Private BF16 ZeroGPU model bring-up

README.md

@@ -1,13 +1,19 @@
 ---
-title: RAP
-emoji: 🌍
-colorFrom: indigo
-colorTo: pink
+title: RAP — Paper to Poster
+emoji: 📰
+colorFrom: purple
+colorTo: blue
 sdk: gradio
-sdk_version: 6.30.0
-python_version: '3.12'
+sdk_version: 5.50.0
+python_version: 3.10
 app_file: app.py
 pinned: false
+startup_duration_timeout: 1h
+short_description: Turn a research paper into an editable academic poster
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# RAP — Paper to Poster
+
+Deployment validation in progress. RL checkpoint-240, BF16, no quantization.
+The model repository remains private; its access token is a server-side Secret.
+ZeroGPU xlarge is requested per generation. No dedicated paid GPU is used.

app.py

@@ -0,0 +1,45 @@
+"""Private bring-up app. Replaced by the paper-to-poster UI after validation."""
+import spaces
+import os
+import json
+import shutil
+import time
+import torch
+import gradio as gr
+import psutil
+from transformers import AutoModelForImageTextToText, AutoProcessor
+
+REPO = "VincentHancoder/RAP-Qwen3.8-27B-SFT-RL"
+REVISION = "50f3adc4ebb9b19e53ef9c21c293993a77bf3fd7"
+print("STARTUP", json.dumps({"ram_gib": psutil.virtual_memory().total / 2**30,
+                            "disk_free_gib": shutil.disk_usage(".").free / 2**30,
+                            "torch": torch.__version__}), flush=True)
+processor = AutoProcessor.from_pretrained(REPO, revision=REVISION, token=os.environ.get("HF_TOKEN"))
+model = AutoModelForImageTextToText.from_pretrained(
+    REPO, revision=REVISION, token=os.environ.get("HF_TOKEN"), dtype=torch.bfloat16,
+    device_map="cuda:0", attn_implementation="sdpa",
+)
+model.eval()
+model.config.use_cache = True
+model.config.text_config.use_cache = True
+print("RAP_MODEL_READY", flush=True)
+
+@spaces.GPU(size="xlarge", duration=120)
+def probe(prompt):
+    start = time.monotonic()
+    messages = [{"role": "user", "content": [{"type": "text", "text": prompt[:1000]}]}]
+    inputs = processor.apply_chat_template(messages, tokenize=True, add_generation_prompt=True,
+        enable_thinking=False, return_dict=True, return_tensors="pt").to("cuda")
+    with torch.inference_mode():
+        result = model.generate(**inputs, max_new_tokens=64, do_sample=False, use_cache=True,
+                                repetition_penalty=1.1, temperature=None, top_p=None, top_k=None)
+    output = processor.decode(result[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
+    return {"text": output, "seconds": round(time.monotonic()-start, 2),
+            "gpu": torch.cuda.get_device_name(), "dtype": str(next(model.parameters()).dtype)}
+
+with gr.Blocks() as demo:
+    gr.Markdown("# RAP deployment validation\nRL checkpoint-240 · BF16 · ZeroGPU xlarge")
+    p = gr.Textbox(value="Write a short HTML heading for a research poster.")
+    out = gr.JSON()
+    gr.Button("Validate GPU").click(probe, p, out, api_name="probe")
+demo.queue(default_concurrency_limit=1, max_size=8).launch(show_error=False)

packages.txt

@@ -0,0 +1,7 @@
+chromium
+fonts-dejavu-core
+fonts-liberation
+fonts-noto-core
+fonts-noto-cjk
+libgl1
+libglib2.0-0

requirements.txt

@@ -0,0 +1,14 @@
+torch==2.9.1
+torchvision==0.24.1
+transformers==5.2.0
+accelerate>=1.10,<2
+flash-linear-attention==0.5.2
+fla-core==0.5.2
+pymupdf==1.27.2.3
+pillow>=11,<13
+beautifulsoup4>=4.13,<5
+playwright==1.58.0
+spaces
+psutil
+timm==0.9.16
+sentencepiece