The codec runs only before and after generation, so it never sits in VRAM next to the model.

Drama 3 is a Qwen3.5-style mixture of experts: 40 layers, each with 256 routed experts of which 8 run per frame, plus attention or DeltaNet, a shared expert and a 13-layer codebook decoder. Everything except the routed experts is about 6 GB and stays on the GPU. With experts in RAM the 32 GB of experts stay in pinned system memory: each frame copies the 8 experts per layer it needs, and a least-recently-used cache in the VRAM left free keeps the ones that come back (85–95% of reads hit a 13 GB cache). The prompt uses nearly every expert, so its pass streams each layer's experts once and keeps the ones the prompt used most. Decoding one frame at a time is CPU-bound, so the cache size moves the speed less than you would expect.

With experts in VRAM the whole file sits on the GPU; two 24 GB cards can split the layers (second_gpu).

• temperature 1.0, top_p 0.9, top_k 30 are the defaults of Fish Audio's own Drama 3 server. Lower values make takes slower, flatter and more likely to fall silent.
• The model copies the reference's delivery: a flat clip gives a flat take, a lively one a lively take.
• Speakers: start each line with <|speaker:N|> and give each speaker a voice whose transcript starts with the same tag (no tag means speaker 0).
• max_frames caps each chunk (21.5 frames per second); generation stops earlier at the end of speech.

From the official checkpoints, one GPU visible (CUDA_VISIBLE_DEVICES=N), with ComfyUI on PYTHONPATH:

python tools/convert.py model <drama-3 folder> drama3_fp8experts_bf16.safetensors --format fp8
python tools/convert.py model <drama-3 folder> drama3_int8experts_bf16.safetensors --format int8_convrot
python tools/convert.py model <drama-3 folder> drama3_w6a8experts_bf16.safetensors --format w6a8
python tools/convert.py model <drama-3 folder> drama3_w4a8experts_bf16.safetensors --format w4a8
python tools/convert.py codec <drama-3 folder>/codec.pth drama3_codec.safetensors
python tools/convert.py asr <Qwen3-ASR-1.7B folder> qwen3_asr_1.7b_bf16.safetensors
python tools/convert.py asr <Qwen3-ASR-0.6B folder> qwen3_asr_0.6b_bf16.safetensors

• Drama 3: against Fish Audio's own PyTorch code, with the same frames forced, the main token's distribution differs by a KL of 0.004–0.007 nats and the top token agrees 93–94% of the time, the noise of bf16 arithmetic; the FP8 and int8 files measure the same. Prompt rows match the builders of Fish Audio's server and of fish-speech token for token, and the codec produces 99.94% of the same codes.
• Qwen3-ASR: against Qwen's qwen-asr package, the prompt ids are identical, the encoder output matches within 1e-6 in fp32 and the first token is the same; transcripts match on clips up to 8 s. On longer clips the encoder attends in Qwen's 8 s windows, as their vLLM backend does, which can move a comma.

Check the voice first: the transcript must cover every word of the clip, and the clip should be clean speech cut between sentences. Then queue again for a new seed. Some clips make the model fall silent in Fish Audio's own code too (known issue); try another stretch of the recording. A take that goes silent is stopped and its silent tail cut, with a warning in the console.

Keep temperature at 1.0 and top_p at 0.9 or above, and use a reference with the delivery you want.

Lower expert_cache_gb, or pick a smaller file if system RAM is the limit. The first run reads the whole model file from disk once; later runs start in seconds.