[WWB] Additional processing of native phi4mm

tools/who_what_benchmark/whowhatbench/model_loaders.py

@@ -234,9 +234,35 @@ def load_visual_text_model(
                     model_id, trust_remote_code=True, device_map=device.lower()
                 )
             except ValueError:
+                from_pretrained_kwargs = {}
+                if config.model_type == "phi4mm":
+                    if "activation_checkpointing" in config.audio_processor["config"]:
+                        config.audio_processor["config"]["activation_checkpointing"] = ""
+                    config._attn_implementation = "sdpa"
+                    from_pretrained_kwargs["config"] = config
+
                 model = AutoModelForCausalLM.from_pretrained(
-                    model_id, trust_remote_code=True, device_map=device.lower(), _attn_implementation="eager", use_flash_attention_2=False
+                    model_id,
+                    trust_remote_code=True,
+                    device_map=device.lower(),
+                    use_flash_attention_2=False,
+                    **from_pretrained_kwargs,
                 )
+
+                if config.model_type == "phi4mm":
+                    use_lora = False
+                    if hasattr(config, "vision_lora") and config.vision_lora is not None:
+                        model.set_lora_adapter("vision")
+                        use_lora = True
+                    if hasattr(config, "speech_lora") and config.speech_lora is not None:
+                        model.set_lora_adapter("speech")
+                        use_lora = True
+                    if use_lora:
+                        model.unset_lora_adapter = lambda: None
+                        model.set_lora_adapter = lambda _: None
+                    if hasattr(model.model, "_require_grads_hook"):
+                        model.model.disable_input_require_grads()
+
         model.eval()
         try:
             model.get_vision_tower().load_model()