_len,
max_seqlen_k=prefill_meta.max_seq_len,
softmax_scale=self.scale,
causal=True,
)
# common code for prefill
assert output[:num_prefill_tokens].shape == out.shape
output[:num_prefill_tokens] = out
else:
# prefix-enabled attention
output[:num_prefill_tokens] = PagedAttention.forward_prefix(
query,
key,
value,
key_cache,
value_cache,
prefill_meta.block_tables,
prefill_meta.subquery_start_loc,
prefill_meta.seq_lens_tensor,
prefill_meta.context_lens_tensor,
prefill_meta.max_query_len,
self.alibi_slopes,
self.sliding_window[0],
)
if decode_meta := attn_metadata.decode_metadata:
# Decoding run.
output[num_prefill_tokens:] = PagedAttention.forward_decode(
decode_query,
key_cache,
value_cache,
decode_meta.block_tables,
decode_meta.seq_lens_tensor,
decode_meta.max_seq_len,
attn_metadata.kv_cache_dtype,
self.num_kv_heads,
self.scale,
self.alibi_slopes,
kv_scale,
)
# Reshape the output tensor.
return output.view(num_tokens, hidden_size)
def _naive_attention(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
seq_lens: List[int],
scale: float,
) -> torch.Tensor:
output = torch.empty_like(query)
start = 0
for _, seq_len in enumerate(seq_lens):
end = start + seq_len
out = _naive_masked_attention(
query[start:end],
key[start:end],
value[start:end],
scale,
)
# TODO(woosuk): Unnecessary copy. Optimize.
output[start:end].copy_(out)
start += seq_len
return output
def _naive_masked_attention(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
scale: float,
) -> torch.Tensor:
seq_len, head_size, head_dim = query.shape
attn_mask = torch.triu(torch.ones(seq_len,
seq_len,
dtype=query.dtype,
device=query.device),
diagonal=1)
attn_mask = attn_mask * torch.finfo(query.dtype).min
attn_weights = scale * torch.einsum("qhd,khd->hqk", query, key).float()
attn_weights = attn_weights + attn_mask.float()
attn_weights = torch.softmax(attn_weights, dim=-1).to(value.dtype)
out = torch.einsum("hqk,khd->qhd", attn_weights, value)
return out

253
vllm/attention/backends/torch_sdpa.py Normal file
View File

@@ -0,0 +1,253 @@
""" Attention layer with torch scaled_dot_product_attention
and PagedAttention."""
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple, Type
import torch
import torch_musa
from torch.nn.functional import scaled_dot_product_attention
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
AttentionMetadata,
AttentionMetadataPerStage)
from vllm.attention.ops.paged_attn import (PagedAttention,
PagedAttentionMetadata)
class TorchSDPABackend(AttentionBackend):
@staticmethod
def get_impl_cls() -> Type["TorchSDPABackendImpl"]:
return TorchSDPABackendImpl
@staticmethod
def make_metadata(*args, **kwargs) -> "TorchSDPAMetadata":
return TorchSDPAMetadata(*args, **kwargs)
@staticmethod
def get_kv_cache_shape(
num_blocks: int,
block_size: int,
num_kv_heads: int,
head_size: int,
) -> Tuple[int, ...]:
return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
num_kv_heads, head_size)
@staticmethod
def swap_blocks(
src_kv_cache: torch.Tensor,
dst_kv_cache: torch.Tensor,
src_to_dst: Dict[int, int],
) -> None:
PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
@staticmethod
def copy_blocks(
kv_caches: List[torch.Tensor],
src_to_dists: Dict[int, List[int]],
) -> None:
PagedAttention.copy_blocks(kv_caches, src_to_dists)
@dataclass
class TorchSDPAMetadata(PagedAttentionMetadata, AttentionMetadata):
"""Metadata for TorchSDPABackend.
"""
# Currently, input sequences can only contain all prompts
# or all decoding. True if all sequences are prompts.
is_prompt: bool
slot_mapping: torch.Tensor
seq_lens: Optional[List[int]]
def __post_init__(self):
# Set during the execution of the first attention op.
# It is a list because it is needed to set per prompt
# when alibi slopes is used. It is because of the limitation
# from xformer API.
# will not appear in the __repr__ and __init__
self.attn_bias: Optional[List[torch.Tensor]] = None
class TorchSDPABackendImpl(AttentionImpl):
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: Optional[int] = None,
alibi_slopes: Optional[List[float]] = None,
sliding_window: Optional[int] = None,
) -> None:
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)
self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
self.sliding_window = sliding_window
if alibi_slopes is not None:
assert len(alibi_slopes) == num_heads
alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
self.alibi_slopes = alibi_slopes
self.need_mask = (self.alibi_slopes is not None
or self.sliding_window is not None)
assert self.num_heads % self.num_kv_heads == 0
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
suppored_head_sizes = PagedAttention.get_supported_head_sizes()
if head_size not in suppored_head_sizes:
raise ValueError(
f"Head size {head_size} is not supported by PagedAttention. "
f"Supported head sizes are: {suppored_head_sizes}.")
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
kv_cache: Optional[torch.Tensor],
attn_metadata: TorchSDPAMetadata, # type: ignore
kv_scale: float,
) -> torch.Tensor:
"""Forward pass with torch SDPA and PagedAttention.
Args:
query: shape = [num_tokens, num_heads * head_size]
key: shape = [num_tokens, num_kv_heads * head_size]
value: shape = [num_tokens, num_kv_heads * head_size]
kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
attn_metadata: Metadata for attention.
Returns:
shape = [num_tokens, num_heads * head_size]
"""
num_tokens, hidden_size = query.shape
# Reshape the query, key, and value tensors.
query = query.view(-1, self.num_heads, self.head_size)
key = key.view(-1, self.num_kv_heads, self.head_size)
value = value.view(-1, self.num_kv_heads, self.head_size)
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_math_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp(True)
if kv_cache is not None:
key_cache, value_cache = PagedAttention.split_kv_cache(
kv_cache, self.num_kv_heads, self.head_size)
PagedAttention.write_to_paged_cache(key, value, key_cache,
value_cache,
attn_metadata.slot_mapping,
attn_metadata.kv_cache_dtype,
kv_scale)
if attn_metadata.is_prompt:
assert attn_metadata.seq_lens is not None
if (kv_cache is None or attn_metadata.block_tables.numel() == 0):
if self.num_kv_heads != self.num_heads:
key = key.repeat_interleave(self.num_queries_per_kv, dim=1)
value = value.repeat_interleave(self.num_queries_per_kv,
dim=1)
if attn_metadata.attn_bias is None:
if self.alibi_slopes is not None:
att_masks = _make_alibi_bias(
self.alibi_slopes, query.dtype,
attn_metadata.seq_lens) # type: ignore
elif self.sliding_window is not None:
att_masks = _make_sliding_window_bias(
attn_metadata.prefill_metadata.seq_lens, self.sliding_window,
query.dtype) # type: ignore
else:
att_masks = [None] * len(attn_metadata.prefill_metadata.seq_lens)
attn_metadata.prefill_metadata.attn_bias = att_masks
query = query.movedim(0, query.dim() - 2).unsqueeze(0)
key = key.movedim(0, key.dim() - 2).unsqueeze(0)
value = value.movedim(0, value.dim() - 2).unsqueeze(0)
start = 0
output = torch.empty(
(1, num_tokens, self.num_heads, self.head_size),
dtype=query.dtype)
for seq_len, mask in zip(attn_metadata.prefill_metadata.seq_lens,
attn_metadata.prefill_metadata.attn_bias):
end = start + seq_len
sub_out = scaled_dot_product_attention(
query[:, :, start:end, :],
key[:, :, start:end, :],
value[:, :, start:end, :],
attn_mask=mask,
dropout_p=0.0,
is_causal=not self.need_mask,
scale=self.scale).movedim(query.dim() - 2, 0)
output[start:end, :, :] = sub_out[0]
start = end
else:
# prefix-enabled attention
raise RuntimeError(
"Torch SDPA backend doesn't support prefix decoding.")
else:
# Decoding run.
output = PagedAttention.forward_decode(
query,
key_cache,
value_cache,
attn_metadata.block_tables,
attn_metadata.seq_lens_tensor,
attn_metadata.max_seq_len,
attn_metadata.kv_cache_dtype,
self.num_kv_heads,
self.scale,
self.alibi_slopes,
kv_scale,
)
# Reshape the output tensor.
return output.view(-1, self.num_heads * self.head_size)
def _make_alibi_bias(
alibi_slopes: torch.Tensor,
dtype: torch.dtype,
seq_lens: List[int],
) -> List[torch.Tensor]:
attn_biases = []
for seq_len in seq_lens:
bias = torch.arange(seq_len, dtype=dtype)
# NOTE(zhuohan): HF uses
# `bias = bias[None, :].repeat(seq_len, 1)`
# here. We find that both biases give the same results, but
# the bias below more accurately follows the original ALiBi
# paper.
bias = bias[None, :] - bias[:, None]
num_heads = alibi_slopes.shape[0]
bias = bias[None, :].repeat((num_heads, 1, 1))
bias.mul_(alibi_slopes[:, None, None])
inf_mask = torch.empty(
(1, seq_len, seq_len),
dtype=bias.dtype).fill_(-torch.inf).triu_(diagonal=1)
attn_biases.append((bias + inf_mask).to(dtype))
return attn_biases
def _make_sliding_window_bias(
seq_lens: List[int],
window_size: Optional[int],
dtype: torch.dtype,
) -> List[torch.Tensor]:
attn_biases = []
for seq_len in seq_lens:
tensor = torch.full(
(1, seq_len, seq_len),
dtype=dtype,
fill_value=1,
)
shift = 0
mask = torch.tril(tensor, diagonal=shift).to(dtype) # type: ignore
if window_size is not None:
mask = torch.triu(mask, diagonal=shift - window_size + 1)
mask = torch.log(mask)
attn_biases.append(mask.to(dtype))
return attn_biases

