[Model] Gemma3n MM (#20495)

Signed-off-by: ShriKode <shrikode@gmail.com>
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Roger Wang <hey@rogerw.me>
Co-authored-by: ShriKode <shrikode@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.me>

docs/models/supported_models.md

@@ -349,7 +349,7 @@ th {
 | `GemmaForCausalLM` | Gemma | `google/gemma-2b`, `google/gemma-1.1-2b-it`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Gemma2ForCausalLM` | Gemma 2 | `google/gemma-2-9b`, `google/gemma-2-27b`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Gemma3ForCausalLM` | Gemma 3 | `google/gemma-3-1b-it`, etc. | ✅︎ | ✅︎ | ✅︎ |
-| `Gemma3nForConditionalGeneration` | Gemma 3n | `google/gemma-3n-E2B-it`, `google/gemma-3n-E4B-it`, etc. | | | ✅︎ |
+| `Gemma3nForCausalLM` | Gemma 3n | `google/gemma-3n-E2B-it`, `google/gemma-3n-E4B-it`, etc. | | | ✅︎ |
 | `GlmForCausalLM` | GLM-4 | `zai-org/glm-4-9b-chat-hf`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Glm4ForCausalLM` | GLM-4-0414 | `zai-org/GLM-4-32B-0414`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Glm4MoeForCausalLM` | GLM-4.5 | `zai-org/GLM-4.5`, etc. | ✅︎ | ✅︎ | ✅︎ |
@@ -412,9 +412,6 @@ th {
 !!! note
     Currently, the ROCm version of vLLM supports Mistral and Mixtral only for context lengths up to 4096.
 
-!!! note
-    Only text inputs are currently supported for `Gemma3nForConditionalGeneration`. To use this model, please upgrade Hugging Face Transformers to version 4.53.0.
-
 ### Pooling Models
 
 See [this page](./pooling_models.md) for more information on how to use pooling models.
@@ -608,6 +605,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `Florence2ForConditionalGeneration` | Florence-2 | T + I | `microsoft/Florence-2-base`, `microsoft/Florence-2-large`, etc. | | | |
 | `FuyuForCausalLM` | Fuyu | T + I | `adept/fuyu-8b`, etc. | | ✅︎ | ✅︎ |
 | `Gemma3ForConditionalGeneration` | Gemma 3 | T + I<sup>+</sup> | `google/gemma-3-4b-it`, `google/gemma-3-27b-it`, etc. | ✅︎ | ✅︎ | ⚠️ |
+| `Gemma3nForConditionalGeneration` | Gemma 3n | T + I + A | `google/gemma-3n-E2B-it`, `google/gemma-3n-E4B-it`, etc. | | | ✅︎ |
 | `GLM4VForCausalLM`<sup>^</sup> | GLM-4V | T + I | `zai-org/glm-4v-9b`, `zai-org/cogagent-9b-20241220`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Glm4vForConditionalGeneration` | GLM-4.1V-Thinking | T + I<sup>E+</sup> + V<sup>E+</sup> | `zai-org/GLM-4.1V-9B-Thinking`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Glm4vMoeForConditionalGeneration` | GLM-4.5V | T + I<sup>E+</sup> + V<sup>E+</sup> | `zai-org/GLM-4.5V`, etc. | ✅︎ | ✅︎ | ✅︎ |
@@ -677,6 +675,15 @@ Some models are supported only via the [Transformers backend](#transformers). Th
 
     This limitation exists because the model's mixed attention pattern (bidirectional for images, causal otherwise) is not yet supported by vLLM's attention backends.
 
