Recipes

For a first real decision, follow the complete CPU walkthrough (中文) and its recorded LAYA demo.

• Laya text worker: start the external Python worker, connect the Rust frontend and compare direct and proxied responses.

• Laya native CUDA worker: build the Hopper bundle and serve English text decisions with Rust and CUDA.

• Laya on Apple Silicon: serve Laya on the Mac GPU with the Laya worker, put the frontend in front of it and run the benchmarks.

• Cua-S1 4B 0.2 text worker: download the pinned weights, start the worker and connect the Rust frontend.

• Valen-Preview-0923 reference worker: prepare the pinned Preview checkpoint and Qwen base, start the text-or-image choice/noul/score worker and verify the frontend.

• Valen-Preview-0923 checkpoint export: export the pinned checkpoint and base to a manifest-described safetensors artifact set and run the export regression.

• Cua-S1 4B 0.2 native text worker: build the CUDA library and the Rust worker, export BF16 weights with the original FP32 adapter and start the worker.

• Decider-2B v11 native text worker: load verified pinned weights, build the Rust/CUDA worker and configure optional batching, Graph, prefix reuse, request/token capacity and HTTP deadlines; validation.

• Cua-S1 native screenshot worker: export the multimodal language checkpoint and serve one-image decisions with Rust/CUDA.

• Open-Jev native text worker: export the merged text backbone and trained decision head of Open-Jev-27B-v1.1 or Open-Jev-9B, then serve with Rust and CUDA; the 27B path supports compatible candidate batching and within-question prefix sharing. Open-Jev-9B validation compares the 9B worker with the reference.

• StartLux Instant 4B native worker: build the Ada sm_89 CUDA library, serve the pinned CC BY-NC Instant checkpoint, and run the L20X closed-loop latency recipe. AutoThink and 27B are not in this recipe. GPU parity is unverified.

• JEMM native worker: pinned text/image decisions and export; known multi-image numerical limitation is documented in the model contract.

• JEV-27B-VL experimental worker: native text and preencoded image decisions; bounded H800 validation with explicit deployment limits.

• OmniJev-4B v1.1 preparation: export the checkpoint and regenerate the CPU fixtures; native worker: serve it on CUDA and compare it with the reference.

• CLM behind the frontend: run CLM's own server behind the frontend on CPU with a stub encoder, and what the response comparison has to allow for.

• pplx-decider reference worker: pin perplexity-ai/pplx-decider-v1.1-27b, serve choice/noul/score through the checkpoint runtime. Head 1ba07d1 has a reported L20X primary. Later heads have no measured sample.

• D3 reference worker: download the pinned vllm-sr/d3 package, verify its runtime manifest, and serve choice/noul/score decisions over text, image and video state through the Python worker and the Rust frontend (text serving and parity validated on A100; media budgets and caps in the recipe).

• Vela 2.0 reference workers (0.3B and the 0.8B hybrid): download the pinned vllm-sr/Vela-2.0-0.3B package, pass its full-hash package gate, and serve all five SystemOne question kinds (choice/noul/score plus the span/set extensions) through the Python worker and the Rust frontend on CPU FP32; or vllm-sr/Vela-2.0-0.8B at revision 326b01d8 behind its own 10-file gate.

• Clef reference worker: build omni-clef, load pinned Clef-flash (9B) or Clef (27B), and serve text choice/noul/score. The joint head stays in the checkpoint's Python module, on one persistent executor thread. Image, video, and a native joint head are not served. Operator-reported flash Public231 concurrency-1 timing on the L20X-labeled lab host is the previous Python HTTP server. The primary is tip 594528a9: p50 164.969 ms, p95 216.755 ms, /health fla+causal_conv1d, versus 9030065 run1 p50 247.518 ms, p95 413.025 ms (Δp50 −82.549 ms, Δp95 −196.270 ms). Operator-reported omni-clef tip f9127a0 is a serve-path paste, not a latency KEEP: R1 p50 169.605 ms, p95 220.003 ms (231/231) does not pass the dual gate versus 164.969/216.755; R2 53.580/97.031 and R3 52.950/97.608 are not a new primary. Smoke 4/4, /health fla+causal_conv1d. Start omni-clef with CLEF_PYTHON set to the recipe venv; system python3 lacks causal-conv1d.

• LFM2.5-350M choice worker: score text decisions with independent questions, hybrid cache branching and configurable candidate batches.

Recipes contain setup, launch commands and examples. Reusable implementation code belongs under src/.

compare_with_backend.py checks that the frontend returns what the worker returned, for any recipe; test_compare_with_backend.py covers it without a model.