[Qwen3 VL] Add LoRA support for Qwen 3 VL (#12165)

python/sglang/srt/models/qwen3_vl.py

@@ -14,6 +14,7 @@
 # ==============================================================================
 """Inference-only Qwen3-VL model compatible with HuggingFace weights."""
 import logging
+import re
 from functools import lru_cache, partial
 from typing import Callable, Iterable, List, Optional, Tuple, Union
 
@@ -691,6 +692,13 @@ class Qwen3VLForConditionalGeneration(nn.Module):
     def get_input_embeddings(self):
         return self.model.embed_tokens
 
+    _lora_pattern = re.compile(
+        r"^model\.layers\.(\d+)\.(?:self_attn|mlp)\.(?:qkv_proj|o_proj|down_proj|gate_up_proj)$"
+    )
+
+    def should_apply_lora(self, module_name: str) -> bool:
+        return bool(self._lora_pattern.match(module_name))
+
     def forward(
         self,
         input_ids: torch.Tensor,

python/sglang/srt/models/qwen3_vl_moe.py

@@ -14,6 +14,7 @@
 # ==============================================================================
 """Inference-only Qwen3-VL model compatible with HuggingFace weights."""
 import logging
+import re
 from functools import lru_cache
 from typing import Iterable, Optional, Tuple, Union
 
@@ -167,6 +168,14 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
     ):
         super().__init__(config, quant_config, prefix, language_model_cls)
 
+    # Only allow LoRA on attention projections within text layers for MoE.
+    _lora_pattern_moe = re.compile(
+        r"^model\.layers\.(\d+)\.self_attn\.(?:qkv_proj|o_proj)$"
+    )
+
+    def should_apply_lora(self, module_name: str) -> bool:
+        return bool(self._lora_pattern_moe.match(module_name))
+
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)