[Attention][DSA] Take the native decode path for MTP=3 on SM90 (#52164)

Signed-off-by: zobinHuang <zobin1999@gmail.com>
Co-authored-by: nodeeeeee <zhangkai.nodeee@gmail.com>
This commit is contained in:
Zhuobin Huang
2026-08-14 10:09:51 +00:00
committed by GitHub
co-authored by nodeeeeee
parent aa31003573
commit 63a9a5010a
4 changed files with 229 additions and 104 deletions
@@ -13,6 +13,7 @@ from vllm.utils.deep_gemm import (
fp8_fp4_paged_mqa_logits,
get_num_sms,
get_paged_mqa_logits_metadata,
native_next_n_supported,
)
from vllm.utils.import_utils import has_deep_gemm
from vllm.utils.math_utils import cdiv
@@ -205,7 +206,12 @@ def _ref_fp8_fp4_paged_mqa_logits(
@pytest.mark.skipif(
not current_platform.has_device_capability(90), reason="SM90 and SM100 only"
)
def test_deepgemm_fp8_fp4_paged_mqa_logits():
# next_n = 1 + num_speculative_tokens, so next_n=4 is MTP=3 (issue #35878).
@pytest.mark.parametrize("batch_size,next_n", [(4, 1), (2, 2), (2, 4)])
def test_deepgemm_fp8_fp4_paged_mqa_logits(batch_size: int, next_n: int):
if not native_next_n_supported(next_n):
pytest.skip(f"next_n={next_n} has no native kernel on this architecture")
# NOTE: clean_logits=True is incompatible with the 2D context_lens
# required by csrc/apis/attention.hpp; only the False path is exercised.
clean_logits = False
@@ -213,98 +219,97 @@ def test_deepgemm_fp8_fp4_paged_mqa_logits():
random.seed(0)
max_model_len = 4096
for batch_size, next_n in [(4, 1), (2, 2)]:
for heads, index_dim in [(32, 128)]:
for avg_kv in (2048,):
num_blocks, blocksize = max_model_len * 2, 64
for heads, index_dim in [(32, 128)]:
for avg_kv in (2048,):
num_blocks, blocksize = max_model_len * 2, 64
q = torch.randn(
(batch_size, next_n, heads, index_dim),
device="cuda",
dtype=torch.bfloat16,
)
kv_cache = torch.randn(
(num_blocks, blocksize, 1, index_dim),
device="cuda",
dtype=torch.bfloat16,
)
weights = torch.randn(
(batch_size * next_n, heads),
device="cuda",
dtype=torch.float32,
)
q = torch.randn(
(batch_size, next_n, heads, index_dim),
device="cuda",
dtype=torch.bfloat16,
)
kv_cache = torch.randn(
(num_blocks, blocksize, 1, index_dim),
device="cuda",
dtype=torch.bfloat16,
)
weights = torch.randn(
(batch_size * next_n, heads),
device="cuda",
dtype=torch.float32,
)
context_lens = (
torch.randint(int(0.8 * avg_kv), int(1.2 * avg_kv), (batch_size,))
.cuda()
.to(torch.int32)
)
max_block_len = (
(context_lens.max().item() + blocksize - 1) // blocksize * blocksize
)
block_tables = torch.zeros(
(batch_size, max_block_len),
device="cuda",
dtype=torch.int32,
)
context_lens = (
torch.randint(int(0.8 * avg_kv), int(1.2 * avg_kv), (batch_size,))
.cuda()
.to(torch.int32)
)
max_block_len = (
(context_lens.max().item() + blocksize - 1) // blocksize * blocksize
)
block_tables = torch.zeros(
(batch_size, max_block_len),
device="cuda",
dtype=torch.int32,
)
counter = 0
block_idx_pool = list(range(num_blocks))
random.shuffle(block_idx_pool)
for i in range(batch_size):
ctx_len = int(context_lens[i].item())
for j in range((ctx_len + blocksize - 1) // blocksize):
block_tables[i][j] = block_idx_pool[counter]
counter += 1
counter = 0
block_idx_pool = list(range(num_blocks))
random.shuffle(block_idx_pool)
for i in range(batch_size):
ctx_len = int(context_lens[i].item())
for j in range((ctx_len + blocksize - 1) // blocksize):
block_tables[i][j] = block_idx_pool[counter]
counter += 1
q_fp8 = q.to(torch.float8_e4m3fn)
kv_cache_fp8 = kv_cache_cast_to_fp8(kv_cache)
q_fp8 = q.to(torch.float8_e4m3fn)
kv_cache_fp8 = kv_cache_cast_to_fp8(kv_cache)