393
vllm/attention/backends/xformers.py Normal file
View File

@@ -0,0 +1,393 @@
"""Attention layer with xFormers and PagedAttention."""
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple, Type
import torch
from xformers import ops as xops
from xformers.ops.fmha.attn_bias import (AttentionBias,
BlockDiagonalCausalMask,
LowerTriangularMaskWithTensorBias)
from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
AttentionMetadata,
AttentionMetadataPerStage)
from vllm.attention.ops.paged_attn import (PagedAttention,
PagedAttentionMetadata)
from vllm.logger import init_logger
logger = init_logger(__name__)
class XFormersBackend(AttentionBackend):
@staticmethod
def get_impl_cls() -> Type["XFormersImpl"]:
return XFormersImpl
@staticmethod
def make_metadata(*args, **kwargs) -> "XFormersMetadata":
return XFormersMetadata(*args, **kwargs)
@staticmethod
def get_kv_cache_shape(
num_blocks: int,
block_size: int,
num_kv_heads: int,
head_size: int,
) -> Tuple[int, ...]:
return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
num_kv_heads, head_size)
@staticmethod
def swap_blocks(
src_kv_cache: torch.Tensor,
dst_kv_cache: torch.Tensor,
src_to_dst: Dict[int, int],
) -> None:
PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
@staticmethod
def copy_blocks(
kv_caches: List[torch.Tensor],
src_to_dists: Dict[int, List[int]],
) -> None:
PagedAttention.copy_blocks(kv_caches, src_to_dists)
@dataclass
class XFormersMetadata(AttentionMetadataPerStage, PagedAttentionMetadata):
"""Metadata for XFormersbackend.
NOTE: Any python object stored here is not updated when it is
cuda-graph replayed. If you have values that need to be changed
dynamically, it should be stored in tensor. The tensor has to be
updated from `CUDAGraphRunner.forward` API.
"""
# Currently, input sequences can only contain all prompts
# or all decoding. True if all sequences are prompts.
is_prompt: bool
# (batch_size,). The sequence length per sequence. Sequence length means
# the computed tokens + new tokens None if it is a decoding.
seq_lens: Optional[List[int]]
# seq_lens stored as a tensor.
seq_lens_tensor: Optional[torch.Tensor]
# |---------- N-1 iteration --------|
# |---------------- N iteration ---------------------|
# |- tokenA -|......................|-- newTokens ---|
# |---------- context_len ----------|
# |-------------------- seq_len ----------------------|
# |-- query_len ---|
# Maximum query length in the batch.
max_query_len: Optional[int]
# FIXME: It is for flash attn.
# Maximum sequence length in the batch.
max_seq_len: Optional[int]
# (batch_size + 1,). The cumulative subquery lengths of the sequences in
# the batch, used to index into subquery. E.g., if the subquery length
# is [4, 6], it is [0, 4, 10].
subquery_start_loc: Optional[torch.Tensor]
# FIXME: It is for flash attn.
# (batch_size + 1,). The cumulative sequence lengths of the sequences in
# the batch, used to index into sequence. E.g., if the sequence length is
# [4, 6], it is [0, 4, 10].
seq_start_loc: Optional[torch.Tensor]
# (batch_size,) A tensor of context lengths (tokens that are computed
# so far).
context_lens_tensor: Optional[torch.Tensor]
# Whether or not if cuda graph is enabled.
# Cuda-graph is currently enabled for decoding only.
# TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
use_cuda_graph: bool
def __post_init__(self):
# Set during the execution of the first attention op.
# It is a list because it is needed to set per prompt
# when alibi slopes is used. It is because of the limitation
# from xformer API.
# will not appear in the __repr__ and __init__
self.attn_bias: Optional[List[AttentionBias]] = None
class XFormersImpl(AttentionImpl):
"""
If the input tensors contain prompt tokens, the layout is as follows:
|<--------------- num_prefill_tokens ----------------->|
|<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
Otherwise, the layout is as follows:
|<----------------- num_decode_tokens ------------------>|
|<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
Generation tokens can contain padding when cuda-graph is used.
Currently, prompt tokens don't contain any padding.
The prompts might have different lengths, while the generation tokens
always have length 1.
If chunked prefill is enabled, prefill tokens and decode tokens can be
batched together in a flattened 1D query.
|<----- num_prefill_tokens ---->|<------- num_decode_tokens --------->|
|<-prefill_0->|...|<-prefill_N-1->|<--decode_0-->|...|<--decode_M-1-->|
Currently, cuda graph is disabled for chunked prefill, meaning there's no
padding between prefill and decode tokens.
"""
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: Optional[int] = None,
alibi_slopes: Optional[List[float]] = None,
sliding_window: Optional[int] = None,
) -> None:
self.num_heads = num_heads
self.head_size = head_size
self.scale = float(scale)
self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
self.sliding_window = sliding_window
if alibi_slopes is not None:
alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
self.alibi_slopes = alibi_slopes
assert self.num_heads % self.num_kv_heads == 0
self.num_queries_per_kv = self.num_heads // self.num_kv_heads
suppored_head_sizes = PagedAttention.get_supported_head_sizes()
if head_size not in suppored_head_sizes:
raise ValueError(
f"Head size {head_size} is not supported by PagedAttention. "
f"Supported head sizes are: {suppored_head_sizes}.")
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
kv_cache: Optional[torch.Tensor],
attn_metadata: AttentionMetadata[XFormersMetadata],
kv_scale: float,
) -> torch.Tensor:
"""Forward pass with xFormers and PagedAttention.
Args:
query: shape = [num_tokens, num_heads * head_size]
key: shape = [num_tokens, num_kv_heads * head_size]
value: shape = [num_tokens, num_kv_heads * head_size]
kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
attn_metadata: Metadata for attention.
Returns:
shape = [num_tokens, num_heads * head_size]
"""
num_tokens, hidden_size = query.shape
query = query.view(-1, self.num_heads, self.head_size)
key = key.view(-1, self.num_kv_heads, self.head_size)
value = value.view(-1, self.num_kv_heads, self.head_size)
if kv_cache is not None:
key_cache, value_cache = PagedAttention.split_kv_cache(
kv_cache, self.num_kv_heads, self.head_size)
# Reshape the input keys and values and store them in the cache.
# If kv_cache is not provided, the new key and value tensors are
# not cached. This happens during the initial memory profiling run.
PagedAttention.write_to_paged_cache(key, value, key_cache,
value_cache,
attn_metadata.slot_mapping,
attn_metadata.kv_cache_dtype,
kv_scale)
num_prefill_tokens = attn_metadata.num_prefill_tokens
num_decode_tokens = attn_metadata.num_decode_tokens
assert key.shape[0] == num_prefill_tokens + num_decode_tokens
assert value.shape[0] == num_prefill_tokens + num_decode_tokens
output = torch.empty_like(query)
# Query for decode. KV is not needed because it is already cached.
decode_query = query[num_prefill_tokens:]
# QKV for prefill.
query = query[:num_prefill_tokens]
key = key[:num_prefill_tokens]
value = value[:num_prefill_tokens]
assert query.shape[0] == num_prefill_tokens
assert decode_query.shape[0] == num_decode_tokens
if prefill_meta := attn_metadata.prefill_metadata:
# Prompt run.
if kv_cache is None or prefill_meta.block_tables.numel() == 0:
# normal attention.
# block tables are empty if the prompt does not have a cached
# prefix.
out = self._run_memory_efficient_xformers_forward(
query, key, value, prefill_meta)
assert out.shape == output[:num_prefill_tokens].shape
output[:num_prefill_tokens] = out
else:
# prefix-enabled attention
# TODO(Hai) this triton kernel has regression issue (broke) to
# deal with different data types between KV and FP8 KV cache,
# to be addressed separately.
out = PagedAttention.forward_prefix(
query,
key,
value,
key_cache,
value_cache,
prefill_meta.block_tables,
prefill_meta.subquery_start_loc,
prefill_meta.seq_lens_tensor,
prefill_meta.context_lens_tensor,
prefill_meta.max_query_len,
self.alibi_slopes,
self.sliding_window,
)
assert output[:num_prefill_tokens].shape == out.shape
output[:num_prefill_tokens] = out
if decode_meta := attn_metadata.decode_metadata:
output[num_prefill_tokens:] = PagedAttention.forward_decode(
decode_query,
key_cache,
value_cache,
decode_meta.block_tables,
decode_meta.seq_lens_tensor,
decode_meta.max_seq_len,
attn_metadata.kv_cache_dtype,
self.num_kv_heads,
self.scale,
self.alibi_slopes,
kv_scale,
)
# Reshape the output tensor.
return output.view(-1, self.num_heads * self.head_size)
def _run_memory_efficient_xformers_forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
attn_metadata: XFormersMetadata,
) -> torch.Tensor:
"""Attention for 1D query of multiple prompts. Multiple prompt
tokens are flattened in to `query` input.
See https://facebookresearch.github.io/xformers/components/ops.html
for API spec.
Args:
output: shape = [num_prefill_tokens, num_heads, head_size]
query: shape = [num_prefill_tokens, num_heads, head_size]
key: shape = [num_prefill_tokens, num_kv_heads, head_size]
value: shape = [num_prefill_tokens, num_kv_heads, head_size]
attn_metadata: Metadata for attention.
"""
assert attn_metadata.seq_lens is not None
original_query = query
if self.num_kv_heads != self.num_heads:
# GQA/MQA requires the shape [B, M, G, H, K].
# Note that the output also has the same shape (which is different
# from a spec from the doc).
query = query.view(query.shape[0], self.num_kv_heads,
self.num_queries_per_kv, query.shape[-1])
key = key[:, :,
None, :].expand(key.shape[0], self.num_kv_heads,
self.num_queries_per_kv, key.shape[-1])
value = value[:, :,
None, :].expand(value.shape[0], self.num_kv_heads,
self.num_queries_per_kv,
value.shape[-1])
# Set attention bias if not provided. This typically happens at
# the very attention layer of every iteration.
# FIXME(woosuk): This is a hack.
if attn_metadata.attn_bias is None:
if self.alibi_slopes is None:
attn_bias = BlockDiagonalCausalMask.from_seqlens(
attn_metadata.seq_lens)
if self.sliding_window is not None:
attn_bias = attn_bias.make_local_attention(
self.sliding_window)
attn_metadata.attn_bias = [attn_bias]
else:
attn_metadata.attn_bias = _make_alibi_bias(
self.alibi_slopes, self.num_kv_heads, query.dtype,
attn_metadata.seq_lens)
# No alibi slopes.
# TODO(woosuk): Too many view operations. Let's try to reduce
# them in the future for code readability.
if self.alibi_slopes is None:
# Add the batch dimension.
query = query.unsqueeze(0)
key = key.unsqueeze(0)
value = value.unsqueeze(0)
out = xops.memory_efficient_attention_forward(
query,
key,
value,
attn_bias=attn_metadata.attn_bias[0],
p=0.0,
scale=self.scale)
return out.view_as(original_query)
# Attention with alibi slopes.
# FIXME(woosuk): Because xformers does not support dynamic sequence
# lengths with custom attention bias, we process each prompt one by
# one. This is inefficient, especially when we have many short prompts.
output = torch.empty_like(original_query)
start = 0
for i, seq_len in enumerate(attn_metadata.seq_lens):
end = start + seq_len
out = xops.memory_efficient_attention_forward(
query[None, start:end],
key[None, start:end],
value[None, start:end],
attn_bias=attn_metadata.attn_bias[i],
p=0.0,
scale=self.scale)
# TODO(woosuk): Unnecessary copy. Optimize.
output[start:end].copy_(out.view_as(original_query[start:end]))
start += seq_len
return output
def _make_alibi_bias(
alibi_slopes: torch.Tensor,
num_kv_heads: int,
dtype: torch.dtype,
seq_lens: List[int],
) -> LowerTriangularMaskWithTensorBias:
attn_biases = []
for seq_len in seq_lens:
bias = torch.arange(seq_len, dtype=dtype)
# NOTE(zhuohan): HF uses
# `bias = bias[None, :].repeat(seq_len, 1)`
# here. We find that both biases give the same results, but
# the bias below more accurately follows the original ALiBi
# paper.
# Calculate a matrix where each element represents ith element- jth
# element.
bias = bias[None, :] - bias[:, None]
padded_len = (seq_len + 7) // 8 * 8
num_heads = alibi_slopes.shape[0]
bias = torch.empty(
1, # batch size
num_heads,
seq_len,
padded_len,
device=alibi_slopes.device,
dtype=dtype,
)[:, :, :, :seq_len].copy_(bias)
bias.mul_(alibi_slopes[:, None, None])
if num_heads != num_kv_heads:
bias = bias.unflatten(1, (num_kv_heads, num_heads // num_kv_heads))
attn_biases.append(LowerTriangularMaskWithTensorBias(bias))
return attn_biases

