Add StartLux Instant 4B native worker (#160)

Cargo.toml

@@ -1,3 +1,3 @@
 [workspace]
-members = ["src/frontend", "src/runtime", "src/models/clm", "src/models/cua_s1/native", "src/models/qwen3_5/native", "src/models/open_jev/native", "src/models/jev_vl/native", "src/models/omnijev/native", "src/models/laya", "src/backends/cuda", "src/models/decider/native", "src/models/jemm/native"]
+members = ["src/frontend", "src/runtime", "src/models/clm", "src/models/cua_s1/native", "src/models/qwen3_5/native", "src/models/open_jev/native", "src/models/jev_vl/native", "src/models/omnijev/native", "src/models/laya", "src/backends/cuda", "src/models/decider/native", "src/models/jemm/native", "src/models/startlux/native"]
 resolver = "3"

README.md

@@ -66,8 +66,11 @@ The LFM2.5-350M Transformers worker owns candidate scoring and hybrid cache bran
   processing and executor modules, with shared FIFO admission and blocking
   dispatch in the [native runtime](src/runtime/README.md).
 - **Native CUDA workers.** The Cua-S1 4B 0.2 `text` adapter,
-  Open-Jev-27B-v1.1 and Open-Jev-9B run as native workers with shared CUDA
-  kernels. Cua-S1 also has a Python worker that serves as the correctness reference.
+  Open-Jev-27B-v1.1, Open-Jev-9B, and StartLux-Decision v1.1 4B Instant run
+  as native workers with shared CUDA kernels. Cua-S1 also has a Python
+  worker that serves as the correctness reference. The StartLux worker is
+  text-only Instant; its weights are CC BY-NC 4.0 and its GPU parity is
+  unverified. AutoThink is not served.
 - **LAYA text serving.** LAYA runs as an external CPU Python worker, the
   in-repository Python MPS/CPU worker, or a native Rust/CUDA worker on Hopper.
 - **CUDA backend and planned Metal backend.** High-performance GPU operations
@@ -176,6 +179,7 @@ pinned runtime for text, image and video state:
 | Valen-Preview-0923 | [Python reference worker](recipe/valen/reference.md); text state or one inline PNG/JPEG image, `choice`; author-run on RTX 5060 Laptop GPU (sm_120, WSL2), owner: community |
 | Open-Jev-27B-v1.1 | [Native Rust/CUDA worker](recipe/open_jev/native.md); eager independent text candidates; [H200 validation](recipe/open_jev/validation.md) |
 | Open-Jev-9B | The same [native Rust/CUDA worker](recipe/open_jev/native.md); eager independent text candidates; [reference comparison on sm_89](recipe/open_jev/validation-9b.md) |
+| StartLux-Decision v1.1 4B Instant | [Native Rust/CUDA text worker](recipe/startlux/README.md); `choice` / `noul` / `score`; weights CC BY-NC 4.0; Instant only (no AutoThink, no 27B); GPU parity unverified; [L20X latency recipe](recipe/startlux/latency.md) |
 | autotrust/JEV-27B-VL | [Experimental Rust/CUDA worker](recipe/jev_vl/README.md); single-question text, online inline PNG/JPEG and offline-preencoded image inputs; [validation scope and limits](recipe/jev_vl/validation.md) |
 | CLM-v0.1-8B | [External worker with a CPU stub encoder](recipe/clm/README.md); contract checks only, real Qwen3-8B decisions unverified by this recipe |
 | LFM2.5-350M | [Transformers worker](recipe/lfm2/README.md); text state and independent choice questions |

docs/architecture.md

@@ -7,9 +7,12 @@ design. Concrete input/output types follow each executor's supported layout.
 ## Implementation status
 