+!!! note
+    `Gemma3nForConditionalGeneration` is only supported on V1 due to shared KV caching and it depends on `timm>=1.0.17` to make use of its
+    MobileNet-v5 vision backbone.
+  
+    Performance is not yet fully optimized mainly due to:
+  
+    - Both audio and vision MM encoders use `transformers.AutoModel` implementation.  
+    - There's no PLE caching or out-of-memory swapping support, as described in [Google's blog](https://developers.googleblog.com/en/introducing-gemma-3n/). These features might be too model-specific for vLLM, and swapping in particular may be better suited for constrained setups.
+
 !!! note
     Only `InternVLChatModel` with Qwen2.5 text backbone (`OpenGVLab/InternVL3-2B`, `OpenGVLab/InternVL2.5-1B` etc) has video inputs support currently.
 

examples/offline_inference/audio_language.py

@@ -96,6 +96,25 @@ def run_voxtral(question: str, audio_count: int) -> ModelRequestData:
     )
 
 
+# Gemma3N
+def run_gemma3n(question: str, audio_count: int) -> ModelRequestData:
+    model_name = "google/gemma-3n-E2B-it"
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=2048,
+        max_num_batched_tokens=2048,
+        max_num_seqs=2,
+        limit_mm_per_prompt={"audio": audio_count},
+        enforce_eager=True,
+    )
+    prompt = f"<start_of_turn>user\n<audio_soft_token>{question}"
+    "<end_of_turn>\n<start_of_turn>model\n"
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+    )
+
+
 # Granite Speech
 def run_granite_speech(question: str, audio_count: int) -> ModelRequestData:
     # NOTE - the setting in this example are somehat different than what is
@@ -331,6 +350,7 @@ def run_whisper(question: str, audio_count: int) -> ModelRequestData:
 
 model_example_map = {
     "voxtral": run_voxtral,
+    "gemma3n": run_gemma3n,
     "granite_speech": run_granite_speech,
     "minicpmo": run_minicpmo,
     "phi4_mm": run_phi4mm,

examples/offline_inference/vision_language.py

@@ -211,7 +211,33 @@ def run_gemma3(questions: list[str], modality: str) -> ModelRequestData:
         )
         for question in questions
     ]
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
 
+# Gemma3N
+def run_gemma3n(questions: list[str], modality: str) -> ModelRequestData:
+    assert modality == "image"
+    model_name = "google/gemma-3n-E2B-it"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=2048,
+        max_num_seqs=2,
+        limit_mm_per_prompt={modality: 1},
+        enforce_eager=True,
+    )
+
+    prompts = [
+        (
+            "<start_of_turn>user\n"
+            f"<image_soft_token>{question}<end_of_turn>\n"
+            "<start_of_turn>model\n"
+        )
+        for question in questions
+    ]
     return ModelRequestData(
         engine_args=engine_args,
         prompts=prompts,
@@ -1395,6 +1421,7 @@ model_example_map = {
     "florence2": run_florence2,
     "fuyu": run_fuyu,
     "gemma3": run_gemma3,
+    "gemma3n": run_gemma3n,
     "glm4v": run_glm4v,
     "glm4_1v": run_glm4_1v,
     "h2ovl_chat": run_h2ovl,

requirements/test.in

@@ -21,7 +21,7 @@ ray[cgraph,default]>=2.48.0 # Ray Compiled Graph, required by pipeline paralleli
 sentence-transformers # required for embedding tests
 soundfile # required for audio tests
 jiwer # required for audio tests
-timm # required for internvl test
+timm >=1.0.17 # required for internvl and gemma3n-mm test
 torch==2.7.1
 torchaudio==2.7.1
 torchvision==0.22.1

requirements/test.txt

@@ -1051,7 +1051,7 @@ tiktoken==0.7.0
     # via
     #   lm-eval
     #   mistral-common
-timm==1.0.15
+timm==1.0.17
     # via
     #   -r requirements/test.in
     #   open-clip-torch

tests/models/multimodal/processing/test_common.py

@@ -271,6 +271,7 @@ def _test_processing_correctness_one(
     "microsoft/Florence-2-base",
     "adept/fuyu-8b",
     "google/gemma-3-4b-it",
+    "google/gemma-3n-E2B-it",
     "zai-org/glm-4v-9b",
     "zai-org/GLM-4.1V-9B-Thinking",
     "ibm-granite/granite-speech-3.3-2b",
@@ -315,7 +316,7 @@ def _test_processing_correctness_one(
     "fixie-ai/ultravox-v0_5-llama-3_2-1b",
     "openai/whisper-large-v3",
     "omni-research/Tarsier-7b",
-    "omni-research/Tarsier2-Recap-7b"
+    "omni-research/Tarsier2-Recap-7b",
 ])
 @pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
 @pytest.mark.parametrize("num_batches", [32])
@@ -327,6 +328,8 @@ def test_processing_correctness(
     num_batches: int,
     simplify_rate: float,
 ):
+    if model_id == "google/gemma-3n-E2B-it":
+        pytest.skip("Skipping gemma-3n-E2B-it due to transformers #39911 bug.")
     _test_processing_correctness(
         model_id,
         hit_rate=hit_rate,

tests/models/registry.py

@@ -186,7 +186,7 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
     "GemmaForCausalLM": _HfExamplesInfo("google/gemma-1.1-2b-it"),
     "Gemma2ForCausalLM": _HfExamplesInfo("google/gemma-2-9b"),
     "Gemma3ForCausalLM": _HfExamplesInfo("google/gemma-3-1b-it"),
-    "Gemma3nForConditionalGeneration": _HfExamplesInfo("google/gemma-3n-E2B-it",    # noqa: E501
+    "Gemma3nForCausalLM": _HfExamplesInfo("google/gemma-3n-E2B-it",
                                           min_transformers_version="4.53"),
     "GlmForCausalLM": _HfExamplesInfo("zai-org/glm-4-9b-chat-hf"),
     "Glm4ForCausalLM": _HfExamplesInfo("zai-org/GLM-4-9B-0414"),
@@ -391,6 +391,8 @@ _MULTIMODAL_EXAMPLE_MODELS = {
     "Emu3ForConditionalGeneration": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
     "FuyuForCausalLM": _HfExamplesInfo("adept/fuyu-8b"),
     "Gemma3ForConditionalGeneration": _HfExamplesInfo("google/gemma-3-4b-it"),
+    "Gemma3nForConditionalGeneration": _HfExamplesInfo("google/gemma-3n-E2B-it",    # noqa: E501
+                                        min_transformers_version="4.53"),
     "GraniteSpeechForConditionalGeneration": _HfExamplesInfo("ibm-granite/granite-speech-3.3-2b"),  # noqa: E501
     "GLM4VForCausalLM": _HfExamplesInfo("zai-org/glm-4v-9b",
                                         trust_remote_code=True,

tests/test_test.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import pytest
+
+from vllm import LLM, envs
+from vllm.sampling_params import SamplingParams
+
+if not envs.VLLM_USE_V1:
+    pytest.skip(
+        "Skipping V1 tests. Rerun with `VLLM_USE_V1=1` to test.",
+        allow_module_level=True,
+    )
+
+
+@pytest.mark.parametrize("model_name", ["Qwen/Qwen2.5-1.5B-Instruct"])
+# TODO TPU will appear busy if we fan-out test params here
+@pytest.mark.parametrize("n_prompts", [1])
+def test_logprobs(model_name: str, n_prompts: int):
+    """
+    Request top logprobs with different sampling settings and check
+    that results contains the requested number, ordered ascendingly.  
+    """
+
+    def check_num_logprobs(logprobs, expected_num: int):
+        for step in logprobs:
+            prev_logp = 1.0
+            # order by rank
+            sorted_step = dict(
+                sorted(step.items(), key=lambda item: item[1].rank))
+
+            if len(step) != expected_num:
+                print("watch out", sorted_step)
+
+            # check results are ordered by prob value
+            # assert len(step) == expected_num
+            for rankno, (tid, logp) in enumerate(sorted_step.items()):
+                assert logp.logprob <= prev_logp
+                prev_logp = logp.logprob
+                assert logp.rank == rankno + 1
+
+    llm = LLM(model_name,
+              enforce_eager=False,
+              max_num_seqs=1,
+              max_model_len=128,
+              max_num_batched_tokens=128)
+    prompts = [
+        "Write a short story about a robot that dreams for the first time."
+    ] * n_prompts
+    greedy_sampling_params = SamplingParams(temperature=0.0, max_tokens=64,\
+         logprobs=4)
+    regular_sampling_params = SamplingParams(temperature=0.4, max_tokens=64,\
+         logprobs=4)
+    topkp_sampling_params = SamplingParams(temperature=0.4, max_tokens=64,\
+         logprobs=4, top_k=12, top_p=0.5)
+
+    for sp in [greedy_sampling_params, regular_sampling_params, \
+               topkp_sampling_params]:
+        output = llm.generate(prompts, sp)
+        for o in output:
+            check_num_logprobs(o.outputs[0].logprobs, 4)

vllm/model_executor/models/gemma3n.py

@@ -331,14 +331,15 @@ class Gemma3nAttention(nn.Module):
                                      config.num_kv_shared_layers)
         self.is_kv_shared = layer_idx >= first_kv_shared_layer_idx
 
