Models / qwen3_tts

Qwen3-TTS-12Hz-1.7B-CustomVoice: GPU memory to fine-tune

Fine-tuning Qwen3-TTS-12Hz-1.7B-CustomVoice at batch 1 and 512 tokens takes about 3.15 GB of GPU memory for a full fine-tune (fits one T4 16 GB), and 0.60 GB with LoRA r=16 on a bf16 base (fits one T4 16 GB). Estimated from the model's structure, for the towers the graph reconstructs, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 2.6M downloads on Hugging Face.

Our count against the checkpoint

The left number comes from the graph. The right one is the number of scalars in the published weight files. Nothing on this page was tuned to make them agree.

Derived from structure

152M

151,862,016 parameters

In the published checkpoint

1.92B

1,916,676,352 scalars · safetensors.total, read 2026-09-06

Delta

-92.1%

multi-tower The config declares 2 sub-models (talker_config, code_predictor_config). The published checkpoint carries all of them; the graph below carries the towers the reader reconstructs. A gap here is a statement about what we read, not about the checkpoint.

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32, 512 tokens over 50,000 samples (assumed; no dataset attached).

Layers

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$2.30

Memory: what fits on which card

Largest batch per card. Training is sized at 512 tokens per sample, serving at 512 tokens per sequence.

Full fine-tune, batch 1 at 512 tokens

3.15 GB

LoRA r=16, batch 1 at 512 tokens

0.60 GB

Serve: bf16 weights

0.30 GB

Serve: KV cache, 512 tokens

19 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 512 tokens
T4 16 GB	batch 47	batch 58	batch 721
L4 24 GB	batch 78	batch 89	batch 1089
A100 40 GB	batch 138	batch 149	batch 1826
A100 / H100 80 GB	batch 289	batch 301	batch 3669
H200 141 GB	batch 520	batch 531	batch 6479

Training: 512 tokens per sample (2048, or the model's context when shorter), AdamW with bf16 autocast, SDPA attention (no score matrix stored), 1.15x headroom on every card. Full fine-tune holds weights, gradients and optimizer state (1.2 x 16 bytes per parameter) plus activations. LoRA is r=16 on every linear layer but the head, base frozen in bf16 (2 bytes per parameter), adapters at full AdamW state, same activations. A two-card split is a contiguous layer cut (pipeline parallel), computed only where one card cannot hold batch 1. Serving: bf16 weights plus the KV cache of 512 tokens per sequence; prefill working set, CUDA context and framework reserve not counted.

These are estimates from the graph, not measurements. Against measured runs, held out by architecture, the estimate flagged 113 of the 131 jobs that did not fit their card and said "fits" on the other 18. Every measured run.

Structure

76 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 512 tokens: the config's maximum context.

