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

[Dist][EP] Remove ETP/EP maintained in vllm-ascend (#1681)

Browse Source 
### What this PR does / why we need it?
Remove ETP/EP maintained in branch main. We drop this as there is no
relevant scenarios to use ETP now, and we may subsequently advocate
implementing expert tensor parallelism in vLLM to support scenarios
where the expert is needed to be sliced

This is a part of #1422 backport.

Fixes https://github.com/vllm-project/vllm-ascend/issues/1396
https://github.com/vllm-project/vllm-ascend/issues/1154

### Does this PR introduce _any_ user-facing change?
We'll not maintain etp/ep in vllm-ascend anymore, and use the tp/ep in
vllm instead.

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


- vLLM version: v0.9.2
- vLLM main:
fe8a2c544a

Signed-off-by: MengqingCao <cmq0113@163.com>
This commit is contained in:
Mengqing Cao
2025-07-21 09:08:04 +08:00
committed by GitHub
parent a8b316ac5b
commit 8cfd257992
24 changed files with 66 additions and 548 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
vllm_ascend/distributed/parallel_state.py
View File

@@ -1,77 +0,0 @@
from typing import Optional
import torch
from vllm.distributed.parallel_state import (GroupCoordinator, get_world_group,
init_model_parallel_group)
# vllm-ascend will maintain its own EP GroupCoordinator and ETP GroupCoordinator for
# customize parallel solution
_EP: Optional[GroupCoordinator] = None
_ETP: Optional[GroupCoordinator] = None
def get_ep_group() -> GroupCoordinator:
assert _EP is not None, ("expert model parallel group is not initialized")
return _EP
def get_etp_group() -> GroupCoordinator:
assert _ETP is not None, (
"expert tensor parallel group is not initialized")
return _ETP
def model_parallel_initialized():
return (_ETP is not None and _EP is not None)
def init_ascend_model_parallel(
expert_parallel_size: int = 1,
expert_tensor_parallel_size: int = 1,
world_size: Optional[int] = None,
backend: Optional[str] = None,
):
if model_parallel_initialized():
return
assert torch.distributed.is_initialized()
world_size = world_size or torch.distributed.get_world_size()
backend = backend or torch.distributed.get_backend(
get_world_group().device_group)
num_expert_parallel_groups = expert_tensor_parallel_size
num_expert_tensor_parallel_groups = expert_parallel_size
global _EP
group_ranks = []
for i in range(num_expert_parallel_groups):
ranks = list(range(i, world_size, num_expert_parallel_groups))
group_ranks.append(ranks)
_EP = init_model_parallel_group(group_ranks,
get_world_group().local_rank,
backend,
group_name="ep")
group_ranks = []
global _ETP
for i in range(num_expert_tensor_parallel_groups):
ranks = list(
range(i * expert_tensor_parallel_size,
(i + 1) * expert_tensor_parallel_size))
group_ranks.append(ranks)
_ETP = init_model_parallel_group(group_ranks,
get_world_group().local_rank,
backend,
group_name="etp")
def destory_ascend_model_parallel():
global _EP
if _EP:
_EP.destroy()
_EP = None
global _ETP
if _ETP:
_ETP.destroy()
_ETP = None