56
vllm/attention/layer.py Normal file
View File

@@ -0,0 +1,56 @@
"""Attention layer."""
from typing import List, Optional
import torch
import torch.nn as nn
from vllm.attention.backends.abstract import (AttentionMetadata,
AttentionMetadataPerStage)
from vllm.attention.selector import get_attn_backend
class Attention(nn.Module):
"""Attention layer.
This class takes query, key, and value tensors as input. The input tensors
can either contain prompt tokens or generation tokens.
The class does the following:
1. Store the input key and value tensors in the KV cache.
2. Perform (multi-head/multi-query/grouped-query) attention.
3. Return the output tensor.
"""
def __init__(
self,
num_heads: int,
head_size: int,
scale: float,
num_kv_heads: Optional[int] = None,
alibi_slopes: Optional[List[float]] = None,
sliding_window: Optional[int] = None,
) -> None:
super().__init__()
self.backend = get_attn_backend(torch.get_default_dtype())
impl_cls = self.backend.get_impl_cls()
self.impl = impl_cls(num_heads, head_size, scale, num_kv_heads,
alibi_slopes, sliding_window)
def forward(
self,
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
kv_cache: Optional[torch.Tensor],
attn_metadata: AttentionMetadata[AttentionMetadataPerStage],
kv_scale: float = 1.0,
) -> torch.Tensor:
return self.impl.forward(query, key, value, kv_cache, attn_metadata,
kv_scale)
def extra_repr(self) -> str:
s = f"head_size={self.impl.head_size}" # type: ignore
s += f", num_heads={self.impl.num_heads}" # type: ignore
s += f", num_kv_heads={self.impl.num_kv_heads}" # type: ignore
s += f", scale={self.impl.scale}" # type: ignore
return s

0
vllm/attention/ops/__init__.py Normal file
View File

216
vllm/attention/ops/paged_attn.py Normal file
View File

@@ -0,0 +1,216 @@
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
import torch
from vllm import _custom_ops as ops
from vllm.attention.ops.prefix_prefill import context_attention_fwd
# Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
_PARTITION_SIZE = 512
@dataclass
class PagedAttentionMetadata:
"""Metadata for PagedAttention."""
# (batch_size,). The length of sequences (entire tokens seen so far) per
# sequence.
seq_lens_tensor: Optional[torch.Tensor]
# Maximum sequence length in the batch.
max_seq_len: Optional[int]
# (batch_size, max_blocks_per_seq).
# Block addresses per sequence. (Seq id -> list of physical block)
# E.g., [0, 1, 2] means tokens are stored in 0th, 1st, and 2nd blocks
# in the kv cache. Each block can contain up to block_size tokens.
# 2nd dimensions are padded up to max_blocks_per_seq if it is cuda-graph
# captured.
block_tables: Optional[torch.Tensor]
class PagedAttention:
@staticmethod
def get_supported_head_sizes() -> List[int]:
return [64, 80, 96, 112, 128, 256]
@staticmethod
def get_kv_cache_shape(
num_blocks: int,
block_size: int,
num_kv_heads: int,
head_size: int,
) -> Tuple[int, ...]:
return (2, num_blocks, block_size * num_kv_heads * head_size)
@staticmethod
def split_kv_cache(
kv_cache: torch.Tensor,
num_kv_heads: int,
head_size: int,
) -> Tuple[torch.Tensor, torch.Tensor]:
x = 16 // kv_cache.element_size()
num_blocks = kv_cache.shape[1]
key_cache = kv_cache[0]
key_cache = key_cache.view(num_blocks, num_kv_heads, head_size // x,
-1, x)
value_cache = kv_cache[1]
value_cache = value_cache.view(num_blocks, num_kv_heads, head_size, -1)
return key_cache, value_cache
@staticmethod
def write_to_paged_cache(
key: torch.Tensor,
value: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
slot_mapping: torch.Tensor,
kv_cache_dtype: str,
kv_scale: float,
) -> None:
ops.reshape_and_cache(
key,
value,
key_cache,
value_cache,
slot_mapping.flatten(),
kv_cache_dtype,
kv_scale,
)
@staticmethod
def forward_decode(
query: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
block_tables: torch.Tensor,
seq_lens: torch.Tensor,
max_seq_len: int,
kv_cache_dtype: str,
num_kv_heads: int,
scale: float,
alibi_slopes: Optional[torch.Tensor],
kv_scale: float,
) -> torch.Tensor:
output = torch.empty_like(query)
block_size = value_cache.shape[3]
num_seqs, num_heads, head_size = query.shape
max_num_partitions = ((max_seq_len + _PARTITION_SIZE - 1) //
_PARTITION_SIZE)
# NOTE(woosuk): We use a simple heuristic to decide whether to use
# PagedAttention V1 or V2. If the number of partitions is 1, we use
# V1 to avoid the overhead of reduction. Also, if the number of
# sequences or heads is large, we use V1 since there is enough work
# to parallelize.
# TODO(woosuk): Tune this heuristic.
# For context len > 8192, use V2 kernel to avoid shared memory shortage.
use_v1 = (max_seq_len <= 8192
and (max_num_partitions == 1 or num_seqs * num_heads > 512))
if use_v1:
# Run PagedAttention V1.
ops.paged_attention_v1(
output,
query,
key_cache,
value_cache,
num_kv_heads,
scale,
block_tables,
seq_lens,
block_size,
max_seq_len,
alibi_slopes,
kv_cache_dtype,
kv_scale,
)
else:
# Run PagedAttention V2.
assert _PARTITION_SIZE % block_size == 0
tmp_output = torch.empty(
size=(num_seqs, num_heads, max_num_partitions, head_size),
dtype=output.dtype,
device=output.device,
)
exp_sums = torch.empty(
size=(num_seqs, num_heads, max_num_partitions),
dtype=torch.float32,
device=output.device,
)
max_logits = torch.empty_like(exp_sums)
ops.paged_attention_v2(
output,
exp_sums,
max_logits,
tmp_output,
query,
key_cache,
value_cache,
num_kv_heads,
scale,
block_tables,
seq_lens,
block_size,
max_seq_len,
alibi_slopes,
kv_cache_dtype,
kv_scale,
)
return output
@staticmethod
def forward_prefix(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
key_cache: torch.Tensor,
value_cache: torch.Tensor,
block_tables: torch.Tensor,
subquery_start_loc: torch.Tensor,
seq_lens_tensor: torch.Tensor,
context_lens: torch.Tensor,
max_query_len: int,
alibi_slopes: Optional[torch.Tensor],
sliding_window: Optional[int],
) -> torch.Tensor:
output = torch.empty_like(query)
context_attention_fwd(
query,
key,
value,
output,
key_cache,
value_cache,
block_tables,
# subquery_start_loc is (batch_size + 1,)
subquery_start_loc[:-1],
seq_lens_tensor,
context_lens,
max_query_len,
alibi_slopes,
sliding_window,
)
return output
@staticmethod
def swap_blocks(
src_kv_cache: torch.Tensor,
dst_kv_cache: torch.Tensor,
src_to_dst: Dict[int, int],
) -> None:
src_key_cache = src_kv_cache[0]
dst_key_cache = dst_kv_cache[0]
ops.swap_blocks(src_key_cache, dst_key_cache, src_to_dst)
src_value_cache = src_kv_cache[1]
dst_value_cache = dst_kv_cache[1]
ops.swap_blocks(src_value_cache, dst_value_cache, src_to_dst)
@staticmethod
def copy_blocks(
kv_caches: List[torch.Tensor],
src_to_dists: Dict[int, List[int]],
) -> None:
key_caches = [kv_cache[0] for kv_cache in kv_caches]
value_caches = [kv_cache[1] for kv_cache in kv_caches]
ops.copy_blocks(key_caches, value_caches, src_to_dists)

