cadfa5ddc1
### What this PR does / why we need it?
This PR add `qkv_rmsnorm_rope` operator and introduces a graph fusion
pass for `qknorm_rope` operations. The implementation includes a new
configuration flag, a pattern matching pass using
`torch._inductor.pattern_matcher`, and a custom Triton kernel for the
fused operation.
Co-authored-by: Angazenn
[supperccell@163.com](mailto:supperccell@163.com)
### Does this PR introduce _any_ user-facing change?
Yes, add new additional_config
- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c
---------
Signed-off-by: wxsIcey <1790571317@qq.com>
63 lines
2.3 KiB
Python
63 lines
2.3 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
#
|
|
# 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.
|
|
# This file is a part of the vllm-ascend project.
|
|
#
|
|
|
|
import torch
|
|
from vllm.triton_utils import HAS_TRITON
|
|
|
|
import vllm_ascend.ops.fused_moe.fused_moe # noqa
|
|
import vllm_ascend.ops.layernorm # noqa
|
|
import vllm_ascend.ops.register_custom_ops # noqa
|
|
|
|
if HAS_TRITON:
|
|
import vllm_ascend.ops.triton.linearnorm.split_qkv_rmsnorm_rope # noqa
|
|
|
|
import vllm_ascend.ops.vocab_parallel_embedding # noqa
|
|
from vllm_ascend.ops.activation import AscendQuickGELU, AscendSiluAndMul
|
|
from vllm_ascend.ops.rotary_embedding import (
|
|
AscendDeepseekScalingRotaryEmbedding, AscendRotaryEmbedding)
|
|
|
|
|
|
class dummyFusionOp:
|
|
default = None
|
|
|
|
def __init__(self, name=""):
|
|
self.name = name
|
|
|
|
|
|
def register_dummy_fusion_op() -> None:
|
|
torch.ops._C_ascend.rms_norm = dummyFusionOp(name="rms_norm")
|
|
torch.ops._C_ascend.fused_add_rms_norm = dummyFusionOp(
|
|
name="fused_add_rms_norm")
|
|
torch.ops._C_ascend.static_scaled_fp8_quant = dummyFusionOp(
|
|
name="static_scaled_fp8_quant")
|
|
torch.ops._C_ascend.dynamic_scaled_fp8_quant = dummyFusionOp(
|
|
name="dynamic_scaled_fp8_quant")
|
|
torch.ops._C_ascend.dynamic_per_token_scaled_fp8_quant = dummyFusionOp(
|
|
name="dynamic_per_token_scaled_fp8_quant")
|
|
torch.ops._C_ascend.rms_norm_static_fp8_quant = dummyFusionOp(
|
|
name="rms_norm_static_fp8_quant")
|
|
torch.ops._C_ascend.fused_add_rms_norm_static_fp8_quant = dummyFusionOp(
|
|
name="fused_add_rms_norm_static_fp8_quant")
|
|
torch.ops._C_ascend.rms_norm_dynamic_per_token_quant = dummyFusionOp(
|
|
name="rms_norm_dynamic_per_token_quant")
|
|
|
|
|
|
__all__ = [
|
|
"AscendQuickGELU", "AscendSiluAndMul", "AscendRotaryEmbedding",
|
|
"AscendDeepseekScalingRotaryEmbedding"
|
|
]
|