Models / ilama

Ilama-3.2-1B: GPU memory to fine-tune

Fine-tuning Ilama-3.2-1B at batch 1 and 2048 tokens takes about 26.7 GB of GPU memory for a full fine-tune (borderline on one A100 40 GB), and 5.68 GB with LoRA r=16 on a bf16 base (fits one T4 16 GB). Estimated from the model's structure, reading the config under transformers semantics, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 112K 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

1.24B

1,235,812,352 parameters

In the published checkpoint

1.24B

1,235,814,400 scalars · safetensors.total, read 2025-07-22

Delta

-0.00%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_ilama.py`), so `config.json` names a class in the repo rather than an architecture `transformers` defines. The graph below is what those config keys mean under `transformers` semantics, which is not necessarily what the repo's own file builds. 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, 2048 tokens over 50,000 samples (assumed; no dataset attached).

Layers

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$79.79

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

26.7 GB

LoRA r=16, batch 1 at 2048 tokens

5.68 GB

Serve: bf16 weights

2.47 GB

Serve: KV cache, 4096 tokens

0.13 GB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	2 cards at batch 1 (42 + 56 layers)	batch 3	batch 85
L4 24 GB	2 cards at batch 1 (72 + 26 layers)	batch 6	batch 137
A100 40 GB	batch 3, borderline	batch 10	batch 240
A100 / H100 80 GB	batch 15	batch 22	batch 499
H200 141 GB	batch 33	batch 40	batch 895

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 131072), 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. "Borderline" means the card holds batch 1 on the estimate but not on the conservative peak (the estimate x 1.43, at or above the measured peak on 95% of held-out runs): measure one step before reserving it. 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