792
vllm/attention/ops/prefix_prefill.py Normal file
View File

@@ -0,0 +1,792 @@
# The kernels in this file are adapted from LightLLM's context_attention_fwd:
# https://github.com/ModelTC/lightllm/blob/main/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py
import torch
import triton
import triton.language as tl
if triton.__version__ >= "2.1.0":
@triton.jit
def _fwd_kernel(
Q,
K,
V,
K_cache,
V_cache,
B_Loc,
sm_scale,
B_Start_Loc,
B_Seqlen,
B_Ctxlen,
block_size,
x,
Out,
stride_b_loc_b,
stride_b_loc_s,
stride_qbs,
stride_qh,
stride_qd,
stride_kbs,
stride_kh,
stride_kd,
stride_vbs,
stride_vh,
stride_vd,
stride_obs,
stride_oh,
stride_od,
stride_k_cache_bs,
stride_k_cache_h,
stride_k_cache_d,
stride_k_cache_bl,
stride_k_cache_x,
stride_v_cache_bs,
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr, # head size
BLOCK_DMODEL_PADDED: tl.constexpr, # head size padded to a power of 2
BLOCK_N: tl.constexpr,
SLIDING_WINDOW: tl.constexpr,
):
cur_batch = tl.program_id(0)
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
cur_batch_query_len = cur_batch_seq_len - cur_batch_ctx_len
# start position inside of the query
# generally, N goes over kv, while M goes over query_len
block_start_loc = BLOCK_M * start_m
# initialize offsets
# [N]; starts at 0
offs_n = tl.arange(0, BLOCK_N)
# [D]; starts at 0
offs_d = tl.arange(0, BLOCK_DMODEL_PADDED)
# [M]; starts at current position in query
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
# [M,D]
off_q = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
cur_head * stride_qh + offs_d[None, :] * stride_qd)
dim_mask = tl.where(
tl.arange(0, BLOCK_DMODEL_PADDED) < BLOCK_DMODEL, 1,
0).to(tl.int1) # [D]
q = tl.load(Q + off_q,
mask=dim_mask[None, :] &
(offs_m[:, None] < cur_batch_query_len),
other=0.0) # [M,D]
# initialize pointer to m and l
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf") # [M]
l_i = tl.zeros([BLOCK_M], dtype=tl.float32) # [M]
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_PADDED],
dtype=tl.float32) # [M,D]
# compute query against context (no causal mask here)
for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
((start_n + offs_n) // block_size) * stride_b_loc_s,
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0) # [N]
# [D,N]
off_k = (bn[None, :] * stride_k_cache_bs +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
# [N,D]
off_v = (
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k = tl.load(K_cache + off_k,
mask=dim_mask[:, None] &
((start_n + offs_n[None, :]) < cur_batch_ctx_len),
other=0.0) # [D,N]
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) # [M,N]
qk += tl.dot(q, k)
qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
float("-inf"))
qk *= sm_scale
if SLIDING_WINDOW > 0:
# (cur_batch_ctx_len + offs_m[:, None]) are the positions of
# Q entries in sequence
# (start_n + offs_n[None, :]) are the positions of
# KV entries in sequence
# So the condition makes sure each entry in Q only attends
# to KV entries not more than SLIDING_WINDOW away.
#
# We can't use -inf here, because the
# sliding window may lead to the entire row being masked.
# This then makes m_ij contain -inf, which causes NaNs in
# exp().
qk = tl.where((cur_batch_ctx_len + offs_m[:, None]) -
(start_n + offs_n[None, :]) < SLIDING_WINDOW, qk,
-10000)
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1) # [M]
p = tl.exp(qk - m_ij[:, None]) # [M,N]
l_ij = tl.sum(p, 1) # [M]
# -- update m_i and l_i
m_i_new = tl.maximum(m_i, m_ij) # [M]
alpha = tl.exp(m_i - m_i_new) # [M]
beta = tl.exp(m_ij - m_i_new) # [M]
l_i_new = alpha * l_i + beta * l_ij # [M]
# -- update output accumulator --
# scale p
p_scale = beta / l_i_new
p = p * p_scale[:, None]
# scale acc
acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(V_cache + off_v,
mask=dim_mask[None, :] &
((start_n + offs_n[:, None]) < cur_batch_ctx_len),
other=0.0) # [N,D]
p = p.to(v.dtype)
acc += tl.dot(p, v)
# # update m_i and l_i
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
# block_mask is 0 when we're already past the current query length
block_mask = tl.where(block_start_loc < cur_batch_query_len, 1, 0)
# compute query against itself (with causal mask)
for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
k = tl.load(k_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_kbs,
mask=dim_mask[:, None] &
((start_n + offs_n[None, :]) < cur_batch_query_len),
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k)
qk *= sm_scale
# apply causal mask
qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
float("-inf"))
if SLIDING_WINDOW > 0:
qk = tl.where(
offs_m[:, None] -
(start_n + offs_n[None, :]) < SLIDING_WINDOW, qk, -10000)
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
p = tl.exp(qk - m_ij[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
m_i_new = tl.maximum(m_i, m_ij)
alpha = tl.exp(m_i - m_i_new)
beta = tl.exp(m_ij - m_i_new)
l_i_new = alpha * l_i + beta * l_ij
# -- update output accumulator --
# scale p
p_scale = beta / l_i_new
p = p * p_scale[:, None]
# scale acc
acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(v_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_vbs,
mask=dim_mask[None, :] &
((start_n + offs_n[:, None]) < cur_batch_query_len),
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
# initialize pointers to output
off_o = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
cur_head * stride_oh + offs_d[None, :] * stride_od)
out_ptrs = Out + off_o
tl.store(out_ptrs,
acc,
mask=dim_mask[None, :] &
(offs_m[:, None] < cur_batch_query_len))
return
@triton.jit
def _fwd_kernel_flash_attn_v2(
Q,
K,
V,
K_cache,
V_cache,
B_Loc,
sm_scale,
B_Start_Loc,
B_Seqlen,
B_Ctxlen,
block_size,
x,
Out,
stride_b_loc_b,
stride_b_loc_s,
stride_qbs,
stride_qh,
stride_qd,
stride_kbs,
stride_kh,
stride_kd,
stride_vbs,
stride_vh,
stride_vd,
stride_obs,
stride_oh,
stride_od,
stride_k_cache_bs,
stride_k_cache_h,
stride_k_cache_d,
stride_k_cache_bl,
stride_k_cache_x,
stride_v_cache_bs,
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
):
cur_batch = tl.program_id(0)
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
block_start_loc = BLOCK_M * start_m
# initialize offsets
offs_n = tl.arange(0, BLOCK_N)
offs_d = tl.arange(0, BLOCK_DMODEL)
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
off_q = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
cur_head * stride_qh + offs_d[None, :] * stride_qd)
q = tl.load(
Q + off_q,
mask=offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
# # initialize pointer to m and l
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
((start_n + offs_n) // block_size) * stride_b_loc_s,
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0)
off_k = (bn[None, :] * stride_k_cache_bs +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
off_v = (
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k = tl.load(K_cache + off_k,
mask=(start_n + offs_n[None, :]) < cur_batch_ctx_len,
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k)
qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
float("-inf"))
qk *= sm_scale
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
m_i_new = tl.maximum(m_i, m_ij)
p = tl.math.exp(qk - m_i_new[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
alpha = tl.math.exp(m_i - m_i_new)
l_i_new = alpha * l_i + l_ij
# -- update output accumulator --
# scale p
# scale acc
acc_scale = alpha
# acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(V_cache + off_v,
mask=(start_n + offs_n[:, None]) < cur_batch_ctx_len,
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
block_mask = tl.where(
block_start_loc < cur_batch_seq_len - cur_batch_ctx_len, 1, 0)
for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
k = tl.load(k_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_kbs,
mask=(start_n + offs_n[None, :]) <
cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k)
qk *= sm_scale
qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
float("-inf"))
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
m_i_new = tl.maximum(m_i, m_ij)
p = tl.math.exp(qk - m_i_new[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
alpha = tl.math.exp(m_i - m_i_new)
l_i_new = alpha * l_i + l_ij
