EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[CI] optimize lint term (#5986)

Browse Source 
### What this PR does / why we need it?
This patch purpose to optimize the lint check term. The main idea is to
reduce unnecessary installation time.
1. The installation of vllm is not must, only append the path of vllm
src to the `PATHONPATH` is effective
2. This installation of `requirements-dev.txt` is not must, we have a
pre-built image `quay.io/ascend-ci/vllm-ascend:lint` with all the
requirements installed in advance.
**NOTE**: the conditions for triggering image builds are: 1).Daily
scheduled build; 2) Build when requirements are modified; 3) Manual
build. This ensures that the dependencies in our image are up-to-date to
the greatest extent possible.
3. The `mypy` was separated from the `pre-commit` hook for performance
reasons; we found that integrating `mypy` into the `pre-commit` hook
resulted in poor performance.
4. Reduce the CPU core consumption from 16 -> 8

### Does this PR introduce _any_ user-facing change?
The end-to-end lint time was optimized from 20min/per PR to 8min/per PR
### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
2c24bc6996

---------

Signed-off-by: wangli <wangli858794774@gmail.com>
This commit is contained in:
Li Wang
2026-01-22 15:46:59 +08:00
committed by GitHub
parent 9bba0a2a68
commit 484e7c59dc
10 changed files with 196 additions and 77 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
vllm_ascend/_310p/attention/attention_v1.py
View File

@@ -15,6 +15,7 @@
# This file is a part of the vllm-ascend project.
#
from typing import Any
import torch
import torch_npu
@@ -23,7 +24,7 @@ from vllm_ascend._310p.attention.attention_mask import AttentionMaskBuilder, bui
from vllm_ascend._310p.attention.metadata_builder import AscendAttentionMetadataBuilder310P
from vllm_ascend.attention.attention_v1 import AscendAttentionBackend as _BaseBackend
from vllm_ascend.attention.attention_v1 import AscendAttentionBackendImpl as _BaseImpl
from vllm_ascend.attention.attention_v1 import AscendAttentionMetadataBuilder, AscendAttentionState
from vllm_ascend.attention.attention_v1 import AscendAttentionMetadataBuilder, AscendAttentionState, AscendMetadata
from vllm_ascend.utils import ACL_FORMAT_FRACTAL_NZ, aligned_16, nd_to_nz_2d
@@ -47,9 +48,17 @@ class AscendAttentionBackend310(_BaseBackend):
class AscendAttentionBackendImpl310(_BaseImpl):
def forward_paged_attention(self, query, attn_metadata, output):
def forward_paged_attention(
self,
query: Any,
attn_metadata: AscendMetadata,
output: Any | None = None,
) -> Any:
if attn_metadata.seq_lens.device != query.device:
attn_metadata.seq_lens = attn_metadata.seq_lens.to(device=query.device, non_blocking=True)
attn_metadata.seq_lens = attn_metadata.seq_lens.to(
device=query.device,
non_blocking=True,
)
return super().forward_paged_attention(query, attn_metadata, output)
def _forward_prefill_310p_fallback(self, query, key, value, attn_metadata, output):

7
vllm_ascend/attention/attention_v1.py
View File

@@ -46,9 +46,7 @@ from vllm_ascend.device.device_op import DeviceOperator
from vllm_ascend.ops.flashcomm2_oshard_manager import flashcomm2_oshard_manager
from vllm_ascend.utils import vllm_version_is, weak_ref_tensors
# isort: off
if vllm_version_is("0.13.0"):
from vllm.v1.attention.backends.utils import AttentionCGSupport, AttentionMetadataBuilder
from vllm.attention.backends.abstract import ( # type: ignore
AttentionBackend,
AttentionImpl,
@@ -59,20 +57,21 @@ if vllm_version_is("0.13.0"):
AttentionBackendEnum,
register_backend,
)
from vllm.v1.attention.backends.utils import AttentionCGSupport, AttentionMetadataBuilder
else:
from vllm.v1.attention.backend import ( # type: ignore
AttentionBackend,
AttentionCGSupport,
AttentionImpl,
AttentionLayer,
AttentionType,
AttentionMetadataBuilder,
AttentionType,
)
from vllm.v1.attention.backends.registry import ( # type: ignore
AttentionBackendEnum,
register_backend,
)
# isort: on
# default max value of sliding window size
SWA_INT_MAX = 2147483647

