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

[DP][V1] Fix rank set in DP scenario & Bump torch-npu version to 2.5.1.post1.dev20250528 (#1235)

Browse Source 
### What this PR does / why we need it?
1. Fix rank set in DP scenario. The new poc version of torch-npu support
setting `ASCEND_RT_VISIBLE_DEVICES` dynamically, thus we could use the
rank set in `DPEngineCoreProc` directly instead of calculating local
rank across dp by hand in the patched `_init_data_parallel`

Closes: https://github.com/vllm-project/vllm-ascend/issues/1170

2. Bump torch-npu version to 2.5.1.post1.dev20250528

Closes: https://github.com/vllm-project/vllm-ascend/pull/1242
Closes: https://github.com/vllm-project/vllm-ascend/issues/1232


### How was this patch tested?
CI passed with new added test.

---------

Signed-off-by: MengqingCao <cmq0113@163.com>
Signed-off-by: Icey <1790571317@qq.com>
Co-authored-by: Icey <1790571317@qq.com>
This commit is contained in:
Mengqing Cao
2025-06-16 23:09:53 +08:00
committed by GitHub
parent f5404dc650
commit 96fa7ff63b
19 changed files with 114 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm_ascend/patch/platform/patch_common/patch_distributed.py
View File

@@ -21,10 +21,9 @@ import vllm
import vllm.distributed
import vllm.envs as envs
from torch.distributed import ProcessGroup
from vllm.config import ParallelConfig, VllmConfig
from vllm.config import ParallelConfig
from vllm.distributed.utils import \
stateless_init_torch_distributed_process_group
from vllm.v1.engine.core import DPEngineCoreProc
def ascend_destroy_model_parallel():
@@ -79,21 +78,6 @@ def stateless_init_dp_group(self) -> "ProcessGroup":
return dp_group
def _init_data_parallel(self, vllm_config: VllmConfig):
# Configure NPUs and stateless process group for data parallel.
dp_rank = vllm_config.parallel_config.data_parallel_rank
dp_size = vllm_config.parallel_config.data_parallel_size
local_dp_rank = vllm_config.parallel_config.data_parallel_rank_local
assert dp_size > 1
assert 0 <= local_dp_rank <= dp_rank < dp_size
self.local_dp_rank = local_dp_rank
self.dp_group = vllm_config.parallel_config.stateless_init_dp_group()
self.current_wave = 0
vllm.distributed.parallel_state.destroy_model_parallel = ascend_destroy_model_parallel
DPEngineCoreProc._init_data_parallel = _init_data_parallel
ParallelConfig.get_next_dp_init_port = parallel_config_get_dp_port
ParallelConfig.stateless_init_dp_group = stateless_init_dp_group