torch-type-nn 0.1.0.dev0__tar.gz → 0.1.1.dev0__tar.gz

torch_type_nn-0.1.1.dev0/BOARD.CUDA.md

@@ -0,0 +1,183 @@
+# CUDA board
+
+Numbers on this page come from `benchmarks/board.py --write-cuda`. They are
+**not** bit-identical to C `bench.c`; C stays the number-for-number backend
+on [BOARD.md](BOARD.md). This file is the ragged device store
+(`CudaTypeNN`: packed `W (R, n_in)`, fused `X @ W.T`).
+
+`nix flake check --impure` skips `checks.cuda` unless the Nix daemon enables
+the `cuda` system feature. A GPU machine without that feature still runs the
+suite on the host:
+
+```sh
+nix run .#test-cuda          # pytest + xor on the device (prints the probe)
+nix run .#board-cuda         # 5-seed CUDA board + scale, then this file
+```
+
+or step by step:
+
+```sh
+nix run .#bench-cuda -- all cuda-type-nn,cuda-type-nn-overfit,mlp \
+    --seeds 5 --device cuda --jobs 0 \
+    > benchmarks/results/board-cuda.jsonl
+nix run .#scale-cuda -- all --models cuda-type-nn,cuda-type-nn-overfit,mlp-16,mlp-64 \
+    --seeds 5 --device cuda --jobs 0 \
+    > benchmarks/results/scale-cuda.jsonl
+nix run .#board -- \
+    --board benchmarks/results/board-cuda.jsonl \
+    --scale benchmarks/results/scale-cuda.jsonl \
+    --write-cuda BOARD.CUDA.md
+```
+
+Look for `======== torch-type-nn CUDA probe ========` and
+`torch-type-nn device: CUDA …` in the log. A green run with only
+`torch-type-nn device: CPU` did not use a GPU.
+
+## Why C rows disagreed across the two boards
+
+Two different gaps, not one:
+
+1. **C type-nn on BOARD.md vs C type-nn on BOARD.CUDA.md.**
+   `board.py --write-cuda` used to merge `benchmarks/reference/c-type-nn-7333bf7-30seeds.jsonl`
+   (C binary, 30 seeds) into this file. BOARD.md's C rows are
+   **5-seed NativeTypeNN**. Those are not the same experiment. The CUDA
+   writer no longer reads that 30-seed file. Pass
+   `benchmarks/results/board-per-sample.jsonl` next to `board-cuda.jsonl`
+   so both markdown files show the same 5-seed C cells.
+
+2. **C type-nn vs cuda-type-nn (and the overfit pair).**
+   `CudaTypeNN` is the C *layout* (ragged `W (R, n_in)`, fused GEMM) with
+   a *thinner scaler*. It is not `type_nn_scale.c` / `type_nn_overfit_scale.c`.
+   Missing or simplified relative to C: pair-cache BIC, width probes
+   (new unit + consumer column at 0), the fold when a depth probe is
+   inserted, and C's exact identity-layer metric. So CUDA stays closer to
+   birth depth, grows fewer params, and can lose on xor-train / Friedman
+   while C grows. Forward+Adam on a *fixed* graph match C; the structure
+   trajectory does not. C remains the number-for-number backend.
+
+## Per-sample tasks
+
+Same tasks, split, epochs and lr as [BOARD.md](BOARD.md). CUDA rows are
+per-sample Adam on the device (C protocol, fused GEMV). MLP rows in the
+same jsonl were trained with `--device cuda`.
+
+<!-- board:per-sample -->
+Generated by `benchmarks/board.py --write-cuda` from `benchmarks/results/board-cuda.jsonl`, `benchmarks/results/board-per-sample.jsonl` (5 seeds, batch 1, cpu; 5 seeds, batch 1, cuda; cuda type-nn rows from CudaTypeNN).
+
+| task | model | hold MSE | ± sd | hold acc | train MSE | params | ± sd | layers | train s | µs/inf |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| xor | torch ref-type-nn | n/a | n/a | n/a | 8.40e-06 | 24.8 | 1.8 | 1→2.0 | 55.15 | 482.6 |
+| xor | torch ref-type-nn-overfit | n/a | n/a | n/a | 4.90e-06 | 56.0 | 17.0 | 1→2.4 | 39.42 | 245.0 |
+| xor | **torch type-nn** | n/a | n/a | n/a | 5.11e-06 | 63.0 | 23.7 | 1→2.4 | 35.06 | 232.7 |
+| xor | torch type-nn-bic | n/a | n/a | n/a | 8.40e-06 | 24.8 | 1.8 | 1→2.0 | 41.22 | 238.3 |
+| xor | torch MLP | n/a | n/a | n/a | 9.61e-29 | 33.0 | 0.0 | 2→2.0 | 13.40 | 68.1 |
+| xor | torch MLP, scaled | n/a | n/a | n/a | 8.34e-02 | 39.0 | 37.1 | 2→3.2 | 20.45 | 99.3 |
+| xor | cuda type-nn | n/a | n/a | n/a | 6.74e-06 | 24.8 | 1.8 | 1→2.0 | 411.38 | 7308.3 |
+| xor | cuda type-nn-overfit | n/a | n/a | n/a | 1.04e-06 | 73.6 | 36.0 | 1→2.6 | 288.21 | 10705.1 |
+| iris | torch ref-type-nn | 0.0331 | 0.0142 | 0.947 | 6.29e-03 | 144.6 | 27.0 | 3→4.0 | 201.94 | 614.1 |
+| iris | torch ref-type-nn-overfit | 0.0272 | 0.0034 | 0.956 | 8.28e-04 | 159.6 | 33.6 | 3→4.0 | 191.22 | 1070.9 |
+| iris | **torch type-nn** | 0.0258 | 0.0019 | 0.956 | 6.23e-04 | 180.0 | 15.1 | 3→4.0 | 206.52 | 501.9 |