 The [Rust frontend](../src/frontend/README.md) currently forwards HTTP requests
-to separately running workers. Cua-S1, Open-Jev and Decider have native Rust/CUDA workers
-that share the [Qwen3.5/3.8 executor](../src/models/qwen3_5/native/), which accepts
-single prompts and bounded packed prefill. Cua-S1 and Open-Jev-9B use
+to separately running workers. Cua-S1, Open-Jev, Decider, and StartLux Instant
+4B have native Rust/CUDA workers that share the
+[Qwen3.5/3.8 executor](../src/models/qwen3_5/native/), which accepts
+single prompts and bounded packed prefill. StartLux Instant uses a
+bidirectional full-attention call on each prompt; its GDN stays causal, and
+it does not pack prompts or cache prefixes. Cua-S1 and Open-Jev-9B use
 single-prompt calls; Open-Jev-27B-v1.1 packs candidates within one request for
 input and gate/up GEMMs while preserving per-sequence mixers and output/down GEMM
 shapes. Decider optionally packs up to four complete rows and 4096 tokens within
@@ -72,14 +75,15 @@ and candidate identity, usage, and response metadata outside the executor.
 | Decider | Complete independent token-ID rows in question order; Score expands one no/yes row per level. | One BF16-projection-rounded FP32 selected-label logit vector per row. | Per-type calibration, whole-Score normalization, ordered typed answers and unique-prefix token usage. |
 | Laya | One padded request: token IDs, true lengths, question types and ordered option markers; at most 16 questions, 512 tokens per row and 2048 markers. | Per-question FP32 option logits and two action logits copied back after GPU heads. | Calibrate and decode ordered `choice`, `score` and `noul` answers, usage and metadata. |
 | JEMM | Ordered token IDs, three-axis positions, candidate counts and optional shared image patches for one complete request. | Per-question selected BF16 label logits copied back as FP32. | Apply checksum-covered text/image calibration and reconstruct ordered `choice`, `score` and `noul` answers and usage. |
+| StartLux Instant 4B | One unpadded token-ID prompt per question, or per knockout group and then a final prompt, in request order. | FP32 letter logits from a CPU dot of the tied embedding rows at the last position. | Per-type temperature, none-check split, wide-choice residual, and ordered `choice`, `noul`, and `score` answers with token usage. |
 
 Cua-S1 and Open-Jev-9B input collections are serial work. For Open-Jev-27B-v1.1,
 Open-Jev's model-specific batch adapter packs up to 16 independent candidates and
 4096 tokens per group; longer prompts
 execute alone. It restores question/candidate grouping before normalization.
 Laya batches questions within one request. Shared runtime
 admission precedes blocking dispatch: Cua-S1 admits one question forward at a
-time; Open-Jev, Decider and Laya admit one complete request. Cua-S1's CPU letter projection
+time; Open-Jev, Decider, Laya, and StartLux Instant admit one complete request. StartLux runs each prompt inside that admission one at a time. Cua-S1's CPU letter projection
 stays outside admission; Open-Jev's scalar heads and Laya's GPU heads and
 synchronized readback remain inside their request unit. Decider includes every
 backbone row and CUDA head projection, synchronizes both streams and retires

