fix(MODEL-MUSIC-MUSIC3): the e2e gate asked for 60 s of music, not 0.1 s — one JSON key, and a weight-load stall that was never real (#852, #925)

[localai-bot]

MiniMax-Music3 generates a song end to end, observed

POST /v1/audio/speech -> 200 audio/wav, 12332 bytes
request shape: 2 AR frames -> 6 latent frames -> 3072 samples/channel (0.069660 s)
waveform:      6144 int16 samples, 6144 non-zero, 0 clipped, range [42, 156],
               RMS 0.00267716, 2818 of 3072 positions differ between L and R

Re-read independently of the gate at build/music3/minimax_music3_e2e_http.wav:
44100 Hz, 2 channels, 16-bit, 3072 frames, 0.069660 s, peak 156, 67 distinct
left-channel values.

The defect is one JSON key

The request body posted "audio_duration_s": 0.1. That is the name of the
field (speech_api.h:55), not the wire key. ParseSpeechRequest reads
audio_duration (or duration) — what docs/USAGE.md:1235 documents and what
every other test in the tree uses. The key was dropped silently,
audio_duration_s stayed at its 0.0 sentinel, and Music3ResolveRequest
substituted kMusic3DefaultDurationSeconds = 60.0.

So the gate asked for a sixty-second song:

	asked for	actually ran
duration	0.1 s	60 s
AR frames	2	1500
denoise windows	1	8
vocoder latents	6	5167
samples/channel	3072	2645504

A ~750x job — and one that could never have passed anyway, because the case
asserts the payload length 0.1 s implies.

The recorded diagnosis was wrong, and how it was wrong generalizes

Four runs (85 min, 34 min, 1 h 49 m, and one of mine at 15 h 19 m) were
placed "inside LoadQwen3ForCausalLMWeights" on the evidence that the four
language_model/*.safetensors fds were open and the depth decoder's was not.

Neither fact says that. LoadBf16Direct borrows the mapping when it can
(BorrowStTensorBytes), and a borrowed OwnedTensor holds its own
shared_ptr to it — so those fds stay open for the whole request, long after the
loader returned. The depth decoder's tensors are copied, so its
SafetensorsFile dies and its fd closes on return. The fd pattern is a
copy-versus-borrow artifact, not a program counter.

What resolves it is a symbol-resolved profile — perf record -g on the live
process, 173K samples, symbols intact because the binary was not rebuilt under
it:

phase	span	signature
LM weight load	0 → 180 s	1 thread, state D, ~92 MB/s off the NAS, 3.7 s user CPU
AR generate	180 → 19154 s	20 threads, LinearNoBias 42-57%, Threadpool::Barrier 25%, Bt16Avx512 12-14%
acoustic	19154 s → killed at 15 h 19 m	vocoder1d::ConvTranspose1d 88.5%, Conv1d 7.7%

The load is 3 minutes, in both binaries — the same ~3 minutes
test_minimax_music3_llm_real takes — and it is I/O bound, not spinning. The
__sched_yield frames in the earlier profile are Threadpool::PollForWork's
bounded hybrid poll around a genuinely compute-bound run.

Two supporting measurements, taken rather than argued: the 200000 x 4096
lm_head scalar transpose that was suspected costs 1.388 s, and the vocoder
at this request's real shape costs 5.4 s (against 84 s at the capture's
L=86).

The guard

A silently dropped knob is what made this invisible for four runs, so
audio_duration_s is now refused and named — like every other unsupported
field in that function, whose own comment already said ignoring one "returns a
200 carrying audio the caller did not ask for".

RED first: the two CHECK_THROWS in test_speech_api.cpp did not throw.

test_speech_api.cpp:144: ERROR: CHECK_THROWS( ... "audio_duration_s":0.1 ... ) did NOT throw at all!
test_speech_api.cpp:145: ERROR: CHECK_THROWS( ... extra_params ... ) did NOT throw at all!
[doctest] test cases: 1 | 0 passed | 1 failed | assertions: 21 | 19 passed | 2 failed

The guard is narrow on purpose: refusing every unknown key would break the
extra_params bag's forward compatibility, while this one near-miss is the
spelling the struct itself invites (#925).

Not changed

LinearNoBias keeps its reduction order and its -ffp-contract=off scalar
triple loop — W2/W3 gate it against torch. It is 42-57% of the AR profile and
parallelising it over output rows would not move a gated number, but that is a
separate, measured change with its own evidence. No speed number is claimed:
CPU only, dgx.casa down.

Gates

All on this box (20-core x86_64, CPU only), CASE and assertion counts:

gate	cases	assertions
test_minimax_music3_e2e_real (both env vars)	5	535
test_speech_api	5	56
test_minimax_music3_loader	21	1413
test_minimax_music3_ar	25	338
test_minimax_music3_ar_real	4	894
test_minimax_music3_acoustic	27	265
test_minimax_music3_acoustic_real	6	985
test_minimax_music3_speech	9	223
test_minimax_music3_llm_real	4	220
test_minimax_music3_quant	29	125
test_minimax_music3_quant_real	6	319
test_minimax_h3	79	57395
test_indextts2_family	7	22
test_openai_api_server	61	709
test_capi	63	617
test_qwen3_paged_engine	2	0 — SKIP, Qwen3-0.6B/4B snapshots are dgx-only

scripts/agent-preflight.sh --no-require-role is green except
test_cpu_x86_llamacpp_floor, the known load-dependent #618 (4 != 2,
NO_QUIET_WINDOW instead of GIVING_UP), which the same tree passed 25
minutes earlier at lower load.

Records

.agents/model-matrix.md, docs/FEATURES.md and docs/STATUS.md conflicted on
merge and were rebuilt from main's version with only my rows reapplied; every
unrelated key was diffed against origin/main and is byte-identical (2 rows, 2
rows and 1 row changed respectively). The tree was grepped explicitly for
conflict markers.

Issue: #852, #925, #672

🤖 Generated with Claude Code