100 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 131072 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 131072
2	Embedding	Embedding	1 × 131072 × 2048
3	RoPE	RoPE	1 × 131072 × 2048
4	RMSNorm_1_1	RMSNorm	1 × 131072 × 2048
5	Attention_1	Grouped Query Attn	1 × 131072 × 2048
6	Add_1_attn	Add	1 × 131072 × 2048
7	RMSNorm_1_2	RMSNorm	1 × 131072 × 2048
8	FFN_1	SwiGLU	1 × 131072 × 2048
9	Add_1_ffn	Add	1 × 131072 × 2048
10	RMSNorm_2_1	RMSNorm	1 × 131072 × 2048
11	Attention_2	Grouped Query Attn	1 × 131072 × 2048
12	Add_2_attn	Add	1 × 131072 × 2048
13	RMSNorm_2_2	RMSNorm	1 × 131072 × 2048
14	FFN_2	SwiGLU	1 × 131072 × 2048
15	Add_2_ffn	Add	1 × 131072 × 2048
16	RMSNorm_3_1	RMSNorm	1 × 131072 × 2048
17	Attention_3	Grouped Query Attn	1 × 131072 × 2048
18	Add_3_attn	Add	1 × 131072 × 2048
19	RMSNorm_3_2	RMSNorm	1 × 131072 × 2048
20	FFN_3	SwiGLU	1 × 131072 × 2048
21	Add_3_ffn	Add	1 × 131072 × 2048
22	RMSNorm_4_1	RMSNorm	1 × 131072 × 2048
23	Attention_4	Grouped Query Attn	1 × 131072 × 2048
24	Add_4_attn	Add	1 × 131072 × 2048
25	RMSNorm_4_2	RMSNorm	1 × 131072 × 2048
26	FFN_4	SwiGLU	1 × 131072 × 2048
27	Add_4_ffn	Add	1 × 131072 × 2048
28	RMSNorm_5_1	RMSNorm	1 × 131072 × 2048
29	Attention_5	Grouped Query Attn	1 × 131072 × 2048
30	Add_5_attn	Add	1 × 131072 × 2048
31	RMSNorm_5_2	RMSNorm	1 × 131072 × 2048
32	FFN_5	SwiGLU	1 × 131072 × 2048
33	Add_5_ffn	Add	1 × 131072 × 2048
34	RMSNorm_6_1	RMSNorm	1 × 131072 × 2048
35	Attention_6	Grouped Query Attn	1 × 131072 × 2048
36	Add_6_attn	Add	1 × 131072 × 2048
37	RMSNorm_6_2	RMSNorm	1 × 131072 × 2048
38	FFN_6	SwiGLU	1 × 131072 × 2048
39	Add_6_ffn	Add	1 × 131072 × 2048
40	RMSNorm_7_1	RMSNorm	1 × 131072 × 2048
41	Attention_7	Grouped Query Attn	1 × 131072 × 2048
42	Add_7_attn	Add	1 × 131072 × 2048
43	RMSNorm_7_2	RMSNorm	1 × 131072 × 2048
44	FFN_7	SwiGLU	1 × 131072 × 2048
45	Add_7_ffn	Add	1 × 131072 × 2048
46	RMSNorm_8_1	RMSNorm	1 × 131072 × 2048
47	Attention_8	Grouped Query Attn	1 × 131072 × 2048
48	Add_8_attn	Add	1 × 131072 × 2048
49	RMSNorm_8_2	RMSNorm	1 × 131072 × 2048
50	FFN_8	SwiGLU	1 × 131072 × 2048
51	Add_8_ffn	Add	1 × 131072 × 2048
52	RMSNorm_9_1	RMSNorm	1 × 131072 × 2048
53	Attention_9	Grouped Query Attn	1 × 131072 × 2048
54	Add_9_attn	Add	1 × 131072 × 2048
55	RMSNorm_9_2	RMSNorm	1 × 131072 × 2048
56	FFN_9	SwiGLU	1 × 131072 × 2048
57	Add_9_ffn	Add	1 × 131072 × 2048
58	RMSNorm_10_1	RMSNorm	1 × 131072 × 2048
59	Attention_10	Grouped Query Attn	1 × 131072 × 2048
60	Add_10_attn	Add	1 × 131072 × 2048
61	RMSNorm_10_2	RMSNorm	1 × 131072 × 2048
62	FFN_10	SwiGLU	1 × 131072 × 2048
63	Add_10_ffn	Add	1 × 131072 × 2048
64	RMSNorm_11_1	RMSNorm	1 × 131072 × 2048
65	Attention_11	Grouped Query Attn	1 × 131072 × 2048
66	Add_11_attn	Add	1 × 131072 × 2048
67	RMSNorm_11_2	RMSNorm	1 × 131072 × 2048
68	FFN_11	SwiGLU	1 × 131072 × 2048
69	Add_11_ffn	Add	1 × 131072 × 2048
70	RMSNorm_12_1	RMSNorm	1 × 131072 × 2048
71	Attention_12	Grouped Query Attn	1 × 131072 × 2048
72	Add_12_attn	Add	1 × 131072 × 2048
73	RMSNorm_12_2	RMSNorm	1 × 131072 × 2048
74	FFN_12	SwiGLU	1 × 131072 × 2048
75	Add_12_ffn	Add	1 × 131072 × 2048
76	RMSNorm_13_1	RMSNorm	1 × 131072 × 2048
77	Attention_13	Grouped Query Attn	1 × 131072 × 2048
78	Add_13_attn	Add	1 × 131072 × 2048
79	RMSNorm_13_2	RMSNorm	1 × 131072 × 2048
80	FFN_13	SwiGLU	1 × 131072 × 2048
81	Add_13_ffn	Add	1 × 131072 × 2048
82	RMSNorm_14_1	RMSNorm	1 × 131072 × 2048
83	Attention_14	Grouped Query Attn	1 × 131072 × 2048
84	Add_14_attn	Add	1 × 131072 × 2048
85	RMSNorm_14_2	RMSNorm	1 × 131072 × 2048
86	FFN_14	SwiGLU	1 × 131072 × 2048
87	Add_14_ffn	Add	1 × 131072 × 2048
88	RMSNorm_15_1	RMSNorm	1 × 131072 × 2048
89	Attention_15	Grouped Query Attn	1 × 131072 × 2048
90	Add_15_attn	Add	1 × 131072 × 2048
91	RMSNorm_15_2	RMSNorm	1 × 131072 × 2048
92	FFN_15	SwiGLU	1 × 131072 × 2048
93	Add_15_ffn	Add	1 × 131072 × 2048
94	RMSNorm_16_1	RMSNorm	1 × 131072 × 2048
95	Attention_16	Grouped Query Attn	1 × 131072 × 2048
96	Add_16_attn	Add	1 × 131072 × 2048
97	RMSNorm_16_2	RMSNorm	1 × 131072 × 2048
98	FFN_16	SwiGLU	1 × 131072 × 2048
99	Add_16_ffn	Add	1 × 131072 × 2048
100	Output	Output	1 × 131072 × 2048

What the verifier says

infoAt 16 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 hmellor/Ilama-3.2-1B: 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 hmellor/Ilama-3.2-1B --plan --share

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