Models / gpt_neox

pythia-160m: GPU memory to fine-tune

Fine-tuning pythia-160m at batch 1 and 2048 tokens takes about 4.46 GB of GPU memory for a full fine-tune (fits one T4 16 GB), and 1.69 GB with LoRA r=16 on a bf16 base (fits one T4 16 GB). Estimated from the model's structure, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 3.4M 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

164M

163,946,112 parameters

In the published checkpoint

213M

212,654,688 scalars · safetensors.total, read 2026-09-06

Delta

-22.9%

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

$10.96

Memory: what fits on which card

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

Full fine-tune, batch 1 at 2048 tokens

4.46 GB

LoRA r=16, batch 1 at 2048 tokens

1.69 GB

Serve: bf16 weights

0.33 GB

Serve: KV cache, 2048 tokens

75 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 2048 tokens
T4 16 GB	batch 8	batch 10	batch 179
L4 24 GB	batch 13	batch 15	batch 272
A100 40 GB	batch 24	batch 26	batch 456
A100 / H100 80 GB	batch 50	batch 52	batch 917
H200 141 GB	batch 91	batch 93	batch 1619

Training: 2048 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 2048 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

78 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 2048 tokens: the config's maximum context.

	Layer	Type	Output shape
1	Input	Input	1 × 2048
2	Embedding	Embedding	1 × 2048 × 768
3	Positional_Embedding	Learned Pos Embed	1 × 2048 × 768
4	LayerNorm_1_1	LayerNorm	1 × 2048 × 768
5	Attention_1	Multi-Head Attention	1 × 2048 × 768
6	Add_1_attn	Add	1 × 2048 × 768
7	LayerNorm_1_2	LayerNorm	1 × 2048 × 768
8	FFN_1	Feed Forward	1 × 2048 × 768
9	Add_1_ffn	Add	1 × 2048 × 768
10	LayerNorm_2_1	LayerNorm	1 × 2048 × 768
11	Attention_2	Multi-Head Attention	1 × 2048 × 768
12	Add_2_attn	Add	1 × 2048 × 768
13	LayerNorm_2_2	LayerNorm	1 × 2048 × 768
14	FFN_2	Feed Forward	1 × 2048 × 768
15	Add_2_ffn	Add	1 × 2048 × 768
16	LayerNorm_3_1	LayerNorm	1 × 2048 × 768
17	Attention_3	Multi-Head Attention	1 × 2048 × 768
18	Add_3_attn	Add	1 × 2048 × 768
19	LayerNorm_3_2	LayerNorm	1 × 2048 × 768
20	FFN_3	Feed Forward	1 × 2048 × 768
21	Add_3_ffn	Add	1 × 2048 × 768
22	LayerNorm_4_1	LayerNorm	1 × 2048 × 768
23	Attention_4	Multi-Head Attention	1 × 2048 × 768
24	Add_4_attn	Add	1 × 2048 × 768
25	LayerNorm_4_2	LayerNorm	1 × 2048 × 768
26	FFN_4	Feed Forward	1 × 2048 × 768
27	Add_4_ffn	Add	1 × 2048 × 768
28	LayerNorm_5_1	LayerNorm	1 × 2048 × 768
29	Attention_5	Multi-Head Attention	1 × 2048 × 768
30	Add_5_attn	Add	1 × 2048 × 768
31	LayerNorm_5_2	LayerNorm	1 × 2048 × 768
32	FFN_5	Feed Forward	1 × 2048 × 768
33	Add_5_ffn	Add	1 × 2048 × 768
34	LayerNorm_6_1	LayerNorm	1 × 2048 × 768
35	Attention_6	Multi-Head Attention	1 × 2048 × 768
36	Add_6_attn	Add	1 × 2048 × 768
37	LayerNorm_6_2	LayerNorm	1 × 2048 × 768
38	FFN_6	Feed Forward	1 × 2048 × 768
39	Add_6_ffn	Add	1 × 2048 × 768
40	LayerNorm_7_1	LayerNorm	1 × 2048 × 768
41	Attention_7	Multi-Head Attention	1 × 2048 × 768
42	Add_7_attn	Add	1 × 2048 × 768
43	LayerNorm_7_2	LayerNorm	1 × 2048 × 768
44	FFN_7	Feed Forward	1 × 2048 × 768
45	Add_7_ffn	Add	1 × 2048 × 768
46	LayerNorm_8_1	LayerNorm	1 × 2048 × 768
47	Attention_8	Multi-Head Attention	1 × 2048 × 768
48	Add_8_attn	Add	1 × 2048 × 768
49	LayerNorm_8_2	LayerNorm	1 × 2048 × 768
50	FFN_8	Feed Forward	1 × 2048 × 768
51	Add_8_ffn	Add	1 × 2048 × 768
52	LayerNorm_9_1	LayerNorm	1 × 2048 × 768
53	Attention_9	Multi-Head Attention	1 × 2048 × 768
54	Add_9_attn	Add	1 × 2048 × 768
55	LayerNorm_9_2	LayerNorm	1 × 2048 × 768
56	FFN_9	Feed Forward	1 × 2048 × 768
57	Add_9_ffn	Add	1 × 2048 × 768
58	LayerNorm_10_1	LayerNorm	1 × 2048 × 768
59	Attention_10	Multi-Head Attention	1 × 2048 × 768
60	Add_10_attn	Add	1 × 2048 × 768
61	LayerNorm_10_2	LayerNorm	1 × 2048 × 768
62	FFN_10	Feed Forward	1 × 2048 × 768
63	Add_10_ffn	Add	1 × 2048 × 768
64	LayerNorm_11_1	LayerNorm	1 × 2048 × 768
65	Attention_11	Multi-Head Attention	1 × 2048 × 768
66	Add_11_attn	Add	1 × 2048 × 768
67	LayerNorm_11_2	LayerNorm	1 × 2048 × 768
68	FFN_11	Feed Forward	1 × 2048 × 768
69	Add_11_ffn	Add	1 × 2048 × 768
70	LayerNorm_12_1	LayerNorm	1 × 2048 × 768
71	Attention_12	Multi-Head Attention	1 × 2048 × 768
72	Add_12_attn	Add	1 × 2048 × 768
73	LayerNorm_12_2	LayerNorm	1 × 2048 × 768
74	FFN_12	Feed Forward	1 × 2048 × 768
75	Add_12_ffn	Add	1 × 2048 × 768
76	Final_LayerNorm	LayerNorm	1 × 2048 × 768
77	LM_Head	Linear	1 × 2048 × 50304
78	Output	Output	1 × 2048 × 50304

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 EleutherAI/pythia-160m: 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 EleutherAI/pythia-160m --plan --share

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