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

[long_seq] remove long_seq env (#4660)

Browse Source 
### What this PR does / why we need it?
remove env VLLM_ASCEND_ENABLE_CONTEXT_PARALLEL 

- vLLM version: v0.12.0

---------

Signed-off-by: LookAround <lixushi@huawei.com>
Signed-off-by: ZhangMingWei716 <2894054457@qq.com>
Co-authored-by: ZhangMingWei716 <2894054457@qq.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
LookAround0301
2025-12-05 10:31:49 +08:00
committed by GitHub
parent ea54388e19
commit b32ef53b3b
16 changed files with 230 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
vllm_ascend/worker/model_runner_v1.py
View File

@@ -52,7 +52,8 @@ from vllm.distributed.kv_transfer import (get_kv_transfer_group,
has_kv_transfer_group)
from vllm.distributed.kv_transfer.kv_connector.v1 import KVConnectorBase_V1
from vllm.distributed.parallel_state import (get_dcp_group, get_dp_group,
get_pp_group, get_tp_group,
get_pcp_group, get_pp_group,
get_tp_group,
is_global_first_rank)
from vllm.forward_context import get_forward_context
from vllm.logger import logger
@@ -145,16 +146,9 @@ from vllm_ascend.torchair.torchair_mtp_proposer import TorchairMtpProposer
from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, ACL_FORMAT_FRACTAL_NZ,
AscendDeviceType, ProfileExecuteDuration,
enable_sp, get_ascend_device_type, is_enable_nz,
is_moe_model, lmhead_tp_enable,
prefill_context_parallel_enable)
is_moe_model, lmhead_tp_enable)
from vllm_ascend.worker.npu_input_batch import CachedRequestState, InputBatch
if prefill_context_parallel_enable():
from vllm.distributed import get_pcp_group
from vllm.distributed.parallel_state import (
get_prefill_context_model_parallel_rank,
get_prefill_context_model_parallel_world_size)
if TYPE_CHECKING:
import xgrammar as xgr # type: ignore[import-untyped]
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
@@ -290,10 +284,9 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
self.dp_rank = vllm_config.parallel_config.data_parallel_rank
self.dcp_size = get_dcp_group().world_size
self.dcp_rank = get_dcp_group().rank_in_group
self.pcp_size = get_prefill_context_model_parallel_world_size(
) if prefill_context_parallel_enable() else 1
self.pcp_rank = get_prefill_context_model_parallel_rank(
) if self.pcp_size > 1 else 0
self.pcp_size = get_pcp_group().world_size
self.pcp_rank = get_pcp_group(
).rank_in_group if self.pcp_size > 1 else 0
decode_max_num_seqs = getattr(self.scheduler_config,
'decode_max_num_seqs', 0)
self.max_num_reqs = max(self.scheduler_config.max_num_seqs,
@@ -602,8 +595,7 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
if self.vllm_config.speculative_config else 0),
kernel_block_sizes=[[self.vllm_config.cache_config.block_size]],
cp_kv_cache_interleave_size=self.parallel_config.
cp_kv_cache_interleave_size
if prefill_context_parallel_enable() else 1,
cp_kv_cache_interleave_size,
)
self.num_accepted_tokens = self._make_buffer(self.max_num_reqs,
dtype=torch.int64)
@@ -2742,8 +2734,7 @@ class NPUModelRunner(LoRAModelRunnerMixin, ECConnectorModelRunnerMixin):
device=self.device)
long_seq_metadata = self._generate_pcp_metadata(num_tokens)
if long_seq_metadata is not None:
pcp_world_size = get_pcp_group(
).world_size if prefill_context_parallel_enable() else 1
pcp_world_size = get_pcp_group().world_size
dcp_world_size = get_dcp_group().world_size
num_computed_tokens_of_pcp_dcp = [[
[0] * dcp_world_size for _ in range(pcp_world_size)