[Model] Qwen3 Dense FP8 Compat Fixes (vllm-project#17318)

Signed-off-by: simon-mo <[email protected]>

Author: simon-mo

vllm/model_executor/layers/linear.py

@@ -929,6 +929,15 @@ def weight_loader_v2(self,
         shard_offset = self._get_shard_offset_mapping(loaded_shard_id)
         shard_size = self._get_shard_size_mapping(loaded_shard_id)
 
+        # Note(simon): This is needed for Qwen3's fp8 quantization.
+        if isinstance(param, BlockQuantScaleParameter):
+            assert self.quant_method is not None
+            assert hasattr(self.quant_method, "quant_config")
+            weight_block_size = self.quant_method.quant_config.weight_block_size
+            block_n, _ = weight_block_size[0], weight_block_size[1]
+            shard_offset = (shard_offset + block_n - 1) // block_n
+            shard_size = (shard_size + block_n - 1) // block_n
+
         param.load_qkv_weight(loaded_weight=loaded_weight,
                               num_heads=self.num_kv_head_replicas,
                               shard_id=loaded_shard_id,