+| iris | torch type-nn-bic | 0.0250 | 0.0062 | 0.964 | 1.37e-04 | 134.8 | 41.1 | 3→4.0 | 221.54 | 975.8 |
+| iris | torch MLP | 0.0309 | 0.0043 | 0.956 | 1.48e-02 | 67.0 | 0.0 | 2→2.0 | 36.89 | 70.0 |
+| iris | torch MLP, scaled | 0.0262 | 0.0044 | 0.956 | 6.08e-03 | 71.8 | 16.3 | 2→3.0 | 63.95 | 178.1 |
+| iris | cuda type-nn | 0.0234 | 0.0040 | 0.960 | 1.99e-04 | 139.8 | 40.5 | 3→4.0 | 1398.00 | 15537.4 |
+| iris | cuda type-nn-overfit | 0.0404 | 0.0354 | 0.956 | 1.57e-02 | 189.8 | 38.3 | 3→4.0 | 1265.22 | 15649.2 |
+| wine | torch ref-type-nn | 0.0198 | 0.0070 | 0.959 | 3.20e-05 | 214.6 | 34.6 | 4→5.0 | 236.12 | 602.9 |
+| wine | torch ref-type-nn-overfit | 0.0364 | 0.0107 | 0.933 | 2.72e-05 | 357.0 | 14.6 | 4→5.0 | 239.31 | 592.6 |
+| wine | **torch type-nn** | 0.0189 | 0.0066 | 0.967 | 6.97e-05 | 486.2 | 15.5 | 4→5.0 | 240.25 | 1301.7 |
+| wine | torch type-nn-bic | 0.0264 | 0.0072 | 0.952 | 4.31e-05 | 194.2 | 72.0 | 4→4.8 | 220.96 | 359.4 |
+| wine | torch MLP | 0.0259 | 0.0032 | 0.956 | 3.87e-03 | 275.0 | 0.0 | 2→2.0 | 39.39 | 144.4 |
+| wine | torch MLP, scaled | 0.0177 | 0.0078 | 0.967 | 3.23e-04 | 124.6 | 32.7 | 2→3.0 | 71.08 | 107.6 |
+| wine | cuda type-nn | 0.0208 | 0.0043 | 0.959 | 2.67e-05 | 214.6 | 44.8 | 4→5.0 | 1641.16 | 19566.2 |
+| wine | cuda type-nn-overfit | 0.0220 | 0.0094 | 0.963 | 1.21e-05 | 470.2 | 52.3 | 4→5.0 | 1476.59 | 19374.9 |
+| wdbc | torch ref-type-nn | 0.0439 | 0.0056 | 0.952 | 4.88e-04 | 175.8 | 36.3 | 3→4.0 | 221.42 | 501.3 |
+| wdbc | torch ref-type-nn-overfit | 0.0435 | 0.0059 | 0.950 | 5.14e-04 | 184.8 | 55.6 | 3→3.6 | 206.28 | 129.5 |
+| wdbc | **torch type-nn** | 0.0473 | 0.0037 | 0.949 | 4.94e-04 | 220.0 | 40.7 | 3→3.6 | 230.46 | 480.6 |
+| wdbc | torch type-nn-bic | 0.0436 | 0.0052 | 0.950 | 1.64e-03 | 210.0 | 83.4 | 3→4.0 | 238.39 | 479.6 |
+| wdbc | torch MLP | 0.0471 | 0.0040 | 0.944 | 5.70e-03 | 513.0 | 0.0 | 2→2.0 | 35.73 | 23.4 |
+| wdbc | torch MLP, scaled | 0.0539 | 0.0079 | 0.943 | 1.84e-03 | 132.0 | 23.3 | 2→3.0 | 91.83 | 85.3 |
+| wdbc | cuda type-nn | 0.0435 | 0.0028 | 0.952 | 9.64e-05 | 207.8 | 34.3 | 3→4.0 | 1529.38 | 4255.6 |
+| wdbc | cuda type-nn-overfit | 0.0475 | 0.0035 | 0.945 | 4.42e-04 | 211.2 | 41.6 | 3→3.4 | 1115.29 | 1265.5 |
+| diabetes | torch ref-type-nn | 0.0328 | 0.0005 | n/a | 2.73e-02 | 18.6 | 4.9 | 2→2.2 | 273.17 | 501.6 |
+| diabetes | torch ref-type-nn-overfit | 0.0366 | 0.0030 | n/a | 2.43e-02 | 31.8 | 7.8 | 2→2.4 | 261.55 | 261.2 |
+| diabetes | **torch type-nn** | 0.0501 | 0.0145 | n/a | 1.75e-02 | 100.0 | 42.6 | 2→3.0 | 244.53 | 788.5 |
+| diabetes | torch type-nn-bic | 0.0330 | 0.0006 | n/a | 2.78e-02 | 18.0 | 2.1 | 2→2.6 | 251.05 | 161.2 |
+| diabetes | torch MLP | 0.0478 | 0.0038 | n/a | 1.47e-02 | 193.0 | 0.0 | 2→2.0 | 69.91 | 32.2 |
+| diabetes | torch MLP, scaled | 0.0410 | 0.0017 | n/a | 2.05e-02 | 61.6 | 10.3 | 2→3.0 | 134.33 | 82.6 |
+| diabetes | cuda type-nn | 0.0328 | 0.0006 | n/a | 2.78e-02 | 17.4 | 1.3 | 2→2.8 | 1680.93 | 11631.7 |
+| diabetes | cuda type-nn-overfit | 0.0453 | 0.0053 | n/a | 1.74e-02 | 98.4 | 22.1 | 2→3.0 | 1703.39 | 11891.0 |
+| ionosphere | torch ref-type-nn | 0.0919 | 0.0212 | 0.900 | 7.40e-03 | 157.2 | 59.2 | 4→5.0 | 261.99 | 629.1 |
+| ionosphere | torch ref-type-nn-overfit | 0.0899 | 0.0199 | 0.898 | 1.61e-03 | 319.6 | 71.2 | 4→4.6 | 243.87 | 880.4 |
+| ionosphere | **torch type-nn** | 0.0854 | 0.0209 | 0.909 | 7.06e-04 | 389.6 | 66.3 | 4→4.8 | 233.66 | 1876.2 |
+| ionosphere | torch type-nn-bic | 0.0837 | 0.0137 | 0.909 | 5.08e-03 | 212.8 | 26.7 | 4→4.8 | 272.71 | 1002.3 |
+| ionosphere | torch MLP | 0.1356 | 0.0128 | 0.908 | 1.57e-02 | 577.0 | 0.0 | 2→2.0 | 46.17 | 122.2 |
+| ionosphere | torch MLP, scaled | 0.1149 | 0.0224 | 0.887 | 8.43e-03 | 167.8 | 17.1 | 2→3.0 | 81.57 | 108.7 |
+| ionosphere | cuda type-nn | 0.1001 | 0.0113 | 0.877 | 2.13e-02 | 167.2 | 124.5 | 4→4.2 | 1711.80 | 11845.0 |
+| ionosphere | cuda type-nn-overfit | 0.1028 | 0.0333 | 0.887 | 1.24e-02 | 336.2 | 35.9 | 4→4.8 | 1755.88 | 10559.1 |
+<!-- /board:per-sample -->
+
+## Comparisons
+
+Hold-out gap is **clear** when it exceeds two standard errors of the
+difference, `2 sqrt(sd_a²/n_a + sd_b²/n_b)`.
+
+<!-- board:pairs -->
+| task | comparison | Δ hold MSE | 2se | verdict | params ratio |
+|---|---|---:|---:|---|---:|
