Models / qwen3_bidirec

Giga-Embeddings-instruct-480M-0826: GPU memory to fine-tune

Fine-tuning Giga-Embeddings-instruct-480M-0826 at batch 1 and 2048 tokens takes about 11.5 GB of GPU memory for a full fine-tune (borderline on one T4 16 GB), and 3.34 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. 105K 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

484M

483,713,024 parameters

In the published checkpoint

484M

483,717,632 scalars · safetensors.total, read 2026-08-27

Delta

-0.00%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_gigarembed.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

170

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$31.56

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

11.5 GB

LoRA r=16, batch 1 at 2048 tokens

3.34 GB

Serve: bf16 weights

0.97 GB

Serve: KV cache, 4096 tokens

0.23 GB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	batch 2, borderline	batch 5	batch 55
L4 24 GB	batch 5	batch 8	batch 84
A100 40 GB	batch 11	batch 15	batch 143
A100 / H100 80 GB	batch 27	batch 30	batch 292
H200 141 GB	batch 51	batch 54	batch 517

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 8192), 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

172 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 8192 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 8192
2	Embedding	Embedding	1 × 8192 × 1024
3	RoPE	RoPE	1 × 8192 × 1024
4	RMSNorm_1_1	RMSNorm	1 × 8192 × 1024
5	Attention_1	Grouped Query Attn	1 × 8192 × 1024
6	Add_1_attn	Add	1 × 8192 × 1024
7	RMSNorm_1_2	RMSNorm	1 × 8192 × 1024
8	FFN_1	SwiGLU	1 × 8192 × 1024
9	Add_1_ffn	Add	1 × 8192 × 1024
10	RMSNorm_2_1	RMSNorm	1 × 8192 × 1024
11	Attention_2	Grouped Query Attn	1 × 8192 × 1024
12	Add_2_attn	Add	1 × 8192 × 1024
13	RMSNorm_2_2	RMSNorm	1 × 8192 × 1024
14	FFN_2	SwiGLU	1 × 8192 × 1024
15	Add_2_ffn	Add	1 × 8192 × 1024
16	RMSNorm_3_1	RMSNorm	1 × 8192 × 1024
17	Attention_3	Grouped Query Attn	1 × 8192 × 1024
18	Add_3_attn	Add	1 × 8192 × 1024
19	RMSNorm_3_2	RMSNorm	1 × 8192 × 1024
20	FFN_3	SwiGLU	1 × 8192 × 1024
21	Add_3_ffn	Add	1 × 8192 × 1024
22	RMSNorm_4_1	RMSNorm	1 × 8192 × 1024
23	Attention_4	Grouped Query Attn	1 × 8192 × 1024
24	Add_4_attn	Add	1 × 8192 × 1024
25	RMSNorm_4_2	RMSNorm	1 × 8192 × 1024
26	FFN_4	SwiGLU	1 × 8192 × 1024
27	Add_4_ffn	Add	1 × 8192 × 1024
28	RMSNorm_5_1	RMSNorm	1 × 8192 × 1024
29	Attention_5	Grouped Query Attn	1 × 8192 × 1024
30	Add_5_attn	Add	1 × 8192 × 1024
31	RMSNorm_5_2	RMSNorm	1 × 8192 × 1024
32	FFN_5	SwiGLU	1 × 8192 × 1024
33	Add_5_ffn	Add	1 × 8192 × 1024
34	RMSNorm_6_1	RMSNorm	1 × 8192 × 1024
35	Attention_6	Grouped Query Attn	1 × 8192 × 1024