# deep_gemm paged MQA logits requires 2D context_lens of
# shape (B, next_n) (csrc/apis/attention.hpp:332-335);
# see indexer.py:607-608. For each batch/next_n token, the
# effective context length is context_lens[b] - next_n + j + 1.
next_n_arange = torch.arange(next_n, device="cuda", dtype=torch.int32)
context_lens_2d = (
context_lens.unsqueeze(-1) - next_n + 1 + next_n_arange
).contiguous()
schedule_metadata = get_paged_mqa_logits_metadata(
context_lens_2d, blocksize, get_num_sms()
)
logits = fp8_fp4_paged_mqa_logits(
(q_fp8, None),
kv_cache_fp8,
weights,
context_lens_2d,
block_tables,
schedule_metadata,
max_model_len,
clean_logits=clean_logits,
)
# deep_gemm paged MQA logits requires 2D context_lens of
# shape (B, next_n) (csrc/apis/attention.hpp:332-335);
# see indexer.py:607-608. For each batch/next_n token, the
# effective context length is context_lens[b] - next_n + j + 1.
next_n_arange = torch.arange(next_n, device="cuda", dtype=torch.int32)
context_lens_2d = (
context_lens.unsqueeze(-1) - next_n + 1 + next_n_arange
).contiguous()
schedule_metadata = get_paged_mqa_logits_metadata(
context_lens_2d,
blocksize,
get_num_sms(),
)
logits = fp8_fp4_paged_mqa_logits(
(q_fp8, None),
kv_cache_fp8,
weights,
context_lens_2d,
block_tables,
schedule_metadata,
max_model_len,
clean_logits=clean_logits,
)
ref_logits = _ref_fp8_fp4_paged_mqa_logits(
q,
kv_cache,
weights,
context_lens,
block_tables,
max_model_len,
)
ref_logits = _ref_fp8_fp4_paged_mqa_logits(
q,
kv_cache,
weights,
context_lens,
block_tables,
max_model_len,
)
positions = (
torch.arange(max_model_len, device="cuda")
.unsqueeze(0)
.expand(batch_size * next_n, -1)
)
row_indices = torch.arange(batch_size * next_n, device="cuda") // next_n
next_n_offset = (
torch.arange(batch_size * next_n, device="cuda") % next_n
)
mask = positions <= (
context_lens[row_indices] - next_n + next_n_offset
).unsqueeze(1)
positions = (
torch.arange(max_model_len, device="cuda")
.unsqueeze(0)
.expand(batch_size * next_n, -1)
)
row_indices = torch.arange(batch_size * next_n, device="cuda") // next_n
next_n_offset = torch.arange(batch_size * next_n, device="cuda") % next_n
mask = positions <= (
context_lens[row_indices] - next_n + next_n_offset
).unsqueeze(1)
logits = logits.masked_fill(~mask, 0)
ref_logits = ref_logits.masked_fill(~mask, 0)
diff = calc_diff(logits, ref_logits)
assert diff < 1e-3, f"{diff=}"
logits = logits.masked_fill(~mask, 0)
ref_logits = ref_logits.masked_fill(~mask, 0)
diff = calc_diff(logits, ref_logits)
assert diff < 1e-3, f"{diff=}"
@@ -0,0 +1,75 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Which next_n the DSA indexer decode path may hand to DeepGEMM unflattened.
Getting this wrong is not a slow path but a crash: `fp8_fp4_paged_mqa_logits`
asserts both that the architecture implements the requested `next_n` and that
the schedule metadata was sized for the matching slot count.
