Models / falcon

falcon-7b: GPU memory to fine-tune

Fine-tuning falcon-7b at batch 1 and 512 tokens takes about 75.2 GB of GPU memory for a full fine-tune (fits one H200 141 GB), and 9.66 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. 362K 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

3.84B

3,837,851,648 parameters

In the published checkpoint

7.22B

7,217,189,760 scalars · safetensors.total, read 2024-10-12

Delta

-42.7%

stored twice The checkpoint stores the tied embedding twice: +295.5M (`lm_head.weight` and `transformer.word_embeddings.weight`, 65024 x 4544). The model has one such matrix and the count above has it once; the delta compares that count plus the second copy against the file. Header read 2026-10-18.

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_falcon.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, 512 tokens over 50,000 samples (assumed; no dataset attached).

Layers

194

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$69.25

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

75.2 GB

LoRA r=16, batch 1 at 512 tokens

9.66 GB

Serve: bf16 weights

7.68 GB

Serve: KV cache, 512 tokens

0.30 GB

Card	Full fine-tune	LoRA r=16	Serve bf16, 512 tokens
T4 16 GB	does not fit at batch 1	batch 3	batch 20
L4 24 GB	does not fit at batch 1	batch 8	batch 44
A100 40 GB	does not fit at batch 1	batch 18	batch 91
A100 / H100 80 GB	2 cards at batch 1 (177 + 17 layers)	batch 41	batch 207
H200 141 GB	batch 33	batch 77	batch 385

