Upload 2 files

app.py

@@ -4,8 +4,7 @@ import gradio as gr
 from infer import MiMoVLInfer
 import spaces
 
-# infer = MiMoVLInfer(checkpoint_path="XiaomiMiMo/MiMo-VL-7B-RL")
-infer = MiMoVLInfer(checkpoint_path="XiaomiMiMo/MiMo-VL-7B-RL-2508")
+infer = MiMoVLInfer(checkpoint_path=os.environ.get('CKPT_PATH'))
 
 label_translations = {
     "gr_chatinterface_ofl": {
@@ -153,24 +152,6 @@ with gr.Blocks() as demo:
                                     "text": "Who are you?",
                                     "files": []
                                 },
-                                {
-                                    "text": "OCR and return markdown",
-                                    "files": ["examples/24-25-pl.png"]
-                                },
-                                {
-                                    "text":
-                                    """describe the video""",
-                                    "files":
-                                    ["examples/hitting_baseball.mp4"]
-                                },
-                                {
-                                    "text":
-                                    "For the model ranked first on WebSRC, what is its score on MathVision?",
-                                    "files": [
-                                        "examples/mimovl_gui.png",
-                                        "examples/mimovl_reason.png"
-                                    ]
-                                },
                             ],
                             inputs=[gr_chatinterface_ofl.textbox],
                         )

infer.py

@@ -14,7 +14,6 @@ class MiMoVLInfer:
             checkpoint_path,
             torch_dtype=dtype,
             device_map={"": "cpu"},
-            attn_implementation="eager",
             trust_remote_code=True,
         ).eval()
         self.processor = AutoProcessor.from_pretrained(checkpoint_path, trust_remote_code=True)
@@ -42,15 +41,20 @@ class MiMoVLInfer:
         streamer = TextIteratorStreamer(tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
 
         max_new = int(os.getenv("MAX_NEW_TOKENS", "1024"))
+        temp = float(temperature or 0.0)
+        do_sample = temp > 1e-3
+        if do_sample:
+            samp_args = {"do_sample": True, "temperature": max(temp, 0.01), "top_p": 0.95}
+        else:
+            samp_args = {"do_sample": False}
+
         gen_kwargs = {
-            'max_new_tokens': max_new,
-            'do_sample': True,
-            'temperature': max(0.0, float(temperature)),
-            'top_p': 0.95,
-            'streamer': streamer,
-            'stopping_criteria': StoppingCriteriaList([EosTokenCriteria(eos_token_id=self.model.config.eos_token_id)]),
-            'pad_token_id': self.model.config.eos_token_id,
-            **model_inputs
+            "max_new_tokens": 1024,
+            "streamer": streamer,
+            "stopping_criteria": StoppingCriteriaList([EosTokenCriteria(eos_token_id=self.model.config.eos_token_id)]),
+            "pad_token_id": self.model.config.eos_token_id,
+            **model_inputs,
+            **samp_args,
         }
 
         thread = Thread(target=self.model.generate, kwargs=gen_kwargs, daemon=True)