[drowzeys]

What this changes

Adds families/lfm3_omni, LiquidAI's LFM3-Omni-2B (speech in, streamed text and 24 kHz speech out) as a standalone family in Drama 3's shape, with its Metal backend, host tests, kernel-contract registrations, recipe and the checks' clients under tools/lfm3_omni/. Part of #1627 (the Metal half; the CUDA backend is a separate branch stacked on this one). It's a draft because the gate in #1627 isn't agreed yet.

• zig build lfm3-omni-generate (macOS): one turn from a WAV, stage dumps (--dump, --force, --depth-check) against a reference's, and --serve HOST:PORT with POST /v1/audio/chat (SSE text deltas and 16-bit 24 kHz PCM, or one JSON) and GET /health, one request at a time.
• The family needs only main: the core, Drama 3's WAV module, and the prefill GEMM's tensor-unit blocks. It does not depend on the LFM3 text series (lfm3: serve LFM3-1.7B on CUDA: NVFP4 and FP8 weight-only, MLX 4-bit (stacked on #1513, #1492 and #1503) #1508, lfm3: admit official MLX text and VL checkpoints #1513, lfm3: serve the MLX text decoder on Metal #1514).
• Weights: the bf16 release as stored, revision 5797d07c125503bd4f4c0e7f915ae2d3d703e8de, LFM Open License v1.0 (commercial use only under US$10M annual revenue; the recipe says so).

Receipt

• Environment: this branch's head; M5 Ultra (256 GB), macOS 27.0.1, Xcode 27.1, Zig 0.17.0; reference transformers 5.21.0 with the checkpoint's own code, PyTorch 2.14.1, MPS bf16, same Mac. The 16 synthetic turns of tools/lfm3_omni/utterances.txt (macOS voices), a voice-assistant system text.
• Stage parity (greedy, identity codebook order, tools/lfm3_omni/parity.sh):

Stage	Metal vs MPS bf16
Log-mel, cosine (min), max abs	1.000000, 3.1e-4
Encoder / projector, minimum cosine	0.99999 / 0.99990
Prompt token ids	16/16 identical
Backbone prefill rows, median cosine (min of medians)	0.99982 (0.99977)
Text top-1, teacher-forced	520/524 (99.2%)
Depth decoder on the reference's own states, greedy agreement; logits cosine median / min	95.5%; 0.999994 / 0.99989
Codes, teacher-forced through the whole loop	89.2%
Codec on the reference's codes, waveform SNR median (min)	36.6 dB (25.1 dB)

• The port refactor (files split under 600 lines, one-line comments, the attention fix below) changes no output: greedy dumps and teacher-forced runs of this branch and of the pre-split build are byte-identical on all 16 turns (208 dump files), and the served replies at the card's settings are byte-identical too.
• Latency, the same 16 turns through the server with tools/lfm3_omni/speech_chat_client.py (temperature 0.4, penalty 1.2, seeds 1000 + i); the reference through tools/lfm3_omni/reference_speech.py (first audio = first sampled frame plus that frame's codec decode). Two rounds, this branch, the pre-split build and the reference alternated; other services were resident on the Mac but idle:

Median	This branch	Pre-split build	Reference (MPS bf16)
First text	0.034 s	0.034 s	n/a
First audio	0.063 s	0.063 s	0.172 / 0.191 s
Whole reply (6.5 s of speech)	0.682 / 0.670 s	0.645 / 0.695 s	2.535 / 2.409 s

• Quality of the replies (UTMOS22 strong via SpeechMOS; WER = the checkpoint's own ASR on each reply's audio against its streamed text): TensorFold 4.43 median (4.35 min), WER 1.5%; reference 4.44 (4.15), WER 1.2%.
• Prompt speed: n/a against a previous release (a new family). Prefill is inside "first text" above.