# -- update output accumulator --
# scale p
# scale acc
acc_scale = alpha
# acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(v_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_vbs,
mask=(start_n + offs_n[:, None]) <
cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
# acc /= l_i[:, None]
# initialize pointers to output
off_o = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
cur_head * stride_oh + offs_d[None, :] * stride_od)
out_ptrs = Out + off_o
tl.store(out_ptrs,
acc,
mask=offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len)
return
@triton.jit
def _fwd_kernel_alibi(
Q,
K,
V,
K_cache,
V_cache,
B_Loc,
sm_scale,
B_Start_Loc,
B_Seqlen,
B_Ctxlen,
Alibi_slopes,
block_size,
x,
Out,
stride_b_loc_b,
stride_b_loc_s,
stride_qbs,
stride_qh,
stride_qd,
stride_kbs,
stride_kh,
stride_kd,
stride_vbs,
stride_vh,
stride_vd,
stride_obs,
stride_oh,
stride_od,
stride_k_cache_bs,
stride_k_cache_h,
stride_k_cache_d,
stride_k_cache_bl,
stride_k_cache_x,
stride_v_cache_bs,
stride_v_cache_h,
stride_v_cache_d,
stride_v_cache_bl,
num_queries_per_kv: int,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
):
# attn_bias[]
cur_batch = tl.program_id(0)
cur_head = tl.program_id(1)
start_m = tl.program_id(2)
cur_kv_head = cur_head // num_queries_per_kv
# cur_batch_seq_len: the length of prompts
# cur_batch_ctx_len: the length of prefix
# cur_batch_in_all_start_index: the start id of the dim=0
cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
block_start_loc = BLOCK_M * start_m
# initialize offsets
offs_n = tl.arange(0, BLOCK_N)
offs_d = tl.arange(0, BLOCK_DMODEL)
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
off_q = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
cur_head * stride_qh + offs_d[None, :] * stride_qd)
q = tl.load(
Q + off_q,
mask=offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
# # initialize pointer to m and l
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
alibi_slope = tl.load(Alibi_slopes + cur_head)
alibi_start_q = tl.arange(
0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
alibi_start_k = 0
for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
((start_n + offs_n) // block_size) * stride_b_loc_s,
mask=(start_n + offs_n) < cur_batch_ctx_len,
other=0)
off_k = (bn[None, :] * stride_k_cache_bs +
cur_kv_head * stride_k_cache_h +
(offs_d[:, None] // x) * stride_k_cache_d +
((start_n + offs_n[None, :]) % block_size) *
stride_k_cache_bl +
(offs_d[:, None] % x) * stride_k_cache_x)
off_v = (
bn[:, None] * stride_v_cache_bs +
cur_kv_head * stride_v_cache_h +
offs_d[None, :] * stride_v_cache_d +
(start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
k = tl.load(K_cache + off_k,
mask=(start_n + offs_n[None, :]) < cur_batch_ctx_len,
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k)
qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
float("-inf"))
qk *= sm_scale
# load alibi
alibi = (tl.arange(0, BLOCK_N)[None, :] + alibi_start_k -
alibi_start_q[:, None]) * alibi_slope
alibi = tl.where(
(alibi <= 0) & (alibi_start_q[:, None] < cur_batch_seq_len),
alibi, float("-inf"))
qk += alibi
alibi_start_k += BLOCK_N
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
m_i_new = tl.maximum(m_i, m_ij)
p = tl.math.exp(qk - m_i_new[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
alpha = tl.math.exp(m_i - m_i_new)
l_i_new = alpha * l_i + l_ij
# -- update output accumulator --
# scale p
# scale acc
acc_scale = alpha
# acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(V_cache + off_v,
mask=(start_n + offs_n[:, None]) < cur_batch_ctx_len,
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v, allow_tf32=False)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
offs_d[:, None] * stride_kd)
off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
offs_d[None, :] * stride_vd)
k_ptrs = K + off_k
v_ptrs = V + off_v
block_mask = tl.where(
block_start_loc < cur_batch_seq_len - cur_batch_ctx_len, 1, 0)
# init alibi
alibi_slope = tl.load(Alibi_slopes + cur_head)
alibi_start_q = tl.arange(
0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
alibi_start_k = cur_batch_ctx_len
# # init debugger
# offset_db_q = tl.arange(0, BLOCK_M) + block_start_loc
# offset_db_k = tl.arange(0, BLOCK_N)
# calc q[BLOCK_M, BLOCK_MODEL] mul k[prefix_len: , BLOCK_DMODEL]
for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
start_n = tl.multiple_of(start_n, BLOCK_N)
# -- compute qk ----
k = tl.load(k_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_kbs,
mask=(start_n + offs_n[None, :]) <
cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
qk += tl.dot(q, k, allow_tf32=False)
qk *= sm_scale
qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
float("-inf"))
# load alibi
alibi = (tl.arange(0, BLOCK_N)[None, :] + alibi_start_k -
alibi_start_q[:, None]) * alibi_slope
alibi = tl.where(
(alibi <= 0) & (alibi_start_q[:, None] < cur_batch_seq_len),
alibi, float("-inf"))
qk += alibi
alibi_start_k += BLOCK_N
# -- compute m_ij, p, l_ij
m_ij = tl.max(qk, 1)
m_i_new = tl.maximum(m_i, m_ij)
p = tl.math.exp(qk - m_i_new[:, None])
l_ij = tl.sum(p, 1)
# -- update m_i and l_i
alpha = tl.math.exp(m_i - m_i_new)
l_i_new = alpha * l_i + l_ij
# -- update output accumulator --
# scale p
# scale acc
acc_scale = alpha
# acc_scale = l_i / l_i_new * alpha
acc = acc * acc_scale[:, None]
# update acc
v = tl.load(v_ptrs +
(cur_batch_in_all_start_index + start_n) * stride_vbs,
mask=(start_n + offs_n[:, None]) <
cur_batch_seq_len - cur_batch_ctx_len,
other=0.0)
p = p.to(v.dtype)
acc += tl.dot(p, v, allow_tf32=False)
# update m_i and l_i
l_i = l_i_new
m_i = m_i_new
acc = acc / l_i[:, None]
# initialize pointers to output
off_o = (
(cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
cur_head * stride_oh + offs_d[None, :] * stride_od)
out_ptrs = Out + off_o
tl.store(out_ptrs,
acc,
mask=offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len)
return
@torch.inference_mode()
def context_attention_fwd(q,
k,
v,
o,
k_cache,
v_cache,
b_loc,
b_start_loc,
b_seq_len,
b_ctx_len,
max_input_len,
alibi_slopes=None,
sliding_window=None):
cap = torch.musa.get_device_capability()
BLOCK = 128 if cap[0] >= 8 else 64
# shape constraints
Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
assert Lq == Lk and Lk == Lv
# round up Lk to a power of 2 - this is required for Triton block size
Lk_padded = triton.next_power_of_2(Lk)
sm_scale = 1.0 / (Lq**0.5)
batch, head = b_seq_len.shape[0], q.shape[1]
num_queries_per_kv = q.shape[1] // k.shape[1]
grid = (batch, head, triton.cdiv(max_input_len, BLOCK)) # batch, head,
num_warps = 8 if Lk <= 64 else 8
if alibi_slopes is not None:
assert Lk == Lk_padded
_fwd_kernel_alibi[grid](
q,
k,
v,
k_cache,
v_cache,
b_loc,
sm_scale,
b_start_loc,
b_seq_len,
b_ctx_len,
alibi_slopes,
v_cache.shape[3],
8,
o,
b_loc.stride(0),
b_loc.stride(1),
q.stride(0),
q.stride(1),
q.stride(2),
k.stride(0),
k.stride(1),
k.stride(2),
v.stride(0),
v.stride(1),
v.stride(2),
o.stride(0),
o.stride(1),
o.stride(2),
k_cache.stride(0),
k_cache.stride(1),
k_cache.stride(2),
k_cache.stride(3),
k_cache.stride(
4
), #[num_blocks, num_kv_heads, head_size/x, block_size, x]
v_cache.stride(0),
v_cache.stride(1),
v_cache.stride(2),
v_cache.stride(
3), #[num_blocks, num_kv_heads, head_size, block_size]
num_queries_per_kv=num_queries_per_kv,
BLOCK_M=BLOCK,
BLOCK_DMODEL=Lk,
BLOCK_N=BLOCK,
num_warps=num_warps,
num_stages=1,
)
return
_fwd_kernel[grid](
q,
k,
v,
k_cache,
v_cache,
b_loc,
sm_scale,
b_start_loc,
b_seq_len,
b_ctx_len,
v_cache.shape[3],
8,
o,
b_loc.stride(0),
b_loc.stride(1),
q.stride(0),
q.stride(1),
q.stride(2),
k.stride(0),
k.stride(1),
k.stride(2),
v.stride(0),
v.stride(1),
v.stride(2),
o.stride(0),
o.stride(1),
o.stride(2),
k_cache.stride(0),
k_cache.stride(1),
k_cache.stride(2),
k_cache.stride(3),
k_cache.stride(
4), #[num_blocks, num_kv_heads, head_size/x, block_size, x]
v_cache.stride(0),
v_cache.stride(1),
v_cache.stride(2),
v_cache.stride(
3), #[num_blocks, num_kv_heads, head_size, block_size]
num_queries_per_kv=num_queries_per_kv,
BLOCK_M=BLOCK,
BLOCK_DMODEL=Lk,
BLOCK_DMODEL_PADDED=Lk_padded,
BLOCK_N=BLOCK,
SLIDING_WINDOW=sliding_window if sliding_window is not None else 0,
num_warps=num_warps,
num_stages=1,
)
return

