[Feature] Add AllGather-KV sequence-parallel attention str… (#4968)

Signed-off-by: zhuhh97 <zhuhonghan@huawei.com>
Signed-off-by: dph97 <nenbaying@163.com>
Signed-off-by: zhuhh97 <zhh@unknown>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: dengyunyang <584797741@qq.com>
This commit is contained in:
authored and GitHub committed 2026-07-24 14:45:26 +00:00
1 parent 8927ec057c
commit 84bbd57e4d
33 files changed
+1181 -135

No files matched your search

@@ -0,0 +1,40 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Runtime kernel-selection helpers for attention backends."""
import torch
def can_sdpa_use_fused_gqa(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
attention_mask: torch.Tensor | None,
causal: bool,
) -> bool:
"""Return whether PyTorch SDPA can use a fused kernel for this GQA call.
This check is specific to the SDPA backend. Native GQA backends such as
FlashAttention should keep compressed K/V heads and do not use this helper.
"""
if not query.is_cuda:
return False
try:
params = torch.backends.cuda.SDPAParams(
query,
key,
value,
attention_mask,
0.0,
causal,
True,
)
can_use_flash = torch.backends.cuda.can_use_flash_attention(params)
can_use_cudnn = getattr(torch.backends.cuda, "can_use_cudnn_attention", None)
return can_use_flash or (can_use_cudnn is not None and can_use_cudnn(params))
except (AttributeError, RuntimeError, TypeError):
# Older PyTorch versions may not expose GQA-aware SDPA capability checks.
# Returning False lets the SDPA backend expand K/V and use the mature
# equal-head fused-kernel path instead.
return False
@@ -9,17 +9,24 @@ follows FlashAttention's bottom-right convention (``K[:e]`` is attended by
Per segment:
- causal segment ``[s, e)``: ``attn(Q[:, s:e], K[:, :e], V[:, :e], causal=True)``
- full-attn span ``[a, e)``: ``attn(Q[:, a:e], K[:, :e], V[:, :e], causal=False)``
- full-attn span ``[a, b)`` intersecting the query range at ``[s, e)``:
``attn(Q[:, s:e], K[:, :b], V[:, :b], causal=False)``
"""
from __future__ import annotations
from typing import Literal, NamedTuple
from typing import TYPE_CHECKING, Literal, NamedTuple
import torch
if TYPE_CHECKING:
from vllm_omni.diffusion.attention.backends.abstract import QueryRange
class Segment(NamedTuple):
start: int
end: int
q_start: int
q_end: int
kv_end: int
mode: Literal["causal", "full"]
@@ -30,30 +37,33 @@ def build_segments(full_attn_spans, query_offset, query_len):
query_len: length of the query
return:
List[Segment] in global coordinates, clipped to [query_offset, query_offset + query_len)
List[Segment] in global coordinates, clipped to
[query_offset, query_offset + query_len). Full-attention segments retain
the original span end as kv_end so a local query shard can attend past
its own boundary.
"""
q_start = query_offset
q_end = query_offset + query_len
segs: list[Segment] = []
segments: list[Segment] = []
cur = q_start
for a, e in full_attn_spans:
for span_start, span_end in full_attn_spans:
# clip span to query range
a_clipped = max(a, q_start)
e_clipped = min(e, q_end)
if a_clipped >= e_clipped:
overlap_start = max(span_start, q_start)
overlap_end = min(span_end, q_end)
if overlap_start >= overlap_end:
continue
if cur < a_clipped:
segs.append(Segment(cur, a_clipped, "causal"))
segs.append(Segment(a_clipped, e_clipped, "full"))
cur = e_clipped
if cur < overlap_start:
segments.append(Segment(cur, overlap_start, overlap_start, "causal"))
segments.append(Segment(overlap_start, overlap_end, span_end, "full"))
cur = overlap_end
if cur < q_end:
segs.append(Segment(cur, q_end, "causal"))
segments.append(Segment(cur, q_end, q_end, "causal"))
return segs
return segments
def _check_homogeneous(
@@ -76,23 +86,40 @@ def piecewise_attn(
full_attn_spans: list[list[tuple[int, int]]],
softmax_scale: float,
attn_func,
query_ranges: tuple[QueryRange, ...] | None = None,
):
B, Sq, H, D = query.shape
_check_homogeneous(full_attn_spans)
query_offset = key.shape[1] - Sq
spans = full_attn_spans[0]
out = query.new_zeros(B, Sq, H, D)
if query_ranges is None:
query_len = query.shape[1]
ranges = ((0, query_len, key.shape[1] - query_len),)
else:
ranges = tuple((r.local_start, r.local_end, r.global_start) for r in query_ranges)
for s, e, mode in build_segments(spans, query_offset, Sq):
q_s = s - query_offset
q_e = e - query_offset
out_seg = attn_func(
query[:, q_s:q_e],
key[:, :e],
value[:, :e],
causal=(mode == "causal"),
softmax_scale=softmax_scale,
)
out[:, q_s:q_e] = out_seg
return out
outputs = []
covered = 0
for local_start, local_end, global_start in ranges:
query_len = local_end - local_start
if local_start != covered or query_len < 0:
raise ValueError("query_ranges must cover local query contiguously")
for segment in build_segments(spans, global_start, query_len):
q_start = local_start + segment.q_start - global_start
q_end = local_start + segment.q_end - global_start
outputs.append(
attn_func(
query[:, q_start:q_end],
key[:, : segment.kv_end],
value[:, : segment.kv_end],
causal=(segment.mode == "causal"),
softmax_scale=softmax_scale,
)
)
covered = local_end
if covered != query.shape[1]:
raise ValueError("query_ranges must cover the full local query")
if not outputs:
return torch.empty_like(query)
if len(outputs) == 1:
return outputs[0]
return torch.cat(outputs, dim=1)