36	Add_6_attn	Add	1 × 8192 × 1024
37	RMSNorm_6_2	RMSNorm	1 × 8192 × 1024
38	FFN_6	SwiGLU	1 × 8192 × 1024
39	Add_6_ffn	Add	1 × 8192 × 1024
40	RMSNorm_7_1	RMSNorm	1 × 8192 × 1024
41	Attention_7	Grouped Query Attn	1 × 8192 × 1024
42	Add_7_attn	Add	1 × 8192 × 1024
43	RMSNorm_7_2	RMSNorm	1 × 8192 × 1024
44	FFN_7	SwiGLU	1 × 8192 × 1024
45	Add_7_ffn	Add	1 × 8192 × 1024
46	RMSNorm_8_1	RMSNorm	1 × 8192 × 1024
47	Attention_8	Grouped Query Attn	1 × 8192 × 1024
48	Add_8_attn	Add	1 × 8192 × 1024
49	RMSNorm_8_2	RMSNorm	1 × 8192 × 1024
50	FFN_8	SwiGLU	1 × 8192 × 1024
51	Add_8_ffn	Add	1 × 8192 × 1024
52	RMSNorm_9_1	RMSNorm	1 × 8192 × 1024
53	Attention_9	Grouped Query Attn	1 × 8192 × 1024
54	Add_9_attn	Add	1 × 8192 × 1024
55	RMSNorm_9_2	RMSNorm	1 × 8192 × 1024
56	FFN_9	SwiGLU	1 × 8192 × 1024
57	Add_9_ffn	Add	1 × 8192 × 1024
58	RMSNorm_10_1	RMSNorm	1 × 8192 × 1024
59	Attention_10	Grouped Query Attn	1 × 8192 × 1024
60	Add_10_attn	Add	1 × 8192 × 1024
61	RMSNorm_10_2	RMSNorm	1 × 8192 × 1024
62	FFN_10	SwiGLU	1 × 8192 × 1024
63	Add_10_ffn	Add	1 × 8192 × 1024
64	RMSNorm_11_1	RMSNorm	1 × 8192 × 1024
65	Attention_11	Grouped Query Attn	1 × 8192 × 1024
66	Add_11_attn	Add	1 × 8192 × 1024
67	RMSNorm_11_2	RMSNorm	1 × 8192 × 1024
68	FFN_11	SwiGLU	1 × 8192 × 1024
69	Add_11_ffn	Add	1 × 8192 × 1024
70	RMSNorm_12_1	RMSNorm	1 × 8192 × 1024
71	Attention_12	Grouped Query Attn	1 × 8192 × 1024
72	Add_12_attn	Add	1 × 8192 × 1024
73	RMSNorm_12_2	RMSNorm	1 × 8192 × 1024
74	FFN_12	SwiGLU	1 × 8192 × 1024
75	Add_12_ffn	Add	1 × 8192 × 1024
76	RMSNorm_13_1	RMSNorm	1 × 8192 × 1024
77	Attention_13	Grouped Query Attn	1 × 8192 × 1024
78	Add_13_attn	Add	1 × 8192 × 1024
79	RMSNorm_13_2	RMSNorm	1 × 8192 × 1024
80	FFN_13	SwiGLU	1 × 8192 × 1024
81	Add_13_ffn	Add	1 × 8192 × 1024
82	RMSNorm_14_1	RMSNorm	1 × 8192 × 1024
83	Attention_14	Grouped Query Attn	1 × 8192 × 1024
84	Add_14_attn	Add	1 × 8192 × 1024
85	RMSNorm_14_2	RMSNorm	1 × 8192 × 1024
86	FFN_14	SwiGLU	1 × 8192 × 1024
87	Add_14_ffn	Add	1 × 8192 × 1024
88	RMSNorm_15_1	RMSNorm	1 × 8192 × 1024
89	Attention_15	Grouped Query Attn	1 × 8192 × 1024
90	Add_15_attn	Add	1 × 8192 × 1024
91	RMSNorm_15_2	RMSNorm	1 × 8192 × 1024
92	FFN_15	SwiGLU	1 × 8192 × 1024
93	Add_15_ffn	Add	1 × 8192 × 1024
94	RMSNorm_16_1	RMSNorm	1 × 8192 × 1024
95	Attention_16	Grouped Query Attn	1 × 8192 × 1024
96	Add_16_attn	Add	1 × 8192 × 1024
97	RMSNorm_16_2	RMSNorm	1 × 8192 × 1024
98	FFN_16	SwiGLU	1 × 8192 × 1024
99	Add_16_ffn	Add	1 × 8192 × 1024
100	RMSNorm_17_1	RMSNorm	1 × 8192 × 1024
101	Attention_17	Grouped Query Attn	1 × 8192 × 1024
102	Add_17_attn	Add	1 × 8192 × 1024
103	RMSNorm_17_2	RMSNorm	1 × 8192 × 1024
104	FFN_17	SwiGLU	1 × 8192 × 1024
