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 and the VRAM left free caches the most used ones (83% of reads hit a 15 GB cache); the cache lasts a whole Generate run.

The pack's CUDA kernels, compiled at run time by PyTorch's NVRTC (no compiler or package to install), look each expert up in the cache on the GPU, run the cached ones in place and read the missed ones straight from RAM over PCIe, filling free cache slots as they go. With int8 experts the CPU computes part of the misses at the same time, and the most used of them are copied into the cache between frames. Since nothing in a frame waits on the CPU, every layer and the codebook decoder replay as CUDA graphs. The prompt pass groups its rows by expert and splits the experts between the GPU and the CPU the same way. Where PyTorch can't compile the kernels the nodes copy the experts instead, at the previous version's speed.

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

• temperature 1.0, top_p 0.9, top_k 0 (no cap) are the defaults of Fish Audio's Drama 3 Space. Their server's top_k 30 makes takes fall silent at pauses: 11 of 40 takes in Fish Audio's test, 2 of 4 cloned takes here, each losing a sentence. Lower temperature or top_p 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.
• max_frames caps each chunk (21.5 frames per second); generation stops earlier at the end of speech.
• speakers sets how a dialogue is generated. one take per line (default) generates each line on its own, with only its speaker's voices and earlier takes in the prompt, as Fish Audio's Space voices a scene: voices cannot bleed into each other, but lines do not hear each other. one take per line, hearing the line before adds the take before as context, so lines react again but that voice can bleed in. one take per speaker says all of a speaker's lines in one take (in 300-byte chunks with split_long_text), cuts it at the line breaks it says and puts the lines back in script order. together is one script in which every line hears the others. Decode joins separate takes with 0.3 s of silence. Pause cues are riskier in a solo take: a line that opens with [long pause] can stay silent.
• As in Fish Audio's Space: Encode Voice trims the clip's silent edges and normalizes its peak; Generate takes a chunk once more when it has no sound or stalls in its first 2.5 s; Decode's shorten_long_pauses (on) shortens pauses over 0.75 s to 0.75 s, cuts the silence before the first sound and keeps 0.15 s at the end.

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
python tools/convert.py asr <Qwen3-ForcedAligner-0.6B folder> qwen3_forced_aligner_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.

Set top_k to 0: workflows saved before this fix keep the old 30, which makes takes fall silent at pauses. Then check the voice: the transcript must cover every word of the clip, and the clip should be clean speech cut between sentences. 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; a new seed usually helps.

Give every speaker a voice. A long dialogue is generated in chunks, and an earlier chunk's take in the prompt hands the voice it ends in to the next chunk: with two cloned voices, 6 of 16 lines after the first chunk came out in the other voice. When every speaker has a voice, Generate leaves earlier takes out of dialogue chunks and the references keep the voices apart (1 of 18). A speaker without a voice still needs them to keep the voice the model made up.

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.