Models / qwen2

Qwen2.5-0.5B-Instruct: GPU memory to fine-tune

Fine-tuning Qwen2.5-0.5B-Instruct at batch 1 and 2048 tokens takes about 13.7 GB of GPU memory for a full fine-tune (borderline on one T4 16 GB), and 5.39 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. 6.4M 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

494M

494,004,224 parameters

In the published checkpoint

494M

494,032,768 scalars · safetensors.total, read 2024-09-25

Delta

-0.01%

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

146

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$30.86

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

13.7 GB

LoRA r=16, batch 1 at 2048 tokens

5.39 GB

Serve: bf16 weights

0.99 GB

Serve: KV cache, 4096 tokens

50 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	batch 1, borderline	batch 3	batch 256
L4 24 GB	batch 2	batch 4	batch 395
A100 40 GB	batch 5	batch 7	batch 671
A100 / H100 80 GB	batch 14	batch 16	batch 1362
H200 141 GB	batch 26	batch 28	batch 2416

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 32768), 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

148 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 32768 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 32768
2	Embedding	Embedding	1 × 32768 × 896
3	RoPE	RoPE	1 × 32768 × 896
4	RMSNorm_1_1	RMSNorm	1 × 32768 × 896
5	Attention_1	Grouped Query Attn	1 × 32768 × 896
6	Add_1_attn	Add	1 × 32768 × 896
7	RMSNorm_1_2	RMSNorm	1 × 32768 × 896
8	FFN_1	SwiGLU	1 × 32768 × 896
9	Add_1_ffn	Add	1 × 32768 × 896
10	RMSNorm_2_1	RMSNorm	1 × 32768 × 896
11	Attention_2	Grouped Query Attn	1 × 32768 × 896
12	Add_2_attn	Add	1 × 32768 × 896
13	RMSNorm_2_2	RMSNorm	1 × 32768 × 896
14	FFN_2	SwiGLU	1 × 32768 × 896
15	Add_2_ffn	Add	1 × 32768 × 896
16	RMSNorm_3_1	RMSNorm	1 × 32768 × 896
17	Attention_3	Grouped Query Attn	1 × 32768 × 896
18	Add_3_attn	Add	1 × 32768 × 896
19	RMSNorm_3_2	RMSNorm	1 × 32768 × 896
20	FFN_3	SwiGLU	1 × 32768 × 896
21	Add_3_ffn	Add	1 × 32768 × 896
22	RMSNorm_4_1	RMSNorm	1 × 32768 × 896
23	Attention_4	Grouped Query Attn	1 × 32768 × 896
24	Add_4_attn	Add	1 × 32768 × 896
25	RMSNorm_4_2	RMSNorm	1 × 32768 × 896
26	FFN_4	SwiGLU	1 × 32768 × 896
27	Add_4_ffn	Add	1 × 32768 × 896
28	RMSNorm_5_1	RMSNorm	1 × 32768 × 896
29	Attention_5	Grouped Query Attn	1 × 32768 × 896
30	Add_5_attn	Add	1 × 32768 × 896
31	RMSNorm_5_2	RMSNorm	1 × 32768 × 896
32	FFN_5	SwiGLU	1 × 32768 × 896
33	Add_5_ffn	Add	1 × 32768 × 896
34	RMSNorm_6_1	RMSNorm	1 × 32768 × 896
