dkappe/needle3-gpu

needle3 logits (Gradio SDK + ZeroGPU)

A Jev-shaped endpoint backed by Cactus Needle 3, reading its JAX path so real next-token logits are available. Runs on ZeroGPU.

Verified working

• —ZeroGPU allocated — hardware zero-a10g, Space RUNNING.
• —JAX on ZeroGPU works — the GPU probe tab reports backend: gpu, devices: ["cuda:0"], and completes a real computation. This is not a documented-supported combination (ZeroGPU is PyTorch-shaped), so it was verified rather than assumed.
• —Real distributions — choice returns genuine probabilities, e.g. a billing ticket: billing 0.877 / technical 0.004 / account 0.120.

Honest status

The plumbing works. Accuracy does not yet.

• —Second test case ("I cannot log in, my password is rejected") returns billing 0.709 / account 0.285 — wrong, should be account. So choice is roughly 2/3 on unambiguous cases, not production quality.
• —noul (yes/no) did not discriminate on matched positive/negative pairs.
• —Probabilities are uncalibrated — nothing shows a stated 0.8 is right 80% of the time. Do not build a threshold on them.

Why the JAX path

Needle 3 ships two runtimes, and only one yields probabilities:

The native engine's confidence head returned a constant 1.0 in local testing.

ZeroGPU implementation notes

Four things were required, each of which failed first:

1. 1.A NEW Space. ZeroGPU requires the Gradio SDK, and a Space's SDK is immutable — a Docker Space cannot be converted. This replaced an earlier Docker build (dkappe/needle3-logits), which was deleted once this Gradio version was verified. Its source is still in the repo beside this one as reference, but nothing runs it.
2. 2.Hardware request via REST. hf spaces settings <id> --hardware zero-a10g errors on this CLI version; a PUT /api/spaces/<id>/settings with {"hardware":"zero-a10g"} works.
3. 3.`config.dtype = "float32"` before building the network. The checkpoint declares dtype=bfloat16, and every layer is constructed with it. On this GPU that aborts with Unsupported conversion from bf16 to f16 / LLVM ERROR: Unsupported rounding mode for conversion. Casting the parameters alone is not enough — the model's own dtype must change too. NEEDLE_CAST=bfloat16 reverts this.
4. 4.Checkpoint fetch at module scope, inside `if __name__ == "__main__"`. Gradio executes app_file as __main__, so a bootstrap in an else: branch never runs; and network access inside a @spaces.GPU function is restricted.

Also: XLA_PYTHON_CLIENT_PREALLOCATE=false (JAX otherwise claims ~75% of a shared GPU slice), and the health check deliberately does not import JAX — outside a @spaces.GPU function there is no GPU attached, so it would fail with No visible GPU devices.

Endpoints

Caveats

• —Experimental; not affiliated with TypeSafe AI or Cactus Compute.
• —Free ZeroGPU quota is small (~5 min/day for a free account), so heavy experimentation will exhaust it.
• —Accuracy figures come from three hand-chosen cases — a smoke test, not a benchmark.
• —Apache-2.0, matching the model.