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

[BugFix]Fix eplb problems when using dynamic eplb. (#3364)

Browse Source 
### What this PR does / why we need it?
When using dynamic eplb,it will be blocking by nz tensor.We fix these
prolems by clone src tensor and recv tensor.

### Does this PR introduce any user-facing change?

### How was this patch tested?
Qwen3_moe in A3.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: offline0806 <3337230449@qq.com>
Co-authored-by: offline0806 <3337230449@qq.com>
This commit is contained in:
offline893
2025-10-11 14:04:02 +08:00
committed by GitHub
parent ca05f7d632
commit 82b6c846ca
8 changed files with 58 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm_ascend/eplb/adaptor/vllm_adaptor.py
View File

@@ -80,15 +80,15 @@ class VllmEplbAdaptor(EplbAdaptor):
self.all_topk_ids = []
def init_buffer_tensor(self, num_buffer_tensor):
for name in self.expert_weight_names:
complete_name = "model.layers." + str(
self.num_dense_layers) + ".mlp.experts." + name
expert_tensor = self.param_dict[complete_name].data[
0:num_buffer_tensor]
buffer_tensors = torch.empty_like(expert_tensor)
for buffer_id in range(num_buffer_tensor):
self.buffer_tensor_list[buffer_id].append(
buffer_tensors[buffer_id])
for buffer_id in range(num_buffer_tensor):
for name in self.expert_weight_names:
complete_name = "model.layers." + str(
self.num_dense_layers) + ".mlp.experts." + name
expert_tensor = self.param_dict[complete_name].data[0]
if name in ["w13_weight", "w2_weight"]:
expert_tensor = expert_tensor.clone()
buffer_tensor = torch.empty_like(expert_tensor)
self.buffer_tensor_list[buffer_id].append(buffer_tensor)
def init_expert_param_per_layer(self):
num_local_expert = self.param_dict["model.layers." + str(self.num_dense_layers) + \

3
vllm_ascend/eplb/core/eplb_device_transfer_loader.py
View File

@@ -45,7 +45,7 @@ class D2DExpertWeightLoader:
layer_id):
# When current send/recv and weight.expert_map update tasks are not finished, cannot accept new d2d task
if self.state != ExpertWeightUpdateState.WAITING:
logger.error(
logger.warning_once(
"current d2d weight update tasks are on-going, cannot accept new weight update task"
)
return
@@ -64,6 +64,7 @@ class D2DExpertWeightLoader:
layer_id][global_expert_id_to_send].item()
for src_tensor in self.eplb_adaptor.expert_param_per_layer[
layer_id][local_expert_id]:
src_tensor = src_tensor.clone()
self.comm_op_list.append(
dist.P2POp(dist.isend, src_tensor, dst_rank))