[New Model] Support Command-A-Vision (#22660)

Signed-off-by: donglu <[email protected]>

Author: dongluw

docs/models/supported_models.md

@@ -331,7 +331,7 @@ th {
 | `BloomForCausalLM` | BLOOM, BLOOMZ, BLOOMChat | `bigscience/bloom`, `bigscience/bloomz`, etc. | | ✅︎ | |
 | `BartForConditionalGeneration` | BART | `facebook/bart-base`, `facebook/bart-large-cnn`, etc. | | | |
 | `ChatGLMModel`, `ChatGLMForConditionalGeneration` | ChatGLM | `zai-org/chatglm2-6b`, `zai-org/chatglm3-6b`, `ShieldLM-6B-chatglm3`, etc. | ✅︎ | ✅︎ | ✅︎ |
-| `CohereForCausalLM`, `Cohere2ForCausalLM` | Command-R | `CohereForAI/c4ai-command-r-v01`, `CohereForAI/c4ai-command-r7b-12-2024`, etc. | ✅︎ | ✅︎ | ✅︎ |
+| `CohereForCausalLM`, `Cohere2ForCausalLM` | Command-R | `CohereLabs/c4ai-command-r-v01`, `CohereLabs/c4ai-command-r7b-12-2024`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `DbrxForCausalLM` | DBRX | `databricks/dbrx-base`, `databricks/dbrx-instruct`, etc. | | ✅︎ | ✅︎ |
 | `DeciLMForCausalLM` | DeciLM | `nvidia/Llama-3_3-Nemotron-Super-49B-v1`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `DeepseekForCausalLM` | DeepSeek | `deepseek-ai/deepseek-llm-67b-base`, `deepseek-ai/deepseek-llm-7b-chat`, etc. | | ✅︎ | ✅︎ |
@@ -601,6 +601,7 @@ These models primarily accept the [`LLM.generate`](./generative_models.md#llmgen
 | `AyaVisionForConditionalGeneration` | Aya Vision | T + I<sup>+</sup> | `CohereForAI/aya-vision-8b`, `CohereForAI/aya-vision-32b`, etc. | | ✅︎ | ✅︎ |
 | `Blip2ForConditionalGeneration` | BLIP-2 | T + I<sup>E</sup> | `Salesforce/blip2-opt-2.7b`, `Salesforce/blip2-opt-6.7b`, etc. | | ✅︎ | ✅︎ |
 | `ChameleonForConditionalGeneration` | Chameleon | T + I | `facebook/chameleon-7b`, etc. | | ✅︎ | ✅︎ |
+| `Cohere2VisionForConditionalGeneration` | Command A Vision | T + I<sup>+</sup> | `CohereLabs/command-a-vision-07-2025`, etc. | | ✅︎ | ✅︎ |
 | `DeepseekVLV2ForCausalLM`<sup>^</sup> | DeepSeek-VL2 | T + I<sup>+</sup> | `deepseek-ai/deepseek-vl2-tiny`, `deepseek-ai/deepseek-vl2-small`, `deepseek-ai/deepseek-vl2`, etc. | | ✅︎ | ✅︎ |
 | `Florence2ForConditionalGeneration` | Florence-2 | T + I | `microsoft/Florence-2-base`, `microsoft/Florence-2-large`, etc. | | | |
 | `FuyuForCausalLM` | Fuyu | T + I | `adept/fuyu-8b`, etc. | | ✅︎ | ✅︎ |

examples/offline_inference/vision_language.py

@@ -126,6 +126,29 @@ def run_chameleon(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+def run_command_a_vision(questions: list[str], modality: str) -> ModelRequestData:
+    assert modality == "image"
+
+    model_name = "CohereLabs/command-a-vision-07-2025"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=32768,
+        tensor_parallel_size=4,
+        limit_mm_per_prompt={modality: 1},
+    )
+
+    prompts = [
+        f"<|START_OF_TURN_TOKEN|><|USER_TOKEN|><|IMG_PATCH|>{question}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
 # Deepseek-VL2
 def run_deepseek_vl2(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -1417,6 +1440,7 @@ def run_tarsier2(questions: list[str], modality: str) -> ModelRequestData:
     "aya_vision": run_aya_vision,
     "blip-2": run_blip2,
     "chameleon": run_chameleon,
+    "command_a_vision": run_command_a_vision,
     "deepseek_vl_v2": run_deepseek_vl2,
     "florence2": run_florence2,
     "fuyu": run_fuyu,

examples/offline_inference/vision_language_multi_image.py

@@ -107,6 +107,42 @@ def load_aya_vision(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+def load_command_a_vision(question: str, image_urls: list[str]) -> ModelRequestData:
+    model_name = "CohereLabs/command-a-vision-07-2025"
+
+    # NOTE: This model is 122B parameters and requires tensor parallelism
+    # Recommended to use tp=4 on H100 GPUs
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=32768,
+        tensor_parallel_size=4,
+        limit_mm_per_prompt={"image": len(image_urls)},
+    )
+
+    placeholders = [{"type": "image", "image": url} for url in image_urls]
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                *placeholders,
+                {"type": "text", "text": question},
+            ],
+        }
+    ]
+
+    processor = AutoProcessor.from_pretrained(model_name)
+
+    prompt = processor.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=[fetch_image(url) for url in image_urls],
+    )
+
+
 def load_deepseek_vl2(question: str, image_urls: list[str]) -> ModelRequestData:
     model_name = "deepseek-ai/deepseek-vl2-tiny"
 
@@ -1031,6 +1067,7 @@ def load_tarsier2(question: str, image_urls: list[str]) -> ModelRequestData:
 model_example_map = {
     "aria": load_aria,
     "aya_vision": load_aya_vision,
+    "command_a_vision": load_command_a_vision,
     "deepseek_vl_v2": load_deepseek_vl2,
     "gemma3": load_gemma3,
     "h2ovl_chat": load_h2ovl,

tests/models/registry.py

@@ -383,6 +383,7 @@ def check_available_online(
     "Blip2ForConditionalGeneration": _HfExamplesInfo("Salesforce/blip2-opt-2.7b",  # noqa: E501
                                                      extras={"6b": "Salesforce/blip2-opt-6.7b"}),  # noqa: E501
     "ChameleonForConditionalGeneration": _HfExamplesInfo("facebook/chameleon-7b"),  # noqa: E501
+    "Cohere2VisionForConditionalGeneration": _HfExamplesInfo("CohereLabs/command-a-vision-07-2025"), # noqa: E501
     "DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-tiny",  # noqa: E501
                                                 extras={"fork": "Isotr0py/deepseek-vl2-tiny"},  # noqa: E501
                                                 max_transformers_version="4.48",  # noqa: E501