Set per-event Gradio concurrency limits and ZeroGPU queue size

- detect (ZeroGPU @spaces.GPU): concurrency_limit=2 — each call forks its
own worker/GPU, so VRAM isn't shared; higher mostly waits on the pool and
burns quota faster
- annotate (.change() calling the pilbox Space): concurrency_limit=4 —
network-bound, requests overlap; at the default of 1 it choked after
every detection
- demo.queue(max_size=12) — ZeroGPU otherwise forces max_size=1, so a
second waiting request got HTTP 503 "Queue is full"
- Document the limits in CLAUDE.md

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

CLAUDE.md

@@ -20,3 +20,4 @@ Gradio launches with `mcp_server=True` (exposes an MCP endpoint) and `app_kwargs
 - Loads `vikhyatk/moondream2` at revision `2025-04-14` with `trust_remote_code=True`. Pinning the revision is intentional — the model repo changes its API surface across revisions, so bumping it can break `model.point(...)` / `model.detect(...)` return shapes.
 - `@spaces.GPU(duration=30)` decorators are HF Spaces ZeroGPU directives — they only activate on the Space; locally they are no-ops requiring a CUDA device (the `device_map={"": "cuda"}` hard-codes GPU). Local CPU runs need that removed.
 - `detect()` returns the inner list (`["points"]` or `["objects"]`) rather than the full dict — callers (including the MCP tool surface) depend on that shape.
+- Concurrency (Gradio queue): `detect` runs at most 2 calls at once (`concurrency_limit=2` on the Interface; ZeroGPU forks a worker per call), the pilbox `annotate` `.change()` event at most 4, and `demo.queue(max_size=12)` caps waiters (ZeroGPU otherwise forces `max_size=1`, so a second waiter gets a 503 "Queue is full"). Limits are per event; work inside a single request is still sequential.

app.py

@@ -118,6 +118,8 @@ demo = gr.Interface(
         ),
     ],
     outputs=gr.JSON(label="Output JSON"),
+    # ZeroGPU: each call forks its own worker/GPU (no shared VRAM); higher just waits on the pool and burns quota faster
+    concurrency_limit=2,
 )
 with demo, demo.output_components[0].parent:  # render inside the output column
     demo.output_components[0].change(
@@ -125,7 +127,10 @@ with demo, demo.output_components[0].parent:  # render inside the output column
         [demo.input_components[0], demo.output_components[0], demo.input_components[2]],
         gr.Image(label="Annotated Image"),
         api_name=False,
+        # network-bound call to the pilbox Space: requests overlap; at 1 this would choke after every detection
+        concurrency_limit=4,
     )
+demo.queue(max_size=12)  # ZeroGPU forces max_size=1 otherwise -> 2nd waiter gets HTTP 503 "Queue is full"; ~2x sum of limits
 demo.launch(
     mcp_server=True, app_kwargs={"docs_url": "/docs"}  # add FastAPI Swagger API Docs
 )