docs/supported-models.md

@@ -19,6 +19,7 @@ Models that are being added are also tracked in issues labeled [new model](https
 | Valen-Preview-0923 | [Python reference worker](../recipe/valen/reference.md); text state or one inline PNG/JPEG image; `choice` questions | Not supported | Author-run on RTX 5060 Laptop (sm_120, WSL2): one text and one 2×2 PNG request; input and readout parity ≤1e-6 against reference (model outputs not compared). Owner: community. | Not supported | Python 3.12, pinned Valen source/checkpoint, Qwen3.5-2B base, BF16 CUDA |
 | Open-Jev-27B-v1.1 | [Native Rust/CUDA worker](../recipe/open_jev/native.md) on the shared Qwen3.5/3.8 executor | Not supported | Validated on H200 (sm_90) for the [74 single-candidate workload](../recipe/open_jev/validation.md) | Not supported | Compute capability 8.0 or newer, CUDA toolkit to build, exported merged weights and trained head |
 | Open-Jev-9B | The same [native Rust/CUDA worker](../recipe/open_jev/native.md) | Not supported | Validated on compute capability 8.9 against the reference for [253 requests](../recipe/open_jev/validation-9b.md) | Not supported | Compute capability 8.0 or newer, CUDA toolkit to build, exported merged weights and trained head |
+| StartLux-Decision v1.1 4B Instant | [Native Rust/CUDA worker](../recipe/startlux/README.md) | Contract checks only; no CPU inference | L20X / Ada sm_89 Public231 C1 at `5bb25beb1c23f72497be189156d82f62ffb974da`: 231/231, p50 6.785 ms, p95 51.432 ms ([latency.md](../recipe/startlux/latency.md)). Scratch reserve `2b4a586e` failed that gate: 231/231, p50 7.811 ms, p95 51.947 ms. Grouped KV load `136ddf89` failed it too: 231/231, p50 8.148 ms, p95 53.053 ms. Packed prompt rows `1b5b1ee4` failed it too: 231/231, p50 7.980 ms, p95 52.417 ms. Pinned token-id copy `d139b228` failed it too: 231/231, p50 8.207 ms, p95 53.011 ms. Launch-error query `d3ba9d5` failed it too: 231/231, p50 8.326 ms, p95 53.868 ms. cuBLASLt heuristic sweep `6c268cd` failed it too: 231/231, p50 6.893 ms, p95 52.146 ms. FP8 E4M3×E4M3→BF16 `725c86ac` never became ready (cuBLAS status 7, 0/231, no percentiles). Symmetric int8 `867ed78` failed it too: 231/231, p50 9.848 ms, p95 83.150 ms. Load-time BF16 plan cache `f40f9b58` failed it too: 231/231, p50 7.953 ms, p95 52.484 ms. In-place GEMM layout retarget `d44e634` never finished warmup (cuBLAS status 7, 0/231, no percentiles). Numerical parity unverified. | Not supported | Pinned CC BY-NC 4.0 Instant 4B checkpoint; rebuilt ABI 9 library with `cs1_attention_gated_bidirectional` and `cs1_gemm_shape`. Instant only: AutoThink and 27B are refused |
 | Decider-2B v11 | [Native Rust/CUDA worker](../recipe/decider/README.md) | Contract checks only; no CPU inference | RTX 4090 (`sm_89`) synthetic parity corpus; [limits and provenance](../recipe/decider/validation.md) | Not supported | Pinned original BF16 checkpoint; ABI 9 library and consumers rebuilt together |
 | JEMM / Qwen3.8-27B | [Native Rust/CUDA worker](../recipe/jemm/README.md) | Host checks only; no CPU inference | A800 (`sm_80`) [text and 1–4-image frozen corpus](../src/models/jemm/README.md#validation) | Not supported | Pinned `jemm-native/2` export with BF16 base and unmerged FP32 LoRA, selected untied LM head, cuDNN 9 and additive reference CUDA symbols |
 | autotrust/JEV-27B-VL | [Experimental native worker](../recipe/jev_vl/README.md), single-question API | Contract tests only; no CPU inference | H800 frozen-corpus check; [experimental scope and limits](../recipe/jev_vl/validation.md) | Not supported | Pinned merged checkpoint, rebuilt ABI 9 CUDA library; image inputs require offline Transformers preencoding |
@@ -34,7 +35,8 @@ Models that are being added are also tracked in issues labeled [new model](https
 - **Planned:** not implemented yet; the linked issue tracks it.
 
 The Cua-S1 workers answer `choice` questions only.
-LAYA's English worker, Open-Jev and CLM support `choice`, `score`, and `noul` text questions.
+LAYA's English worker, Open-Jev, CLM, and StartLux Instant 4B support `choice`, `score`, and `noul` text questions.
+StartLux serves Instant only; AutoThink is not implemented, and the weights are CC BY-NC 4.0.
 D3's reference worker serves the same three kinds for text, image and video
 state (inline base64 media under the pinned runtime's per-item budgets);
 its media serving evidence is author-run on the A100 above.