35	Attention_6	Grouped Query Attn	1 × 32768 × 896
36	Add_6_attn	Add	1 × 32768 × 896
37	RMSNorm_6_2	RMSNorm	1 × 32768 × 896
38	FFN_6	SwiGLU	1 × 32768 × 896
39	Add_6_ffn	Add	1 × 32768 × 896
40	RMSNorm_7_1	RMSNorm	1 × 32768 × 896
41	Attention_7	Grouped Query Attn	1 × 32768 × 896
42	Add_7_attn	Add	1 × 32768 × 896
43	RMSNorm_7_2	RMSNorm	1 × 32768 × 896
44	FFN_7	SwiGLU	1 × 32768 × 896
45	Add_7_ffn	Add	1 × 32768 × 896
46	RMSNorm_8_1	RMSNorm	1 × 32768 × 896
47	Attention_8	Grouped Query Attn	1 × 32768 × 896
48	Add_8_attn	Add	1 × 32768 × 896
49	RMSNorm_8_2	RMSNorm	1 × 32768 × 896
50	FFN_8	SwiGLU	1 × 32768 × 896
51	Add_8_ffn	Add	1 × 32768 × 896
52	RMSNorm_9_1	RMSNorm	1 × 32768 × 896
53	Attention_9	Grouped Query Attn	1 × 32768 × 896
54	Add_9_attn	Add	1 × 32768 × 896
55	RMSNorm_9_2	RMSNorm	1 × 32768 × 896
56	FFN_9	SwiGLU	1 × 32768 × 896
57	Add_9_ffn	Add	1 × 32768 × 896
58	RMSNorm_10_1	RMSNorm	1 × 32768 × 896
59	Attention_10	Grouped Query Attn	1 × 32768 × 896
60	Add_10_attn	Add	1 × 32768 × 896
61	RMSNorm_10_2	RMSNorm	1 × 32768 × 896
62	FFN_10	SwiGLU	1 × 32768 × 896
63	Add_10_ffn	Add	1 × 32768 × 896
64	RMSNorm_11_1	RMSNorm	1 × 32768 × 896
65	Attention_11	Grouped Query Attn	1 × 32768 × 896
66	Add_11_attn	Add	1 × 32768 × 896
67	RMSNorm_11_2	RMSNorm	1 × 32768 × 896
68	FFN_11	SwiGLU	1 × 32768 × 896
69	Add_11_ffn	Add	1 × 32768 × 896
70	RMSNorm_12_1	RMSNorm	1 × 32768 × 896
71	Attention_12	Grouped Query Attn	1 × 32768 × 896
72	Add_12_attn	Add	1 × 32768 × 896
73	RMSNorm_12_2	RMSNorm	1 × 32768 × 896
74	FFN_12	SwiGLU	1 × 32768 × 896
75	Add_12_ffn	Add	1 × 32768 × 896
76	RMSNorm_13_1	RMSNorm	1 × 32768 × 896
77	Attention_13	Grouped Query Attn	1 × 32768 × 896
78	Add_13_attn	Add	1 × 32768 × 896
79	RMSNorm_13_2	RMSNorm	1 × 32768 × 896
80	FFN_13	SwiGLU	1 × 32768 × 896
81	Add_13_ffn	Add	1 × 32768 × 896
82	RMSNorm_14_1	RMSNorm	1 × 32768 × 896
83	Attention_14	Grouped Query Attn	1 × 32768 × 896
84	Add_14_attn	Add	1 × 32768 × 896
85	RMSNorm_14_2	RMSNorm	1 × 32768 × 896
86	FFN_14	SwiGLU	1 × 32768 × 896
87	Add_14_ffn	Add	1 × 32768 × 896
88	RMSNorm_15_1	RMSNorm	1 × 32768 × 896
89	Attention_15	Grouped Query Attn	1 × 32768 × 896
90	Add_15_attn	Add	1 × 32768 × 896
91	RMSNorm_15_2	RMSNorm	1 × 32768 × 896
92	FFN_15	SwiGLU	1 × 32768 × 896
93	Add_15_ffn	Add	1 × 32768 × 896
94	RMSNorm_16_1	RMSNorm	1 × 32768 × 896
95	Attention_16	Grouped Query Attn	1 × 32768 × 896
96	Add_16_attn	Add	1 × 32768 × 896
97	RMSNorm_16_2	RMSNorm	1 × 32768 × 896
98	FFN_16	SwiGLU	1 × 32768 × 896
99	Add_16_ffn	Add	1 × 32768 × 896
