c980e68d40
### What this PR does / why we need it?
This PR aims to support aclgraph for model runner v2, please see RFC
#5208. The PR contains these modifications:
- adapt to newest commit of vllm main branch.
- supply a unified interface of extra forward context for both model
runner v1 and model runner v2.
- implement graph mode for main model.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: Ronald1995 <ronaldautomobile@163.com>
104 lines
3.6 KiB
Python
104 lines
3.6 KiB
Python
# Adapt from https://github.com/vllm-project/vllm/blob/main/vllm/v1/worker/gpu/input_batch.py
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
# 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.
|
|
#
|
|
from dataclasses import asdict, dataclass
|
|
|
|
import numpy as np
|
|
import torch
|
|
from vllm.v1.worker.gpu.input_batch import InputBatch, InputBuffers
|
|
|
|
from vllm_ascend.attention.attention_v1 import AscendAttentionState
|
|
|
|
|
|
class AscendInputBuffers(InputBuffers):
|
|
"""Input buffers for Ascend NPUs."""
|
|
|
|
def __init__(
|
|
self,
|
|
max_num_reqs: int,
|
|
max_num_tokens: int,
|
|
device: torch.device,
|
|
):
|
|
super().__init__(
|
|
max_num_reqs,
|
|
max_num_tokens,
|
|
device,
|
|
)
|
|
del self.query_start_loc
|
|
|
|
# NOTE: For FULL mode we change +1 to +2 to reserve extra space for padding.
|
|
# See _pad_query_start_loc_for_fia.
|
|
self.query_start_loc: torch.Tensor = torch.zeros(
|
|
max_num_reqs + 2,
|
|
dtype=torch.int32,
|
|
device=device,
|
|
)
|
|
|
|
# Create seq_lens_cpu and seq_lens_np.
|
|
# npu's attention backend still needs seq_lens on CPU side.
|
|
self.seq_lens_cpu: torch.Tensor = torch.zeros(
|
|
max_num_reqs,
|
|
dtype=torch.int32,
|
|
device="cpu",
|
|
)
|
|
# seq_len_np and seq_lens_cpu share the same memory.
|
|
# define seq_lens_np for easier calculation with numpy.
|
|
self.seq_lens_np: np.ndarray = self.seq_lens_cpu.numpy()
|
|
|
|
|
|
@dataclass
|
|
class AscendInputBatch(InputBatch):
|
|
"""Input batch for Ascend NPUs."""
|
|
|
|
# Create seq_lens_np.
|
|
# npu's attention backend still needs seq_lens on CPU side.
|
|
seq_lens_np: np.ndarray
|
|
# attn_state is used to build attention metadata.
|
|
attn_state: AscendAttentionState | None = None
|
|
|
|
@classmethod
|
|
def make_dummy(
|
|
cls,
|
|
num_reqs: int,
|
|
num_tokens: int,
|
|
input_buffers: AscendInputBuffers,
|
|
device: torch.device,
|
|
) -> "AscendInputBatch":
|
|
"""Override the make_dummy method to calculate seq_lens_np."""
|
|
input_batch = InputBatch.make_dummy(
|
|
num_reqs,
|
|
num_tokens,
|
|
input_buffers,
|
|
device,
|
|
)
|
|
# seq_len equals to query_len
|
|
input_buffers.seq_lens_np[:num_reqs] = num_tokens // num_reqs
|
|
input_buffers.seq_lens_np[num_reqs - 1] += num_tokens % num_reqs
|
|
# Pad for full CUDA graph mode.
|
|
input_buffers.seq_lens_np[num_reqs:] = 0
|
|
seq_lens_np = input_buffers.seq_lens_np[:num_reqs]
|
|
input_batch.seq_lens_np = seq_lens_np
|
|
# A dummy run for dp or memory profiling.
|
|
# When dummy run for dp, num_tokens is set to 1,
|
|
# so attn_state is set to DecodeOnly.
|
|
# when dummy run for memory profiling,
|
|
# attention metadata isn't needed,
|
|
# we can also set attn_state to AscendAttentionState.DecodeOnly.
|
|
input_batch.attn_state = AscendAttentionState.DecodeOnly
|
|
return cls(**asdict(input_batch), seq_lens_np=seq_lens_np)
|