[diffusion] Fix native LingBot-Video text encoding (#36542)

python/sglang/multimodal_gen/runtime/models/encoders/qwen3vl.py

@@ -1221,6 +1221,12 @@ class Qwen3VLForConditionalGeneration(TextEncoder):
         self.lm_head = nn.Linear(
             config.text_config.hidden_size, config.text_config.vocab_size, bias=False
         )
+        if getattr(config, "tie_word_embeddings", False) or getattr(
+            config.text_config, "tie_word_embeddings", False
+        ):
+            # Tied checkpoints may omit lm_head.weight, as Hugging Face does.
+            # Keep one registered parameter so strict native loading stays safe.
+            self.lm_head.weight = self.model.get_input_embeddings().weight
 
     @torch.no_grad()
     def forward(

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/lingbot_video_moe/text_encoding.py

@@ -107,9 +107,12 @@ class LingBotVideoTextEncodingStage(TextEncodingStage):
 
         inputs = self._build_prompt_inputs(prompt)
         inputs = inputs.to(device)
-        outputs = text_encoder(
+        outputs = self._forward_text_encoder(
+            text_encoder,
+            {
                 **inputs,
-            output_hidden_states=self.hidden_state_skip_layer is not None,
+                "output_hidden_states": self.hidden_state_skip_layer is not None,
+            },
         )
         if self.hidden_state_skip_layer is not None:
             prompt_embeds = outputs.hidden_states[-(self.hidden_state_skip_layer + 1)]

python/sglang/multimodal_gen/test/unit/test_lingbot_video_moe.py

@@ -21,6 +21,9 @@ from sglang.multimodal_gen.runtime.layers.moe import (
     LingBotVideoGroupedExperts,
     LingBotVideoRouter,
 )
+from sglang.multimodal_gen.runtime.managers.forward_context import (
+    get_forward_context,
+)
 from sglang.multimodal_gen.runtime.models.dits import (
     lingbot_video_moe as dits_lingbot_video_moe,
 )
@@ -262,6 +265,31 @@ def test_text_encoding_crops_template_then_trims_padding():
     assert stage._compute_crop_start() == prefix_width
 
 
+def test_text_encoding_sets_forward_context_for_native_encoder():
+    prompt_width, prefix_width, true_len, channels = 6, 2, 5, 4
+    hidden = torch.zeros(1, prompt_width, channels)
+
+    class NativeEncoder:
+        uses_sglang_forward_context = True
+
+        def __call__(self, **kwargs):
+            context = get_forward_context()
+            assert context.current_timestep == 0
+            assert context.attn_metadata is None
+            assert kwargs["output_hidden_states"] is True
+            return SimpleNamespace(hidden_states=[hidden])
+
+    stage = _text_encoding_stage(
+        _FakeQwenProcessor(prompt_width, prefix_width, true_len), NativeEncoder()
+    )
+    embeds, mask = stage._encode_prompt(
+        "a structured caption", torch.device("cpu"), torch.float32
+    )
+
+    assert tuple(embeds.shape) == (1, true_len - prefix_width, channels)
+    assert tuple(mask.shape) == (1, true_len - prefix_width)
+
+
 def test_check_inputs_enforces_frame_and_size_contract():
     check = LingBotVideoTextEncodingStage.check_inputs
     check(480, 832, 1)

python/sglang/multimodal_gen/test/unit/test_qwen3vl_vision.py

@@ -170,6 +170,52 @@ def test_native_vision_keeps_position_math_in_fp32():
     assert block.position_embedding_dtypes == (torch.float32, torch.float32)
 
 
+def test_qwen3vl_ties_lm_head_to_input_embeddings():
+    vision_config = SimpleNamespace(
+        hidden_size=16,
+        intermediate_size=24,
+        hidden_act="gelu_pytorch_tanh",
+        num_heads=2,
+        depth=0,
+        patch_size=2,
+        temporal_patch_size=1,
+        in_channels=3,
+        num_position_embeddings=16,
+        spatial_merge_size=2,
+        out_hidden_size=16,
+        deepstack_visual_indexes=[],
+    )
+    text_config = SimpleNamespace(
+        hidden_size=16,
+        vocab_size=32,
+        pad_token_id=0,
+        num_hidden_layers=0,
+        rms_norm_eps=1e-6,
+        tie_word_embeddings=True,
+    )
+    arch_config = SimpleNamespace(
+        vision_config=vision_config,
+        text_config=text_config,
+        tie_word_embeddings=True,
+        _fsdp_shard_conditions=[],
+        stacked_params_mapping=[],
+    )
+    config = SimpleNamespace(arch_config=arch_config)
+
+    with get_parallel().override(tp_size=1, tp_rank=0):
+        model = Qwen3VLForConditionalGeneration(config)
+
+    assert model.lm_head.weight is model.model.get_input_embeddings().weight
+    parameters = dict(model.named_parameters())
+    parameters_with_duplicates = dict(model.named_parameters(remove_duplicate=False))
+    assert "model.language_model.embed_tokens.weight" in parameters
+    assert "lm_head.weight" not in parameters
+    assert (
+        parameters_with_duplicates["lm_head.weight"]
+        is parameters["model.language_model.embed_tokens.weight"]
+    )
+
+
 def test_qwen3_multimodal_encoders_layerwise_offload_vision_blocks():
     assert "model.visual.blocks" in Qwen3VLForConditionalGeneration.layer_names
     assert "model.visual.blocks" in MiniMaxH3Qwen3VLEncoder.layer_names