"""
import pytest
from vllm.platforms import current_platform
from vllm.utils.deep_gemm import _paged_mqa_logits_schedule_slots
from vllm.v1.attention.backends.mla import indexer
NUM_SMS = 114 # H100 PCIe
def _set_arch(monkeypatch, family: int, *, cuda: bool = True, deep_gemm: bool = True):
monkeypatch.setattr(current_platform, "is_cuda", lambda: cuda)
monkeypatch.setattr(
current_platform,
"is_device_capability_family",
lambda capability, device_id=0: capability // 10 == family,
)
monkeypatch.setattr(indexer, "has_deep_gemm", lambda: deep_gemm)
@pytest.mark.parametrize(
"family,expected_native",
[
# SM90 gained next_n=4 (MTP=3) via 2-CTA multicast, but never 3.
(9, {1, 2, 4}),
# SM100 schedules any next_n with multi-atom tiles.
(10, {1, 2, 3, 4, 5, 8}),
# SM120 advertises multi-atom too but is unvalidated on hardware, so
# it stays on the conservative gate. Loosen it only with measurements.
(12, {1, 2}),
],
)
def test_native_decode_gate_per_architecture(monkeypatch, family, expected_native):
_set_arch(monkeypatch, family)
for next_n in (1, 2, 3, 4, 5, 8):
assert indexer._supports_native_decode(next_n) == (next_n in expected_native), (
f"family={family} next_n={next_n}"
)
@pytest.mark.parametrize(
"cuda,deep_gemm", [(False, True), (True, False), (False, False)]
)
def test_native_decode_gate_without_deepgemm(monkeypatch, cuda, deep_gemm):
"""Without the DeepGEMM kernels only the shapes every backend handles."""
_set_arch(monkeypatch, 9, cuda=cuda, deep_gemm=deep_gemm)
assert [indexer._supports_native_decode(n) for n in (1, 2, 3, 4)] == [
True,
True,
False,
False,
]
def test_sm90_next_n_4_halves_the_schedule_slots(monkeypatch):
"""SM90 next_n=4 runs one scheduler task per 2-CTA cluster, not per SM."""
_set_arch(monkeypatch, 9)
assert _paged_mqa_logits_schedule_slots(NUM_SMS, 4) == NUM_SMS // 2
for next_n in (1, 2, 3):
assert _paged_mqa_logits_schedule_slots(NUM_SMS, next_n) == NUM_SMS
@pytest.mark.parametrize("family", [10, 12])
def test_multicast_is_sm90_only(monkeypatch, family):
_set_arch(monkeypatch, family)
for next_n in (1, 2, 3, 4):
assert _paged_mqa_logits_schedule_slots(NUM_SMS, next_n) == NUM_SMS
+31 -5
View File
@@ -552,6 +552,29 @@ def fp8_fp4_mqa_logits(
)
def native_next_n_supported(next_n: int) -> bool:
"""Whether the paged MQA logits kernel takes `next_n` Q rows per request.
SM90 implements only {1, 2, 4}; SM100 and SM120 schedule any `next_n` via
multi-atom tiles. Unsupported values must be flattened to one row per query.
"""
if current_platform.is_device_capability_family(90):
return next_n in (1, 2, 4)
return True
def _paged_mqa_logits_schedule_slots(num_sms: int, next_n: int) -> int:
"""Scheduler tasks the paged MQA logits kernel launches.
SM90 `next_n=4` runs one task per 2-CTA multicast cluster rather than per
SM, and `fp8_fp4_paged_mqa_logits` asserts its metadata is sized to match.
"""
num_kv_multicast = (
2 if next_n == 4 and current_platform.is_device_capability_family(90) else 1
)
return num_sms // num_kv_multicast
def get_paged_mqa_logits_metadata(
context_lens: torch.Tensor,
block_size: int,
@@ -561,22 +584,24 @@ def get_paged_mqa_logits_metadata(
"""Build scheduling metadata for paged MQA logits.
Args:
context_lens: Tensor of shape [B], dtype int32; effective context length
per batch element.
context_lens: Tensor of shape [B, next_n], dtype int32; effective
context length per Q row.
block_size: KV-cache block size in tokens (e.g., 64).
num_sms: Number of SMs available. 132 for Hopper
indices: Optional request index for each varlen row.
Returns:
Backend-specific tensor consumed by `fp8_fp4_paged_mqa_logits` to
schedule work across SMs.
Tensor of shape [slots + 1, 2] consumed by `fp8_fp4_paged_mqa_logits`
to schedule work across SMs.
"""
_lazy_init()
if _get_paged_mqa_logits_metadata_impl is None:
return _missing()
next_n = context_lens.shape[1] if context_lens.dim() == 2 else 1
num_slots = _paged_mqa_logits_schedule_slots(num_sms, next_n)
kwargs = {} if indices is None else {"indices": indices}
return _get_paged_mqa_logits_metadata_impl(
context_lens, block_size, num_sms, **kwargs
context_lens, block_size, num_slots, **kwargs
)
@@ -752,6 +777,7 @@ __all__ = [
"fp8_fp4_mqa_logits",
"fp8_fp4_paged_mqa_logits",
"get_paged_mqa_logits_metadata",
"native_next_n_supported",
"per_block_cast_to_fp8",
"is_deep_gemm_e8m0_used",
"is_deep_gemm_supported",
+30 -11
View File
@@ -21,6 +21,7 @@ from vllm.triton_utils import tl, triton
from vllm.utils.deep_gemm import (
get_paged_mqa_logits_metadata,
has_deep_gemm,
native_next_n_supported,
)
from vllm.utils.platform_utils import num_compute_units
from vllm.v1.attention.backend import (
@@ -467,6 +468,20 @@ def _supports_varlen_paged_mqa_logits() -> bool:
)
def _supports_native_decode(next_n: int) -> bool:
"""Whether decode can pass `next_n` Q rows per request to the kernel
instead of flattening to one single-token row per query, which re-reads
the KV tile once per row.
"""
if not (current_platform.is_cuda() and has_deep_gemm()):
return next_n in (1, 2)
if current_platform.is_device_capability_family(100):
return True
if current_platform.is_device_capability_family(90):
return native_next_n_supported(next_n)
return next_n in (1, 2)
class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
# The indexer opts out of the shared reorder-threshold vote (see __init__),
# so this is None; its own split uses self.decode_threshold.
@@ -524,14 +539,7 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
next_n = self.num_speculative_tokens + 1
self.decode_threshold = next_n
self.reorder_batch_threshold = None
# NOTE: SM100 datacenter GPUs support any next_n natively via the
# multi-atom paged MQA logits kernels (FP8 and FP4 indexer
# caches). Outside the SM100 family the FP8
# paged MQA logits kernel only supports next_n in (1, 2)
# (deepgemm smxx_fp8_fp4_paged_mqa_logits.hpp:233), so flatten there.
self.use_flattening = not current_platform.is_device_capability_family(
100
) and next_n not in (1, 2)
self.use_flattening = not _supports_native_decode(next_n)
self.supports_varlen = _supports_varlen_paged_mqa_logits()
logger.info_once(
"DSA indexer decode path: use_flattening=%s supports_varlen=%s "
@@ -585,7 +593,8 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
device=self.device,
)
# See: DeepGMM/csrc/apis/attention.hpp
# See: DeepGMM/csrc/apis/attention.hpp. Sized for one slot per SM;
# build() narrows it to whatever the kernel actually schedules.
self.scheduler_metadata_buffer = torch.empty(
(self.num_sms + 1, 2), dtype=torch.int32, device=self.device
)
@@ -924,9 +933,16 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
max_decode_len = int(decode_lens_cpu.max().item())
next_n = 1 + self.num_speculative_tokens
# The kernel sees max_decode_len Q rows, not the configured next_n,
# so legality is per-step: on SM90 a uniformly 3-deep batch has no
# native kernel. max_decode_len <= 1 always has one.
step_next_n_ok = max_decode_len <= 1 or _supports_native_decode(
max_decode_len
)
use_native = (
not (self.use_flattening or self.supports_varlen)
and max_decode_len <= next_n
and step_next_n_ok
)
global_seq_lens_for_decode = self._prepare_global_decode_seq_lens(
@@ -992,20 +1008,23 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
seq_lens = seq_lens.unsqueeze(-1)
# DeepGEMM is required for the paged MQA logits on CUDA devices
schedule_metadata = self.scheduler_metadata_buffer
if current_platform.is_cuda() and has_deep_gemm():
self.scheduler_metadata_buffer[:] = get_paged_mqa_logits_metadata(
metadata = get_paged_mqa_logits_metadata(
seq_lens,
self.kv_cache_spec.storage_block_size,
self.num_sms,
indices=decode_indices,
)
schedule_metadata = self.scheduler_metadata_buffer[: metadata.shape[0]]
schedule_metadata[:] = metadata
decode_metadata = DeepSeekV32IndexerDecodeMetadata(
block_table=block_table,
seq_lens=seq_lens,
decode_lens=decode_lens,
requires_padding=requires_padding,
schedule_metadata=self.scheduler_metadata_buffer,
schedule_metadata=schedule_metadata,
indices=decode_indices,
global_seq_lens=global_seq_lens_for_decode,
)