Adaptive ZeroGPU step length with a quota explanation

- Ask for a 600 s xlarge step first (one weight transfer, most posters finish in one step) and fall
back to 300/120/60 s when ZeroGPU answers 'larger than the maximum allowed'.
- Generation budget derives from the step length; status shows the step length.
- Quota hint on failures: 27B BF16 needs the full GPU (2x quota), ~15-25 min per poster; PRO 40 min,
free 5 min, anonymous 2 min per day. README says the same.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

README.md

@@ -19,8 +19,10 @@ Upload a research paper PDF and get an editable academic poster. A 27B vision-la
 and writes the poster as standalone HTML; the Space renders it to PNG.
 
 - Model: RL checkpoint-240, BF16, no quantization, 16K context.
-- Generation runs in ZeroGPU steps of up to 120 s; a full poster usually takes 3–6 steps and the
-  demo chains them automatically. Sign in to Hugging Face to use your GPU quota.
+- The model runs in BF16 on a full 96 GB ZeroGPU slot (`xlarge`, counted 2x against quota). A poster
+  takes about 15–25 minutes of ZeroGPU quota; PRO accounts have 40 min/day, free accounts 5 min and
+  anonymous visitors 2 min, so sign in with a PRO account (on huggingface.co, not the bare hf.space URL)
+  to finish a poster. The demo asks for a 10-minute GPU step first and falls back to shorter steps.
 - API: call `/analyze` once (PDF + settings), then `/step` repeatedly until the status starts with
   `Done`, `Paused` or `Stopped`; each step is its own request so the ZeroGPU token stays valid.
 - The examples on the page were generated by the same checkpoint with the default instruction and settings.

demo_ui.py

@@ -14,7 +14,10 @@ from poster_render import render
 ROOT = Path(__file__).resolve().parent
 ARTIFACTS = Path(tempfile.mkdtemp(prefix='rap_demo_'))
 WINDOW = 16384
-MAX_STEPS = 10   # ZeroGPU steps of <=120 s per run
+MAX_STEPS = 10   # chained ZeroGPU steps per run
+GPU_LADDER = (600, 300, 120, 60)   # requested step length; ZeroGPU refuses what the visitor's quota cannot cover
+QUOTA_HINT = (' — this 27B BF16 model needs the full 96 GB ZeroGPU (2x quota) and a poster takes roughly 15–25 min of quota. '
+              'Daily ZeroGPU quota: PRO 40 min, free account 5 min, anonymous 2 min, so sign in with a PRO account to finish a poster.')
 
 CSS = '''
 .gradio-container {max-width: 1360px !important; margin: 0 auto;}
@@ -64,8 +67,12 @@ def build(model, processor):
         p['input_tokens'] = n; p['continuation_ids'] = []; p['raw'] = ''; p['done'] = False
         return p
 
-    @spaces.GPU(size='xlarge', duration=120)
+    def gpu_seconds(p):
+        return int(p.get('gpu_seconds', 120))
+
+    @spaces.GPU(size='xlarge', duration=gpu_seconds)   # 27B BF16 needs the full 96 GB GPU (2x quota)
     def gpu_chunk(p):
+        budget = max(30, gpu_seconds(p) - 90)   # leave room for moving the weights to the GPU and the prefill
         inputs = processor.apply_chat_template(chat(p), tokenize=True, add_generation_prompt=True,
                                                enable_thinking=False, return_dict=True, return_tensors='pt')
         prefix = p.get('continuation_ids', [])
@@ -79,7 +86,7 @@ def build(model, processor):
         inputs = inputs.to('cuda')
         start = time.monotonic()
         with torch.inference_mode():
-            output = model.generate(**inputs, max_new_tokens=min(8192, remaining), max_time=100, do_sample=False,
+            output = model.generate(**inputs, max_new_tokens=min(8192, remaining), max_time=budget, do_sample=False,
                                     use_cache=True, temperature=None, top_p=None, top_k=None, repetition_penalty=1.1)
         new_ids = output[0, length:].detach().cpu().tolist()
         all_ids = prefix + new_ids
@@ -122,13 +129,21 @@ def build(model, processor):
             return (dict(p, paused=True), f"Stopped after {MAX_STEPS} steps without a complete poster.", None, HIDE, HIDE,
                     gallery, p['spec'], p['raw'], gr.update(visible=True))
         progress(min(.1 + .12 * k, .9), desc=f'Generating poster HTML (step {k})…')
-        try:
-            r = gpu_chunk(p)
-        except Exception as e:                                             # noqa: BLE001  (quota exhausted, timeout, …)
-            return (dict(p, paused=True), f"Paused after step {k - 1}: {str(e)[:200]}", None, HIDE, HIDE,
-                    gallery, p['spec'], p['raw'], gr.update(visible=bool(p.get('continuation_ids'))))
+        r = None
+        for secs in [d for d in GPU_LADDER if d <= p.get('gpu_seconds', GPU_LADDER[0])]:
+            p = dict(p, gpu_seconds=secs)
+            try:
+                r = gpu_chunk(p)
+                break
+            except Exception as e:                                         # noqa: BLE001  (quota, duration limit, timeout, …)
+                msg = str(e)
+                if 'larger than the maximum allowed' in msg and secs != GPU_LADDER[-1]:
+                    continue                                               # ask for a shorter GPU slot
+                hint = QUOTA_HINT if ('maximum allowed' in msg or 'quota' in msg.lower()) else ''
+                return (dict(p, paused=True), f"Paused after step {k - 1}: {msg[:200]}{hint}", None, HIDE, HIDE,
+                        gallery, p['spec'], p['raw'], gr.update(visible=bool(p.get('continuation_ids'))))
         p = dict(p, continuation_ids=r['ids'], raw=r['raw'], done=r['done'], step=k)
-        tokens = f"{len(r['ids']):,} tokens · {r['new_tokens'] / max(r['seconds'], .01):.0f} tok/s"
+        tokens = f"{len(r['ids']):,} tokens · {r['new_tokens'] / max(r['seconds'], .01):.0f} tok/s · {p['gpu_seconds']} s GPU step"
         if not (r['complete'] or p['done']):
             return (p, f"Step {k} done · {tokens} · continuing (step {k + 1})…", None, HIDE, HIDE,
                     gallery, p['spec'], p['raw'], gr.update(visible=False))