xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-24 09:36:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix GGUF loader for Qwen3 MoE. (#22785)

Browse Source 
Signed-off-by: Gh0u1L5 <Gh0u1L5@outlook.com>
This commit is contained in:
Gh0u1L5 2025-08-13 21:08:23 +08:00 committed by GitHub
parent 6772bb0f7d
commit b159c0a67a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 12 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/model_executor/model_loader/gguf_loader.py
View File

@ -74,6 +74,17 @@ class GGUFModelLoader(BaseModelLoader):
f"model.layers.{idx}.mlp.experts.0.gate_proj.weight" f"model.layers.{idx}.mlp.experts.0.gate_proj.weight"
gguf_to_hf_name_map[f"blk.{idx}.ffn_up_exps.weight"] = \ gguf_to_hf_name_map[f"blk.{idx}.ffn_up_exps.weight"] = \
f"model.layers.{idx}.mlp.experts.0.up_proj.weight" f"model.layers.{idx}.mlp.experts.0.up_proj.weight"
if model_type in ("qwen2_moe", "qwen3_moe"):
model_type = model_type.replace("_", "")
# GGUF layer map assumes that we will have a merged expert weights
# so we need to map them manually
for idx in range(config.num_hidden_layers):
gguf_to_hf_name_map[f"blk.{idx}.ffn_down_exps.weight"] = \
f"model.layers.{idx}.mlp.experts.0.down_proj.weight"
gguf_to_hf_name_map[f"blk.{idx}.ffn_gate_exps.weight"] = \
f"model.layers.{idx}.mlp.experts.0.gate_proj.weight"
gguf_to_hf_name_map[f"blk.{idx}.ffn_up_exps.weight"] = \
f"model.layers.{idx}.mlp.experts.0.up_proj.weight"
arch = None arch = None
for key, value in gguf.MODEL_ARCH_NAMES.items(): for key, value in gguf.MODEL_ARCH_NAMES.items():

1
vllm/model_executor/models/qwen3_moe.py
View File

@ -375,6 +375,7 @@ class Qwen3MoeModel(nn.Module):
self.embed_tokens = VocabParallelEmbedding( self.embed_tokens = VocabParallelEmbedding(
config.vocab_size, config.vocab_size,
config.hidden_size, config.hidden_size,
quant_config=quant_config,
prefix=f"{prefix}.embed_tokens") prefix=f"{prefix}.embed_tokens")
self.start_layer, self.end_layer, self.layers = make_layers( self.start_layer, self.end_layer, self.layers = make_layers(
config.num_hidden_layers, config.num_hidden_layers,