[Perf] Add fused DFlash2 grouped convolution (#55960)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Misha Goin
2026-09-17 21:38:59 +00:00
committed by GitHub
parent 668d6c3a77
commit db7a24c230
4 changed files with 199 additions and 0 deletions
+4
View File
@@ -283,6 +283,8 @@ steps:
- vllm/v1/spec_decode/
- vllm/v1/worker/gpu/spec_decode/
- vllm/model_executor/models/qwen3_dflash.py
- vllm/model_executor/models/qwen3_dflash2.py
- vllm/model_executor/models/qwen3_5.py
- vllm/model_executor/models/laguna_dflash.py
- tests/v1/e2e/spec_decode/
commands:
@@ -293,6 +295,8 @@ steps:
dind: false
device: mi300_1
timeout_in_minutes: 120
commands:
- pytest -v -s v1/e2e/spec_decode/acceptance_rates/dflash/ -k "not dflash2"
depends_on:
- image-build-amd
source_file_dependencies:
@@ -26,6 +26,8 @@ class DFlashCorrectnessConfig:
use_chat_completions: bool = False
enforce_eager: bool = False
disable_flashinfer_sampler: bool = False
language_model_only: bool = False
chat_template_kwargs: dict[str, object] | None = None
QWEN3_DFLASH = DFlashCorrectnessConfig(
@@ -35,6 +37,18 @@ QWEN3_DFLASH = DFlashCorrectnessConfig(
expected_acceptance_len=3.5,
)
QWEN3_8_DFLASH2_NVFP4 = DFlashCorrectnessConfig(
model="nvidia/Qwen3.8-27B-NVFP4",
draft_model="z-lab/Qwen3.8-27B-DFlash2",
expected_accuracy=0.90,
expected_acceptance_len=6.14 * 0.95,
num_speculative_tokens=7,
max_num_seqs=32,
use_chat_completions=True,
language_model_only=True,
chat_template_kwargs={"enable_thinking": False},
)
LAGUNA_DFLASH_NVFP4 = DFlashCorrectnessConfig(
model="poolside/Laguna-XS-2.1-NVFP4",
draft_model="poolside/Laguna-XS-2.1-DFlash-NVFP4",
@@ -104,6 +118,11 @@ def test_dflash_reference_acceptance_lengths(
True,
id="laguna-nvfp4-mrv2",
),
pytest.param(
QWEN3_8_DFLASH2_NVFP4,
True,
id="qwen3.8-dflash2-nvfp4-mrv2",
),
],
)
def test_dflash_correctness(
@@ -134,12 +153,14 @@ def test_dflash_correctness(
disable_log_stats=False,
enable_chunked_prefill=None,
compilation_config=CompilationConfig(),
language_model_only=config.language_model_only,
) as spec_runner:
spec_llm = spec_runner.llm
results = evaluate_gsm8k_offline(
spec_llm,
num_questions=config.num_questions,
use_chat_completions=config.use_chat_completions,
chat_template_kwargs=config.chat_template_kwargs,
)
accuracy = results["accuracy"]
acceptance_len = compute_acceptance_len(spec_llm.get_metrics())
+50
View File
@@ -7,6 +7,7 @@ import pytest
import torch
from vllm.model_executor.models.qwen3_dflash2 import _grouped_conv, _score_edges
from vllm.platforms import current_platform
from vllm.v1.worker.gpu.spec_decode.dflash.speculator import DFlashSpeculator
from vllm.v1.worker.gpu.spec_decode.dflash2.speculator import DFlash2Speculator
@@ -35,6 +36,55 @@ def test_grouped_conv_matches_reference(block_size: int):
torch.testing.assert_close(actual, expected.flatten(0, 1).flatten(-2))
@pytest.mark.skipif(not current_platform.is_cuda(), reason="This test requires CUDA")
@pytest.mark.parametrize("dtype", [torch.float32, torch.bfloat16])
@pytest.mark.parametrize(
"batch,num_groups,group_size,block_size,taps",
[
(7, 13, 11, 5, 1),
(7, 13, 11, 5, 3),
(7, 13, 11, 8, 2),
(16, 64, 16, 8, 2),
(16, 160, 16, 8, 2),
],
)
def test_grouped_conv_triton_matches_reference(
dtype: torch.dtype,
batch: int,
num_groups: int,
group_size: int,
block_size: int,
taps: int,
):
torch.manual_seed(0)
rows = batch * block_size
hidden = torch.randn(rows, num_groups * group_size, device="cuda", dtype=dtype)
base = torch.randn(taps, num_groups * group_size, device="cuda", dtype=dtype)
projected = torch.randn(rows, 2, taps, num_groups, device="cuda", dtype=dtype)
delta = projected[:, 1]
actual = _grouped_conv(
hidden, delta, base, block_size, num_groups, group_size, taps
)
hidden_blocks = hidden.float().view(batch, block_size, num_groups, group_size)
expected = torch.zeros_like(hidden_blocks)
base_blocks = base.float().view(taps, num_groups, group_size)
delta_blocks = delta.float().view(batch, block_size, taps, num_groups)
for position in range(block_size):
for tap in range(min(taps, position + 1)):
