fix: use MiniCPM-V-4.5 to resolve unsupported architecture on Space

MiniCPM-V-4.6 has model_type 'minicpmv4_6' and requires transformers>=5.7.0,
which the Space's transformers 4.x can't load (ValueError: architecture not
recognized). Switch to MiniCPM-V-4.5 (model_type 'minicpmv'), which uses
trust_remote_code + model.chat() on transformers 4.x.

- redac/vision.py: MODEL_ID -> openbmb/MiniCPM-V-4_5; load/download weights
on CPU before entering the @spaces.GPU window so the 16GB download can't
exhaust the ZeroGPU time budget; deterministic decoding (sampling=False).
- requirements.txt: pin transformers>=4.49,<5; add timm.
- app.py: make local port configurable via GRADIO_SERVER_PORT (xray-anything
occupies 7860-7863).

Co-authored-by: OpenAI Codex <codex@openai.com>

app.py

@@ -104,7 +104,7 @@ with gr.Blocks(title="Redac") as demo:
         with gr.Tab("Image"):
             gr.Markdown(
                 "Upload a document or ID image. A local vision model "
-                "(MiniCPM-V-4.6) extracts the fields, then Redac redacts them."
+                "(MiniCPM-V-4.5) extracts the fields, then Redac redacts them."
             )
             i_map = gr.State({})
             with gr.Row():
@@ -136,4 +136,7 @@ with gr.Blocks(title="Redac") as demo:
 
 
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860, theme=gr.themes.Soft())
+    import os
+
+    port = int(os.environ.get("GRADIO_SERVER_PORT", "7860"))
+    demo.launch(server_name="0.0.0.0", server_port=port, theme=gr.themes.Soft())

redac/vision.py

@@ -1,9 +1,14 @@
-"""Image -> text extraction with a local vision model (MiniCPM-V-4.6).
+"""Image -> text extraction with a local vision model (MiniCPM-V-4.5).
 
 Runs in-Space on ZeroGPU via the @spaces.GPU decorator. The model reads a
 document/ID image and returns the fields as plain text, which then flows
 through the same PII detection + redaction core as typed text.
 
+MiniCPM-V-4.5 (model_type "minicpmv") uses trust_remote_code + model.chat()
+and runs on transformers 4.x. The weights are downloaded/loaded on CPU
+*before* entering the GPU window so the 16GB download can't blow the
+ZeroGPU time budget; the GPU call only moves the model to CUDA and generates.
+
 Local dev: set REDAC_MOCK=1 to skip the model entirely and return a canned
 extraction, so the Gradio UI runs on a laptop with no GPU.
 """
@@ -13,7 +18,7 @@ from __future__ import annotations
 import os
 from functools import lru_cache
 
-MODEL_ID = "openbmb/MiniCPM-V-4.6"
+MODEL_ID = "openbmb/MiniCPM-V-4_5"
 
 EXTRACTION_PROMPT = (
     "You are a document data extractor. Read this image and transcribe ALL "
@@ -52,6 +57,7 @@ except Exception:  # pragma: no cover - local fallback
 
 @lru_cache(maxsize=1)
 def _load_model():
+    """Download + load weights on CPU. Cached. Called outside the GPU window."""
     import torch
     from transformers import AutoModel, AutoTokenizer
 
@@ -67,10 +73,10 @@ def _load_model():
 
 @_gpu
 def _chat(image, prompt: str) -> str:
-    model, tokenizer = _load_model()
+    model, tokenizer = _load_model()  # cached -> instant inside GPU window
     model = model.to("cuda")
     msgs = [{"role": "user", "content": [image, prompt]}]
-    return model.chat(image=None, msgs=msgs, tokenizer=tokenizer)
+    return model.chat(image=None, msgs=msgs, tokenizer=tokenizer, sampling=False)
 
 
 def extract_text_from_image(image, prompt: str | None = None) -> str:
@@ -79,4 +85,5 @@ def extract_text_from_image(image, prompt: str | None = None) -> str:
         return ""
     if _is_mock():
         return _MOCK_EXTRACTION
+    _load_model()  # warm the cache on CPU before claiming the GPU
     return _chat(image.convert("RGB"), prompt or EXTRACTION_PROMPT).strip()

requirements.txt

@@ -2,9 +2,10 @@ gradio==6.18.0
 gliner>=0.2.13
 torch
 torchvision
-transformers>=4.44
+transformers>=4.49,<5
 accelerate
 sentencepiece
+timm
 pillow
 huggingface_hub
 spaces