[Model] Expose Phi3v num_crops as a mm_processor_kwarg (#8658)

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>

examples/offline_inference_vision_language.py

@@ -83,10 +83,24 @@ def run_phi3v(question, modality):
 
     # In this example, we override max_num_seqs to 5 while
     # keeping the original context length of 128k.
+
+    # num_crops is an override kwarg to the multimodal image processor;
+    # For some models, e.g., Phi-3.5-vision-instruct, it is recommended
+    # to use 16 for single frame scenarios, and 4 for multi-frame.
+    #
+    # Generally speaking, a larger value for num_crops results in more
+    # tokens per image instance, because it may scale the image more in
+    # the image preprocessing. Some references in the model docs and the
+    # formula for image tokens after the preprocessing
+    # transform can be found below.
+    #
+    # https://huggingface.co/microsoft/Phi-3.5-vision-instruct#loading-the-model-locally
+    # https://huggingface.co/microsoft/Phi-3.5-vision-instruct/blob/main/processing_phi3_v.py#L194
     llm = LLM(
         model="microsoft/Phi-3-vision-128k-instruct",
         trust_remote_code=True,
         max_num_seqs=5,
+        mm_processor_kwargs={"num_crops": 16},
     )
     stop_token_ids = None
     return llm, prompt, stop_token_ids