expected[:, position] += (
base_blocks[tap] + delta_blocks[:, position, tap, :, None]
) * hidden_blocks[:, position - tap]
torch.testing.assert_close(
actual,
expected.flatten(0, 1).flatten(-2).to(dtype),
rtol=1e-2 if dtype is torch.bfloat16 else 1e-5,
atol=1e-2 if dtype is torch.bfloat16 else 1e-5,
)
def test_selector_edges_match_sequential_reference():
torch.manual_seed(1)
batch, steps, top_k, rank = 2, 4, 3, 5
+124
View File
@@ -11,6 +11,8 @@ from vllm.config import CacheConfig, VllmConfig
from vllm.model_executor.layers.linear import ReplicatedLinear
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
from vllm.triton_utils import tl, triton
from vllm.utils.torch_utils import direct_register_custom_op
from .qwen3_dflash import (
DFlashQwen3DecoderLayer,
@@ -20,6 +22,125 @@ from .qwen3_dflash import (
from .utils import maybe_prefix
@triton.jit
def _dflash2_grouped_conv_kernel(
x_ptr,
delta_ptr,
base_ptr,
output_ptr,
x_stride_row,
delta_stride_row,
delta_stride_tap,
base_stride_tap,
output_stride_row,
NUM_CHANNELS: tl.constexpr,
BLOCK_SIZE: tl.constexpr,
GROUP_SIZE: tl.constexpr,
TAPS: tl.constexpr,
ELEMENT_BLOCK: tl.constexpr,
) -> None:
row = tl.program_id(0) // triton.cdiv(NUM_CHANNELS, ELEMENT_BLOCK)
col_block = tl.program_id(0) % triton.cdiv(NUM_CHANNELS, ELEMENT_BLOCK)
channels = col_block * ELEMENT_BLOCK + tl.arange(0, ELEMENT_BLOCK)
mask = channels < NUM_CHANNELS
groups = channels // GROUP_SIZE
position = row % BLOCK_SIZE
delta_row = delta_ptr + row * delta_stride_row
x_row = x_ptr + row * x_stride_row
accumulator = (
tl.load(base_ptr + channels, mask=mask, other=0.0).to(tl.float32)
+ tl.load(delta_row + groups, mask=mask, other=0.0).to(tl.float32)
) * tl.load(x_row + channels, mask=mask, other=0.0).to(tl.float32)
for tap in tl.static_range(1, TAPS):
tap_mask = mask & (position >= tap)
coefficient = tl.load(
base_ptr + tap * base_stride_tap + channels,
mask=tap_mask,
other=0.0,
).to(tl.float32) + tl.load(
delta_row + tap * delta_stride_tap + groups,
mask=tap_mask,
other=0.0,
).to(tl.float32)
x = tl.load(
x_ptr + (row - tap) * x_stride_row + channels,
mask=tap_mask,
other=0.0,
).to(tl.float32)
accumulator += coefficient * x
tl.store(
output_ptr + row * output_stride_row + channels,
accumulator,
mask=mask,
)
def dflash2_grouped_conv_impl(
x: torch.Tensor,
delta: torch.Tensor,
base: torch.Tensor,
block_size: int,
group_size: int,
) -> torch.Tensor:
"""DFlash2 grouped convolution adapted from LightSeek TokenSpeed PR #1399."""
num_rows, num_channels = x.shape
output = torch.empty_like(x)
if num_rows == 0:
return output
element_block = 1024 if num_rows >= 128 and num_channels % 1024 == 0 else 512
grid = (num_rows * triton.cdiv(num_channels, element_block),)
_dflash2_grouped_conv_kernel[grid](
x,
delta,
base,
output,
x.stride(0),
delta.stride(0),
delta.stride(1),
base.stride(0),
output.stride(0),
NUM_CHANNELS=num_channels,
BLOCK_SIZE=block_size,
GROUP_SIZE=group_size,
TAPS=base.shape[0],
ELEMENT_BLOCK=element_block,
num_warps=4,
)
return output
def dflash2_grouped_conv_fake(
x: torch.Tensor,
delta: torch.Tensor,
base: torch.Tensor,
block_size: int,
group_size: int,
) -> torch.Tensor:
return torch.empty_like(x)
direct_register_custom_op(
op_name="dflash2_grouped_conv",
op_func=dflash2_grouped_conv_impl,
fake_impl=dflash2_grouped_conv_fake,
tags=(torch.Tag.needs_fixed_stride_order,),
)
def dflash2_grouped_conv(
x: torch.Tensor,
delta: torch.Tensor,
base: torch.Tensor,
block_size: int,
group_size: int,
) -> torch.Tensor:
return torch.ops.vllm.dflash2_grouped_conv(x, delta, base, block_size, group_size)
def _grouped_conv(
hidden_states: torch.Tensor,
delta: torch.Tensor,
@@ -29,6 +150,9 @@ def _grouped_conv(
group_size: int,
taps: int,
) -> torch.Tensor:
if hidden_states.is_cuda:
return dflash2_grouped_conv(hidden_states, delta, base, block_size, group_size)
blocks = hidden_states.unflatten(-1, (num_groups, group_size))
coefficients = base.view(1, taps, num_groups, group_size) + delta.unsqueeze(-1)
output = coefficients[:, 0] * blocks