45
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py
View File

@@ -13,7 +13,7 @@ from collections import defaultdict, deque
from collections.abc import Iterator
from concurrent.futures import ThreadPoolExecutor
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, List, Optional, OrderedDict, Tuple
from typing import TYPE_CHECKING, Any, List, Optional, OrderedDict, Tuple, TypedDict
import msgspec
import numpy as np
@@ -60,6 +60,11 @@ GET_META_MSG = b"get_meta_msg"
DONE_RECVING_MSG = b"done_recving_msg"
class RemotePortInfo(TypedDict):
num: int
host: str
class MooncakeAgentMetadata(msgspec.Struct, omit_defaults=True, dict=True):
engine_id: str
te_rpc_port: int
@@ -384,7 +389,7 @@ class KVCacheRecvingThread(threading.Thread):
remote_handshake_port: int,
offset: int,
tp_num_need_pulls: int,
remote_port_send_num: dict[int, dict[str, int | str]] = {},
remote_port_send_num: dict[int, RemotePortInfo] = {},
all_task_done: bool = False):
"""Add a new request to the queue for processing."""
logger.debug(f"Adding request {request_id} to the queue.")
@@ -458,8 +463,9 @@ class KVCacheRecvingThread(threading.Thread):
self._send_done_signal_to_free_remote_port(remote_request_id, remote_host,
remote_port_send_num)
def _send_done_signal_to_free_remote_port(self, request_id, remote_host,
remote_port_send_num):
def _send_done_signal_to_free_remote_port(
self, request_id: str, remote_host: str,
remote_port_send_num: dict[int, RemotePortInfo]):
if self.side_channel_port != self.local_handshake_port \
or not remote_port_send_num:
return
@@ -708,9 +714,10 @@ class KVCacheRecvingThread(threading.Thread):
logger.debug("Returned socket to pool for %s:%d", remote_host,
remote_handshake_port)
def _send_done_recv_signal(self, request_id: str, remote_host: str,
remote_handshake_port: int,
remote_port_send_num: dict[int, dict[str, int | str]]):
def _send_done_recv_signal(
self, request_id: str, remote_host: str,
remote_handshake_port: int,
remote_port_send_num: dict[int, RemotePortInfo]):
logger.debug("Sending done recving signal for request %s to %s:%d",
request_id, remote_host, remote_handshake_port)
sock: Optional[zmq.Socket] = None # type: ignore
@@ -1177,7 +1184,7 @@ class MooncakeConnectorWorker:
self.tp_num_need_pulls = num_d_block_heads // num_p_block_heads
self.local_remote_block_port_mapping: dict[
str, Optional[List[List[int]]]] = {}
self.remote_port_send_num: dict[str, dict[int, dict[str, int | str]]] = {}
self.remote_port_send_num: dict[str, dict[int, RemotePortInfo]] = {}
def _get_prefill_decode_size(self, vllm_config: VllmConfig):
# get prefill tp and dp size from extra config
@@ -1463,16 +1470,20 @@ class MooncakeConnectorWorker:
return local_remote_block_port_mappings
def get_remote_port_send_num(local_remote_block_port_mappings):
remote_port_send_num: dict[int, dict[str, int | str]] = {}
def get_remote_port_send_num(
local_remote_block_port_mappings: dict[int, list[list[int]]]
) -> dict[int, RemotePortInfo]:
remote_port_send_num: dict[int, RemotePortInfo] = {}
for port in range(self._prefill_tp_size * meta.remote_pcp_size):
remote_host = meta.remote_multi_nodes_meta_mapping[str(port)]['host']
remote_port_send_num[meta.remote_port + port] = {}
remote_port_send_num[meta.remote_port + port]['num'] = 0
remote_port_send_num[meta.remote_port + port]['host'] = remote_host
for local_port in local_remote_block_port_mappings.keys():
remote_port_head_list = local_remote_block_port_mappings[
local_port]
remote_host = str(meta.remote_multi_nodes_meta_mapping[str(
port)]['host'])
remote_port_send_num[meta.remote_port + port] = {
'num': 0,
'host': remote_host
}
for remote_port_head_list in local_remote_block_port_mappings.values(
):
for remote_port_list in remote_port_head_list:
for remote_port in remote_port_list:
remote_port_send_num[remote_port]['num'] += 1

8
vllm_ascend/xlite/xlite.py
View File

@@ -25,7 +25,7 @@ from vllm.distributed import (get_ep_group,
from vllm.forward_context import get_forward_context
from vllm.logger import logger
from vllm.sequence import IntermediateTensors
from xlite._C import (AttnMHA, Model, ModelAttnMeta, ModelConfig, Runtime,
from xlite._C import (AttnMHA, Model, ModelAttnMeta, ModelConfig, Runtime, # type: ignore[attr-defined]
ScoringFuncSoftmax)
import vllm_ascend.envs as envs_ascend
@@ -214,10 +214,10 @@ class QwenMoeXliteModel(LlamaXliteModel):
config.def_dp_size = vllm_config.parallel_config.data_parallel_size
config.moe_ep_size = ep_group.world_size if vllm_config.parallel_config.enable_expert_parallel else 1
config.moe_tp_size = 1 if vllm_config.parallel_config.enable_expert_parallel else ep_group.world_size
config.experts_weight_transpose = True
config.experts_weight_transpose = True # type: ignore
config.moe_intermediate_size = hf_config.moe_intermediate_size
config.norm_topk_prob = hf_config.norm_topk_prob
config.scoring_func = ScoringFuncSoftmax
config.norm_topk_prob = hf_config.norm_topk_prob # type: ignore
config.scoring_func = ScoringFuncSoftmax # type: ignore
return config
def _build_model(self, runnable: nn.Module, vllm_config: VllmConfig,