Models / nemotron_h_puzzle

NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4: GPU memory to fine-tune

Fine-tuning NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 at batch 1 and 2048 tokens takes about 8.52 TB of GPU memory for a full fine-tune (at least 70 x H200 sharded), and 893 GB with LoRA r=16 on a bf16 base (at least 8 x H200 sharded). Estimated from the model's structure, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 156K 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

443B

443,337,265,152 parameters

In the published checkpoint

44.54B

44,539,891,200 scalars · safetensors.total, read 2026-07-07

Delta

+895%

quantized This checkpoint is stored quantized (modelopt). The tensor count in the file counts stored elements under a packing scheme, not logical parameters, so the two numbers below are not measuring the same thing in either direction.

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32, 2048 tokens over 50,000 samples (assumed; no dataset attached).

Layers

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$1933.04

Memory: what fits on which card

Largest batch per card. Training is sized at 2048 tokens per sample, serving at 4096 tokens per sequence.

Full fine-tune, batch 1 at 2048 tokens

8.52 TB

LoRA r=16, batch 1 at 2048 tokens

893 GB

Serve: bf16 weights

887 GB

Serve: KV cache, 4096 tokens

50 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
L4 24 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
A100 40 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
A100 / H100 80 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
H200 141 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
Sharded, by memory alone	at least 70 x H200 or 123 x 80 GB	at least 8 x H200 or 13 x 80 GB	at least 8 x H200 or 13 x 80 GB

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 262144), 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. "Sharded" is a floor: the batch-1 peak with headroom divided across cards as if it split perfectly evenly (FSDP / ZeRO-3 for training, tensor parallel for serving), with no communication buffers counted. Serving: bf16 weights plus the KV cache of 4096 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