810
vllm/attention/ops/triton_flash_attention.py Normal file
View File

@@ -0,0 +1,810 @@
#!/usr/bin/env python
"""
Fused Attention
===============
This is a Triton implementation of the Flash Attention v2 algorithm from Tri Dao
(https://tridao.me/publications/flash2/flash2.pdf)
Credits: OpenAI kernel team, AMD ML Frameworks Triton team
Features supported:
1) Fwd with causal masking
2) Any sequence lengths without padding (currently fwd kernel only)
3) Support for different sequence lengths for q and k
4) Nested tensor API currently does not support dropout or bias.
Not currently supported:
1) Non power of two head dims
"""
import torch
import triton
import triton.language as tl
torch_dtype: tl.constexpr = torch.float16
@triton.jit
def cdiv_fn(x, y):
return (x + y - 1) // y
@triton.jit
def max_fn(x, y):
return tl.math.max(x, y)
@triton.jit
def dropout_offsets(philox_seed, philox_offset, dropout_p, m, n, stride):
ms = tl.arange(0, m)
ns = tl.arange(0, n)
return philox_offset + ms[:, None] * stride + ns[None, :]
@triton.jit
def dropout_rng(philox_seed, philox_offset, dropout_p, m, n, stride):
rng_offsets = dropout_offsets(philox_seed, philox_offset, dropout_p, m, n,
stride).to(tl.uint32)
# TODO: use tl.randint for better performance
return tl.rand(philox_seed, rng_offsets)
@triton.jit
def dropout_mask(philox_seed, philox_offset, dropout_p, m, n, stride):
rng_output = dropout_rng(philox_seed, philox_offset, dropout_p, m, n,
stride)
rng_keep = rng_output > dropout_p
return rng_keep
@triton.jit
def load_fn(block_ptr, first, second, pad):
if first and second:
tensor = tl.load(block_ptr, boundary_check=(0, 1), padding_option=pad)
elif first:
tensor = tl.load(block_ptr, boundary_check=(0, ), padding_option=pad)
elif second:
tensor = tl.load(block_ptr, boundary_check=(1, ), padding_option=pad)
else:
tensor = tl.load(block_ptr)
return tensor
@triton.jit
def _attn_fwd_inner(
acc,
l_i,
m_i,
q,
K_block_ptr,
V_block_ptr,
start_m,
actual_seqlen_k,
dropout_p,
philox_seed,
batch_philox_offset,
encoded_softmax_block_ptr,
block_min,
block_max,
offs_n_causal,
masked_blocks,
n_extra_tokens,
bias_ptr,
IS_CAUSAL: tl.constexpr,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
OFFS_M: tl.constexpr,
OFFS_N: tl.constexpr,
PRE_LOAD_V: tl.constexpr,
MASK_STEPS: tl.constexpr,
ENABLE_DROPOUT: tl.constexpr,
RETURN_ENCODED_SOFTMAX: tl.constexpr,
PADDED_HEAD: tl.constexpr,
):
# loop over k, v, and update accumulator
for start_n in range(block_min, block_max, BLOCK_N):
# For padded blocks, we will overrun the tensor size if
# we load all BLOCK_N. For others, the blocks are all within range.
k = load_fn(
K_block_ptr,
PADDED_HEAD,
MASK_STEPS and (n_extra_tokens != 0),
"zero",
)
if PRE_LOAD_V:
v = load_fn(
V_block_ptr,
MASK_STEPS and (n_extra_tokens != 0),
PADDED_HEAD,
"zero",
)
qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
# We start from end of seqlen_k so only the first iteration would need
# to be checked for padding if it is not a multiple of block_n
# TODO: This can be optimized to only be true for the padded block.
if MASK_STEPS: # noqa: SIM102
# If this is the last block / iteration, we want to
# mask if the sequence length is not a multiple of block size
# a solution is to always do BLOCK_M // BLOCK_N + 1 steps
# if not is_modulo_mn. last step might get wasted but that is okay.
# check if this masking works for that case.
if (start_n + BLOCK_N == block_max) and (n_extra_tokens != 0):
boundary_m = tl.full([BLOCK_M],
actual_seqlen_k,
dtype=tl.int32)
size_n = start_n + OFFS_N[None, :]
mask = size_n < boundary_m[:, None]
qk = tl.where(mask, qk, float("-inf"))
if IS_CAUSAL:
causal_boundary = start_n + offs_n_causal
causal_mask = OFFS_M[:, None] >= causal_boundary[None, :]
qk = tl.where(causal_mask, qk, float("-inf"))
# -- compute qk ----
qk += tl.dot(q, k)
if bias_ptr is not None:
bias = load_fn(bias_ptr, False, MASK_STEPS
and (n_extra_tokens != 0), "zero")
# While bias is added after multiplying qk with sm_scale, our
# optimization to use 2^x instead of e^x results in an additional
# scale factor of log2(e) which we must also multiply the bias with.
qk += bias * 1.44269504089
m_ij = tl.maximum(m_i, tl.max(qk, 1))
qk = qk - m_ij[:, None]
p = tl.math.exp2(qk)
# CAVEAT: Must update l_ij before applying dropout
l_ij = tl.sum(p, 1)
if ENABLE_DROPOUT:
philox_offset = (batch_philox_offset +
start_m * BLOCK_M * actual_seqlen_k + start_n -
BLOCK_N)
keep = dropout_mask(
philox_seed,
philox_offset,
dropout_p,
BLOCK_M,
BLOCK_N,
actual_seqlen_k,
)
if RETURN_ENCODED_SOFTMAX:
tl.store(
encoded_softmax_block_ptr,
tl.where(keep, p,
-p).to(encoded_softmax_block_ptr.type.element_ty),
)
p = tl.where(keep, p, 0.0)
elif RETURN_ENCODED_SOFTMAX:
tl.store(
encoded_softmax_block_ptr,
p.to(encoded_softmax_block_ptr.type.element_ty),
)
# -- update output accumulator --
alpha = tl.math.exp2(m_i - m_ij)
acc = acc * alpha[:, None]
if not PRE_LOAD_V:
v = load_fn(
V_block_ptr,
MASK_STEPS and (n_extra_tokens != 0),
PADDED_HEAD,
"zero",
)
# -- update m_i and l_i
l_i = l_i * alpha + l_ij
# update m_i and l_i
m_i = m_ij
acc += tl.dot(p.to(V_block_ptr.type.element_ty), v)
V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
K_block_ptr = tl.advance(K_block_ptr, (0, BLOCK_N))
if bias_ptr is not None:
bias_ptr = tl.advance(bias_ptr, (0, BLOCK_N))
if RETURN_ENCODED_SOFTMAX:
encoded_softmax_block_ptr = tl.advance(encoded_softmax_block_ptr,
(0, BLOCK_N))
return acc, l_i, m_i
@triton.autotune(
configs=[
triton.Config(
{
"BLOCK_M": 256,
"BLOCK_N": 64,
"waves_per_eu": 2,
"PRE_LOAD_V": False,
},
num_stages=1,
num_warps=8,
),
triton.Config(
{
"BLOCK_M": 128,
"BLOCK_N": 128,
"waves_per_eu": 2,
"PRE_LOAD_V": False,
},
num_stages=1,
num_warps=4,
),
triton.Config(
{
"BLOCK_M": 256,
"BLOCK_N": 128,
"waves_per_eu": 2,
"PRE_LOAD_V": False,
},
num_stages=1,
num_warps=8,
),
triton.Config(
{
"BLOCK_M": 128,
"BLOCK_N": 64,
"waves_per_eu": 3,
"PRE_LOAD_V": True,
},
num_stages=1,
num_warps=4,
),
triton.Config(
{
"BLOCK_M": 128,
"BLOCK_N": 64,
"waves_per_eu": 3,
"PRE_LOAD_V": False,
},
num_stages=1,
num_warps=4,
),
triton.Config(
{
"BLOCK_M": 64,
"BLOCK_N": 64,
"waves_per_eu": 4,
"PRE_LOAD_V": False,
},
num_stages=1,
num_warps=8,
),
triton.Config(
{
"BLOCK_M": 32,
"BLOCK_N": 32,
"waves_per_eu": 4,
"PRE_LOAD_V": False,
},
num_stages=1,
num_warps=8,
),
# TODO: This config fails with head_size not pow2 with data mismatches.
# triton.Config({'BLOCK_M': 32, 'BLOCK_N': 16, 'waves_per_eu': 1,
# 'PRE_LOAD_V': False}, num_stages=1, num_warps=4),
triton.Config(
{
"BLOCK_M": 16,
"BLOCK_N": 16,
"waves_per_eu": 1,
"PRE_LOAD_V": False,
},
num_stages=1,
num_warps=4,
),
],
key=['IS_CAUSAL', 'dropout_p', 'BLOCK_DMODEL'],
)
@triton.jit
def attn_fwd(
Q,
K,
V,
bias,
sm_scale,
L,
Out,
stride_qz,
stride_qh,
stride_qm,
stride_qk,
stride_kz,
stride_kh,
stride_kn,
stride_kk,
stride_vz,
stride_vh,
stride_vk,
stride_vn,
stride_oz,
stride_oh,
stride_om,
stride_on,
stride_bz,
stride_bh,
stride_bm,
stride_bn,
cu_seqlens_q,
cu_seqlens_k,
dropout_p,
philox_seed,
philox_offset_base,
encoded_softmax,
HQ: tl.constexpr,
HK: tl.constexpr,
ACTUAL_BLOCK_DMODEL: tl.constexpr,
MAX_SEQLENS_Q: tl.constexpr,
MAX_SEQLENS_K: tl.constexpr,
VARLEN: tl.constexpr,
IS_CAUSAL: tl.constexpr,
BLOCK_M: tl.constexpr,
BLOCK_DMODEL: tl.constexpr,
BLOCK_N: tl.constexpr,
PRE_LOAD_V: tl.constexpr,
BIAS_TYPE: tl.constexpr,
ENABLE_DROPOUT: tl.constexpr,
RETURN_ENCODED_SOFTMAX: tl.constexpr,
):
start_m = tl.program_id(0)
off_h_q = tl.program_id(1)
off_z = tl.program_id(2)
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
offs_n = tl.arange(0, BLOCK_N)
if VARLEN:
cu_seqlens_q_start = tl.load(cu_seqlens_q + off_z)
cu_seqlens_q_end = tl.load(cu_seqlens_q + off_z + 1)
seqlen_q = cu_seqlens_q_end - cu_seqlens_q_start
# We have a one-size-fits-all grid in id(0). Some seqlens might be too
# small for all start_m so for those we return early.
if start_m * BLOCK_M > seqlen_q:
return
cu_seqlens_k_start = tl.load(cu_seqlens_k + off_z)
cu_seqlens_k_end = tl.load(cu_seqlens_k + off_z + 1)
seqlen_k = cu_seqlens_k_end - cu_seqlens_k_start
else:
cu_seqlens_q_start = 0
cu_seqlens_k_start = 0
seqlen_q = MAX_SEQLENS_Q
seqlen_k = MAX_SEQLENS_K
# Now we compute whether we need to exit early due to causal masking.
# This is because for seqlen_q > seqlen_k, M rows of the attn scores
# are completely masked, resulting in 0s written to the output, and
# inf written to LSE. We don't need to do any GEMMs in this case.
# This block of code determines what N is, and if this WG is operating
# on those M rows.
n_blocks = cdiv_fn(seqlen_k, BLOCK_N)
if IS_CAUSAL:
# If seqlen_q == seqlen_k, the attn scores are a square matrix.
# If seqlen_q != seqlen_k, attn scores are rectangular which means
# the causal mask boundary is bottom right aligned, and ends at either
# the top edge (seqlen_q < seqlen_k) or left edge.
# This captures the decrease in n_blocks if we have a rectangular attn
# matrix
n_blocks_seqlen = cdiv_fn(
(start_m + 1) * BLOCK_M + seqlen_k - seqlen_q, BLOCK_N)
# This is what adjusts the block_max for the current WG, only
# if IS_CAUSAL. Otherwise we want to always iterate through all n_blocks
n_blocks = min(n_blocks, n_blocks_seqlen)
# If we have no blocks after adjusting for seqlen deltas, this WG is
# part of the blocks that are all 0. We exit early.
if n_blocks <= 0:
o_offset = (off_z * stride_oz + cu_seqlens_q_start * stride_om +
off_h_q * stride_oh)
O_block_ptr = tl.make_block_ptr(
base=Out + o_offset,
shape=(seqlen_q, BLOCK_DMODEL),
strides=(stride_om, stride_on),
offsets=(start_m * BLOCK_M, 0),
block_shape=(BLOCK_M, BLOCK_DMODEL),
order=(1, 0),
)
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=Out.type.element_ty)
# We still need to write 0s to the result
# tl.store(O_block_ptr,
# acc.to(Out.type.element_ty), boundary_check=(0,1))
# l_ptrs = L + off_z * HQ * MAX_SEQLENS_Q + off_h_q * MAX_SEQLENS_Q
# + offs_m
# We store inf to LSE, not -inf because in the bwd pass,
# we subtract this
# from qk which makes it -inf, such that exp(qk - inf) = 0
# for these masked blocks.
# l = tl.full([BLOCK_M], value=float("inf"), dtype=tl.float32)
# tl.store(l_ptrs, l)
# TODO: Should dropout and return encoded softmax be handled here?
return
# If MQA / GQA, set the K and V head offsets appropriately.
GROUP_SIZE: tl.constexpr = HQ // HK
