train 1112114823: #2673 #2843 (CI only)

CHANGELOG.md

@@ -4,11 +4,13 @@ Each release's page on GitHub has its notes and measurements. See [the 1.0.0 rel
 
 ## Unreleased
 
-- Cortheon (`--cortheon`): DuckDuckGo results are kept, and its ads still skipped, when only the spelling differs (either quote style, a `class` attribute in any letter case, an unquoted `href`, spaces around its `=`, or a line break before it); a duplicate whose scheme or host differs only in letter case is dropped while paths and queries keep their case, and the PyPI, npm and crates.io allow-list takes the host in any case with the path exact (#2750). Thanks to @ppardus.
+- CUDA EXL3 grouped experts launch and size their storage at the caller's tile and pair count, so decode no longer launches three quarters of its grid empty, and Qwen3.6-35B-A3B EXL3 sizes its grouped storage for 36,864 pair rows instead of 163,904 (about 3 GB less); outputs are bit-equal (#2843, for #2823).
 
-- GLM-5.3-Flash on CUDA: with `--vision` and no video library, rank 0 says at startup how much each stream sets aside for `--vision-video-tokens` rows that no request can fill (128 MiB a stream at the default 16,384), and that `--vision-video-tokens 64` on every rank cuts it to 512 KiB (#2813). Thanks to @Bizuayeu.
+- CUDA: dots.mocr's text decoder serves text prompts on a GB10 from the publisher's BF16 weights, on a new shared engine for dense GQA decoders (`zig/src/cuda_dense/`); `--vision`, prompt caching and fp8 caches are refused by name, and served CUDA checkpoints are unchanged (#2673, part of #1927). Thanks to @laudney.
 
+- Cortheon (`--cortheon`): DuckDuckGo results are kept, and its ads still skipped, when only the spelling differs (either quote style, a `class` attribute in any letter case, an unquoted `href`, spaces around its `=`, or a line break before it); a duplicate whose scheme or host differs only in letter case is dropped while paths and queries keep their case, and the PyPI, npm and crates.io allow-list takes the host in any case with the path exact (#2750). Thanks to @ppardus.
 
+- GLM-5.3-Flash on CUDA: with `--vision` and no video library, rank 0 says at startup how much each stream sets aside for `--vision-video-tokens` rows that no request can fill (128 MiB a stream at the default 16,384), and that `--vision-video-tokens 64` on every rank cuts it to 512 KiB (#2813). Thanks to @Bizuayeu.
 
 - CUDA: `Qwen/Qwen3-Embedding-0.6B`, `-4B` and `-8B` load on the native engine from the published BF16 weights and serve `/v1/embeddings` (last-token pooling, unit-normed, `dimensions` crops). On one GB10 the worst cosine distance to float32 is 1.55e-4, 1.34e-4 and 1.08e-4 (transformers' own BF16: 1.88e-4, 1.79e-4 and 1.77e-4), a request's vectors equal each input sent alone, and a reranker folder or a flag the family does not serve is refused by name (#2285). Thanks to @Arminova.
 - Qwen3.8-27B on Metal: the startup line says how many streams draft ("for both streams", "for all 4 streams", "for 8 of 16 streams"), the smaller of the stream count and `TF_Q27_DRAFT_SLOTS`, instead of "for one of them"; log text only (#2806).

README.md

@@ -177,7 +177,7 @@ Conversions are explicitly marked: their upstream weights need the linked conver
 
 | Model / format | Download | Machine and memory scope | Native recipe / limits |
 | --- | --- | --- | --- |
-| dots.mocr, BF16 OCR | Use the local `dots-studio/dots.mocr` checkpoint. | Native HIP; AMD R9700 (`gfx1201`) checked, one GPU. | [Native OCR recipe](docs/recipes/dots-mocr-hip.md); full model-card demo matches HF tokens. No video, MTP head, CUDA, or Metal qualification. |
+| dots.mocr, BF16 OCR | Use the local `dots-studio/dots.mocr` checkpoint. | Native HIP; AMD R9700 (`gfx1201`) checked, one GPU. | [Native OCR recipe](docs/recipes/dots-mocr-hip.md); full model-card demo matches HF tokens. Native CUDA serves text prompts only ([CUDA recipe](docs/recipes/dots-mocr-cuda.md)). No video, MTP head, or Metal qualification. |
 | Qwen3.5-2B, MLX 4-bit | `tensorfold pull mlx-community/Qwen3.5-2B-MLX-4bit@93760be4f1f69842a46bc13dbdc0f19e291392a3` | One Mac (M5 Max checked), or one GB10; no RAM minimum published. | [2B](docs/recipes/qwen3.5-2b-native.md); [CUDA release checks](CHANGELOG.md#110). Text only. |
 | Qwen3.8-27B, MLX 4-bit | `tensorfold pull TensorFold/Qwen3.8-27B-MLX-4bit`<br>`tensorfold pull z-lab/Qwen3.8-27B-DFlash2` | One Mac (M5 Max/M3 Ultra), or one GB10; a 32 GB Mac needs `iogpu.wired_limit_mb` raised (the startup refusal names the value). | [27B](docs/recipes/qwen3.8-27b.md). Add `--drafter z-lab/Qwen3.8-27B-DFlash2`; without it decoding is plain. |
 | Nemotron 3.5 Lightning 30B-A3B, MLX 4-bit | `tensorfold pull TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit` | One Mac (M1-M5), or one GB10; also checked on RTX 3090 / RTX PRO 6000. | [Lightning](docs/recipes/nemotron-3.5.md). Includes MTP. |

