[Bugfix] Fix gemma3 with transformers backend (#23178)

Signed-off-by: raushan <raushan@huggingface.co>
Signed-off-by: Raushan Turganbay <raushan@huggingface.co>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>

tests/models/multimodal/generation/test_common.py

@@ -193,6 +193,20 @@ VLM_TEST_SETTINGS = {
         # when processing the 3rd prompt in vLLM
         marks=[pytest.mark.core_model, pytest.mark.skip(reason="Test hangs")],
     ),
+    # Gemma3 has bidirectional mask on images
+    "gemma3-transformers": VLMTestInfo(
+        models=["google/gemma-3-4b-it"],
+        test_type=VLMTestType.IMAGE,
+        prompt_formatter=lambda vid_prompt: f"<'<bos><start_of_turn>user\n{vid_prompt}<start_of_image><end_of_turn>\n<start_of_turn>model\n",  # noqa: E501
+        max_model_len=4096,
+        auto_cls=AutoModelForImageTextToText,
+        vllm_output_post_proc=model_utils.gemma3_vllm_to_hf_output,
+        image_size_factors=[(0.25, 0.5, 1.0)],
+        vllm_runner_kwargs={
+            "model_impl": "transformers",
+        },
+        marks=[pytest.mark.core_model],
+    ),
     "idefics3-transformers": VLMTestInfo(
         models=["HuggingFaceTB/SmolVLM-256M-Instruct"],
         test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),

tests/models/multimodal/generation/vlm_utils/model_utils.py

@@ -342,6 +342,29 @@ def gemma3_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     return hf_model
 
 
+def gemma3_vllm_to_hf_output(vllm_output: RunnerOutput, model: str) -> RunnerOutput:
+    """Sanitize vllm output [gemma-3] to compare with hf output."""
+    output_ids, output_str, out_logprobs = vllm_output
+
+    config = AutoConfig.from_pretrained(model)
+    image_token_id = config.image_token_id
+
+    tokenizer = AutoTokenizer.from_pretrained(model)
+    eos_token_id = tokenizer.eos_token_id
+
+    hf_output_ids = [
+        token_id
+        for idx, token_id in enumerate(output_ids)
+        if token_id != image_token_id
+    ]
+
+    hf_output_str = output_str
+    if hf_output_ids[-1] == eos_token_id:
+        hf_output_str = hf_output_str + tokenizer.decode(eos_token_id)
+
+    return hf_output_ids, hf_output_str, out_logprobs
+
+
 def glm4v_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
     """Patches and returns an instance of the HfRunner to use for GLM4V."""
     hf_processor = hf_model.processor

vllm/model_executor/models/transformers.py

@@ -68,13 +68,7 @@ from vllm.multimodal.processing import BaseMultiModalProcessor, BaseProcessingIn
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 
-from .interfaces import (
+from .interfaces import SupportsLoRA, SupportsMultiModal, SupportsPP, SupportsQuant
-    MultiModalEmbeddings,
-    SupportsLoRA,
-    SupportsMultiModal,
-    SupportsPP,
-    SupportsQuant,
-)
 from .utils import (
     AutoWeightsLoader,
     PPMissingLayer,
@@ -534,7 +528,10 @@ class TransformersBase(nn.Module, SupportsQuant, SupportsLoRA, SupportsPP):
         self.attention_instances = self.create_attention_instances()
 
         # Input embeddings
-        if not isinstance(self.model.get_input_embeddings(), PPMissingLayer):
+        input_embeddings = self.model.get_input_embeddings()
+        if not isinstance(input_embeddings, PPMissingLayer):
+            # Some models use embedding scales
+            self.embed_scale = getattr(input_embeddings, "embed_scale", None)
             names = ("embedding_size", "hidden_size")
             embedding_dim = getattr_iter(self.text_config, names, None)
             assert embedding_dim is not None
@@ -671,6 +668,7 @@ class TransformersBase(nn.Module, SupportsQuant, SupportsLoRA, SupportsPP):
         num_heads = self.model_config.get_num_attention_heads(self.parallel_config)
         head_size = self.model_config.get_head_size()
         num_kv_heads = self.model_config.get_num_kv_heads(self.parallel_config)
+        logits_soft_cap = getattr(self.text_config, "attn_logit_softcapping", None)
         start, end = get_pp_indices(
             self.text_config.num_hidden_layers,
             self.pp_group.rank_in_group,
@@ -696,6 +694,7 @@ class TransformersBase(nn.Module, SupportsQuant, SupportsLoRA, SupportsPP):
                 num_kv_heads=num_kv_heads,
                 cache_config=self.cache_config,
                 quant_config=self.quant_config,
+                logits_soft_cap=logits_soft_cap,
                 per_layer_sliding_window=per_layer_sliding_window,
                 prefix=f"{i}.attn",
                 attn_type=attn_type,
@@ -735,6 +734,7 @@ class TransformersBase(nn.Module, SupportsQuant, SupportsLoRA, SupportsPP):
         positions: torch.Tensor,
         intermediate_tensors: Optional[IntermediateTensors] = None,
         inputs_embeds: Optional[torch.Tensor] = None,
+        **kwargs,
     ) -> Union[torch.Tensor, IntermediateTensors]:
         if not self.pp_group.is_first_rank:
             assert intermediate_tensors is not None
@@ -758,6 +758,7 @@ class TransformersBase(nn.Module, SupportsQuant, SupportsLoRA, SupportsPP):
             position_ids=position_ids,
             attention_instances=self.attention_instances,
             return_dict=False,
+            **kwargs,
         )[0][0, ...]  # we remove batch dimension for now
 
         if not self.pp_group.is_last_rank:
@@ -819,7 +820,10 @@ class TransformersForCausalLM(TransformersBase):
             self.lm_head = PPMissingLayer()
 