90 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 262144 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 262144
2	Embedding	Embedding	1 × 262144 × 4096
3	Positional_Embedding	Learned Pos Embed	1 × 262144 × 4096
4	LayerNorm_1_1	LayerNorm	1 × 262144 × 4096
5	Attention_1	Grouped Query Attn	1 × 262144 × 4096
6	Add_1_attn	Add	1 × 262144 × 4096
7	LayerNorm_1_2	LayerNorm	1 × 262144 × 4096
8	MoE_1	Shared-Expert MoE	1 × 262144 × 4096
9	Add_1_ffn	Add	1 × 262144 × 4096
10	LayerNorm_2_1	LayerNorm	1 × 262144 × 4096
11	Attention_2	Grouped Query Attn	1 × 262144 × 4096
12	Add_2_attn	Add	1 × 262144 × 4096
13	LayerNorm_2_2	LayerNorm	1 × 262144 × 4096
14	MoE_2	Shared-Expert MoE	1 × 262144 × 4096
15	Add_2_ffn	Add	1 × 262144 × 4096
16	LayerNorm_3_1	LayerNorm	1 × 262144 × 4096
17	Attention_3	Grouped Query Attn	1 × 262144 × 4096
18	Add_3_attn	Add	1 × 262144 × 4096
19	LayerNorm_3_2	LayerNorm	1 × 262144 × 4096
20	MoE_3	Shared-Expert MoE	1 × 262144 × 4096
21	Add_3_ffn	Add	1 × 262144 × 4096
22	LayerNorm_4_1	LayerNorm	1 × 262144 × 4096
23	Attention_4	Grouped Query Attn	1 × 262144 × 4096
24	Add_4_attn	Add	1 × 262144 × 4096
25	LayerNorm_4_2	LayerNorm	1 × 262144 × 4096
26	MoE_4	Shared-Expert MoE	1 × 262144 × 4096
27	Add_4_ffn	Add	1 × 262144 × 4096
28	LayerNorm_5_1	LayerNorm	1 × 262144 × 4096
29	Attention_5	Grouped Query Attn	1 × 262144 × 4096
30	Add_5_attn	Add	1 × 262144 × 4096
31	LayerNorm_5_2	LayerNorm	1 × 262144 × 4096
32	MoE_5	Shared-Expert MoE	1 × 262144 × 4096
33	Add_5_ffn	Add	1 × 262144 × 4096
34	LayerNorm_6_1	LayerNorm	1 × 262144 × 4096
35	Attention_6	Grouped Query Attn	1 × 262144 × 4096
36	Add_6_attn	Add	1 × 262144 × 4096
37	LayerNorm_6_2	LayerNorm	1 × 262144 × 4096
38	MoE_6	Shared-Expert MoE	1 × 262144 × 4096
39	Add_6_ffn	Add	1 × 262144 × 4096
40	LayerNorm_7_1	LayerNorm	1 × 262144 × 4096
41	Attention_7	Grouped Query Attn	1 × 262144 × 4096
42	Add_7_attn	Add	1 × 262144 × 4096
43	LayerNorm_7_2	LayerNorm	1 × 262144 × 4096
44	MoE_7	Shared-Expert MoE	1 × 262144 × 4096
45	Add_7_ffn	Add	1 × 262144 × 4096
46	LayerNorm_8_1	LayerNorm	1 × 262144 × 4096
47	Attention_8	Grouped Query Attn	1 × 262144 × 4096
48	Add_8_attn	Add	1 × 262144 × 4096
49	LayerNorm_8_2	LayerNorm	1 × 262144 × 4096
50	MoE_8	Shared-Expert MoE	1 × 262144 × 4096
51	Add_8_ffn	Add	1 × 262144 × 4096
52	LayerNorm_9_1	LayerNorm	1 × 262144 × 4096
53	Attention_9	Grouped Query Attn	1 × 262144 × 4096
54	Add_9_attn	Add	1 × 262144 × 4096
55	LayerNorm_9_2	LayerNorm	1 × 262144 × 4096
56	MoE_9	Shared-Expert MoE	1 × 262144 × 4096
57	Add_9_ffn	Add	1 × 262144 × 4096
58	LayerNorm_10_1	LayerNorm	1 × 262144 × 4096
59	Attention_10	Grouped Query Attn	1 × 262144 × 4096
60	Add_10_attn	Add	1 × 262144 × 4096
61	LayerNorm_10_2	LayerNorm	1 × 262144 × 4096
62	MoE_10	Shared-Expert MoE	1 × 262144 × 4096
63	Add_10_ffn	Add	1 × 262144 × 4096
64	LayerNorm_11_1	LayerNorm	1 × 262144 × 4096
65	Attention_11	Grouped Query Attn	1 × 262144 × 4096
66	Add_11_attn	Add	1 × 262144 × 4096
67	LayerNorm_11_2	LayerNorm	1 × 262144 × 4096
68	MoE_11	Shared-Expert MoE	1 × 262144 × 4096
69	Add_11_ffn	Add	1 × 262144 × 4096
70	LayerNorm_12_1	LayerNorm	1 × 262144 × 4096
71	Attention_12	Grouped Query Attn	1 × 262144 × 4096
72	Add_12_attn	Add	1 × 262144 × 4096
73	LayerNorm_12_2	LayerNorm	1 × 262144 × 4096
74	MoE_12	Shared-Expert MoE	1 × 262144 × 4096
75	Add_12_ffn	Add	1 × 262144 × 4096
76	MTP_1_HiddenNorm	LayerNorm	1 × 262144 × 4096
77	MTP_1_EmbeddingNorm	LayerNorm	1 × 262144 × 4096
78	MTP_1_Concat	Concatenate	1 × 262144 × 8192
79	MTP_1_Projection	Linear	1 × 262144 × 4096
80	LayerNorm_MTP_1_1	LayerNorm	1 × 262144 × 4096
81	Attention_MTP_1	Grouped Query Attn	1 × 262144 × 4096
82	Add_MTP_1_attn	Add	1 × 262144 × 4096
83	LayerNorm_MTP_1_2	LayerNorm	1 × 262144 × 4096
84	MoE_MTP_1	Shared-Expert MoE	1 × 262144 × 4096
85	Add_MTP_1_ffn	Add	1 × 262144 × 4096
86	MTP_1_Norm	LayerNorm	1 × 262144 × 4096
87	Final_LayerNorm	LayerNorm	1 × 262144 × 4096
88	LM_Head	Linear	1 × 262144 × 131072
89	Output	Output	1 × 262144 × 131072
90	MTP_1_Output	Output	1 × 262144 × 4096

What the verifier says

infoAt 13 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 nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4: 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 nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 --plan --share

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