[VLM] Support HF format Phi-4-MM model

docs/models/supported_models.md

@@ -596,6 +596,7 @@ Specified using `--task generate`.
 | `PaliGemmaForConditionalGeneration` | PaliGemma, PaliGemma 2 | T + I<sup>E</sup> | `google/paligemma-3b-pt-224`, `google/paligemma-3b-mix-224`, `google/paligemma2-3b-ft-docci-448`, etc. | | ✅︎ | ⚠️ |
 | `Phi3VForCausalLM` | Phi-3-Vision, Phi-3.5-Vision | T + I<sup>E+</sup> | `microsoft/Phi-3-vision-128k-instruct`, `microsoft/Phi-3.5-vision-instruct`, etc. | | ✅︎ | ✅︎ |
 | `Phi4MMForCausalLM` | Phi-4-multimodal | T + I<sup>+</sup> / T + A<sup>+</sup> / I<sup>+</sup> + A<sup>+</sup> | `microsoft/Phi-4-multimodal-instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |
+| `Phi4MultimodalForCausalLM` | Phi-4-multimodal (HF Transformers) | T + I<sup>+</sup> / T + A<sup>+</sup> / I<sup>+</sup> + A<sup>+</sup> | `microsoft/Phi-4-multimodal-instruct` (with revision `refs/pr/70`), etc. | ✅︎ | ✅︎ | ✅︎ |
 | `PixtralForConditionalGeneration` | Mistral 3 (Mistral format), Pixtral (Mistral format) | T + I<sup>+</sup> | `mistralai/Mistral-Small-3.1-24B-Instruct-2503`, `mistralai/Pixtral-12B-2409`, etc. | | ✅︎ | ✅︎ |
 | `QwenVLForConditionalGeneration`<sup>^</sup> | Qwen-VL | T + I<sup>E+</sup> | `Qwen/Qwen-VL`, `Qwen/Qwen-VL-Chat`, etc. | ✅︎ | ✅︎ | ✅︎ |
 | `Qwen2AudioForConditionalGeneration` | Qwen2-Audio | T + A<sup>+</sup> | `Qwen/Qwen2-Audio-7B-Instruct` | | ✅︎ | ✅︎ |

examples/offline_inference/audio_language.py

@@ -190,6 +190,37 @@ def run_phi4mm(question: str, audio_count: int) -> ModelRequestData:
     )
 
 
+def run_phi4_multimodal(question: str, audio_count: int) -> ModelRequestData:
+    """
+    Phi-4-multimodal-instruct supports both image and audio inputs. Here, we
+    show how to process audio inputs.
+    """
+    model_path = snapshot_download(
+        "microsoft/Phi-4-multimodal-instruct", revision="refs/pr/70"
+    )
+    # Since the vision-lora and speech-lora co-exist with the base model,
+    # we have to manually specify the path of the lora weights.
+    speech_lora_path = os.path.join(model_path, "speech-lora")
+    placeholders = "<|audio|>" * audio_count
+
+    prompts = f"<|user|>{placeholders}{question}<|end|><|assistant|>"
+
+    engine_args = EngineArgs(
+        model=model_path,
+        max_model_len=12800,
+        max_num_seqs=2,
+        enable_lora=True,
+        max_lora_rank=320,
+        limit_mm_per_prompt={"audio": audio_count},
+    )
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompts,
+        lora_requests=[LoRARequest("speech", 1, speech_lora_path)],
+    )
+
+
 # Qwen2-Audio
 def run_qwen2_audio(question: str, audio_count: int) -> ModelRequestData:
     model_name = "Qwen/Qwen2-Audio-7B-Instruct"
@@ -303,6 +334,7 @@ def run_whisper(question: str, audio_count: int) -> ModelRequestData:
     "granite_speech": run_granite_speech,
     "minicpmo": run_minicpmo,
     "phi4_mm": run_phi4mm,
+    "phi4_multimodal": run_phi4_multimodal,
     "qwen2_audio": run_qwen2_audio,
     "qwen2_5_omni": run_qwen2_5_omni,
     "ultravox": run_ultravox,

