640ecd1b77
### What this PR does / why we need it?
fix: support model-specific routed_scaling_factor in muls_add fusion
Previously, MulsAddFusionPass used a hardcoded scale=1.0, which failed
to match the x * routed_scaling_factor + y pattern in models like GLM5
that use routed_scaling_factor=2.5. This caused the muls_add fusion to
be skipped, leaving unoptimized mul+add operations.
This fix reads routed_scaling_factor from model config (defaulting to
1.0
for backward compatibility) and uses it as the pattern scale, enabling
correct fusion for GLM5 and other models with custom scaling factors.
Fixes: Unoptimized mul+add in GLM5 attention blocks
Tested: GLM5-W8A8 with routed_scaling_factor=2.5
### Does this PR introduce _any_ user-facing change?
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
15d76f74e2
Signed-off-by: liuchenbing <chenliumail@163.com>
Co-authored-by: liuchenbing <chenliumail@163.com>
116 lines
4.2 KiB
Python
116 lines
4.2 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
# This file is a part of the vllm-ascend project.
|
|
#
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
#
|
|
|
|
from __future__ import annotations
|
|
|
|
import torch
|
|
from torch._inductor.pattern_matcher import PatternMatcherPass
|
|
from vllm.config import VllmConfig
|
|
from vllm.config.compilation import Range
|
|
from vllm.logger import logger
|
|
|
|
from vllm_ascend.compilation.passes.base_pattern import BasePattern
|
|
from vllm_ascend.utils import vllm_version_is
|
|
|
|
if vllm_version_is("0.15.0"):
|
|
from vllm.compilation.vllm_inductor_pass import VllmInductorPass # type: ignore
|
|
else:
|
|
from vllm.compilation.passes.vllm_inductor_pass import VllmInductorPass
|
|
|
|
|
|
class MulsAddPattern(BasePattern):
|
|
"""
|
|
Pattern that matches an element-wise mul + add sequence:
|
|
tmp = x * scale
|
|
out = tmp + y
|
|
and replaces it with a call to the muls_add_triton kernel.
|
|
"""
|
|
|
|
def __init__(self, vllm_config: VllmConfig, scale: float = 1.0):
|
|
super().__init__(vllm_config)
|
|
self.scale = scale
|
|
|
|
def get_inputs(self) -> list[torch.Tensor]:
|
|
"""
|
|
Generate example inputs for the MulsAddPattern.
|
|
|
|
The exact shapes are not important for pattern matching; they only
|
|
provide meta information for the pattern matcher.
|
|
"""
|
|
x = torch.randn(2, 2048, device="npu", dtype=self.dtype)
|
|
y = torch.randn(2, 2048, device="npu", dtype=self.dtype)
|
|
# Only tensor inputs are needed here. The scalar scale is stored on the
|
|
# pattern instance (self.scale) instead of being passed as an input.
|
|
return [x, y]
|
|
|
|
def get_pattern(self):
|
|
def pattern(x: torch.Tensor, y: torch.Tensor):
|
|
"""
|
|
Pattern for element-wise x * scale + y.
|
|
"""
|
|
tmp = x * self.scale
|
|
out = tmp + y
|
|
return out
|
|
|
|
return pattern
|
|
|
|
def get_replacement(self):
|
|
def replacement(x: torch.Tensor, y: torch.Tensor):
|
|
"""
|
|
Replacement that calls the muls_add_triton kernel using the
|
|
class-level scalar self.scale.
|
|
"""
|
|
return torch.ops.vllm.muls_add(x, y, self.scale)
|
|
|
|
return replacement
|
|
|
|
|
|
class MulsAddFusionPass(VllmInductorPass):
|
|
"""
|
|
A fusion pass that replaces simple element-wise x * scale + y patterns
|
|
with the Triton-based muls_add_triton kernel on Ascend.
|
|
"""
|
|
|
|
def __init__(self, vllm_config: VllmConfig):
|
|
super().__init__(vllm_config)
|
|
self.pattern_match_passes: PatternMatcherPass = PatternMatcherPass(pass_name="muls_add_fusion_pass")
|
|
|
|
# For now we enable this pass for all floating-point dtypes that the
|
|
# model is configured to use.
|
|
dtype = vllm_config.model_config.dtype
|
|
if dtype not in (torch.float16, torch.bfloat16, torch.float32):
|
|
logger.debug("MulsAdd fusion not enabled: unsupported dtype %s", dtype)
|
|
return
|
|
|
|
routed_scaling_factor = getattr(vllm_config.model_config.hf_text_config, "routed_scaling_factor", 1.0)
|
|
MulsAddPattern(vllm_config, scale=routed_scaling_factor).register(self.pattern_match_passes)
|
|
|
|
def __call__(self, graph: torch.fx.Graph) -> None: # type: ignore[override]
|
|
self.begin()
|
|
self.matched_count = self.pattern_match_passes.apply(graph)
|
|
logger.debug("Fused %s muls_add patterns", self.matched_count)
|
|
self.end_and_log()
|
|
|
|
def is_applicable_for_range(self, compile_range: Range) -> bool:
|
|
"""
|
|
Check if the pass is applicable for the current configuration.
|
|
|
|
For now, muls_add fusion is always allowed for the selected ranges.
|
|
This hook exists so that we can add more fine-grained range control
|
|
in the future if needed.
|
|
"""
|
|
return True
|