[0xShug0]

Summary

Add optional CUDA BF16 reference-encoder arithmetic for Fish Audio and Drama 3:

• fish_audio.codec_encoder_bf16_compute
• fish_drama3.codec_encoder_bf16_compute

Fish Audio defaults to false. Drama 3 defaults to true on its model-local configuration; the mode only applies on CUDA with F16/BF16 codec weights. Set fish_drama3.codec_encoder_bf16_compute=false to restore its previous encoder arithmetic. This intentionally changes Drama 3 reference-conditioned generation. The generic framework default remains false. Align encoder convolution, Snake, transformer and RVQ rounding with the Python codec. Decoder behavior is unchanged. Extend existing Conv1d with an opt-in BF16 path and CUDA im2col with BF16 output dispatch; existing F32/F16 paths remain unchanged. Update specs, help and model documentation, including standalone legacy-contract handling.

Follow-up to #936 / #935; this does not replace the merged CPU layout fix.

Codec parity

Codec-only roundtrips against the native Python reference, not end-to-end TTS:

Codec	Code agreement before / after	WAV SNR before / after	WAV cosine before / after
Fish Audio	40.93% / 84.50%	8.06 / 14.29 dB	0.92258 / 0.98140
Drama 3	39.22% / 85.81%	8.46 / 15.25 dB	0.92879 / 0.98497

Controlled F32 tests using identical GGUF weights achieve 100% code agreement for both codecs. Native BF16 code IDs are not bit-identical to Python. Python cold/warm execution also varies; on the measured same-weight Fish case, common-decoder C++ versus warm-Python waveform drift is comparable to Python cold versus warm drift. This is not a claim of universally identical perceptual quality.

Performance

RTX 5090, Debug build, same approximately six-second reference WAV, sequential measurements. Off/on refers to the new option.

Codec	Warm encode off / on	Warm decode off / on	Peak VRAM off / on
Fish Audio	27.33 / 27.22 ms	36.63 / 36.69 ms	2,644 / 2,676 MiB
Drama 3	30.47 / 28.96 ms	38.67 / 37.36 ms	2,610 / 2,688 MiB

These are codec timings, not end-to-end TTS RTF.

Validation and scope

• Default-off code arrays and WAVs match the frozen baseline byte-for-byte for both codecs.
• Decoding identical Python codes is unchanged with the option enabled.
• Public voice-clone requests pass ASR with existing embedded GGUF contracts.
• Explicit F16 codec storage tested; model spec system test passed.
• No global precision overrides, shared attention changes, or new GGML operation.
• Large graphs remain session-owned and reused.
• New C++ configuration fields require rebuilding external binary consumers.
• CUDA BF16 mode validated; no claim of new BF16-mode validation on other backends. Drama 3 now selects this previously tested mode by default; no additional runtime run was performed for the default-selection change.
• Rebased cleanly onto current main; runtime measurements were collected before the rebase, not rerun afterward.