mkdocs.yml

@@ -70,6 +70,8 @@ nav:
       - Cua-S1 native screenshots: recipe/cua_s1/native_screenshot.md
       - Cua-S1 vision CUDA Graph: recipe/cua_s1/native_vision_graph.md
       - Open-Jev native text worker: recipe/open_jev/native.md
+      - StartLux Instant 4B native worker: recipe/startlux/README.md
+      - StartLux Instant 4B L20X latency: recipe/startlux/latency.md
       - Open-Jev-27B H200 validation: recipe/open_jev/validation.md
       - Open-Jev-9B validation: recipe/open_jev/validation-9b.md
       - Decider native text worker: recipe/decider/README.md

recipe/README.md

@@ -26,6 +26,10 @@ For a first real decision, follow the [complete CPU walkthrough](../docs/getting
   backbone and trained decision head of Open-Jev-27B-v1.1 or Open-Jev-9B, then
   serve with Rust and CUDA. [Open-Jev-9B validation](open_jev/validation-9b.md)
   compares the 9B worker with the reference.
+- [StartLux Instant 4B native worker](startlux/README.md): build the Ada sm_89
+  CUDA library, serve the pinned CC BY-NC Instant checkpoint, and run the
+  [L20X closed-loop latency recipe](startlux/latency.md). AutoThink and 27B
+  are not in this recipe. GPU parity is unverified.
 - [JEMM native worker](jemm/README.md): pinned text/image decisions and export;
   known multi-image numerical limitation is documented in the model contract.
 - [JEV-27B-VL experimental worker](jev_vl/README.md): native text and preencoded