+| xor | cuda type-nn vs torch MLP | n/a | n/a | both fit (no hold-out) | 0.75× |
+| xor | cuda type-nn-overfit vs torch MLP | n/a | n/a | both fit (no hold-out) | 2.23× |
+| xor | cuda type-nn vs cuda type-nn-overfit | n/a | n/a | both fit (no hold-out) | 0.34× |
+| iris | cuda type-nn vs torch MLP | -0.0075 | 0.0053 | lower, **clear** | 2.09× |
+| iris | cuda type-nn-overfit vs torch MLP | +0.0095 | 0.0319 | higher, within noise | 2.83× |
+| iris | cuda type-nn vs cuda type-nn-overfit | -0.0171 | 0.0319 | lower, within noise | 0.74× |
+| wine | cuda type-nn vs torch MLP | -0.0050 | 0.0048 | lower, **clear** | 0.78× |
+| wine | cuda type-nn-overfit vs torch MLP | -0.0039 | 0.0089 | lower, within noise | 1.71× |
+| wine | cuda type-nn vs cuda type-nn-overfit | -0.0011 | 0.0092 | lower, within noise | 0.46× |
+| wdbc | cuda type-nn vs torch MLP | -0.0037 | 0.0044 | lower, within noise | 0.41× |
+| wdbc | cuda type-nn-overfit vs torch MLP | +0.0003 | 0.0047 | higher, within noise | 0.41× |
+| wdbc | cuda type-nn vs cuda type-nn-overfit | -0.0040 | 0.0040 | lower, within noise | 0.98× |
+| diabetes | cuda type-nn vs torch MLP | -0.0150 | 0.0034 | lower, **clear** | 0.09× |
+| diabetes | cuda type-nn-overfit vs torch MLP | -0.0024 | 0.0058 | lower, within noise | 0.51× |
+| diabetes | cuda type-nn vs cuda type-nn-overfit | -0.0125 | 0.0047 | lower, **clear** | 0.18× |
+| ionosphere | cuda type-nn vs torch MLP | -0.0355 | 0.0153 | lower, **clear** | 0.29× |
+| ionosphere | cuda type-nn-overfit vs torch MLP | -0.0328 | 0.0319 | lower, **clear** | 0.58× |
+| ionosphere | cuda type-nn vs cuda type-nn-overfit | -0.0027 | 0.0315 | lower, within noise | 0.50× |
+<!-- /board:pairs -->
+
+## Larger data
+
+Same `scale.py` tasks as [BOARD.md](BOARD.md) (digits, Friedman #1). CUDA
+rows use `CudaTypeNN.epoch` (per-sample on the device). MLP-16 / MLP-64 in
+this file are the batched baseline on the same `--device`.
+
+<!-- board:scale -->
+Generated by `benchmarks/board.py` from `benchmarks/results/scale-cuda.jsonl`, `benchmarks/results/scale-batch32.jsonl` (5 seeds, batch 1, cuda; 5 seeds, batch 32, cpu).
+
+| task | model | hold MSE | ± sd | × noise floor | hold acc | train MSE | params | layers | train s |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| digits | **type-nn** | 0.00620 | 0.00131 | n/a | 0.963 | 0.00023 | 5509 | 7.0 | 32.1 |
+| digits | type-nn-bic | 0.00696 | 0.00028 | n/a | 0.957 | 0.00014 | 3352 | 7.0 | 221.1 |
+| digits | MLP, scaled | 0.00779 | 0.00053 | n/a | 0.964 | 0.00170 | 1323 | 3.0 | 8.5 |
+| digits | MLP-16 | 0.01110 | 0.00074 | n/a | 0.959 | 0.00490 | 1210 | 2.0 | 4.1 |
+| digits | MLP-64 | 0.01089 | 0.00022 | n/a | 0.974 | 0.00343 | 4810 | 2.0 | 5.4 |
+| digits | cuda type-nn | 0.00754 | 0.00059 | n/a | 0.954 | 0.00013 | 3273 | 7.0 | 6584.2 |
+| digits | cuda type-nn-overfit | 0.00663 | 0.00089 | n/a | 0.958 | 0.00011 | 3678 | 6.0 | 5958.5 |
+| friedman | **type-nn** | 0.00660 | 0.00037 | 6.65 | n/a | 0.00666 | 18 | 2.0 | 217.7 |
+| friedman | type-nn-bic | 0.00195 | 0.00089 | 1.97 | n/a | 0.00194 | 80 | 3.0 | 243.6 |
+| friedman | MLP, scaled | 0.01377 | 0.01039 | 13.88 | n/a | 0.01362 | 17 | 2.6 | 96.7 |
+| friedman | MLP-16 | 0.00121 | 0.00013 | 1.22 | n/a | 0.00118 | 193 | 2.0 | 48.1 |
+| friedman | MLP-64 | 0.00114 | 0.00005 | 1.15 | n/a | 0.00110 | 769 | 2.0 | 34.8 |
+| friedman | cuda type-nn | 0.00152 | 0.00032 | 1.53 | n/a | 0.00151 | 79 | 3.0 | 28158.7 |
+| friedman | cuda type-nn-overfit | 0.00670 | 0.00003 | 6.75 | n/a | 0.00677 | 15 | 2.0 | 23214.0 |
+
+friedman: noise floor (best reachable hold-out MSE, `1 / span²`) = 0.00099
+
+| task | comparison | Δ hold MSE | 2se | verdict | params ratio |
+|---|---|---:|---:|---|---:|
+| digits | cuda type-nn vs MLP-16 | -0.0036 | 0.0008 | lower, **clear** | 2.71× |
+| digits | cuda type-nn-overfit vs MLP-16 | -0.0045 | 0.0010 | lower, **clear** | 3.04× |
+| digits | cuda type-nn vs cuda type-nn-overfit | +0.0009 | 0.0010 | higher, within noise | 0.89× |
+| friedman | cuda type-nn vs MLP-16 | +0.0003 | 0.0003 | higher, **clear** | 0.41× |
+| friedman | cuda type-nn-overfit vs MLP-16 | +0.0055 | 0.0001 | higher, **clear** | 0.08× |
+| friedman | cuda type-nn vs cuda type-nn-overfit | -0.0052 | 0.0003 | lower, **clear** | 5.27× |
+<!-- /board:scale -->

{torch_type_nn-0.1.0.dev0 → torch_type_nn-0.1.1.dev0}/BOARD.md

@@ -31,7 +31,33 @@ over the second's.
 | torch ref-type-nn-overfit | C's `type-nn-overfit` exactly: threshold rule, `width_through_depth_probe=False` |
 | torch MLP | the C board's baseline: Linear-ReLU-Linear, 8 hidden units (n ≤ 4) or 16, stock Adam |
 | torch MLP, scaled | `ScalableMLP` grown and pruned by the default (threshold) rule, stock Adam |
-| C type-nn, C type-nn-overfit, C c-mlp | hadilq/type-nn `7333bf7`, run on the same machine as its committed reference (deterministic; reproduces digit for digit) |
+| C type-nn, C type-nn-overfit | `NativeTypeNN` wrapping C `type-nn` / `type-nn-overfit` (`7333bf7`): ragged Ors, C's own scaler; `bench.py` models `c-type-nn` / `c-type-nn-overfit` (`backend=nativetypenn`). One seed matches `bench.c` bit-for-bit. Default board is **5 seeds** (type-nn `BOARD.txt`). `benchmarks/reference/` is a **30-seed** C-binary file and is only the fallback before those cells are run; do not compare 5-seed wrapper means to that 30-seed file. |
+| C c-mlp | hadilq/type-nn `7333bf7` binary board (no Python wrapper) |
+
+### type-nn vs type-nn-overfit (not a wrapper bug)
+
+Both C rows are `NativeTypeNN` calling `7333bf7` `bench.c`. The 5-seed means
+match type-nn `BOARD.txt`. Overfit is **not** "BIC plus extra width so hold-out
+always wins."
+
+- **C type-nn** (`rule="bic"`): a structure stays only if it pays its
+  parameters on the measured training pairs,
+  `n ln(MSE_without/MSE_with) > k ln n`.
+- **C type-nn-overfit** (`rule="threshold"`): probes promote on displacement
+  past `θ(T)=lr T^{3/4}` and drop only when they fall back inside the identity
+  band. That *does* let the graph grow more freely.
+
+On the 5-seed C board that is what happens: overfit has **more params** than
+BIC on xor (108 vs 26), iris (175 vs 125), wine (393 vs 170), diabetes
+(33 vs 20), ionosphere (298 vs 154); wdbc is a tie (~163). Train MSE is
+usually tighter. Hold-out vs BIC is **mixed** (overfit better on wine and
+ionosphere; BIC better on iris, wdbc, diabetes). That is the prior doing
+its job, not a failed grow.
+
+Overfit vs C c-mlp on hold-out: lower on iris, wine, diabetes, ionosphere
+(clear on the last two); within noise on wdbc — with fewer params than the
+MLP except on iris/wine. The type-nn README claim is "fits train ~10× tighter
+than c-mlp at similar hold-out," not "beats BIC on hold-out."
 
 ## Per-sample board
 
@@ -53,57 +79,57 @@ Generated by `benchmarks/board.py` from `benchmarks/reference/c-type-nn-7333bf7-
 | xor | C type-nn | n/a | n/a | n/a | 1.67e-02 | 23.6 | 5.8 | 1→1.9 | 0.01 | 0.1 |
 | xor | C type-nn-overfit | n/a | n/a | n/a | 2.03e-04 | 78.8 | 47.4 | 1→2.7 | 0.01 | 0.2 |
 | xor | C c-mlp | n/a | n/a | n/a | 8.76e-03 | 33.0 | 0.0 | 2→2.0 | 0.00 | 0.0 |
-| xor | torch ref-type-nn | n/a | n/a | n/a | 8.40e-06 | 24.8 | 1.8 | 1→2.0 | 67.24 | 319.7 |
-| xor | torch ref-type-nn-overfit | n/a | n/a | n/a | 4.90e-06 | 56.0 | 17.0 | 1→2.4 | 46.32 | 497.0 |
-| xor | **torch type-nn** | n/a | n/a | n/a | 5.11e-06 | 63.0 | 23.7 | 1→2.4 | 46.60 | 1066.2 |
-| xor | torch type-nn-bic | n/a | n/a | n/a | 8.40e-06 | 24.8 | 1.8 | 1→2.0 | 58.89 | 149.1 |
-| xor | torch MLP | n/a | n/a | n/a | 9.61e-29 | 33.0 | 0.0 | 2→2.0 | 12.39 | 66.9 |
-| xor | torch MLP, scaled | n/a | n/a | n/a | 8.34e-02 | 39.0 | 37.1 | 2→3.2 | 23.44 | 97.2 |
+| xor | torch ref-type-nn | n/a | n/a | n/a | 8.40e-06 | 24.8 | 1.8 | 1→2.0 | 55.15 | 482.6 |
+| xor | torch ref-type-nn-overfit | n/a | n/a | n/a | 4.90e-06 | 56.0 | 17.0 | 1→2.4 | 39.42 | 245.0 |
+| xor | **torch type-nn** | n/a | n/a | n/a | 5.11e-06 | 63.0 | 23.7 | 1→2.4 | 35.06 | 232.7 |
+| xor | torch type-nn-bic | n/a | n/a | n/a | 8.40e-06 | 24.8 | 1.8 | 1→2.0 | 41.22 | 238.3 |
+| xor | torch MLP | n/a | n/a | n/a | 9.61e-29 | 33.0 | 0.0 | 2→2.0 | 13.40 | 68.1 |
+| xor | torch MLP, scaled | n/a | n/a | n/a | 8.34e-02 | 39.0 | 37.1 | 2→3.2 | 20.45 | 99.3 |
 | iris | C type-nn | 0.0248 | 0.0055 | 0.961 | 3.17e-03 | 120.0 | 45.6 | 3→4.0 | 0.38 | 0.6 |
 | iris | C type-nn-overfit | 0.0240 | 0.0053 | 0.961 | 8.43e-04 | 166.2 | 26.7 | 3→4.0 | 0.07 | 0.6 |
 | iris | C c-mlp | 0.0302 | 0.0052 | 0.956 | 1.24e-02 | 67.0 | 0.0 | 2→2.0 | 0.02 | 0.1 |
-| iris | torch ref-type-nn | 0.0331 | 0.0142 | 0.947 | 6.29e-03 | 144.6 | 27.0 | 3→4.0 | 226.46 | 496.2 |
-| iris | torch ref-type-nn-overfit | 0.0272 | 0.0034 | 0.956 | 8.28e-04 | 159.6 | 33.6 | 3→4.0 | 223.01 | 965.7 |
-| iris | **torch type-nn** | 0.0258 | 0.0019 | 0.956 | 6.23e-04 | 180.0 | 15.1 | 3→4.0 | 225.93 | 1105.1 |
-| iris | torch type-nn-bic | 0.0250 | 0.0062 | 0.964 | 1.37e-04 | 134.8 | 41.1 | 3→4.0 | 243.32 | 456.3 |
-| iris | torch MLP | 0.0309 | 0.0043 | 0.956 | 1.48e-02 | 67.0 | 0.0 | 2→2.0 | 41.09 | 65.9 |
-| iris | torch MLP, scaled | 0.0262 | 0.0044 | 0.956 | 6.08e-03 | 71.8 | 16.3 | 2→3.0 | 83.05 | 193.3 |
+| iris | torch ref-type-nn | 0.0331 | 0.0142 | 0.947 | 6.29e-03 | 144.6 | 27.0 | 3→4.0 | 201.94 | 614.1 |
+| iris | torch ref-type-nn-overfit | 0.0272 | 0.0034 | 0.956 | 8.28e-04 | 159.6 | 33.6 | 3→4.0 | 191.22 | 1070.9 |
+| iris | **torch type-nn** | 0.0258 | 0.0019 | 0.956 | 6.23e-04 | 180.0 | 15.1 | 3→4.0 | 206.52 | 501.9 |
+| iris | torch type-nn-bic | 0.0250 | 0.0062 | 0.964 | 1.37e-04 | 134.8 | 41.1 | 3→4.0 | 221.54 | 975.8 |
+| iris | torch MLP | 0.0309 | 0.0043 | 0.956 | 1.48e-02 | 67.0 | 0.0 | 2→2.0 | 36.89 | 70.0 |
+| iris | torch MLP, scaled | 0.0262 | 0.0044 | 0.956 | 6.08e-03 | 71.8 | 16.3 | 2→3.0 | 63.95 | 178.1 |
 | wine | C type-nn | 0.0241 | 0.0070 | 0.956 | 4.21e-04 | 183.2 | 62.2 | 4→5.1 | 0.52 | 0.5 |
 | wine | C type-nn-overfit | 0.0258 | 0.0099 | 0.954 | 2.94e-04 | 378.0 | 47.3 | 4→4.9 | 0.11 | 0.9 |
 | wine | C c-mlp | 0.0243 | 0.0042 | 0.970 | 2.68e-03 | 275.0 | 0.0 | 2→2.0 | 0.03 | 0.2 |
-| wine | torch ref-type-nn | 0.0198 | 0.0070 | 0.959 | 3.20e-05 | 214.6 | 34.6 | 4→5.0 | 249.18 | 589.4 |
-| wine | torch ref-type-nn-overfit | 0.0364 | 0.0107 | 0.933 | 2.72e-05 | 357.0 | 14.6 | 4→5.0 | 239.02 | 1400.0 |
-| wine | **torch type-nn** | 0.0189 | 0.0066 | 0.967 | 6.97e-05 | 486.2 | 15.5 | 4→5.0 | 241.50 | 579.0 |
-| wine | torch type-nn-bic | 0.0264 | 0.0072 | 0.952 | 4.31e-05 | 194.2 | 72.0 | 4→4.8 | 263.42 | 1204.3 |
-| wine | torch MLP | 0.0259 | 0.0032 | 0.956 | 3.87e-03 | 275.0 | 0.0 | 2→2.0 | 37.77 | 71.8 |
-| wine | torch MLP, scaled | 0.0177 | 0.0078 | 0.967 | 3.23e-04 | 124.6 | 32.7 | 2→3.0 | 70.93 | 176.3 |
+| wine | torch ref-type-nn | 0.0198 | 0.0070 | 0.959 | 3.20e-05 | 214.6 | 34.6 | 4→5.0 | 236.12 | 602.9 |
+| wine | torch ref-type-nn-overfit | 0.0364 | 0.0107 | 0.933 | 2.72e-05 | 357.0 | 14.6 | 4→5.0 | 239.31 | 592.6 |
+| wine | **torch type-nn** | 0.0189 | 0.0066 | 0.967 | 6.97e-05 | 486.2 | 15.5 | 4→5.0 | 240.25 | 1301.7 |
+| wine | torch type-nn-bic | 0.0264 | 0.0072 | 0.952 | 4.31e-05 | 194.2 | 72.0 | 4→4.8 | 220.96 | 359.4 |
+| wine | torch MLP | 0.0259 | 0.0032 | 0.956 | 3.87e-03 | 275.0 | 0.0 | 2→2.0 | 39.39 | 144.4 |
+| wine | torch MLP, scaled | 0.0177 | 0.0078 | 0.967 | 3.23e-04 | 124.6 | 32.7 | 2→3.0 | 71.08 | 107.6 |
 | wdbc | C type-nn | 0.0488 | 0.0089 | 0.946 | 2.07e-03 | 153.0 | 36.5 | 3→4.0 | 0.17 | 0.3 |
 | wdbc | C type-nn-overfit | 0.0472 | 0.0077 | 0.948 | 2.85e-03 | 176.6 | 40.5 | 3→3.4 | 0.06 | 0.2 |
 | wdbc | C c-mlp | 0.0459 | 0.0028 | 0.944 | 4.99e-03 | 513.0 | 0.0 | 2→2.0 | 0.09 | 0.3 |
-| wdbc | torch ref-type-nn | 0.0439 | 0.0056 | 0.952 | 4.88e-04 | 175.8 | 36.3 | 3→4.0 | 241.50 | 1055.6 |
-| wdbc | torch ref-type-nn-overfit | 0.0435 | 0.0059 | 0.950 | 5.14e-04 | 184.8 | 55.6 | 3→3.6 | 227.80 | 165.4 |
-| wdbc | **torch type-nn** | 0.0473 | 0.0037 | 0.949 | 4.94e-04 | 220.0 | 40.7 | 3→3.6 | 235.37 | 1032.9 |
-| wdbc | torch type-nn-bic | 0.0436 | 0.0052 | 0.950 | 1.64e-03 | 210.0 | 83.4 | 3→4.0 | 243.94 | 494.9 |
-| wdbc | torch MLP | 0.0471 | 0.0040 | 0.944 | 5.70e-03 | 513.0 | 0.0 | 2→2.0 | 31.36 | 25.5 |
-| wdbc | torch MLP, scaled | 0.0539 | 0.0079 | 0.943 | 1.84e-03 | 132.0 | 23.3 | 2→3.0 | 85.69 | 80.2 |
+| wdbc | torch ref-type-nn | 0.0439 | 0.0056 | 0.952 | 4.88e-04 | 175.8 | 36.3 | 3→4.0 | 221.42 | 501.3 |
+| wdbc | torch ref-type-nn-overfit | 0.0435 | 0.0059 | 0.950 | 5.14e-04 | 184.8 | 55.6 | 3→3.6 | 206.28 | 129.5 |
+| wdbc | **torch type-nn** | 0.0473 | 0.0037 | 0.949 | 4.94e-04 | 220.0 | 40.7 | 3→3.6 | 230.46 | 480.6 |
+| wdbc | torch type-nn-bic | 0.0436 | 0.0052 | 0.950 | 1.64e-03 | 210.0 | 83.4 | 3→4.0 | 238.39 | 479.6 |
+| wdbc | torch MLP | 0.0471 | 0.0040 | 0.944 | 5.70e-03 | 513.0 | 0.0 | 2→2.0 | 35.73 | 23.4 |
+| wdbc | torch MLP, scaled | 0.0539 | 0.0079 | 0.943 | 1.84e-03 | 132.0 | 23.3 | 2→3.0 | 91.83 | 85.3 |
 | diabetes | C type-nn | 0.0331 | 0.0014 | n/a | 2.72e-02 | 20.1 | 5.5 | 2→2.5 | 0.06 | 0.1 |
 | diabetes | C type-nn-overfit | 0.0374 | 0.0027 | n/a | 2.41e-02 | 33.0 | 9.9 | 2→2.5 | 0.03 | 0.1 |
 | diabetes | C c-mlp | 0.0476 | 0.0038 | n/a | 1.38e-02 | 193.0 | 0.0 | 2→2.0 | 0.04 | 0.2 |
-| diabetes | torch ref-type-nn | 0.0328 | 0.0005 | n/a | 2.73e-02 | 18.6 | 4.9 | 2→2.2 | 263.32 | 547.7 |
-| diabetes | torch ref-type-nn-overfit | 0.0366 | 0.0030 | n/a | 2.43e-02 | 31.8 | 7.8 | 2→2.4 | 264.63 | 247.2 |
-| diabetes | **torch type-nn** | 0.0501 | 0.0145 | n/a | 1.75e-02 | 100.0 | 42.6 | 2→3.0 | 311.33 | 367.8 |
-| diabetes | torch type-nn-bic | 0.0330 | 0.0006 | n/a | 2.78e-02 | 18.0 | 2.1 | 2→2.6 | 288.10 | 503.3 |
-| diabetes | torch MLP | 0.0478 | 0.0038 | n/a | 1.47e-02 | 193.0 | 0.0 | 2→2.0 | 71.59 | 33.9 |
-| diabetes | torch MLP, scaled | 0.0410 | 0.0017 | n/a | 2.05e-02 | 61.6 | 10.3 | 2→3.0 | 137.15 | 142.8 |
+| diabetes | torch ref-type-nn | 0.0328 | 0.0005 | n/a | 2.73e-02 | 18.6 | 4.9 | 2→2.2 | 273.17 | 501.6 |
+| diabetes | torch ref-type-nn-overfit | 0.0366 | 0.0030 | n/a | 2.43e-02 | 31.8 | 7.8 | 2→2.4 | 261.55 | 261.2 |
+| diabetes | **torch type-nn** | 0.0501 | 0.0145 | n/a | 1.75e-02 | 100.0 | 42.6 | 2→3.0 | 244.53 | 788.5 |
+| diabetes | torch type-nn-bic | 0.0330 | 0.0006 | n/a | 2.78e-02 | 18.0 | 2.1 | 2→2.6 | 251.05 | 161.2 |
+| diabetes | torch MLP | 0.0478 | 0.0038 | n/a | 1.47e-02 | 193.0 | 0.0 | 2→2.0 | 69.91 | 32.2 |
+| diabetes | torch MLP, scaled | 0.0410 | 0.0017 | n/a | 2.05e-02 | 61.6 | 10.3 | 2→3.0 | 134.33 | 82.6 |
 | ionosphere | C type-nn | 0.0932 | 0.0166 | 0.895 | 1.22e-02 | 146.4 | 60.1 | 4→4.8 | 0.23 | 0.5 |
 | ionosphere | C type-nn-overfit | 0.0811 | 0.0183 | 0.911 | 2.35e-03 | 336.9 | 78.6 | 4→4.4 | 0.10 | 1.0 |
 | ionosphere | C c-mlp | 0.1412 | 0.0172 | 0.899 | 1.59e-02 | 577.0 | 0.0 | 2→2.0 | 0.07 | 0.4 |
-| ionosphere | torch ref-type-nn | 0.0919 | 0.0212 | 0.900 | 7.40e-03 | 157.2 | 59.2 | 4→5.0 | 275.86 | 612.0 |
-| ionosphere | torch ref-type-nn-overfit | 0.0899 | 0.0199 | 0.898 | 1.61e-03 | 319.6 | 71.2 | 4→4.6 | 263.69 | 570.3 |
-| ionosphere | **torch type-nn** | 0.0854 | 0.0209 | 0.909 | 7.06e-04 | 389.6 | 66.3 | 4→4.8 | 263.67 | 1315.4 |
-| ionosphere | torch type-nn-bic | 0.0837 | 0.0137 | 0.909 | 5.08e-03 | 212.8 | 26.7 | 4→4.8 | 280.34 | 474.0 |
-| ionosphere | torch MLP | 0.1356 | 0.0128 | 0.908 | 1.57e-02 | 577.0 | 0.0 | 2→2.0 | 46.46 | 66.3 |
-| ionosphere | torch MLP, scaled | 0.1149 | 0.0224 | 0.887 | 8.43e-03 | 167.8 | 17.1 | 2→3.0 | 82.35 | 88.1 |
+| ionosphere | torch ref-type-nn | 0.0919 | 0.0212 | 0.900 | 7.40e-03 | 157.2 | 59.2 | 4→5.0 | 261.99 | 629.1 |
+| ionosphere | torch ref-type-nn-overfit | 0.0899 | 0.0199 | 0.898 | 1.61e-03 | 319.6 | 71.2 | 4→4.6 | 243.87 | 880.4 |
+| ionosphere | **torch type-nn** | 0.0854 | 0.0209 | 0.909 | 7.06e-04 | 389.6 | 66.3 | 4→4.8 | 233.66 | 1876.2 |
+| ionosphere | torch type-nn-bic | 0.0837 | 0.0137 | 0.909 | 5.08e-03 | 212.8 | 26.7 | 4→4.8 | 272.71 | 1002.3 |
+| ionosphere | torch MLP | 0.1356 | 0.0128 | 0.908 | 1.57e-02 | 577.0 | 0.0 | 2→2.0 | 46.17 | 122.2 |
+| ionosphere | torch MLP, scaled | 0.1149 | 0.0224 | 0.887 | 8.43e-03 | 167.8 | 17.1 | 2→3.0 | 81.57 | 108.7 |
 <!-- /board:per-sample -->
 
 ### Is the port faithful?
@@ -201,23 +227,25 @@ type-nn, stock Adam otherwise), 5 seeds, the same 70/30 xorshift split.
 
 The type-nn and scaled-MLP models are built by the same constructor as the
 per-sample board (`bench.build`); MLP-16 and MLP-64 are fixed one-hidden-layer
-MLPs.
+MLPs. `c-type-nn` and `c-type-nn-overfit` use NativeTypeNN's per-sample C
+epoch on these tasks too (batch size 1, not 32). CUDA rows live on
+[BOARD.CUDA.md](BOARD.CUDA.md).
 
 <!-- board:scale -->
 Generated by `benchmarks/board.py` from `benchmarks/results/scale-batch32.jsonl` (5 seeds, batch 32, cpu).
 
 | task | model | hold MSE | ± sd | × noise floor | hold acc | train MSE | params | layers | train s |
 |---|---|---:|---:|---:|---:|---:|---:|---:|---:|
-| digits | **type-nn** | 0.00620 | 0.00131 | n/a | 0.963 | 0.00023 | 5509 | 7.0 | 38.0 |
-| digits | type-nn-bic | 0.00696 | 0.00028 | n/a | 0.957 | 0.00014 | 3352 | 7.0 | 202.5 |
-| digits | MLP, scaled | 0.00779 | 0.00053 | n/a | 0.964 | 0.00170 | 1323 | 3.0 | 9.0 |
-| digits | MLP-16 | 0.01110 | 0.00074 | n/a | 0.959 | 0.00490 | 1210 | 2.0 | 3.7 |
-| digits | MLP-64 | 0.01089 | 0.00022 | n/a | 0.974 | 0.00343 | 4810 | 2.0 | 4.7 |
-| friedman | **type-nn** | 0.00660 | 0.00037 | 6.65 | n/a | 0.00666 | 18 | 2.0 | 211.7 |
-| friedman | type-nn-bic | 0.00195 | 0.00089 | 1.97 | n/a | 0.00194 | 80 | 3.0 | 254.0 |
-| friedman | MLP, scaled | 0.01377 | 0.01039 | 13.88 | n/a | 0.01362 | 17 | 2.6 | 93.0 |
-| friedman | MLP-16 | 0.00121 | 0.00013 | 1.22 | n/a | 0.00118 | 193 | 2.0 | 56.3 |
-| friedman | MLP-64 | 0.00114 | 0.00005 | 1.15 | n/a | 0.00110 | 769 | 2.0 | 43.2 |
+| digits | **type-nn** | 0.00620 | 0.00131 | n/a | 0.963 | 0.00023 | 5509 | 7.0 | 32.1 |
+| digits | type-nn-bic | 0.00696 | 0.00028 | n/a | 0.957 | 0.00014 | 3352 | 7.0 | 221.1 |
+| digits | MLP, scaled | 0.00779 | 0.00053 | n/a | 0.964 | 0.00170 | 1323 | 3.0 | 8.5 |
+| digits | MLP-16 | 0.01110 | 0.00074 | n/a | 0.959 | 0.00490 | 1210 | 2.0 | 4.1 |
+| digits | MLP-64 | 0.01089 | 0.00022 | n/a | 0.974 | 0.00343 | 4810 | 2.0 | 5.4 |
+| friedman | **type-nn** | 0.00660 | 0.00037 | 6.65 | n/a | 0.00666 | 18 | 2.0 | 217.7 |
+| friedman | type-nn-bic | 0.00195 | 0.00089 | 1.97 | n/a | 0.00194 | 80 | 3.0 | 243.6 |
+| friedman | MLP, scaled | 0.01377 | 0.01039 | 13.88 | n/a | 0.01362 | 17 | 2.6 | 96.7 |
+| friedman | MLP-16 | 0.00121 | 0.00013 | 1.22 | n/a | 0.00118 | 193 | 2.0 | 48.1 |
+| friedman | MLP-64 | 0.00114 | 0.00005 | 1.15 | n/a | 0.00110 | 769 | 2.0 | 34.8 |
 
 friedman: noise floor (best reachable hold-out MSE, `1 / span²`) = 0.00099
 

{torch_type_nn-0.1.0.dev0 → torch_type_nn-0.1.1.dev0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: torch-type-nn
-Version: 0.1.0.dev0
+Version: 0.1.1.dev0
 Summary: type-nn for PyTorch: networks of partition functions (And of Ors) that grow and prune their own width, degree and depth
 Project-URL: Homepage, https://github.com/hadilq/torch-type-nn
 Project-URL: Repository, https://github.com/hadilq/torch-type-nn
@@ -34,7 +34,7 @@ grows and prunes its own width, degree and depth while it trains. The
 design is explained in
 [Train the knowledge](https://hadilq.com/posts/train-the-knowledge/).
 
-> Status: alpha (0.1.0.dev0). Forward, gradients and the per-Or Adam step
+> Status: alpha (0.1.1.dev0). Forward, gradients and the per-Or Adam step
 > agree with the C reference to 1e-12, and both C scaling rules have exact
 > ports on the board (see [Validation](#validation)). Deliberate differences
 > from C are listed in [docs/DIFFERENCES.md](https://github.com/hadilq/torch-type-nn/blob/main/docs/DIFFERENCES.md); the audit
@@ -74,6 +74,34 @@ for epoch in range(E):
 scaler.end()                             # last prune; no probe survives
 ```
 
+### C backend (board rows "C type-nn" / "C type-nn-overfit")
+
+The torch `AndOr` pads every unit to max degree. C keeps one `w[n_in]` per
+live Or. Those models are `NativeTypeNN`, and they are what the board
+calls **C type-nn** and **C type-nn-overfit**:
+
+```python
+from torch_type_nn import NativeTypeNN
+net = NativeTypeNN(4, 3, rule="threshold")   # C type-nn-overfit
+# net = NativeTypeNN(4, 3, rule="bic")       # C type-nn
+```
+
+```sh
+python benchmarks/bench.py iris c-type-nn,c-type-nn-overfit
+```
+
+`CudaTypeNN` is the same ragged layout packed for a fused GEMM on a torch
+device (CUDA when present). Board models `cuda-type-nn` /
+`cuda-type-nn-overfit`:
+
+```sh
+python benchmarks/bench.py xor cuda-type-nn,cuda-type-nn-overfit --device cuda
+```
+
+`nix flake check -L` prints `torch-type-nn device: …` from pytest so a
+GPU run cannot look green without saying which device it used. See
+[docs/BACKENDS.md](docs/BACKENDS.md).
+
 ## Structure learning
 
 One dummy rule on three axes. A **probe** is an identity that back-prop is
@@ -184,7 +212,8 @@ Two independent checks against [hadilq/type-nn](https://github.com/hadilq/type-n
   writes every table of [BOARD.md](https://github.com/hadilq/torch-type-nn/blob/main/BOARD.md) from the result files.
 
 ```sh
-nix run .#board-all      # every result on all cores, then BOARD.md's tables
+nix run .#board-all      # CPU board + scale (now includes C type-nn rows)
+nix run .#board-cuda     # CUDA board → BOARD.CUDA.md (host GPU; see below)
 ```
 
 ## Development
@@ -201,8 +230,17 @@ nix flake check        # build the package, run pytest inside the build (with th
 nix build              # ./result: the installed package
 nix run .#dist         # sdist + wheel in ./dist, checked by twine
 nix run .#publish -- --repository testpypi   # upload (TestPyPI first)
+nix run .#update-deps      # VERSIONS.md; bump patch + PR inputs if a pin moved
 ```
 
+`nix run .#update-deps -- --pr` pushes `deps/update` and opens the PR (needs
+`gh` and `GH_TOKEN`). The same command is the daily cron in
+[`.github/workflows/update-deps.yml`](.github/workflows/update-deps.yml).
+Pins and the resolved nixpkgs versions are in [VERSIONS.md](VERSIONS.md).
+A type-nn rev change copies the vendored C snapshot only (`tnn_py.c` /
+`tnn_py.h` stay); the Python port is not rewritten.
+
+
 The benchmark datasets are pinned (URL + SRI hash) in
 `benchmarks/datasets.json` and never committed; see
 [benchmarks/data/README.md](https://github.com/hadilq/torch-type-nn/blob/main/benchmarks/data/README.md).
@@ -224,6 +262,7 @@ nix flake check --impure         # adds checks.cuda when /dev/nvidiactl exists
 TORCH_TYPE_NN_CUDA=1 nix flake check --impure   # force it (=0: leave it out)
 nix develop .#cuda               # torch-bin (CUDA 13.0) shell
 nix run .#bench-cuda -- all all --seeds 5 --device cuda
+nix run .#board-cuda             # writes BOARD.CUDA.md from device runs
 ```
 
 A plain `nix flake check` is a pure evaluation, which cannot see the host, so
@@ -265,37 +304,7 @@ path that may be missing). With several GPUs add `/dev/nvidia1`, ... Check with
 
 ## Publishing
 
-Releases go to PyPI from GitHub Actions with trusted publishing
-(`.github/workflows/publish.yml`): no API token is stored anywhere.
-
-One-time setup, on PyPI → *Your account* → *Publishing* → *Add a new pending
-publisher* → *GitHub*:
-
-| field | value |
-|---|---|
-| PyPI Project Name | `torch-type-nn` |
-| Owner | `hadilq` |
-| Repository name | `torch-type-nn` |
-| Workflow name | `publish.yml` |
-| Environment name | `pypi` |
-
-For TestPyPI, the same on test.pypi.org (a separate account) with environment
-`testpypi`. GitHub creates both environments on first use; add protection
-rules (e.g. required reviewers) under *Settings → Environments* if wanted.
-
-Release:
-
-```sh
-# 1. set the version in pyproject.toml (e.g. 0.1.0), commit, push
-# 2. rehearse: Actions → publish → Run workflow → testpypi
-# 3. release: the tag must equal the version
-git tag v0.1.0 && git push origin v0.1.0
-```
-
-The workflow builds the sdist and wheel, runs the test-suite against the
-built wheel (CPU torch), and only then uploads. Locally, `nix run .#dist`
-builds and checks the same files; `nix run .#publish` uploads them with
-twine and a token.
+It's in development state, but you can use download it from PyPI [here](https://pypi.org/project/torch-type-nn/).
 
 ## License
 

{torch_type_nn-0.1.0.dev0 → torch_type_nn-0.1.1.dev0}/README.md

@@ -6,7 +6,7 @@ grows and prunes its own width, degree and depth while it trains. The
 design is explained in
 [Train the knowledge](https://hadilq.com/posts/train-the-knowledge/).
 
-> Status: alpha (0.1.0.dev0). Forward, gradients and the per-Or Adam step
+> Status: alpha (0.1.1.dev0). Forward, gradients and the per-Or Adam step
 > agree with the C reference to 1e-12, and both C scaling rules have exact
 > ports on the board (see [Validation](#validation)). Deliberate differences
 > from C are listed in [docs/DIFFERENCES.md](https://github.com/hadilq/torch-type-nn/blob/main/docs/DIFFERENCES.md); the audit
@@ -46,6 +46,34 @@ for epoch in range(E):
 scaler.end()                             # last prune; no probe survives
 ```
 
+### C backend (board rows "C type-nn" / "C type-nn-overfit")
+
+The torch `AndOr` pads every unit to max degree. C keeps one `w[n_in]` per
+live Or. Those models are `NativeTypeNN`, and they are what the board
+calls **C type-nn** and **C type-nn-overfit**:
+
+```python
+from torch_type_nn import NativeTypeNN
+net = NativeTypeNN(4, 3, rule="threshold")   # C type-nn-overfit
+# net = NativeTypeNN(4, 3, rule="bic")       # C type-nn
+```
+
+```sh
+python benchmarks/bench.py iris c-type-nn,c-type-nn-overfit
+```
+
+`CudaTypeNN` is the same ragged layout packed for a fused GEMM on a torch
+device (CUDA when present). Board models `cuda-type-nn` /
+`cuda-type-nn-overfit`:
+
+```sh
+python benchmarks/bench.py xor cuda-type-nn,cuda-type-nn-overfit --device cuda
+```
+
+`nix flake check -L` prints `torch-type-nn device: …` from pytest so a
+GPU run cannot look green without saying which device it used. See
+[docs/BACKENDS.md](docs/BACKENDS.md).
+
 ## Structure learning
 
 One dummy rule on three axes. A **probe** is an identity that back-prop is
@@ -156,7 +184,8 @@ Two independent checks against [hadilq/type-nn](https://github.com/hadilq/type-n
   writes every table of [BOARD.md](https://github.com/hadilq/torch-type-nn/blob/main/BOARD.md) from the result files.
 
 ```sh
-nix run .#board-all      # every result on all cores, then BOARD.md's tables
+nix run .#board-all      # CPU board + scale (now includes C type-nn rows)
+nix run .#board-cuda     # CUDA board → BOARD.CUDA.md (host GPU; see below)
 ```
 
 ## Development
@@ -173,8 +202,17 @@ nix flake check        # build the package, run pytest inside the build (with th
 nix build              # ./result: the installed package
 nix run .#dist         # sdist + wheel in ./dist, checked by twine
 nix run .#publish -- --repository testpypi   # upload (TestPyPI first)
+nix run .#update-deps      # VERSIONS.md; bump patch + PR inputs if a pin moved
 ```
 
+`nix run .#update-deps -- --pr` pushes `deps/update` and opens the PR (needs
+`gh` and `GH_TOKEN`). The same command is the daily cron in
+[`.github/workflows/update-deps.yml`](.github/workflows/update-deps.yml).
+Pins and the resolved nixpkgs versions are in [VERSIONS.md](VERSIONS.md).
+A type-nn rev change copies the vendored C snapshot only (`tnn_py.c` /
+`tnn_py.h` stay); the Python port is not rewritten.
+
+
 The benchmark datasets are pinned (URL + SRI hash) in
 `benchmarks/datasets.json` and never committed; see
 [benchmarks/data/README.md](https://github.com/hadilq/torch-type-nn/blob/main/benchmarks/data/README.md).
@@ -196,6 +234,7 @@ nix flake check --impure         # adds checks.cuda when /dev/nvidiactl exists
 TORCH_TYPE_NN_CUDA=1 nix flake check --impure   # force it (=0: leave it out)
 nix develop .#cuda               # torch-bin (CUDA 13.0) shell
 nix run .#bench-cuda -- all all --seeds 5 --device cuda
+nix run .#board-cuda             # writes BOARD.CUDA.md from device runs
 ```
 
 A plain `nix flake check` is a pure evaluation, which cannot see the host, so
@@ -237,37 +276,7 @@ path that may be missing). With several GPUs add `/dev/nvidia1`, ... Check with
 
 ## Publishing
 
-Releases go to PyPI from GitHub Actions with trusted publishing
-(`.github/workflows/publish.yml`): no API token is stored anywhere.
-
-One-time setup, on PyPI → *Your account* → *Publishing* → *Add a new pending
-publisher* → *GitHub*:
-
-| field | value |
-|---|---|
-| PyPI Project Name | `torch-type-nn` |
-| Owner | `hadilq` |
-| Repository name | `torch-type-nn` |
-| Workflow name | `publish.yml` |
-| Environment name | `pypi` |
-
-For TestPyPI, the same on test.pypi.org (a separate account) with environment
-`testpypi`. GitHub creates both environments on first use; add protection
-rules (e.g. required reviewers) under *Settings → Environments* if wanted.
-
-Release:
-
-```sh
-# 1. set the version in pyproject.toml (e.g. 0.1.0), commit, push
-# 2. rehearse: Actions → publish → Run workflow → testpypi
-# 3. release: the tag must equal the version
-git tag v0.1.0 && git push origin v0.1.0
-```
-
-The workflow builds the sdist and wheel, runs the test-suite against the
-built wheel (CPU torch), and only then uploads. Locally, `nix run .#dist`
-builds and checks the same files; `nix run .#publish` uploads them with
-twine and a token.
+It's in development state, but you can use download it from PyPI [here](https://pypi.org/project/torch-type-nn/).
 
 ## License