[gilby]

What this changes

DeepSeek V4.1's MXFP4 routed experts for decode rows: tf_dsv41_xp_up (fp8 act_quant, clamped SwiGLU, route weight) and the down, plan and combine as a DeepSeek-only copy of Kimi K3's (experts_kimi.metal: tf_dsv41_route_plan, tf_dsv41_xp_down_one/_multi, tf_dsv41_xp_combine), whose down chain is exact over the whole E8M0 scale range. Kimi K3's shipped kernels (kimi/experts.metal, kimi_common.h) are unchanged: git diff main -- zig/kernels/metal/kimi is empty and tf-kernel-contract --only kimi passes. All registered with the kernel contract; docs/kernel-contract.md describes the checks.

Part 3 of 18 of the stack for #1676; stacks on dsv41-metal-pr/02-fp8-kernels (part 2, #1706).

Receipt

• Environment, with the TensorFold commit, MLX or PyTorch version, machine and GPU, checkpoint and revision: TensorFold 61305ae on main 2761c24 (native Zig 0.17.0; no MLX or PyTorch); Apple M3 Ultra (60-core GPU), 96 GB, macOS, for the kernel and device tests; checkpoint deepseek-ai/DeepSeek-V4.1-Flash-1017.
• Exactness, from tools/bench_concurrent.py --alone --serial or token_sha against "draft": false: n/a: kernels only; the kernel contract checks each launch geometry's bits against the others and against float64.
• Decode speed before and after, from tools/bench_openai.py: measured on the PR that turns serving on.
• Prompt speed before and after: the cold first request's ttft_s from tools/bench_cache.py --require-cold on a long prompt, or "n/a" with the reason when the change cannot reach prompt processing: measured on the PR that turns serving on.
• Tests run, with pass, fail and skip counts: zig build test tf-kernel-contract -Dcpu=apple_m1: exit 0, 583/583 steps, 0 fail, 36 skip; kernel contract, tf-kernel-contract --only dsv41 (Metal): SUMMARY 14 kernels: 14 pass gates 1-4, 0 fail, 0 known findings; 20 controls: 20 red as planted, 0 missed; 0 limit problems; gates 40783 ms, limits 129 ms; tf-kernel-contract --only kimi: SUMMARY 1 kernels: 1 pass gates 1-4, 0 fail, 0 known findings; python3 tools/zig/lean_check.py: 4838 files, 424 problems, the same count as main (4807 files, 424). Zig's summary counts only the test binaries it ran in that invocation, so the step count is the comparable figure.
• Not run, and why: CUDA: no CUDA code changes; Metal on M1, M2, M4 and M5: one M3 Ultra only (the contract's limits check covers the M1 to M5 pipeline limits).

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.
• 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.
• 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.
• 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.

[gilby]

On the Kimi K3 point: this part no longer touches Kimi's shipped kernels. DeepSeek has its own copy of the expert plan, down chains and combine (deepseek_v41/experts_kimi.metal, kernels tf_dsv41_route_plan, tf_dsv41_xp_down_one/_multi, tf_dsv41_xp_combine), with the full-range E8M0 scale in the copy only; git diff main -- zig/kernels/metal/kimi is empty. The stack is rebased on current main (2761c24) with main's shared files changed only by one-line hooks, so later parts add to the family's own files. Receipts for #1705, #1706 and this part are updated in their descriptions.

[ashhart]

Thank you @gilby. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @gilby! DeepSeek V4.1's MXFP4 expert kernels on Metal, under the kernel contract, are in 1.2.1: https://github.com/ashhart/TensorFold/releases/tag/v1.2.1