d01fd1d1c3
### What this PR does / why we need it?
This miscellaneous contains several small fixes:
1) fix initialization and forward bugs of DeepseekMTPLayer with
`shared_expert_dp` enabled.
2) fix a tensor shape mismatches after o_proj caused by a work-aroud
change in NPUModelRunner.
3) avoid unnecessary decline of kv_cache memory (default: 64MB) with
`use_cached_kv_cache_bytes` disabled.
4) fall back `fused_moe_state` from `MC2` to `All2All` since the padding
logic of `mc2_mask` is incompatible with input hidden_states when
`shared_expert_dp` enabled.
Once this PR is merged, users can launch disaggregated_prefill
deployments (large_ep) with `deepseek_mtp` and `shared_expert_dp` as
`v0.9.1-dev` branch. The remaining problem of kv_cache tokens decline
compared to `v0.9.1-dev` will be resolved by
https://github.com/vllm-project/vllm-ascend/pull/3073.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E vllm serving about deepseek_mtp with torchair graph mode and
`enable_shared_expert_dp` with eager mode. Large ep deployments are also
tested with this PR.
- vLLM version: v0.10.2
- vLLM main:
5aeb925452
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
64 lines
3.0 KiB
Python
64 lines
3.0 KiB
Python
#
|
|
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
|
# Copyright 2023 The vLLM team.
|
|
#
|
|
# 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.
|
|
import torch
|
|
from vllm.logger import logger
|
|
|
|
import vllm_ascend.envs as envs_ascend
|
|
from vllm_ascend.ascend_config import get_ascend_config
|
|
from vllm_ascend.torchair.torchair_model_runner import NPUTorchairModelRunner
|
|
from vllm_ascend.torchair.utils import (check_kv_cache_bytes_cache_exist,
|
|
delete_torchair_cache_file,
|
|
read_kv_cache_bytes_from_file)
|
|
from vllm_ascend.worker.worker_v1 import NPUWorker
|
|
|
|
|
|
class NPUTorchairWorker(NPUWorker):
|
|
"""Torchair worker bases on NPUWorker. Only torchair specified code should be added in this class."""
|
|
|
|
def determine_available_memory(self) -> int:
|
|
"""Override determine_available_memory to use cached torchair kv_cache_bytes."""
|
|
|
|
available_kv_cache_memory = super().determine_available_memory()
|
|
ascend_config = get_ascend_config()
|
|
if ascend_config.enable_shared_expert_dp:
|
|
return available_kv_cache_memory
|
|
if ascend_config.torchair_graph_config.use_cached_kv_cache_bytes:
|
|
if check_kv_cache_bytes_cache_exist():
|
|
old_kv_cache_bytes = read_kv_cache_bytes_from_file(
|
|
torch.distributed.get_rank())
|
|
if 0 < old_kv_cache_bytes <= available_kv_cache_memory:
|
|
logger.info(
|
|
f"Use cached torchair kv_cache_bytes: {old_kv_cache_bytes}"
|
|
)
|
|
self.model_runner.new_kv_cache_bytes = old_kv_cache_bytes
|
|
return old_kv_cache_bytes
|
|
else:
|
|
logger.info(
|
|
"Cached torchair kv_cache_bytes is too big, invalidate old torchair_cache"
|
|
)
|
|
delete_torchair_cache_file()
|
|
bytes_floating_tolerance = 1024 * 1024 * envs_ascend.VLLM_ASCEND_KV_CACHE_MEGABYTES_FLOATING_TOLERANCE
|
|
available_kv_cache_memory -= bytes_floating_tolerance
|
|
logger.info(f"Use new kv_cache_bytes: {available_kv_cache_memory}")
|
|
self.model_runner.new_kv_cache_bytes = available_kv_cache_memory
|
|
return available_kv_cache_memory
|
|
|
|
def init_device(self):
|
|
"""Override init_device to init torchair model runner"""
|
|
device = self._init_device()
|
|
# Init ModelRunner here, so that we have access to self.device.
|
|
self.model_runner = NPUTorchairModelRunner(self.vllm_config, device)
|