off_h_k = off_h_q // GROUP_SIZE if GROUP_SIZE != 1 else off_h_q
n_extra_tokens = 0
if seqlen_k < BLOCK_N:
n_extra_tokens = BLOCK_N - seqlen_k
elif seqlen_k % BLOCK_N:
n_extra_tokens = seqlen_k % BLOCK_N
padded_head = ACTUAL_BLOCK_DMODEL != BLOCK_DMODEL
# Compute pointers for all the tensors used in this kernel.
q_offset = (off_z * stride_qz + off_h_q * stride_qh +
cu_seqlens_q_start * stride_qm)
Q_block_ptr = tl.make_block_ptr(
base=Q + q_offset,
shape=(seqlen_q, ACTUAL_BLOCK_DMODEL),
strides=(stride_qm, stride_qk),
offsets=(start_m * BLOCK_M, 0),
block_shape=(BLOCK_M, BLOCK_DMODEL),
order=(1, 0),
)
k_offset = (off_z * stride_kz + off_h_k * stride_kh +
cu_seqlens_k_start * stride_kn)
K_block_ptr = tl.make_block_ptr(
base=K + k_offset,
shape=(ACTUAL_BLOCK_DMODEL, seqlen_k),
strides=(stride_kk, stride_kn),
offsets=(0, 0),
block_shape=(BLOCK_DMODEL, BLOCK_N),
order=(0, 1),
)
v_offset = (off_z * stride_vz + off_h_k * stride_vh +
cu_seqlens_k_start * stride_vk)
V_block_ptr = tl.make_block_ptr(
base=V + v_offset,
shape=(seqlen_k, ACTUAL_BLOCK_DMODEL),
strides=(stride_vk, stride_vn),
offsets=(0, 0),
block_shape=(BLOCK_N, BLOCK_DMODEL),
order=(1, 0),
)
if BIAS_TYPE != 0:
bias_ptr = tl.make_block_ptr(
base=bias + off_h_q * stride_bh,
shape=(seqlen_q, seqlen_k),
strides=(stride_bm, stride_bn),
offsets=(start_m * BLOCK_M, 0),
block_shape=(BLOCK_M, BLOCK_N),
order=(1, 0),
)
else:
bias_ptr = None
if ENABLE_DROPOUT:
batch_philox_offset = philox_offset_base \
+ (off_z * HQ + off_h_q) \
* seqlen_q * seqlen_k
else:
batch_philox_offset = 0
# We can ask to return the dropout mask without actually doing any dropout.
# In this case, we return an invalid pointer so indicate the mask is not i
# valid.
# TODO: Fix encoded softmax. It currently uses just h_q in the base offset.
if RETURN_ENCODED_SOFTMAX:
encoded_softmax_block_ptr = tl.make_block_ptr(
base=encoded_softmax + off_h_q * seqlen_q * seqlen_k,
shape=(seqlen_q, seqlen_k),
strides=(seqlen_k, 1),
offsets=(start_m * BLOCK_M, 0),
block_shape=(BLOCK_M, BLOCK_N),
order=(1, 0),
)
else:
encoded_softmax_block_ptr = 0
# initialize pointer to m and l
m_i = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
l_i = tl.full([BLOCK_M], 1.0, dtype=tl.float32)
acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
# scale sm_scale by log_2(e) and use 2^x in the loop as we do not
# have native e^x support in HW.
qk_scale = sm_scale * 1.44269504089
# Q is loaded once at the beginning and shared by all N blocks.
q = load_fn(Q_block_ptr, True, padded_head, "zero")
q = (q * qk_scale).to(Q_block_ptr.type.element_ty)
# Here we compute how many full and masked blocks we have.
padded_block_k = n_extra_tokens != 0
is_modulo_mn = not padded_block_k and (seqlen_q % BLOCK_M == 0)
if IS_CAUSAL:
# There are always at least BLOCK_M // BLOCK_N masked blocks.
# Additionally there might be one more due to dissimilar seqlens.
masked_blocks = BLOCK_M // BLOCK_N + (not is_modulo_mn)
else:
# Padding on Q does not need to be masked in the FA loop.
masked_blocks = padded_block_k
# if IS_CAUSAL, not is_modulo_mn does not always result in an additional
# block. In this case we might exceed n_blocks so pick the min.
masked_blocks = min(masked_blocks, n_blocks)
n_full_blocks = n_blocks - masked_blocks
block_min = 0
block_max = n_blocks * BLOCK_N
# Compute for full blocks. Here we set causal to false regardless of its
# value because there is no masking. Similarly we do not need padding.
if n_full_blocks > 0:
block_max = (n_blocks - masked_blocks) * BLOCK_N
acc, l_i, m_i = _attn_fwd_inner(
acc,
l_i,
m_i,
q,
K_block_ptr,
V_block_ptr,
start_m,
seqlen_k,
dropout_p,
philox_seed,
batch_philox_offset,
encoded_softmax_block_ptr,
# _, _, offs_n_causal, masked_blocks, n_extra_tokens, _
block_min,
block_max,
0,
0,
0,
bias_ptr,
# IS_CAUSAL, ....
False,
BLOCK_M,
BLOCK_DMODEL,
BLOCK_N,
offs_m,
offs_n,
# _, MASK_STEPS, ...
PRE_LOAD_V,
False,
ENABLE_DROPOUT,
RETURN_ENCODED_SOFTMAX,
padded_head,
)
block_min = block_max
block_max = n_blocks * BLOCK_N
tl.debug_barrier()
# Remaining blocks, if any, are full / not masked.
if masked_blocks > 0:
offs_n_causal = offs_n + (seqlen_q - seqlen_k) if IS_CAUSAL else 0
K_block_ptr = tl.advance(K_block_ptr, (0, n_full_blocks * BLOCK_N))
V_block_ptr = tl.advance(V_block_ptr, (n_full_blocks * BLOCK_N, 0))
if bias_ptr is not None:
bias_ptr = tl.advance(bias_ptr, (0, n_full_blocks * BLOCK_N))
if RETURN_ENCODED_SOFTMAX:
encoded_softmax_block_ptr = tl.advance(encoded_softmax_block_ptr,
(0, n_full_blocks))
acc, l_i, m_i = _attn_fwd_inner(
acc,
l_i,
m_i,
q,
K_block_ptr,
V_block_ptr,
start_m,
seqlen_k,
dropout_p,
philox_seed,
batch_philox_offset,
encoded_softmax_block_ptr,
block_min,
block_max,
offs_n_causal,
masked_blocks,
n_extra_tokens,
bias_ptr,
IS_CAUSAL,
BLOCK_M,
BLOCK_DMODEL,
BLOCK_N,
offs_m,
offs_n,
# _, MASK_STEPS, ...
PRE_LOAD_V,
True,
ENABLE_DROPOUT,
RETURN_ENCODED_SOFTMAX,
padded_head,
)
# epilogue
acc = acc / l_i[:, None]
if ENABLE_DROPOUT:
acc = acc / (1 - dropout_p)
# If seqlen_q > seqlen_k but the delta is not a multiple of BLOCK_M,
# then we have one block with a row of all NaNs which come from computing
# softmax over a row of all -infs (-inf - inf = NaN). We check for that here
# and store 0s where there are NaNs as these rows should've been zeroed out.
end_m_idx = (start_m + 1) * BLOCK_M
start_m_idx = start_m * BLOCK_M
causal_start_idx = seqlen_q - seqlen_k
acc = acc.to(Out.type.element_ty)
if IS_CAUSAL: # noqa: SIM102
if causal_start_idx > start_m_idx and causal_start_idx < end_m_idx:
out_mask_boundary = tl.full((BLOCK_DMODEL, ),
causal_start_idx,
dtype=tl.int32)
mask_m_offsets = start_m_idx + tl.arange(0, BLOCK_M)
out_ptrs_mask = (mask_m_offsets[:, None] >=
out_mask_boundary[None, :])
z = 0.0
acc = tl.where(out_ptrs_mask, acc, z.to(acc.type.element_ty))
# write back LSE
# l_ptrs = L + off_z * HQ * MAX_SEQLENS_Q + off_h_q * MAX_SEQLENS_Q + offs_m
# If seqlen_q not multiple of BLOCK_M, we need to mask out the last
# few rows. This is only true for the last M block. For others,
# overflow_size will be -ve
# overflow_size = end_m_idx - seqlen_q
# if overflow_size > 0:
# boundary = tl.full((BLOCK_M,), BLOCK_M - overflow_size, dtype=tl.int32)
# # This is a > check because mask being 0 blocks the store.
# l_ptrs_mask = boundary > tl.arange(0, BLOCK_M)
# tl.store(l_ptrs, m_i + tl.math.log2(l_i), mask=l_ptrs_mask)
# else:
# tl.store(l_ptrs, m_i + tl.math.log2(l_i))
# write back O
o_offset = (off_z * stride_oz + cu_seqlens_q_start * stride_om +
off_h_q * stride_oh)
O_block_ptr = tl.make_block_ptr(
base=Out + o_offset,
shape=(seqlen_q, ACTUAL_BLOCK_DMODEL),
strides=(stride_om, stride_on),
offsets=(start_m * BLOCK_M, 0),
block_shape=(BLOCK_M, BLOCK_DMODEL),
order=(1, 0),
)
# Need boundary check on this to make sure the padding from the
# Q and KV tensors in both dims are not part of what we store back.
# TODO: Do the boundary check optionally.
tl.store(O_block_ptr, acc, boundary_check=(0, 1))
def check_args(
q,
k,
v,
o,
varlen=True,
max_seqlens=None,
cu_seqlens_q=None,
cu_seqlens_k=None,
):
assert q.dim() == k.dim() and q.dim() == v.dim()
if varlen:
assert q.dim() == 3
total_q, nheads_q, head_size = q.shape
total_k, nheads_k, _ = k.shape
assert cu_seqlens_q is not None
assert cu_seqlens_k is not None
assert len(cu_seqlens_q) == len(cu_seqlens_k)
else:
assert q.dim() == 4
batch, nheads_q, seqlen_q, head_size = q.shape
_, nheads_k, seqlen_k, _ = k.shape
assert max_seqlens > 0
assert k.shape == v.shape
assert q.shape[-1] == k.shape[-1] and q.shape[-1] == v.shape[-1]
# TODO: Change assert if we support qkl f8 and v f16
assert q.dtype == k.dtype and q.dtype == v.dtype
assert head_size <= 256
assert o.shape == q.shape
assert (nheads_q % nheads_k) == 0
class _attention(torch.autograd.Function):
@staticmethod
def forward(
ctx,
q,
k,
v,
o,
cu_seqlens_q,
cu_seqlens_k,
max_seqlens_q,
max_seqlens_k,
causal=False,
sm_scale=1.0,
bias=None,
):
if o is None:
o = torch.empty_like(q, dtype=v.dtype)
check_args(
q,
k,
v,
o,
varlen=True,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_k=cu_seqlens_k,
)
if True: # varlen
total_q, nheads_q, head_size = q.shape
total_k, nheads_k, _ = k.shape
batch = len(cu_seqlens_q) - 1
q_strides = (0, q.stride(1), q.stride(0), q.stride(2))
k_strides = (0, k.stride(1), k.stride(0), k.stride(2))
v_strides = (0, v.stride(1), v.stride(0), v.stride(2))
o_strides = (0, o.stride(1), o.stride(0), o.stride(2))
else:
batch, seqlen_q, nheads_q, head_size = q.shape
_, seqlen_k, nheads_k, _ = k.shape
q_strides = (q.stride(0), q.stride(2), q.stride(1), q.stride(3))
k_strides = (k.stride(0), k.stride(2), k.stride(1), k.stride(3))
v_strides = (v.stride(0), v.stride(2), v.stride(1), v.stride(3))
o_strides = (o.stride(0), o.stride(2), o.stride(1), o.stride(3))
# Get closest power of 2 over or equal to 32.
unpadded_head_dims = {32, 64, 128, 256}
if head_size not in unpadded_head_dims:
padded_d_model = None
for i in unpadded_head_dims:
if i > head_size:
padded_d_model = i
break
assert padded_d_model is not None
else:
padded_d_model = head_size
grid = lambda META: (
triton.cdiv(max_seqlens_q, META["BLOCK_M"]),
nheads_q,
batch,
)
encoded_softmax = None
# Seed the RNG so we get reproducible results for testing.
philox_seed = 0x1BF52
philox_offset = 0x1D4B42
if bias is not None:
bias_strides = (
bias.stride(0),
bias.stride(1),
bias.stride(2),
bias.stride(3),
)
else:
bias_strides = (0, 0, 0, 0)
attn_fwd[grid](
q,
k,
v,
bias,
sm_scale,
None,
o,
*q_strides,
*k_strides,
*v_strides,
*o_strides,
*bias_strides,
cu_seqlens_q,
cu_seqlens_k,
dropout_p=0.0,
philox_seed=philox_seed,
philox_offset_base=philox_offset,
encoded_softmax=encoded_softmax,
HQ=nheads_q,
HK=nheads_k,
ACTUAL_BLOCK_DMODEL=head_size,
MAX_SEQLENS_Q=max_seqlens_q,
MAX_SEQLENS_K=max_seqlens_k,
IS_CAUSAL=causal,
VARLEN=True,
BLOCK_DMODEL=padded_d_model,
BIAS_TYPE=0 if bias is None else 1,
ENABLE_DROPOUT=False,
RETURN_ENCODED_SOFTMAX=False,
)
ctx.grid = grid
ctx.sm_scale = sm_scale
ctx.BLOCK_DMODEL = head_size
ctx.causal = causal
ctx.dropout_p = 0.0
ctx.philox_seed = philox_seed
ctx.philox_offset = philox_offset
ctx.encoded_softmax = encoded_softmax
ctx.return_encoded_softmax = False
return o, encoded_softmax
triton_attention = _attention.apply