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
-        return self.model.get_input_embeddings()(input_ids)
+        inputs_embeds = self.model.get_input_embeddings()(input_ids)
+        if self.embed_scale is not None:
+            inputs_embeds *= self.embed_scale
+        return inputs_embeds
 
     def compute_logits(
         self,
@@ -845,6 +849,7 @@ class TransformersForCausalLM(TransformersBase):
     enable_if=can_enable_torch_compile,
 )
 class TransformersForMultimodalLM(TransformersForCausalLM, SupportsMultiModal):
+    supports_multimodal_raw_input_only = True
     merge_by_field_config = True
     # Backwards compatibility for prev released models. State dicts back then
     # had different formats and cannot be loaded with `AutoModel` mapping as is
@@ -883,13 +888,27 @@ class TransformersForMultimodalLM(TransformersForCausalLM, SupportsMultiModal):
         inputs_embeds: Optional[torch.Tensor] = None,
         **kwargs: object,
     ) -> Union[torch.Tensor, IntermediateTensors]:
+        # Gemma3 and PaliGemma needs `token_type_ids` to work correctly
+        # Other models will not have `token_type_ids` in kwargs
+        kwargs = {k: v for k, v in kwargs.items() if k == "token_type_ids"}
         model_output = super().forward(
-            input_ids, positions, intermediate_tensors, inputs_embeds
+            input_ids, positions, intermediate_tensors, inputs_embeds, **kwargs
         )
         return model_output
 
     def get_language_model(self) -> torch.nn.Module:
-        return self.model
+        """`TransformersForMultimodalLM` does not contain a vLLM language model class.
+        Therefore, in order to return a language model vLLM class, we use a wrapper to
+        give `self` the same interface as `TransformersForCausalLM`."""
+
+        class LanguageModelWrapper(TransformersForCausalLM):
+            def __init__(self, multimodal_model):
+                # Don't call super().__init__() to avoid re-initialization
+                self.__dict__.update(multimodal_model.__dict__)
+
+            model = getattr_iter(self.model, ("language_model", "text_model"), None)
+
+        return LanguageModelWrapper(self)
 
     def get_multimodal_embeddings(self, **kwargs):
         pixel_values: Optional[torch.Tensor] = kwargs.pop("pixel_values", None)
@@ -905,6 +924,7 @@ class TransformersForMultimodalLM(TransformersForCausalLM, SupportsMultiModal):
             return None
 
         num_image_patches = kwargs.pop("num_image_patches")
+        kwargs.pop("token_type_ids", None)  # used only in `forward`
         if pixel_values is not None:
             vision_embeddings = self.model.get_image_features(pixel_values, **kwargs)
 
@@ -925,46 +945,4 @@ class TransformersForMultimodalLM(TransformersForCausalLM, SupportsMultiModal):
 
             return vision_embeddings
 
-    def get_input_embeddings(
+    get_input_embeddings = SupportsMultiModal.get_input_embeddings
-        self,
-        input_ids: torch.Tensor,
-        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
-        *,
-        is_multimodal: Optional[torch.Tensor] = None,
-        handle_oov_mm_token: bool = False,
-    ) -> torch.Tensor:
-        """
-        Apply token embeddings to `input_ids`.
-
-        If `multimodal_embeddings` is passed, scatter them into
-        `input_ids` according to the mask `is_multimodal`.
-
-        In case the multi-modal token IDs exceed the vocabulary size of
-        the language model, you can set `handle_oov_mm_token=False`
-        to avoid calling the language model's `get_input_embeddings` method
-        on those tokens.
-        """
-        from .utils import _merge_multimodal_embeddings
-
-        inputs_embeds = self._get_text_embeddings(
-            input_ids,
-            self.model.get_input_embeddings(),
-            is_multimodal=is_multimodal,
-            handle_oov_mm_token=handle_oov_mm_token,
-        )
-
-        if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
-            return inputs_embeds
-
-        if is_multimodal is None:
-            raise ValueError(
-                "`get_input_embeddings` now requires `is_multimodal` arg, "
-                "please update your model runner according to "
-                "https://github.com/vllm-project/vllm/pull/16229."
-            )
-
-        return _merge_multimodal_embeddings(
-            inputs_embeds=inputs_embeds,
-            multimodal_embeddings=multimodal_embeddings,
-            is_multimodal=is_multimodal,
-        )

vllm/model_executor/models/transformers_moe.py

@@ -30,7 +30,7 @@ from vllm.model_executor.layers.fused_moe import FusedMoE
 from vllm.platforms import current_platform
 from vllm.utils import direct_register_custom_op
 
-from .interfaces import MixtureOfExperts
+from .interfaces import MixtureOfExperts, SupportsMultiModal
 from .transformers import (
     TransformersBase,
     TransformersForCausalLM,
@@ -335,7 +335,5 @@ class TransformersMoEForCausalLM(TransformersMoEBase, TransformersForCausalLM):
     },
     enable_if=can_enable_torch_compile,
 )
-class TransformersMoEForMultimodalLM(
+class TransformersMoEForMultimodalLM(TransformersMoEBase, TransformersForMultimodalLM):
-    TransformersMoEForCausalLM, TransformersForMultimodalLM
+    get_input_embeddings = SupportsMultiModal.get_input_embeddings
-):
-    pass