5d12446573
### What this PR does / why we need it?
2nd PR for https://github.com/vllm-project/vllm-ascend/issues/5712,
extend SP to VL MoE models.
### Does this PR introduce _any_ user-facing change?
remove `sp_threshold` in additional config and reuse `sp_min_token_num`
from vLLM.
### How was this patch tested?
- Model: Qwen3-VL-30B-A3B,
- TP4 DP2
- 100 reqs
- max concurrency 1
| Seq length | Mean TTFT (ms) main | Mean TTFT (ms) this PR |
|------------|---------------------|------------------------|
| 4k | 429.40 | 323.3 |
| 16k | 1297.01 | 911.74 |
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
78 lines
3.0 KiB
Python
78 lines
3.0 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 torch import fx as fx
|
|
from vllm.compilation.passes.inductor_pass import get_pass_context
|
|
from vllm.compilation.passes.vllm_inductor_pass import VllmInductorPass
|
|
from vllm.config import VllmConfig
|
|
|
|
|
|
class GraphFusionPassManager:
|
|
"""
|
|
A pass manager for graph fusion passes.
|
|
It handles the configuration and execution of passes.
|
|
The counterpart in vllm is PostGradPassManager. Since torch_npu
|
|
does not support triton for now, we define our own pass manager.
|
|
"""
|
|
|
|
def __init__(self):
|
|
self.passes: list[VllmInductorPass] = []
|
|
|
|
def __call__(self, graph: fx.Graph) -> fx.Graph:
|
|
compile_range = get_pass_context().compile_range
|
|
|
|
for pass_ in self.passes:
|
|
if pass_.is_applicable_for_range(compile_range):
|
|
pass_(graph)
|
|
graph.recompile()
|
|
return graph
|
|
|
|
def add(self, pass_: VllmInductorPass):
|
|
assert isinstance(pass_, VllmInductorPass)
|
|
self.passes.append(pass_)
|
|
|
|
def configure(self, config: VllmConfig):
|
|
# By default, we enable the graph fusion and quantization fusion pass.
|
|
self.ascend_compilation_config: dict = config.additional_config.get("ascend_compilation_config", {})
|
|
if self.ascend_compilation_config.get("fuse_norm_quant", True):
|
|
from .passes.norm_quant_fusion_pass import AddRMSNormQuantFusionPass
|
|
|
|
self.passes.append(AddRMSNormQuantFusionPass(config))
|
|
|
|
if self.ascend_compilation_config.get("fuse_qknorm_rope", True):
|
|
from .passes.qknorm_rope_fusion_pass import QKNormRopeFusionPass
|
|
|
|
self.passes.append(QKNormRopeFusionPass(config))
|
|
|
|
if self.ascend_compilation_config.get("fuse_allreduce_rms", True):
|
|
from .passes.allreduce_rmsnorm_fusion_pass import MatmulAllReduceAddRMSNormPass
|
|
|
|
self.passes.append(MatmulAllReduceAddRMSNormPass(config))
|
|
|
|
if self.ascend_compilation_config.get("fuse_muls_add", True):
|
|
from .passes.muls_add_pass import MulsAddFusionPass
|
|
|
|
self.passes.append(MulsAddFusionPass(config))
|
|
|
|
if config.compilation_config.pass_config.enable_sp:
|
|
from .passes.sequence_parallelism import SequenceParallelismPass
|
|
from .passes.sequence_parallelism_moe import SequenceParallelismMoePass
|
|
|
|
self.passes.append(SequenceParallelismPass(config))
|
|
self.passes.append(SequenceParallelismMoePass(config))
|