Kernel contract (Metal, M5 Ultra)

All 45 pipelines of lfm3_omni.metallib are registered (69 registrations over the stacks' variants): gates 1-4 against a float64 replay of each kernel's bf16 rounding points, stated rules for the picks and word moves, a planted source change per registration, and gate-5 launch records with M1/M2 limits from --offline.

tf-kernel-contract --only "lfm3 omni":
SUMMARY 69 kernels: 69 pass gates 1-4, 0 fail, 0 known findings; 69 controls: 69 red as planted, 0 missed; 0 limit problems; gates 13690 ms, limits 7242 ms
tf-kernel-contract (every family), exit 0:
SUMMARY 378 kernels: 378 pass gates 1-4, 0 fail, 3 known findings; 414 controls: 414 red as planted, 0 missed; 0 limit problems; gates 2444558 ms, limits 1229 ms

The contract found one real bug, fixed here before the receipt: om_attention_tiles multiplied masked keys' zero weights into the value sum, so a non-finite value row a query cannot see turned its output into NaN. The value loop now covers only the row's visible keys. Finite outputs are unchanged, byte for byte. The full suite ran through tf-kernel-contract --rotary-accuracy with a tf-rotary-sweep receipt, because zig build kernel-contract stops first at Flash Next's UnrecordedRotaryCompiler precondition, which this change doesn't touch.

Two findings that no gate fails, noted for review: in sampled mode om_depth_pick maps h >> 40 == 2^24 - 1 to u = 1 (about 2^-24 per candidate), and om_logmel returns the floor for a NaN log (registered as its stated rule).

Lean checker and tests

• python3 tools/zig/lean_check.py: main 4571 files, 425 problems; this branch 4600 files, 425 problems. Zero added; the 425 are inherited.
• zig build test test-golden -Dcpu=apple_m1: 585/585 steps, 8484 passed, 0 failed, 52 skipped (exit 0).
• zig build test-kernel-contract: 139/139 passed.
• zig build dev-tools -Dcpu=apple_m1: 215/215 steps (exit 0).
• zig build lfm3-omni-generate -Doptimize=ReleaseFast -Dcpu=apple_m1: builds.

Not run, and why

• Metal M1 to M4: no such hardware here. Gate 5 covers them through --offline compiles only. On chips without tensor units, products of many rows take the simdgroup-matrix kernels, which are registered and pass on the M5 but are untested on those chips and slow on long prompts.
• Drafts, resume and concurrency contracts: n/a. The tool serves one request at a time outside the lane core (as Drama 3's does). Question 2 in LFM3-Omni-2B on Metal and CUDA: speech in, streamed text and speech out (gate and split) #1627 asks whether it should instead become a route of tensorfold-native serve.
• Streamed vs whole codec decode differ (34 to 43 dB SNR, the one-frame chunk takes a different product): question 3 in LFM3-Omni-2B on Metal and CUDA: speech in, streamed text and speech out (gate and split) #1627.

From elsewhere

• om_common.h's erf (for the exact GELU) uses the Chebyshev erfcc coefficients published in Numerical Recipes. Say if you'd rather have a different erf.
• tools/lfm3_omni/score_replies.py loads UTMOS22 strong from tarepan/SpeechMOS v1.2.0 through torch.hub (a development tool, not engine code).

Checklist

• A new GPU kernel is registered with the kernel contract; the Metal SUMMARY lines are above.
• No precision traded for speed; bf16 rounding points follow the reference op for op.
• Comments and docstrings are one line; no measurements or history in the source.
• No personal data, machine names, internal hosts or local paths.
• Lane-round, drafted, resumed and concurrent contracts: n/a for this one-request tool; see LFM3-Omni-2B on Metal and CUDA: speech in, streamed text and speech out (gate and split) #1627 question 2.

[drowzeys]

atc overlap --json (ATC 0.7.0) against main 3f0a715. Of 72 overlapping open PRs, four conflict, all where both sides append to a shared list:

• minimax_m3 (CUDA): draft several streams' DSpark blocks in one pass #1426: zig/src/core/contract/registry.zig, zig/tests/contract/metal_main.zig and zig/tests/contract/metal_limits.zig, where both register kernels.
• llama: Nanbeige4.3-33B on Metal, the bf16 release served as published, every kernel in the contract #1415: the kernel table in docs/kernel-contract.md and the index row in docs/recipes/README.md.
• Muse Spark Open on CUDA: EXL3, DFlash2, two GPUs #1295 and MiMo-V2.6-Flash on CUDA: EXL3, DFlash, two GPUs #728: the index row in docs/recipes/README.md.

None touches the same code. Each resolves by keeping both entries, whichever lands first. The rest report unknown because of ATC's pullFilesLimit, so I also ran git merge-tree against every open PR that merges cleanly with main: no further conflicts. #1627 is now claimed through ATC.