[Model] Molmo vLLM Integration (#9016)

Co-authored-by: sanghol <sanghol@allenai.org>
Co-authored-by: Roger Wang <136131678+ywang96@users.noreply.github.com>
Co-authored-by: Roger Wang <ywang@roblox.com>

docs/source/models/supported_models.rst

@@ -399,6 +399,12 @@ Text Generation
     - :code:`meta-llama/Llama-3.2-90B-Vision-Instruct`, :code:`meta-llama/Llama-3.2-11B-Vision`, etc.
     -
     -
+  * - :code:`MolmoForCausalLM`
+    - Molmo
+    - Image
+    - :code:`allenai/Molmo-7B-D-0924`, :code:`allenai/Molmo-72B-0924`, etc.
+    -
+    - ✅︎
   * - :code:`NVLM_D_Model`
     - NVLM-D 1.0
     - Image\ :sup:`E+`

examples/offline_inference_vision_language.py

@@ -300,6 +300,23 @@ def run_mllama(question: str, modality: str):
     return llm, prompt, stop_token_ids
 
 
+# Molmo
+def run_molmo(question, modality):
+    assert modality == "image"
+
+    model_name = "allenai/Molmo-7B-D-0924"
+
+    llm = LLM(
+        model=model_name,
+        trust_remote_code=True,
+        dtype="bfloat16",
+    )
+
+    prompt = question
+    stop_token_ids = None
+    return llm, prompt, stop_token_ids
+
+
 # GLM-4v
 def run_glm4v(question: str, modality: str):
     assert modality == "image"
@@ -331,6 +348,7 @@ model_example_map = {
     "qwen_vl": run_qwen_vl,
     "qwen2_vl": run_qwen2_vl,
     "mllama": run_mllama,
+    "molmo": run_molmo,
     "glm4v": run_glm4v,
 }
 

vllm/entrypoints/chat_utils.py

@@ -163,6 +163,8 @@ class BaseMultiModalItemTracker(ABC, Generic[_T]):
                 return "<|image|>"
             if model_type == "qwen2_vl":
                 return "<|vision_start|><|image_pad|><|vision_end|>"
+            if model_type == "molmo":
+                return ""
 
             raise TypeError(f"Unknown model type: {model_type}")
         elif modality == "audio":

vllm/model_executor/models/__init__.py

@@ -20,4 +20,4 @@ __all__ = [
     "supports_multimodal",
     "SupportsPP",
     "supports_pp",
-]
+]

vllm/model_executor/models/qwen2_vl.py

@@ -1167,8 +1167,7 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
                         continue
                     param = params_dict[name]
                 except KeyError:
-                    print(params_dict.keys())
+                    raise ValueError(f"Unexpected weight: {name}") from None
-                    raise
 
                 weight_loader = getattr(param, "weight_loader",
                                         default_weight_loader)

vllm/model_executor/models/registry.py

@@ -104,6 +104,7 @@ _MULTIMODAL_MODELS = {
     "LlavaNextVideoForConditionalGeneration": ("llava_next_video", "LlavaNextVideoForConditionalGeneration"),  # noqa: E501
     "LlavaOnevisionForConditionalGeneration": ("llava_onevision", "LlavaOnevisionForConditionalGeneration"),  # noqa: E501
     "MiniCPMV": ("minicpmv", "MiniCPMV"),
+    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
     "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
     "PaliGemmaForConditionalGeneration": ("paligemma", "PaliGemmaForConditionalGeneration"),  # noqa: E501
     "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),