+        kv_sharing_target_layer_name = None
         if self.is_kv_shared:
             # Last full attention layer is 1 before sharing
             # Last sliding attention layer is 2 before sharing
             offset = 2 if self.sliding_window is not None else 1
             kv_shared_layer_index = first_kv_shared_layer_idx - offset
-            kv_sharing_target_layer_name = f"model.language_model.layers.{kv_shared_layer_index}.self_attn.attn"  # noqa: E501
-        else:
-            kv_sharing_target_layer_name = None
+            if kv_shared_layer_index >= 0:
+                # Only the greater layer is required to specify sharing.
+                kv_sharing_target_layer_name = f"language_model.model.layers.{kv_shared_layer_index}.self_attn.attn"  # noqa: E501
 
         self.rotary_emb = get_rope(
             self.head_dim,
@@ -396,6 +397,7 @@ class Gemma3nDecoderLayer(nn.Module):
         prefix: str = "",
     ) -> None:
         super().__init__()
+        assert isinstance(config, Gemma3nTextConfig)
         self.altup_active_idx = config.altup_active_idx
         assert config.altup_correct_scale
 
@@ -537,7 +539,7 @@ class Gemma3nTextModel(nn.Module, SupportsQuant):
 
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
-        config = vllm_config.model_config.hf_config.text_config
+        config = vllm_config.model_config.hf_config
         cache_config = vllm_config.cache_config
         quant_config = vllm_config.quant_config
         self.config = config
@@ -553,6 +555,7 @@ class Gemma3nTextModel(nn.Module, SupportsQuant):
             config.hidden_size**0.5,
             dtype=self.embed_tokens.weight.dtype,
         )
+        # Additional per-layer embeddings (PLE)
         self.embed_tokens_per_layer = VocabParallelEmbedding(
             config.vocab_size_per_layer_input,
             config.num_hidden_layers * config.hidden_size_per_layer_input,
@@ -636,6 +639,8 @@ class Gemma3nTextModel(nn.Module, SupportsQuant):
         self,
         input_ids: Optional[torch.Tensor],
         positions: torch.Tensor,
+        per_layer_inputs: torch.Tensor,
+        intermediate_tensors: Optional[IntermediateTensors] = None,
         inputs_embeds: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Union[torch.Tensor, IntermediateTensors]:
@@ -644,13 +649,6 @@ class Gemma3nTextModel(nn.Module, SupportsQuant):
         else:
             hidden_states_0 = self.get_input_embeddings(input_ids)
 
-        # Per layer inputs.
-        if input_ids is None:
-            raise ValueError("Passing None for input ids is not supported.")
-        per_layer_inputs = self.get_per_layer_input_embeddings(input_ids)
-        per_layer_inputs = per_layer_inputs.reshape(
-            -1, self.config.num_hidden_layers,
-            self.config.hidden_size_per_layer_input)
         per_layer_projection = self.per_layer_model_projection(hidden_states_0)
         per_layer_projection = per_layer_projection.reshape(
             *hidden_states_0.shape[:-1],
@@ -659,8 +657,13 @@ class Gemma3nTextModel(nn.Module, SupportsQuant):
         )
         per_layer_projection = self.per_layer_projection_norm(
             per_layer_projection)
-        per_layer_inputs = per_layer_projection + per_layer_inputs
-        per_layer_inputs *= self.per_layer_input_scale
+
+        if per_layer_inputs is not None:
+            # Profiling run does not compute per_layer_inputs
+            per_layer_inputs = per_layer_projection + per_layer_inputs
+            per_layer_inputs *= self.per_layer_input_scale
+        else:
+            per_layer_inputs = per_layer_projection
 
         # Altup embed.
         hidden_states = [hidden_states_0] * self.config.altup_num_inputs
@@ -760,29 +763,7 @@ class Gemma3nTextModel(nn.Module, SupportsQuant):
         return loaded_params
 
 
-class Gemma3nModel(nn.Module):
-
-    def __init__(self, vllm_config: VllmConfig, prefix: str = ""):
-        super().__init__()
-        self.language_model = Gemma3nTextModel(vllm_config=vllm_config,
-                                               prefix=maybe_prefix(
-                                                   prefix, "language_model"))
-
-    def forward(
-        self,
-        input_ids: Optional[torch.Tensor],
-        positions: torch.Tensor,
-        intermediate_tensors: Optional[IntermediateTensors] = None,
-        inputs_embeds: Optional[torch.Tensor] = None,
-        **kwargs,
-    ) -> torch.Tensor:
-        return self.language_model(input_ids=input_ids,
-                                   positions=positions,
-                                   inputs_embeds=inputs_embeds,
-                                   **kwargs)
-
-
-class Gemma3nForConditionalGeneration(nn.Module, SupportsQuant):
+class Gemma3nForCausalLM(nn.Module):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
@@ -802,25 +783,33 @@ class Gemma3nForConditionalGeneration(nn.Module, SupportsQuant):
         super().__init__()
         self.config = config
         self.cache_config = vllm_config.cache_config
-        self.model = Gemma3nModel(vllm_config=vllm_config,
-                                  prefix=maybe_prefix(prefix, "model"))
+        self.model = Gemma3nTextModel(vllm_config=vllm_config,
+                                      prefix=maybe_prefix(prefix, "model"))
         self.logits_processor = LogitsProcessor(
-            config.text_config.vocab_size,
-            soft_cap=config.text_config.final_logit_softcapping)
+            config.vocab_size, soft_cap=config.final_logit_softcapping)
 
     def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
-        return self.model.language_model.get_input_embeddings(input_ids)
+        return self.model.get_input_embeddings(input_ids)
 
     def forward(
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
+        *,
+        per_layer_inputs: Optional[torch.Tensor] = None,
         intermediate_tensors: Optional[IntermediateTensors] = None,
         inputs_embeds: Optional[torch.Tensor] = None,
         **kwargs,
     ) -> Union[torch.Tensor, IntermediateTensors]:
-        hidden_states = self.model(input_ids, positions, intermediate_tensors,
-                                   inputs_embeds, **kwargs)
+
+        hidden_states = self.model(
+            input_ids,
+            positions,
+            per_layer_inputs=per_layer_inputs,
+            intermediate_tensors=intermediate_tensors,
+            inputs_embeds=inputs_embeds,
+            **kwargs,
+        )
         return hidden_states
 
     def compute_logits(
@@ -828,8 +817,8 @@ class Gemma3nForConditionalGeneration(nn.Module, SupportsQuant):
         hidden_states: torch.Tensor,
         sampling_metadata: Optional[SamplingMetadata],
     ) -> Optional[torch.Tensor]:
-        logits = self.logits_processor(self.model.language_model.embed_tokens,
-                                       hidden_states, sampling_metadata)
+        logits = self.logits_processor(self.model.embed_tokens, hidden_states,
+                                       sampling_metadata)
         return logits
 
     def load_weights(self, weights: Iterable[tuple[str,

vllm/model_executor/models/registry.py

@@ -69,8 +69,7 @@ _TEXT_GENERATION_MODELS = {
     "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
     "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
     "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
-    #TODO(ywang96): Support multimodal gemma3n
-    "Gemma3nForConditionalGeneration": ("gemma3n", "Gemma3nForConditionalGeneration"),    # noqa: E501
+    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
     "GlmForCausalLM": ("glm", "GlmForCausalLM"),
     "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
     "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
@@ -205,6 +204,7 @@ _MULTIMODAL_MODELS = {
     "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
     "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
     "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
+    "Gemma3nForConditionalGeneration": ("gemma3n_mm", "Gemma3nForConditionalGeneration"),    # noqa: E501
     "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
     "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
     "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501