	Layer	Type	Output shape
1	Input	Input	1 × 512
2	Embedding	Embedding	1 × 512 × 768
3	RoPE	RoPE	1 × 512 × 768
4	RMSNorm_1_1	RMSNorm	1 × 512 × 768
5	Attention_1	Grouped Query Attn	1 × 512 × 768
6	Add_1_attn	Add	1 × 512 × 768
7	RMSNorm_1_2	RMSNorm	1 × 512 × 768
8	FFN_1	SwiGLU	1 × 512 × 768
9	Add_1_ffn	Add	1 × 512 × 768
10	RMSNorm_2_1	RMSNorm	1 × 512 × 768
11	Attention_2	Grouped Query Attn	1 × 512 × 768
12	Add_2_attn	Add	1 × 512 × 768
13	RMSNorm_2_2	RMSNorm	1 × 512 × 768
14	FFN_2	SwiGLU	1 × 512 × 768
15	Add_2_ffn	Add	1 × 512 × 768
16	RMSNorm_3_1	RMSNorm	1 × 512 × 768
17	Attention_3	Grouped Query Attn	1 × 512 × 768
18	Add_3_attn	Add	1 × 512 × 768
19	RMSNorm_3_2	RMSNorm	1 × 512 × 768
20	FFN_3	SwiGLU	1 × 512 × 768
21	Add_3_ffn	Add	1 × 512 × 768
22	RMSNorm_4_1	RMSNorm	1 × 512 × 768
23	Attention_4	Grouped Query Attn	1 × 512 × 768
24	Add_4_attn	Add	1 × 512 × 768
25	RMSNorm_4_2	RMSNorm	1 × 512 × 768
26	FFN_4	SwiGLU	1 × 512 × 768
27	Add_4_ffn	Add	1 × 512 × 768
28	RMSNorm_5_1	RMSNorm	1 × 512 × 768
29	Attention_5	Grouped Query Attn	1 × 512 × 768
30	Add_5_attn	Add	1 × 512 × 768
31	RMSNorm_5_2	RMSNorm	1 × 512 × 768
32	FFN_5	SwiGLU	1 × 512 × 768
33	Add_5_ffn	Add	1 × 512 × 768
34	RMSNorm_6_1	RMSNorm	1 × 512 × 768
35	Attention_6	Grouped Query Attn	1 × 512 × 768
36	Add_6_attn	Add	1 × 512 × 768
37	RMSNorm_6_2	RMSNorm	1 × 512 × 768
38	FFN_6	SwiGLU	1 × 512 × 768
39	Add_6_ffn	Add	1 × 512 × 768
40	RMSNorm_7_1	RMSNorm	1 × 512 × 768
41	Attention_7	Grouped Query Attn	1 × 512 × 768
42	Add_7_attn	Add	1 × 512 × 768
43	RMSNorm_7_2	RMSNorm	1 × 512 × 768
44	FFN_7	SwiGLU	1 × 512 × 768
45	Add_7_ffn	Add	1 × 512 × 768
46	RMSNorm_8_1	RMSNorm	1 × 512 × 768
47	Attention_8	Grouped Query Attn	1 × 512 × 768
48	Add_8_attn	Add	1 × 512 × 768
49	RMSNorm_8_2	RMSNorm	1 × 512 × 768
50	FFN_8	SwiGLU	1 × 512 × 768
51	Add_8_ffn	Add	1 × 512 × 768
52	RMSNorm_9_1	RMSNorm	1 × 512 × 768
53	Attention_9	Grouped Query Attn	1 × 512 × 768
54	Add_9_attn	Add	1 × 512 × 768
55	RMSNorm_9_2	RMSNorm	1 × 512 × 768
56	FFN_9	SwiGLU	1 × 512 × 768
57	Add_9_ffn	Add	1 × 512 × 768
58	RMSNorm_10_1	RMSNorm	1 × 512 × 768
59	Attention_10	Grouped Query Attn	1 × 512 × 768
60	Add_10_attn	Add	1 × 512 × 768
61	RMSNorm_10_2	RMSNorm	1 × 512 × 768
62	FFN_10	SwiGLU	1 × 512 × 768
63	Add_10_ffn	Add	1 × 512 × 768
64	RMSNorm_11_1	RMSNorm	1 × 512 × 768
65	Attention_11	Grouped Query Attn	1 × 512 × 768
66	Add_11_attn	Add	1 × 512 × 768
67	RMSNorm_11_2	RMSNorm	1 × 512 × 768
68	FFN_11	SwiGLU	1 × 512 × 768
69	Add_11_ffn	Add	1 × 512 × 768
70	RMSNorm_12_1	RMSNorm	1 × 512 × 768
71	Attention_12	Grouped Query Attn	1 × 512 × 768
72	Add_12_attn	Add	1 × 512 × 768
73	RMSNorm_12_2	RMSNorm	1 × 512 × 768
74	FFN_12	SwiGLU	1 × 512 × 768
75	Add_12_ffn	Add	1 × 512 × 768
76	Output	Output	1 × 512 × 768

What the verifier says

infoAt 12 stacked attention layers, residual-branch outputs add up; unscaled init lets activation variance grow with depth. GPT-2/LLaMA-family models scale the residual projections by depth (N(0, 0.02 / √(2L))). Fix: Scale residual output projections by depth: nn.init.normal_(w, std=0.02 / math.sqrt(2 * n_layers))
deep-attention-default-init

What changed upstream

The change tracker follows Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice: every revision it has read, what changed in the config, tokenizer, chat template or weights between them, and an OSV advisory per change.

Revision history and changes Badge JSON OSV advisories All tracked changes

Do this to your own model

Same numbers, on a model in your repo, in one command. No account.

pip install neurarch-trace
neurarch-trace Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --plan --share

Paste a model instead Markdown twin Index JSON POST /api/v1/plan