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

[bugfix] dep ineffective (#4416)

Browse Source 
### What this PR does / why we need it?
The expert mapping table and weights of the dynamic EPLB were not
updated, causing the accuracy to be correct but not effective. This bug
has now been fixed.

If it was tested in a way different from regular unit tests, please
clarify how you tested step by step, ideally copy and paste-able, so
that other reviewers can test and check, and descendants can verify in
the future.

Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
This commit is contained in:
LI SHENGYONG
2025-11-29 15:19:11 +08:00
committed by GitHub
parent 71acc8ddeb
commit cd9f5c0611
3 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm_ascend/eplb/adaptor/vllm_adaptor.py
View File

@@ -194,15 +194,15 @@ class VllmEplbAdaptor(EplbAdaptor):
json.dump(record, f, indent=4)
def do_update_expert_map(self, layer_id, updated_expert_map):
self.expert_map_per_layer[layer_id] = updated_expert_map.clone()
self.expert_map_per_layer_cpu[layer_id] = updated_expert_map.clone()
self.expert_map_per_layer[layer_id].copy_(updated_expert_map)
self.expert_map_per_layer_cpu[layer_id].copy_(updated_expert_map)
def do_update_expert_weight(self, layer_id, local_expert_to_replace,
buffer_tensor_id):
for expert_tensor, buffer_tensor in zip(
self.expert_param_per_layer[layer_id][local_expert_to_replace],
self.buffer_tensor_list[buffer_tensor_id]):
expert_tensor = buffer_tensor.clone()
expert_tensor.copy_(buffer_tensor)
logger.debug(f"Expert tensor shape is :{expert_tensor.shape}")
def do_update_log2phy_map(self, layer_id, updated_log2phy_map):

2
vllm_ascend/ops/common_fused_moe.py
View File

@@ -271,7 +271,7 @@ class AscendFusedMoE(FusedMoE):
return self.expert_map
def get_log2phy_map(self):
return self.logical_to_physical_map
return self.log2phy
def clear_moe_load(self):
if self.moe_load is not None:

2
vllm_ascend/torchair/ops/torchair_fused_moe.py
View File

@@ -1392,7 +1392,7 @@ class TorchairAscendFusedMoE(FusedMoE):
return self.expert_map
def get_log2phy_map(self):
return self.logical_to_physical_map
return self.log2phy
def clear_moe_load(self):
if self.moe_load is not None: