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

[Feature]EPLB:Adapt DispatchGmmCombineDecode operator to eplb tensor list and expert token numbers (#5552)

Browse Source 
#### What this PR does / why we need it?
This PR adapt DispatchGmmCombineDecode operator to eplb tensor list and
expert token numbers.

This operator support gmm1, gmm2, gmm1Scale and gmm2Scale in format of
list.
This operator support couting how many token each local expert recieves
by expertTokensNum .


- vLLM version: v0.13.0
- vLLM main:
7157596103

More info about this operator, please refer to RFC: issue
https://github.com/vllm-project/vllm-ascend/issues/5476
This commit is contained in:
wangyibo1005
2026-01-07 11:23:42 +08:00
committed by GitHub
parent 086c093347
commit 25baf6df09
18 changed files with 425 additions and 195 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm_ascend/eplb/adaptor/vllm_adaptor.py
View File

@@ -54,12 +54,15 @@ class VllmEplbAdaptor(EplbAdaptor):
self.model.model.layers[i].mlp.experts.w13_weight_scale_fp32_list
self.param_dict["model.layers." + str(i) + ".mlp.experts." + "w2_weight_scale_list"] = \
self.model.model.layers[i].mlp.experts.w2_weight_scale_list
self.param_dict["model.layers." + str(i) + ".mlp.experts." + "w2_weight_scale_fp32_list"] = \
self.model.model.layers[i].mlp.experts.w2_weight_scale_fp32_list
# TODO: init self.expert_weight_names depending on different model types, only deepseek v3 w8a8 and qwen3-moe is supported here
if self.model.quant_config is not None:
self.expert_weight_names = [
"w13_weight_list", "w2_weight_list",
"w13_weight_scale_fp32_list", "w13_weight_offset",
"w2_weight_scale_list", "w2_weight_offset"
"w2_weight_scale_list", "w2_weight_offset",
"w2_weight_scale_fp32_list"
]
else:
self.expert_weight_names = ["w13_weight", "w2_weight"]
@@ -97,7 +100,8 @@ class VllmEplbAdaptor(EplbAdaptor):
self.num_dense_layers) + ".mlp.experts." + name
if name in [
"w13_weight_list", "w2_weight_list",
"w13_weight_scale_fp32_list", "w2_weight_scale_list"
"w13_weight_scale_fp32_list", "w2_weight_scale_list",
"w2_weight_scale_fp32_list"
]:
expert_tensor = self.param_dict[complete_name][0]
expert_tensor = expert_tensor.clone()
@@ -118,7 +122,7 @@ class VllmEplbAdaptor(EplbAdaptor):
if name in [
"w13_weight_list", "w2_weight_list",
"w13_weight_scale_fp32_list",
"w2_weight_scale_list"
"w2_weight_scale_list", "w2_weight_scale_fp32_list"
]:
per_expert_param.append(
self.param_dict["model.layers." + str(layer_idx) +