recipe/startlux/README.md

@@ -0,0 +1,134 @@
+# StartLux-Decision v1.1 4B Instant
+
+Serve text `choice`, `noul`, and `score` from the pinned Instant 4B
+checkpoint with the native worker `omni-startlux-native`. AutoThink and
+Instant 27B are not part of this recipe.
+
+Weights are **CC BY-NC 4.0**
+(`startlux-models/StartLux-Decision-v1.1-4B-Instant`, revision
+`416088f77cfce09633906a3675c7cd2e5dabcf10`, about 9.4 GB of BF16 files).
+Download them on the machine that will serve. This repository does not
+contain them. Inference code adapted from StartLux-Decision
+`a99977d9332decc9fb27a2488b4a83b533b6cd5f` is Apache-2.0; see
+[NOTICE](../../src/models/startlux/NOTICE).
+
+GPU parity is unverified here. The intended device is one L20X (Ada, sm_89).
+[latency.md](latency.md) records the native Public231 C1 samples and the
+operator checklist. The 8192-token scratch reserve, the 99 KiB grouped KV
+load, packed prompt rows, the pinned token-id copy, skipping the per-launch
+CUDA error query, the load-time cuBLASLt heuristic sweep, FP8
+E4M3×E4M3→BF16 projections, symmetric int8 projections, the load-time
+BF16 plan cache, and the in-place GEMM layout retarget all failed the
+latency gate and are not in the worker. Record the bench
+metadata field `reservation` for the device hold. This recipe does not add
+scheduler flags.
+
+Commands below run from the repository root unless a path says otherwise.
+
+## Build
+
+The CUDA library and the worker must be built together. ABI stays 9.
+`cs1_attention_gated_bidirectional` is an extra symbol; a library built
+before it loads for causal workers and is refused by this one.
+`cs1_gemm_shape` is optional. Instant refuses to load without it. Rebuild
+this library before the next sample. Load keeps one cuBLASLt descriptor set
+per weight shape. Layouts are created at the current length. A different
+length replaces that set. Column counts are not rewritten in place. There is
+no quantization and no load-time sweep.
+
+```sh
+src/backends/cuda/qwen3_5/build.sh target/release 89
+cargo build --release --locked -p omni-startlux-native -p omni-jev
+```
+
+`build.sh` needs the CUDA toolkit (`nvcc`) and writes
+`target/release/libqwen3_5_cuda.so`. The second command builds the worker
+and the Rust frontend. There is no merge/export step.
+
+## Weights
+
+```sh
+hf download startlux-models/StartLux-Decision-v1.1-4B-Instant \
+  --revision 416088f77cfce09633906a3675c7cd2e5dabcf10 \
+  --local-dir "$STARTLUX_MODEL"
+```
+
+`STARTLUX_MODEL` is the folder that contains `config.json`,
+`decision_config.json`, `tokenizer.json`, and the safetensors index. The
+worker checks the Instant 4B shapes and refuses an AutoThink or 27B folder.
+
+## Serve
+
+```sh
+STARTLUX_MODEL=/path/to/StartLux-Decision-v1.1-4B-Instant \
+STARTLUX_CUDA_LIB=$PWD/target/release/libqwen3_5_cuda.so \
+STARTLUX_HOST=127.0.0.1 \
+STARTLUX_PORT=8000 \
+  ./target/release/omni-startlux-native
+```
+
+The process loads the checkpoint, runs one warmup decision, then listens.
+Startup prints the CC BY-NC line. `GET /health` reports `ready`, the model
+id, `prompt_attention: bidirectional`, the license, and the weights pin.
+
+| Variable | Default | Meaning |
+| --- | --- | --- |
+| `STARTLUX_MODEL` | required | Instant 4B folder |
+| `STARTLUX_CUDA_LIB` | `libqwen3_5_cuda.so` beside the binary | Shared CUDA library |
+| `STARTLUX_HOST` | `127.0.0.1` | Listen address |
+| `STARTLUX_PORT` | `8000` | Listen port, so `OMNI_JEV_BACKEND_URL=http://127.0.0.1:8000` matches |
+| `STARTLUX_MAX_MODEL_LEN` | `65536` | Reject longer prompts. `1..=262144` |
+| `STARTLUX_NONE_CHECK` | on | `0` turns off the extra "None of the above" option |
+| `STARTLUX_ALLOW_NONE` | off | `1` allows `choice: null` when that extra option wins |
+
+The body limit is 32 MiB. A larger body is 413. The content type must be
+`application/json`. Bad JSON is 400. Contract errors are 422 with
+`error` and `detail`. The worker does not truncate.
+
+Optional frontend, from the same root, against the worker above:
+
+```sh
+OMNI_JEV_BIND=127.0.0.1:8080 \
+OMNI_JEV_BACKEND_URL=http://127.0.0.1:8000 \
+  ./target/release/omni-jev
+```
+
+## Requests
+
+```sh
+curl -sS http://127.0.0.1:8000/health
+
+curl -sS http://127.0.0.1:8000/v1/systemone \
+  -H 'content-type: application/json' \
+  -d '{"state":"I was charged twice.","questions":{"route":{"type":"choice","instructions":"What should happen?","criteria":{"refund":"Refund the charge","keep":"Keep the charge"}}}}'
+
+curl -sS http://127.0.0.1:8000/v1/systemone \
+  -H 'content-type: application/json' \
+  -d '{"state":"I was charged twice.","questions":{"ask":{"type":"noul","instructions":"Is a refund requested?"}}}'
+
+curl -sS http://127.0.0.1:8000/v1/systemone \
+  -H 'content-type: application/json' \
+  -d '{"state":"I was charged twice.","questions":{"severity":{"type":"score","instructions":"How severe is it?","criteria":["low","high"]}}}'
+```
+
+With the defaults, the choice answer includes `none_probability` and the
+choice stays one of the listed options. `think_mode` omitted or `never` is
+the Instant path. `think_mode` `auto` or `always` is 422.
+
+## Checks without a GPU
+
+```sh
+cargo test --locked -p omni-startlux-native --test contract
+```
+
+That test does not load weights or CUDA. It does not establish numerical
+parity.
+
+## Limits
+
+Text only. No AutoThink thought, no 27B, no images, no prefix cache, no
+CUDA graph, no packed prompts. Two GDN parameters are stored FP32 upstream
+and rounded to BF16 at load. Letter logits are a CPU dot of the tied
+embedding and are unverified against vLLM. A causal vLLM server without the
+StartLux bidirectional plugin is a different model; do not use it as the
+Instant baseline. The baseline command is in [latency.md](latency.md).