docs/recipes/README.md

@@ -8,6 +8,7 @@ For agent-assisted contributions, read the [TensorFold contributor skill](SKILL.
 | Family | Recipe |
 | --- | --- |
 | dots.mocr | [Native HIP OCR, BF16, one GPU](dots-mocr-hip.md) |
+| dots.mocr | [Native CUDA text decoder, BF16, one GPU](dots-mocr-cuda.md) |
 | Kimi K3 / K3.1 | [Native Metal bring-up, synthetic only, resident budget refusal](kimi-k3-metal.md) |
 | Nemotron 3.5 Lightning | [MLX](nemotron-3.5.md), [Intel Arc, experimental](xpu-arc.md) |
 | Nemotron 3.5 Parse | [Native HIP OCR, validation and qualification limits](nemotron-parse.md) |

docs/recipes/dots-mocr-cuda.md

@@ -0,0 +1,37 @@
+# dots.mocr on native CUDA
+
+The native CUDA engine serves dots.mocr's Qwen2 decoder for text prompts
+through the shared dense engine (`zig/src/cuda_dense/`), from the BF16
+`dots-studio/dots.mocr` checkpoint at revision
+`e539fbb52280393adc081b289ec597430a0f9031`. It is qualified on a GB10
+(sm_121) only. Images are not served on CUDA yet: `--vision` is refused
+by name. The HIP engine, with the image tower, is in
+[dots.mocr on native HIP](dots-mocr-hip.md).
+
+```sh
+zig build native -Dnvcc=/usr/local/cuda/bin/nvcc
+zig-out/native/bin/tensorfold-native serve MODEL_DIRECTORY \
+  --context 8192 --parallel 8 --name dots --host 127.0.0.1 --port 8090
+```
+
+Replies carry no reasoning block: all text is answer content. The engine
+keeps no prompt states, so `--prompt-cache-gib` is refused; so are
+`--drafter`, `--slide`, `--kv-cache fp8` and `--prompt-precision int8`.
+The engine's own copy drafts and `--no-drafts` give the same tokens.
+
+| Responsibility | Files |
+|---|---|
+| Lanes, pages, attention, draws, the host's declarations | `zig/src/cuda_dense/` |
+| Geometry, weights, one layer's arithmetic | `zig/src/families/dots_mocr/cuda_recipe.zig`, `cuda_weights.zig` |
+| q, k and v biases with the rotary in Torch's rounding | `zig/kernels/cuda/dots_rope.cuh` (`tf_dots_qkv_rope`) |
+
+The text reference runs the pinned oracle's `text` mode on any HIP or CUDA
+GPU, and `score` classifies every native id by HF's teacher-forced logits
+(exact, one BF16 step, or a failure, which exits nonzero):
+
+```sh
+python3 tools/zig/dots_mocr_reference.py text MODEL_DIRECTORY REFERENCE \
+  --prompt-file PROMPT.txt --pci PCI_ADDRESS
+python3 tools/zig/dots_mocr_reference.py score MODEL_DIRECTORY REFERENCE \
+  NATIVE_IDS.json DIFFERENCES.json
+```

docs/recipes/dots-mocr-hip.md

@@ -3,7 +3,8 @@
 This implementation serves the BF16 `dots-studio/dots.mocr`
 checkpoint at revision `e539fbb52280393adc081b289ec597430a0f9031`.
 It runs the image tower and Qwen2 decoder in native Zig/HIP. Python is
-used only for the validation reference. CUDA and Metal are not qualified.
+used only for the validation reference. CUDA serves the text decoder only
+([dots.mocr on native CUDA](dots-mocr-cuda.md)); Metal is not qualified.
 
 Build for the local GPU architecture. The checked host uses `gfx1201`: