[wesley4546]

What this changes

Makes the native Windows build from docs/WINDOWS.md compile and link again, in two commits. Explicit kernel instantiations now repeat their definitions' __restrict__, and GLM's openRank sets its NCCL defaults through sys.c. Fixes #1210.

• cuda: explicit instantiations repeat their kernels' __restrict__, so nvcc with cl.exe compiles them: 103 instantiations in 20 .cu files, some inside #defines. Only parameter lists change; the lines are rewrapped where they passed 120 columns. exl3_block/x3ld_epi.cu is fixed through its includes.
• glm cuda: openRank's NCCL defaults go through sys.c, so tensorfold-native links on Windows: env is @import("sys").c on Windows (the _putenv_s wrapper in win_clock.zig) and the existing extern "c" fn setenv elsewhere, the pattern heat.zig and flashnext/cuda_lanes.zig already use.

On Linux the change is a no-op: all 125 release fatbins are byte-identical before and after (below).

Receipt

• Environment:
  • TensorFold c9fab624 plus these two commits.
  • Windows: Windows 11 x86_64, RTX 5090 (sm_120), driver 610.88, Zig 0.17.0, MSVC 19.44, CUDA 13.0.2 (V13.0.88). 13.4 (V13.4.59) was also checked for the kernel compile.
  • Linux: DGX Spark (GB10), Ubuntu 24.04 aarch64, GCC 13.3, CUDA 13.0 (V13.0.88), Zig 0.17.0.
  • Checkpoint: mlx-community/Qwen3.5-2B-MLX-4bit@93760be4f1f69842a46bc13dbdc0f19e291392a3.
• Exactness:
  • Linux: zig build fatbins fatbins-dsv41 -Dsm=80,86,120,121 at c9fab624 and at this branch, each in its own worktree and cache. Both give the same 125 fatbins, all byte-identical (cmp), including the 21 built from the changed files.
  • Windows (--context 8192 --no-thinking, greedy, drafts on): for one prose and one code prompt, a repeated request and a "draft": false request return the same text as the drafted one. main can't build on Windows, so there is no before.
• Decode speed: n/a. The Linux fatbins don't change, and on Windows there is no before to compare.
• Prompt speed: n/a, for the same reason.
• Tests run:
  • Windows, zig build test test-kernel-contract test-golden fuzz-cortheon: 283/283 steps, 3290 passed, 0 failed, 62 skipped.
  • Windows, plain zig build: passes.
  • zig fmt --check and git diff --check: clean.
  • tools/zig/lean_check.py: 447 findings on c9fab624 and 447 on this branch, none new.
  • tf-cuda-test on the RTX 5090, all passing:

suite	result
info	sm_120, 170 SMs, kernels embedded
smoke	7/7
graph	3/3
launch-ex	3/3
ptx	1/1
symbols	1/1
cublaslt	1/1
norms, norm-rows	1/1, 1/1
sample	2/2
glue	1,470/1,470 byte-equal
quant-ref	4/4
affine8	1/1
modelopt	5/5
exlr8	7/7

• Not run, and why:
  • kernel-contract on sm_120: not run yet. No kernel is added, and the bits don't change on Linux.
  • Metal: I have no Apple hardware, and no Metal code changes.
  • Speed benchmarks: there is no Windows baseline, and Linux is byte-identical.

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (Linux images unchanged; drafted equals "draft": false checked on Windows.)
• No precision traded for speed. If the bits change, the description says which sums change and why.
• A new or changed GPU kernel is registered with the kernel contract (docs/kernel-contract.md); each backend's SUMMARY line is in the receipt. (No new kernel, and the images don't change. sm_120 not run, see above.)
• Prompt processing is no slower than the last release.
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No new tests; see the note on CI below.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

Notes

• Open PRs touching the same code:
  • qwen CUDA: long-prompt prefill 5.4x faster (4096-row chunks, tensor-core attention, the delta-rule chain) #1080, qwen CUDA: decode attention on tensor cores (fp8 and bf16 caches), decode after long prompts 1.5-2.3x faster #1168 and qwen CUDA: --prompt-precision int8, prompt matmuls on int8 activations (cold prefill 20-27% shorter) #1176 add instantiations of chain_kernel, hpattn_kernel, hpattn8_kernel and prefill_kernel without __restrict__, so they will need the same treatment.
  • glm cuda: rank 0's NCCL id goes over the ranks' lines, so a /32 ring's followers bootstrap on NCCL_SOCKET_IFNAME #1178 adds extern "c" fn unsetenv next to setenv and conflicts textually with this. The resolution is to declare unsetenv inside the non-Windows env struct and call env.unsetenv; sys.c already wraps it on Windows.
• The Windows CI job runs host tests only, so neither break showed there. A job that compiles the fatbins with nvcc and links tensorfold-native would catch both.
• With CUDA 13, cublasLt64_13.dll is in bin\x64, and tf-cuda-test cublaslt reports CublasLtUnavailable until that folder is on PATH.