examples/offline_inference/vision_language.py

@@ -949,6 +949,41 @@ def run_phi4mm(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+# HF format Phi-4-multimodal-instruct
+def run_phi4_multimodal(questions: list[str], modality: str) -> ModelRequestData:
+    """
+    Phi-4-multimodal-instruct supports both image and audio inputs. Here, we
+    show how to process image inputs.
+    """
+    assert modality == "image"
+    model_path = snapshot_download(
+        "microsoft/Phi-4-multimodal-instruct", revision="refs/pr/70"
+    )
+    # Since the vision-lora and speech-lora co-exist with the base model,
+    # we have to manually specify the path of the lora weights.
+    vision_lora_path = os.path.join(model_path, "vision-lora")
+    prompts = [
+        f"<|user|><|image|>{question}<|end|><|assistant|>" for question in questions
+    ]
+    engine_args = EngineArgs(
+        model=model_path,
+        max_model_len=5120,
+        max_num_seqs=2,
+        max_num_batched_tokens=12800,
+        enable_lora=True,
+        max_lora_rank=320,
+        # Note - mm_processor_kwargs can also be passed to generate/chat calls
+        mm_processor_kwargs={"dynamic_hd": 16},
+        limit_mm_per_prompt={"image": 1},
+    )
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+        lora_requests=[LoRARequest("vision", 1, vision_lora_path)],
+    )
+
+
 # Pixtral HF-format
 def run_pixtral_hf(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -1205,6 +1240,7 @@ def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
     "paligemma2": run_paligemma2,
     "phi3_v": run_phi3v,
     "phi4_mm": run_phi4mm,
+    "phi4_multimodal": run_phi4_multimodal,
     "pixtral_hf": run_pixtral_hf,
     "qwen_vl": run_qwen_vl,
     "qwen2_vl": run_qwen2_vl,

examples/offline_inference/vision_language_multi_image.py

@@ -686,6 +686,40 @@ def load_phi4mm(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+def load_phi4_multimodal(question: str, image_urls: list[str]) -> ModelRequestData:
+    """
+    Phi-4-multimodal-instruct supports both image and audio inputs. Here, we
+    show how to process multi images inputs.
+    """
+
+    model_path = snapshot_download(
+        "microsoft/Phi-4-multimodal-instruct", revision="refs/pr/70"
+    )
+    # Since the vision-lora and speech-lora co-exist with the base model,
+    # we have to manually specify the path of the lora weights.
+    vision_lora_path = os.path.join(model_path, "vision-lora")
+    engine_args = EngineArgs(
+        model=model_path,
+        max_model_len=4096,
+        max_num_seqs=2,
+        limit_mm_per_prompt={"image": len(image_urls)},
+        enable_lora=True,
+        max_lora_rank=320,
+        # Note - mm_processor_kwargs can also be passed to generate/chat calls
+        mm_processor_kwargs={"dynamic_hd": 4},
+    )
+
+    placeholders = "<|image|>" * len(image_urls)
+    prompt = f"<|user|>{placeholders}{question}<|end|><|assistant|>"
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=[fetch_image(url) for url in image_urls],
+        lora_requests=[LoRARequest("vision", 1, vision_lora_path)],
+    )
+
+
 def load_qwen_vl_chat(question: str, image_urls: list[str]) -> ModelRequestData:
     model_name = "Qwen/Qwen-VL-Chat"
     engine_args = EngineArgs(
@@ -912,6 +946,7 @@ def load_tarsier2(question: str, image_urls: list[str]) -> ModelRequestData:
     "ovis": load_ovis,
     "phi3_v": load_phi3v,
     "phi4_mm": load_phi4mm,
+    "phi4_multimodal": load_phi4_multimodal,
     "pixtral_hf": load_pixtral_hf,
     "qwen_vl_chat": load_qwen_vl_chat,
     "qwen2_vl": load_qwen2_vl,