94
vllm/attention/selector.py Normal file
View File

@@ -0,0 +1,94 @@
import enum
from functools import lru_cache
from typing import Type
import torch
import vllm.envs as envs
from vllm.attention.backends.abstract import AttentionBackend
from vllm.logger import init_logger
from vllm.utils import is_cpu, is_hip, is_musa
logger = init_logger(__name__)
class _Backend(enum.Enum):
FLASH_ATTN = enum.auto()
XFORMERS = enum.auto()
ROCM_FLASH = enum.auto()
TORCH_SDPA = enum.auto()
FLASHINFER = enum.auto()
@lru_cache(maxsize=None)
def get_attn_backend(dtype: torch.dtype) -> Type[AttentionBackend]:
backend = _which_attn_to_use(dtype)
if backend == _Backend.FLASH_ATTN:
logger.info("Using FlashAttention-2 backend.")
from vllm.attention.backends.flash_attn import ( # noqa: F401
FlashAttentionBackend)
return FlashAttentionBackend
elif backend == _Backend.XFORMERS:
logger.info("Using XFormers backend.")
from vllm.attention.backends.xformers import ( # noqa: F401
XFormersBackend)
return XFormersBackend
elif backend == _Backend.ROCM_FLASH:
logger.info("Using ROCmFlashAttention backend.")
from vllm.attention.backends.rocm_flash_attn import ( # noqa: F401
ROCmFlashAttentionBackend)
return ROCmFlashAttentionBackend
elif backend == _Backend.TORCH_SDPA:
logger.info("Using Torch SDPA backend.")
from vllm.attention.backends.torch_sdpa import TorchSDPABackend
return TorchSDPABackend
elif backend == _Backend.FLASHINFER:
logger.info("Using Flashinfer backend.")
logger.warning("Eager mode is enforced for the Flashinfer backend. ")
from vllm.attention.backends.flashinfer import FlashInferBackend
return FlashInferBackend
else:
raise ValueError("Invalid attention backend.")
def _which_attn_to_use(dtype: torch.dtype) -> _Backend:
"""Returns which flash attention backend to use."""
if is_cpu():
return _Backend.TORCH_SDPA
if is_musa():
return _Backend.FLASH_ATTN
if is_hip():
# AMD GPUs.
if torch.cuda.get_device_capability()[0] != 9:
# not Instinct series GPUs.
logger.info("flash_atten is not supported on NAVI GPUs.")
return _Backend.ROCM_FLASH
# NVIDIA GPUs.
if torch.cuda.get_device_capability()[0] < 8:
# Volta and Turing NVIDIA GPUs.
logger.info("Cannot use FlashAttention-2 backend for Volta and Turing "
"GPUs.")
return _Backend.XFORMERS
if dtype not in (torch.float16, torch.bfloat16):
logger.info("Cannot use FlashAttention-2 backend for dtype other than "
"torch.float16 or torch.bfloat16.")
return _Backend.XFORMERS
try:
import flash_attn # noqa: F401
except ImportError:
logger.info(
"Cannot use FlashAttention-2 backend because the flash_attn "
"package is not found. Please install it for better performance.")
return _Backend.XFORMERS
backend_by_env_var = envs.VLLM_ATTENTION_BACKEND
if backend_by_env_var is not None:
return _Backend[backend_by_env_var]
# Default case.
return _Backend.FLASH_ATTN