biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ROCm][Bugfix] Fix ptpc scale load issue for fused shared expert path in deepseek mtp (#33018)

Browse Source 
Signed-off-by: ganyi <ygan@amd.com>
This commit is contained in:
Pleaplusone
2026-01-26 23:19:04 +08:00
committed by GitHub
parent 9ef3b718d9
commit be6931ee27
1 changed files with 11 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
vllm/model_executor/models/deepseek_mtp.py
View File

@@ -316,7 +316,11 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
# Determine split axis based on op type
# gate/up: ColumnParallel → split along dim 0
# down: RowParallel → split along dim 1
split_dim = 1 if "down_proj.weight" in name else 0
split_dim = (
1
if ("down_proj.weight" in name and loaded_weight.ndim > 1)
else 0
)
total = loaded_weight.shape[split_dim]
assert total % num_chunks == 0, (
f"Shared expert weight dim {total} "
@@ -329,14 +333,13 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
weight_to_load = loaded_weight
if is_fusion_moe_shared_experts_layer:
if split_dim == 0:
weight_to_load = loaded_weight[
j * chunk_size : (j + 1) * chunk_size, :
]
chunk_slice = slice(j * chunk_size, (j + 1) * chunk_size)
if loaded_weight.ndim == 1:
weight_to_load = loaded_weight[chunk_slice]
elif split_dim == 0:
weight_to_load = loaded_weight[chunk_slice, :]
else:
weight_to_load = loaded_weight[
:, j * chunk_size : (j + 1) * chunk_size
]
weight_to_load = loaded_weight[:, chunk_slice]
# Synthesize an expert-style name so expert mapping
# can route it
chunk_name = name.replace(