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

196 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 × 4544
3	Positional_Embedding	Learned Pos Embed	1 × 512 × 4544
4	LayerNorm_1_1	LayerNorm	1 × 512 × 4544
5	Attention_1	Multi-Head Attention	1 × 512 × 4544
6	Add_1_attn	Add	1 × 512 × 4544
7	LayerNorm_1_2	LayerNorm	1 × 512 × 4544
8	FFN_1	Feed Forward	1 × 512 × 4544
9	Add_1_ffn	Add	1 × 512 × 4544
10	LayerNorm_2_1	LayerNorm	1 × 512 × 4544
11	Attention_2	Multi-Head Attention	1 × 512 × 4544
12	Add_2_attn	Add	1 × 512 × 4544
13	LayerNorm_2_2	LayerNorm	1 × 512 × 4544
14	FFN_2	Feed Forward	1 × 512 × 4544
15	Add_2_ffn	Add	1 × 512 × 4544
16	LayerNorm_3_1	LayerNorm	1 × 512 × 4544
17	Attention_3	Multi-Head Attention	1 × 512 × 4544
18	Add_3_attn	Add	1 × 512 × 4544
19	LayerNorm_3_2	LayerNorm	1 × 512 × 4544
20	FFN_3	Feed Forward	1 × 512 × 4544
21	Add_3_ffn	Add	1 × 512 × 4544
22	LayerNorm_4_1	LayerNorm	1 × 512 × 4544
23	Attention_4	Multi-Head Attention	1 × 512 × 4544
24	Add_4_attn	Add	1 × 512 × 4544
25	LayerNorm_4_2	LayerNorm	1 × 512 × 4544
26	FFN_4	Feed Forward	1 × 512 × 4544
27	Add_4_ffn	Add	1 × 512 × 4544
28	LayerNorm_5_1	LayerNorm	1 × 512 × 4544
29	Attention_5	Multi-Head Attention	1 × 512 × 4544
30	Add_5_attn	Add	1 × 512 × 4544
31	LayerNorm_5_2	LayerNorm	1 × 512 × 4544
32	FFN_5	Feed Forward	1 × 512 × 4544
33	Add_5_ffn	Add	1 × 512 × 4544
34	LayerNorm_6_1	LayerNorm	1 × 512 × 4544
35	Attention_6	Multi-Head Attention	1 × 512 × 4544
36	Add_6_attn	Add	1 × 512 × 4544
37	LayerNorm_6_2	LayerNorm	1 × 512 × 4544
38	FFN_6	Feed Forward	1 × 512 × 4544
39	Add_6_ffn	Add	1 × 512 × 4544
40	LayerNorm_7_1	LayerNorm	1 × 512 × 4544
41	Attention_7	Multi-Head Attention	1 × 512 × 4544
42	Add_7_attn	Add	1 × 512 × 4544
43	LayerNorm_7_2	LayerNorm	1 × 512 × 4544
44	FFN_7	Feed Forward	1 × 512 × 4544
45	Add_7_ffn	Add	1 × 512 × 4544
46	LayerNorm_8_1	LayerNorm	1 × 512 × 4544
47	Attention_8	Multi-Head Attention	1 × 512 × 4544
48	Add_8_attn	Add	1 × 512 × 4544
49	LayerNorm_8_2	LayerNorm	1 × 512 × 4544
50	FFN_8	Feed Forward	1 × 512 × 4544
51	Add_8_ffn	Add	1 × 512 × 4544
52	LayerNorm_9_1	LayerNorm	1 × 512 × 4544
53	Attention_9	Multi-Head Attention	1 × 512 × 4544
54	Add_9_attn	Add	1 × 512 × 4544
55	LayerNorm_9_2	LayerNorm	1 × 512 × 4544
56	FFN_9	Feed Forward	1 × 512 × 4544
57	Add_9_ffn	Add	1 × 512 × 4544
58	LayerNorm_10_1	LayerNorm	1 × 512 × 4544
59	Attention_10	Multi-Head Attention	1 × 512 × 4544
60	Add_10_attn	Add	1 × 512 × 4544
61	LayerNorm_10_2	LayerNorm	1 × 512 × 4544
62	FFN_10	Feed Forward	1 × 512 × 4544
63	Add_10_ffn	Add	1 × 512 × 4544
64	LayerNorm_11_1	LayerNorm	1 × 512 × 4544
65	Attention_11	Multi-Head Attention	1 × 512 × 4544
66	Add_11_attn	Add	1 × 512 × 4544
67	LayerNorm_11_2	LayerNorm	1 × 512 × 4544
68	FFN_11	Feed Forward	1 × 512 × 4544
69	Add_11_ffn	Add	1 × 512 × 4544
70	LayerNorm_12_1	LayerNorm	1 × 512 × 4544
71	Attention_12	Multi-Head Attention	1 × 512 × 4544
72	Add_12_attn	Add	1 × 512 × 4544
73	LayerNorm_12_2	LayerNorm	1 × 512 × 4544
74	FFN_12	Feed Forward	1 × 512 × 4544
75	Add_12_ffn	Add	1 × 512 × 4544
76	LayerNorm_13_1	LayerNorm	1 × 512 × 4544
77	Attention_13	Multi-Head Attention	1 × 512 × 4544
78	Add_13_attn	Add	1 × 512 × 4544
79	LayerNorm_13_2	LayerNorm	1 × 512 × 4544
80	FFN_13	Feed Forward	1 × 512 × 4544
81	Add_13_ffn	Add	1 × 512 × 4544
82	LayerNorm_14_1	LayerNorm	1 × 512 × 4544
83	Attention_14	Multi-Head Attention	1 × 512 × 4544
84	Add_14_attn	Add	1 × 512 × 4544
85	LayerNorm_14_2	LayerNorm	1 × 512 × 4544
86	FFN_14	Feed Forward	1 × 512 × 4544
87	Add_14_ffn	Add	1 × 512 × 4544
88	LayerNorm_15_1	LayerNorm	1 × 512 × 4544
89	Attention_15	Multi-Head Attention	1 × 512 × 4544
90	Add_15_attn	Add	1 × 512 × 4544
91	LayerNorm_15_2	LayerNorm	1 × 512 × 4544
92	FFN_15	Feed Forward	1 × 512 × 4544
93	Add_15_ffn	Add	1 × 512 × 4544
94	LayerNorm_16_1	LayerNorm	1 × 512 × 4544
95	Attention_16	Multi-Head Attention	1 × 512 × 4544
96	Add_16_attn	Add	1 × 512 × 4544
97	LayerNorm_16_2	LayerNorm	1 × 512 × 4544
98	FFN_16	Feed Forward	1 × 512 × 4544
99	Add_16_ffn	Add	1 × 512 × 4544
100	LayerNorm_17_1	LayerNorm	1 × 512 × 4544
101	Attention_17	Multi-Head Attention	1 × 512 × 4544
102	Add_17_attn	Add	1 × 512 × 4544
103	LayerNorm_17_2	LayerNorm	1 × 512 × 4544
104	FFN_17	Feed Forward	1 × 512 × 4544
105	Add_17_ffn	Add	1 × 512 × 4544
106	LayerNorm_18_1	LayerNorm	1 × 512 × 4544
107	Attention_18	Multi-Head Attention	1 × 512 × 4544
108	Add_18_attn	Add	1 × 512 × 4544
109	LayerNorm_18_2	LayerNorm	1 × 512 × 4544
110	FFN_18	Feed Forward	1 × 512 × 4544
111	Add_18_ffn	Add	1 × 512 × 4544
112	LayerNorm_19_1	LayerNorm	1 × 512 × 4544
113	Attention_19	Multi-Head Attention	1 × 512 × 4544
114	Add_19_attn	Add	1 × 512 × 4544
115	LayerNorm_19_2	LayerNorm	1 × 512 × 4544
116	FFN_19	Feed Forward	1 × 512 × 4544
117	Add_19_ffn	Add	1 × 512 × 4544
118	LayerNorm_20_1	LayerNorm	1 × 512 × 4544
119	Attention_20	Multi-Head Attention	1 × 512 × 4544
120	Add_20_attn	Add	1 × 512 × 4544
121	LayerNorm_20_2	LayerNorm	1 × 512 × 4544
122	FFN_20	Feed Forward	1 × 512 × 4544
123	Add_20_ffn	Add	1 × 512 × 4544
124	LayerNorm_21_1	LayerNorm	1 × 512 × 4544
125	Attention_21	Multi-Head Attention	1 × 512 × 4544
126	Add_21_attn	Add	1 × 512 × 4544
127	LayerNorm_21_2	LayerNorm	1 × 512 × 4544
128	FFN_21	Feed Forward	1 × 512 × 4544
129	Add_21_ffn	Add	1 × 512 × 4544
130	LayerNorm_22_1	LayerNorm	1 × 512 × 4544
131	Attention_22	Multi-Head Attention	1 × 512 × 4544
132	Add_22_attn	Add	1 × 512 × 4544
133	LayerNorm_22_2	LayerNorm	1 × 512 × 4544
134	FFN_22	Feed Forward	1 × 512 × 4544
135	Add_22_ffn	Add	1 × 512 × 4544
136	LayerNorm_23_1	LayerNorm	1 × 512 × 4544
137	Attention_23	Multi-Head Attention	1 × 512 × 4544
138	Add_23_attn	Add	1 × 512 × 4544
139	LayerNorm_23_2	LayerNorm	1 × 512 × 4544
140	FFN_23	Feed Forward	1 × 512 × 4544
141	Add_23_ffn	Add	1 × 512 × 4544
142	LayerNorm_24_1	LayerNorm	1 × 512 × 4544
143	Attention_24	Multi-Head Attention	1 × 512 × 4544
144	Add_24_attn	Add	1 × 512 × 4544
145	LayerNorm_24_2	LayerNorm	1 × 512 × 4544
146	FFN_24	Feed Forward	1 × 512 × 4544
147	Add_24_ffn	Add	1 × 512 × 4544
148	LayerNorm_25_1	LayerNorm	1 × 512 × 4544
149	Attention_25	Multi-Head Attention	1 × 512 × 4544
150	Add_25_attn	Add	1 × 512 × 4544
151	LayerNorm_25_2	LayerNorm	1 × 512 × 4544
152	FFN_25	Feed Forward	1 × 512 × 4544
153	Add_25_ffn	Add	1 × 512 × 4544
154	LayerNorm_26_1	LayerNorm	1 × 512 × 4544
155	Attention_26	Multi-Head Attention	1 × 512 × 4544
156	Add_26_attn	Add	1 × 512 × 4544
157	LayerNorm_26_2	LayerNorm	1 × 512 × 4544
158	FFN_26	Feed Forward	1 × 512 × 4544
159	Add_26_ffn	Add	1 × 512 × 4544
160	LayerNorm_27_1	LayerNorm	1 × 512 × 4544
161	Attention_27	Multi-Head Attention	1 × 512 × 4544
162	Add_27_attn	Add	1 × 512 × 4544
163	LayerNorm_27_2	LayerNorm	1 × 512 × 4544
164	FFN_27	Feed Forward	1 × 512 × 4544
165	Add_27_ffn	Add	1 × 512 × 4544
166	LayerNorm_28_1	LayerNorm	1 × 512 × 4544
167	Attention_28	Multi-Head Attention	1 × 512 × 4544
168	Add_28_attn	Add	1 × 512 × 4544
169	LayerNorm_28_2	LayerNorm	1 × 512 × 4544
170	FFN_28	Feed Forward	1 × 512 × 4544
171	Add_28_ffn	Add	1 × 512 × 4544
172	LayerNorm_29_1	LayerNorm	1 × 512 × 4544
173	Attention_29	Multi-Head Attention	1 × 512 × 4544
174	Add_29_attn	Add	1 × 512 × 4544
175	LayerNorm_29_2	LayerNorm	1 × 512 × 4544
176	FFN_29	Feed Forward	1 × 512 × 4544
177	Add_29_ffn	Add	1 × 512 × 4544
178	LayerNorm_30_1	LayerNorm	1 × 512 × 4544
179	Attention_30	Multi-Head Attention	1 × 512 × 4544
180	Add_30_attn	Add	1 × 512 × 4544
181	LayerNorm_30_2	LayerNorm	1 × 512 × 4544
182	FFN_30	Feed Forward	1 × 512 × 4544
183	Add_30_ffn	Add	1 × 512 × 4544
184	LayerNorm_31_1	LayerNorm	1 × 512 × 4544
185	Attention_31	Multi-Head Attention	1 × 512 × 4544
186	Add_31_attn	Add	1 × 512 × 4544
187	LayerNorm_31_2	LayerNorm	1 × 512 × 4544
188	FFN_31	Feed Forward	1 × 512 × 4544
189	Add_31_ffn	Add	1 × 512 × 4544
190	LayerNorm_32_1	LayerNorm	1 × 512 × 4544
191	Attention_32	Multi-Head Attention	1 × 512 × 4544
192	Add_32_attn	Add	1 × 512 × 4544
193	LayerNorm_32_2	LayerNorm	1 × 512 × 4544
194	FFN_32	Feed Forward	1 × 512 × 4544
195	Add_32_ffn	Add	1 × 512 × 4544
196	Output	Output	1 × 512 × 4544

What the verifier says

info32 attention layers at embedDim 4544 cache full per-head K/V: about 568 KB per token at fp16, which dominates memory at long context. Grouped-query attention (e.g. 8:1) would cut this ~8×; multi-head latent attention (MLA) shrinks it ~10× or more. This is the move production LLMs make; it does not change the parameter count. Fix: Switch attention to groupedQueryAttention (set numKVHeads below numHeads, e.g. numHeads/4) or mla (a low-rank cached latent).
full-mha-serving-cost

infoAt 32 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

warnAcross 32 attention layers this design caches 568 KB per token, so a single 8,192-token sequence needs ~4.8 GB of KV cache before weights or activations. That exceeds the 4 GB budget this rule assumes for serving headroom. Fix: Cut KV width: raise the GQA ratio (fewer numKVHeads), switch to MLA, reduce depth or embedDim, or accept a shorter serving context.
kv-cache-context-budget

What changed upstream

The change tracker follows tiiuae/falcon-7b: 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 tiiuae/falcon-7b --plan --share

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