100	RMSNorm_17_1	RMSNorm	1 × 32768 × 896
101	Attention_17	Grouped Query Attn	1 × 32768 × 896
102	Add_17_attn	Add	1 × 32768 × 896
103	RMSNorm_17_2	RMSNorm	1 × 32768 × 896
104	FFN_17	SwiGLU	1 × 32768 × 896
105	Add_17_ffn	Add	1 × 32768 × 896
106	RMSNorm_18_1	RMSNorm	1 × 32768 × 896
107	Attention_18	Grouped Query Attn	1 × 32768 × 896
108	Add_18_attn	Add	1 × 32768 × 896
109	RMSNorm_18_2	RMSNorm	1 × 32768 × 896
110	FFN_18	SwiGLU	1 × 32768 × 896
111	Add_18_ffn	Add	1 × 32768 × 896
112	RMSNorm_19_1	RMSNorm	1 × 32768 × 896
113	Attention_19	Grouped Query Attn	1 × 32768 × 896
114	Add_19_attn	Add	1 × 32768 × 896
115	RMSNorm_19_2	RMSNorm	1 × 32768 × 896
116	FFN_19	SwiGLU	1 × 32768 × 896
117	Add_19_ffn	Add	1 × 32768 × 896
118	RMSNorm_20_1	RMSNorm	1 × 32768 × 896
119	Attention_20	Grouped Query Attn	1 × 32768 × 896
120	Add_20_attn	Add	1 × 32768 × 896
121	RMSNorm_20_2	RMSNorm	1 × 32768 × 896
122	FFN_20	SwiGLU	1 × 32768 × 896
123	Add_20_ffn	Add	1 × 32768 × 896
124	RMSNorm_21_1	RMSNorm	1 × 32768 × 896
125	Attention_21	Grouped Query Attn	1 × 32768 × 896
126	Add_21_attn	Add	1 × 32768 × 896
127	RMSNorm_21_2	RMSNorm	1 × 32768 × 896
128	FFN_21	SwiGLU	1 × 32768 × 896
129	Add_21_ffn	Add	1 × 32768 × 896
130	RMSNorm_22_1	RMSNorm	1 × 32768 × 896
131	Attention_22	Grouped Query Attn	1 × 32768 × 896
132	Add_22_attn	Add	1 × 32768 × 896
133	RMSNorm_22_2	RMSNorm	1 × 32768 × 896
134	FFN_22	SwiGLU	1 × 32768 × 896
135	Add_22_ffn	Add	1 × 32768 × 896
136	RMSNorm_23_1	RMSNorm	1 × 32768 × 896
137	Attention_23	Grouped Query Attn	1 × 32768 × 896
138	Add_23_attn	Add	1 × 32768 × 896
139	RMSNorm_23_2	RMSNorm	1 × 32768 × 896
140	FFN_23	SwiGLU	1 × 32768 × 896
141	Add_23_ffn	Add	1 × 32768 × 896
142	RMSNorm_24_1	RMSNorm	1 × 32768 × 896
143	Attention_24	Grouped Query Attn	1 × 32768 × 896
144	Add_24_attn	Add	1 × 32768 × 896
145	RMSNorm_24_2	RMSNorm	1 × 32768 × 896
146	FFN_24	SwiGLU	1 × 32768 × 896
147	Add_24_ffn	Add	1 × 32768 × 896
148	Output	Output	1 × 32768 × 896

What the verifier says

infoAt 24 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/Qwen2.5-0.5B-Instruct: 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/Qwen2.5-0.5B-Instruct --plan --share

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

Other qwen2 checkpoints

gte-Qwen2-1.5B-instruct

1.78B derived · +0.01% against the checkpoint

gte-Qwen2-7B-instruct

7.61B derived · +0.00% against the checkpoint

Qwen2.5-1.5B-Instruct

1.54B derived · -0.00% against the checkpoint

Qwen2.5-3B-Instruct

3.09B derived · -0.00% against the checkpoint