Fix GGUF loader for Qwen3 MoE. (vllm-project#22785)

Signed-off-by: Gh0u1L5 <[email protected]>
Signed-off-by: Diego-Castan <[email protected]>

Author: Gh0u1L5

vllm/model_executor/model_loader/gguf_loader.py

@@ -74,6 +74,17 @@ def _get_gguf_weights_map(self, model_config: ModelConfig):
                         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"
+        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
         for key, value in gguf.MODEL_ARCH_NAMES.items():

vllm/model_executor/models/qwen3_moe.py

@@ -375,6 +375,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.embed_tokens = VocabParallelEmbedding(
             config.vocab_size,
             config.hidden_size,
+            quant_config=quant_config,
             prefix=f"{prefix}.embed_tokens")
         self.start_layer, self.end_layer, self.layers = make_layers(
             config.num_hidden_layers,