[jschmied]

LFM3's backends each kept their own copy of three host-side facts from the family config. This moves them into families/lfm3/config.zig, so every backend reads the same numbers:

• Rotary.inverseFrequencies(): RoPE's inverse frequencies, computed as theta^(2i/128) in float64, rounded to float32, inverted in float32, and zero past the rotated pairs. A test pins both of LFM3's settings to torch 2.14's float32 rows, bit for bit: theta 1e6 with partial 0.5 for the full layers, and theta 1e4 with partial 1 for the sliding layers. A float32 pow, which the Intel GPU copy used, is 1–3 ulp off in 16 of 32 and 28 of 64 entries.
• count(kind): the number of layers of a kind in the admitted schedule (13 conv, 6 full, 5 sliding).
• LayerType.window(): the keys a layer's attention reads back (256 for a sliding layer, 0 meaning all).
• ring(rows): the keys a sliding ring holds when a pass writes its rows before the first of them reads its window.

There are no users on main yet. Two of our open PRs are the first:

• xpu: LFM3-1.7B on Intel GPUs, Iris Xe first (draft, stacked) #1696 (Intel GPU) takes all three. With the exact frequencies, its 64 greedy tokens on the three reference prompts are unchanged, and its float64 layer errors are the same or within 1 %.
• The CUDA side of lfm3: serve LFM3-1.7B on CUDA: NVFP4 and FP8 weight-only, MLX 4-bit (stacked on #1513, #1492 and #1503) #1508 and lfm3 CUDA: serve the block FP8 export weight-only and the MLX 4-bit export as stored (stacked on #1508) #1634 switches to it on its next rebase (local, waiting for lfm2_moe: serve LFM2.5-24B-A2B on CUDA (plain serve, shared kernels) #1492 and lfm2_moe CUDA: attention pages from the shared KV pool, and the prompt cache (stacked on #1492) #1503). There, the served replies of the NVFP4, FP8 and MLX 4-bit exports are byte-identical before and after, and the SASS is unchanged.

Checks on this head: lean has no new findings; zig build test, the contract, golden, security and fuzz steps pass; git diff --check is clean.

[ashhart]

Thank you @jschmied. 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.