feat: fine-tune Moonshine's streaming models

[Rebreda]

Lemonade now serves Moonshine-Medium-Streaming, and the streaming models carry their own neural VAD, which makes them the interesting target for this project. They are a separate architecture in transformers, moonshine_streaming with MoonshineStreamingForConditionalGeneration and MoonshineStreamingProcessor, not a flag on the offline model.

They fit the existing seq2seq path: same raw waveform in input_values, same Wav2Vec2 feature extractor, English-only, and unlike offline Moonshine the tokenizer already carries a pad token.

Two things had to change, and only running it found them

Frame alignment. The encoder reshapes its input to [batch, -1, 80], one 5 ms frame at 16 kHz, and raises on any other length. Padding a batch to its longest clip lands on a multiple of 80 only by luck. Architecture gained pad_to_multiple and the collator honours it.

A synthetic test passed because 16000 and 24000 both divide by 80. The first real batch gave:

RuntimeError: shape '[4, -1, 80]' is invalid for input of size 4246272

Name detection. moonshine-streaming-small resolved to offline Moonshine, because hub ids use hyphens, model_type uses underscores, and moonshine is a prefix of moonshine_streaming. It would have silently trained with the wrong processor. Detection now flattens separators and takes the longest match.

Verified on real hardware

gfx1151, ROCm torch, real MDC dataset:

Architecture: moonshine_streaming
Trainable params: 327,680 / 140,644,025 (0.23%)
torch 2.9.1+rocm7.2.0 (rocm build), device: AMD Radeon Graphics

Training runs and the loss moves. Two honest caveats.

Throughput is roughly 46 s/step at batch 4 on this iGPU, so a real run is hours.

The mid-training eval crashed the GPU queue:

HSA_STATUS_ERROR_EXCEPTION: An HSAIL operation resulted in a hardware exception. code: 0x1016

at batch 34 of 193, during predict_with_generate. That is the gfx1151 instability AMD's own support matrix implies, not something this PR introduces, and Whisper runs on the same box have completed eval fine. Worth knowing before planning a long run: consider --eval-steps beyond --max-steps to skip mid-training eval on this hardware.

Not covered

moonshine-ai/moonshine-streaming, the medium model Lemonade serves, ships as ONNX graphs with no PyTorch weights and cannot be fine-tuned. moonshine-streaming-small is the transformers checkpoint. Serving and fine-tuning are different artifacts here.

406 tests pass.