[Model][Quantization] Fix / Add GGUF support for Qwen2 MoE models (#30307)

Signed-off-by: Tsukasa OI <floss_llm@irq.a4lg.com>

Author: a4lg

vllm/model_executor/models/qwen2_moe.py

@@ -367,6 +367,8 @@ 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,
@@ -512,6 +514,12 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
                             continue
                         else:
                             name = remapped_kv_scale_name
+                    # GGUF: make sure that shared_expert_gate is a 2D tensor.
+                    if (
+                        "mlp.shared_expert_gate" in name
+                        and len(loaded_weight.shape) == 1
+                    ):
+                        loaded_weight = loaded_weight[None, :]
                     param = params_dict[name]
                     weight_loader = getattr(
                         param, "weight_loader", default_weight_loader