Drama 3 VRAM requirements: how much GPU memory does it need?

Drama 3 GPU requirements: published number versus our estimate

Fish Audio's installation docs list at least 48 GB of GPU memory for Drama 3 (FP8) and NVIDIA GPUs only (no CPU, Apple MPS, or AMD ROCm). The SGLang-Omni cookbook reports the engine's weights at 37.0 GB for the official checkpoint and 20.8 GB for the int4 checkpoint. Our calculator, which reads the checkpoint file sizes and config.json, estimates 38.90 GiB–45.38 GiB at 8,192 tokens, one stream, official checkpoint. The top of that range is 48.7 GB in decimal units, just under the published 48 GB floor. Neither number is a measurement of peak usage on a specific GPU.

"Tight" and "Does not fit" are formula verdicts. We have not run Drama 3 on any of these GPUs. DGX Spark is unified memory shared with the system, not discrete VRAM.

Drama 3 memory by context and checkpoint

Verified facts

• ModelFish Audio Drama 3, fishaudio/drama-3
• TaskText-to-speech: multi-speaker (up to 16 speaker tags), voice cloning, natural-language control
• Parameters35,899,643,008 total (Hugging Face safetensors metadata); model card: 35B total, 3B active per token
• PrecisionFP8 (E4M3) MoE experts, 32,212,254,720 parameters; BF16 for the remaining 3,687,388,288
• Files8 safetensors shards, 39,594,895,616 bytes (36.88 GiB / 39.59 GB); codec.pth 1,871,099,728 bytes (1.74 GiB)
• Text backbone40 layers: 30 Gated DeltaNet (linear attention) + 10 full attention; hidden size 2048; 16 query heads, 2 KV heads, head dimension 256
• MoE256 routed experts, 8 per token, plus a shared expert
• Audio decoderCodebook-MTP, 10 codebooks, 4 layers; the codec turns 44.1 kHz audio into frames of 2,048 samples (about 21.5 rows per second)
• LanguagesThe model card lists 83 language codes
• LicenseFish Audio Community License v1.0 (Oct 13, 2026): research and non-commercial use free; commercial use while annual revenue is under USD 500,000; hosted TTS services or APIs need a separate license
• RuntimeSGLang-Omni, Triton attention backend only; NVIDIA GPUs only
• Int4 checkpointUnofficial W4A16 GPTQ-Marlin quantization, fishaudio/drama-3-int4; 22.33 GiB on disk with codec; needs compute capability 8.0 or newer. Not the publisher's own release.

How this page estimates a TTS model

• Weights are the checkpoint files: safetensors plus codec.pth (38.62 GiB official, 22.33 GiB int4). All 256 experts are resident; the 3B active count is compute, not memory.
• KV cache exists only on the 10 full-attention layers: 10 layers × 2 (K and V) × 2 KV heads × 256 head dim × 2 bytes = 20,480 bytes per cached row at FP16, 0.156 GiB at 8,192 rows.
• Gated DeltaNet state for the 30 linear layers is a fixed 0.059 GiB per stream (FP32 recurrent state, from config mamba_ssm_dtype) plus a small convolution state. It does not grow with context.
• Context counts prompt rows (text, reference transcript, and about 21.5 rows per second of reference audio) plus generated audio frames. The cookbook's default cap is 8,192 prompt rows.
• Not modelled: the 4-layer audio decoder's cache (sequence length 11 per frame, small), CUDA graph memory, and many concurrent streams. Runtime extras use the calculator's generic activation and overhead range, not a TTS measurement.
• Engine memory share (our arithmetic, untested): the cookbook gives the engine 0.70 of GPU memory by default, plus 0.10 for the vocoder and a 6 GiB preprocessing reserve. On a 48 GB card that is about 33.6 GB for the engine against 37.0 GB of weights, so you would likely need to raise --tts_engine.gpu_memory_fraction. We have not tried this.

Drama 3 FAQ

Memory layers in this estimate

Artifact bytes, raw KV, engine extras, and total demand are separate. This is not a runtime trace.

Weight artifact

38.62 GiB (41,465,995,344 bytes)

Raw KV cache

0.16 GiB (167,772,160 bytes)

Engine extras / workspace

0.12 GiB–6.61 GiB

Total estimated demand

38.90 GiB–45.38 GiB

Published artifacts

Memory by quantization and context

Concurrent sequences = 1. KV defaults to FP16. The range includes documented runtime extras for the listed engine. These figures use the same core as the interactive calculator.

Raw KV contribution

Related pages

• KV cache calculator
• 4-bit, 8-bit, and FP16 memory
• File size is not runtime VRAM
• Methodology
• All can-i-run checks

Sources

• Drama 3 model card — 35B parameters, 3B active per token; FP8 MoE experts, remaining weights bf16; license: Fish Audio Community License.
• config.json — revision e0a9218aa48888270f06ef1ed9ea67622c0e17d2; 40 text layers (30 linear_attention + 10 full_attention), 16 query / 2 KV heads, head_dim 256, hidden 2048, 256 experts, 8 per token.
• safetensors index — metadata.total_size=39594895616
• Drama 3 SGLang Omni cookbook — Engine weights 37.0 GB (FP8) / 20.8 GB (int4); default engine memory share 0.70; reference audio up to 60 s; prompt up to 8192 rows.
• fishaudio/drama-3-int4 — Unofficial int4 W4A16 checkpoint; GPTQ-Marlin needs compute capability 8.0+.

Assumptions that affect this result

• Requested context is within the recorded native window of 262144 tokens.
• Full-resident weight estimates use total parameters, including inactive MoE experts.
• This MoE lists 35899643008 total parameters and 3000000000 active parameters. Active-parameter compute is not used as the weight-residency figure.
• Weight bytes come from the selected weight files only, not from every file in the repository.
• Download file size is not peak GPU memory. Runtime layout, allocator padding, and KV are extra.
• Weight bytes come from 9 published safetensors shards totaling 41465995344 bytes.
• Context budget is the number of cached tokens per sequence (prompt plus reserved generation).
• Concurrent sequences share one model replica; weights are not multiplied by sequence count.

Calculator