105	Add_17_ffn	Add	1 × 8192 × 1024
106	RMSNorm_18_1	RMSNorm	1 × 8192 × 1024
107	Attention_18	Grouped Query Attn	1 × 8192 × 1024
108	Add_18_attn	Add	1 × 8192 × 1024
109	RMSNorm_18_2	RMSNorm	1 × 8192 × 1024
110	FFN_18	SwiGLU	1 × 8192 × 1024
111	Add_18_ffn	Add	1 × 8192 × 1024
112	RMSNorm_19_1	RMSNorm	1 × 8192 × 1024
113	Attention_19	Grouped Query Attn	1 × 8192 × 1024
114	Add_19_attn	Add	1 × 8192 × 1024
115	RMSNorm_19_2	RMSNorm	1 × 8192 × 1024
116	FFN_19	SwiGLU	1 × 8192 × 1024
117	Add_19_ffn	Add	1 × 8192 × 1024
118	RMSNorm_20_1	RMSNorm	1 × 8192 × 1024
119	Attention_20	Grouped Query Attn	1 × 8192 × 1024
120	Add_20_attn	Add	1 × 8192 × 1024
121	RMSNorm_20_2	RMSNorm	1 × 8192 × 1024
122	FFN_20	SwiGLU	1 × 8192 × 1024
123	Add_20_ffn	Add	1 × 8192 × 1024
124	RMSNorm_21_1	RMSNorm	1 × 8192 × 1024
125	Attention_21	Grouped Query Attn	1 × 8192 × 1024
126	Add_21_attn	Add	1 × 8192 × 1024
127	RMSNorm_21_2	RMSNorm	1 × 8192 × 1024
128	FFN_21	SwiGLU	1 × 8192 × 1024
129	Add_21_ffn	Add	1 × 8192 × 1024
130	RMSNorm_22_1	RMSNorm	1 × 8192 × 1024
131	Attention_22	Grouped Query Attn	1 × 8192 × 1024
132	Add_22_attn	Add	1 × 8192 × 1024
133	RMSNorm_22_2	RMSNorm	1 × 8192 × 1024
134	FFN_22	SwiGLU	1 × 8192 × 1024
135	Add_22_ffn	Add	1 × 8192 × 1024
136	RMSNorm_23_1	RMSNorm	1 × 8192 × 1024
137	Attention_23	Grouped Query Attn	1 × 8192 × 1024
138	Add_23_attn	Add	1 × 8192 × 1024
139	RMSNorm_23_2	RMSNorm	1 × 8192 × 1024
140	FFN_23	SwiGLU	1 × 8192 × 1024
141	Add_23_ffn	Add	1 × 8192 × 1024
142	RMSNorm_24_1	RMSNorm	1 × 8192 × 1024
143	Attention_24	Grouped Query Attn	1 × 8192 × 1024
144	Add_24_attn	Add	1 × 8192 × 1024
145	RMSNorm_24_2	RMSNorm	1 × 8192 × 1024
146	FFN_24	SwiGLU	1 × 8192 × 1024
147	Add_24_ffn	Add	1 × 8192 × 1024
148	RMSNorm_25_1	RMSNorm	1 × 8192 × 1024
149	Attention_25	Grouped Query Attn	1 × 8192 × 1024
150	Add_25_attn	Add	1 × 8192 × 1024
151	RMSNorm_25_2	RMSNorm	1 × 8192 × 1024
152	FFN_25	SwiGLU	1 × 8192 × 1024
153	Add_25_ffn	Add	1 × 8192 × 1024
154	RMSNorm_26_1	RMSNorm	1 × 8192 × 1024
155	Attention_26	Grouped Query Attn	1 × 8192 × 1024
156	Add_26_attn	Add	1 × 8192 × 1024
157	RMSNorm_26_2	RMSNorm	1 × 8192 × 1024
158	FFN_26	SwiGLU	1 × 8192 × 1024
159	Add_26_ffn	Add	1 × 8192 × 1024
160	RMSNorm_27_1	RMSNorm	1 × 8192 × 1024
161	Attention_27	Grouped Query Attn	1 × 8192 × 1024
162	Add_27_attn	Add	1 × 8192 × 1024
163	RMSNorm_27_2	RMSNorm	1 × 8192 × 1024
164	FFN_27	SwiGLU	1 × 8192 × 1024
165	Add_27_ffn	Add	1 × 8192 × 1024
166	RMSNorm_28_1	RMSNorm	1 × 8192 × 1024
167	Attention_28	Grouped Query Attn	1 × 8192 × 1024
168	Add_28_attn	Add	1 × 8192 × 1024
169	RMSNorm_28_2	RMSNorm	1 × 8192 × 1024
170	FFN_28	SwiGLU	1 × 8192 × 1024
171	Add_28_ffn	Add	1 × 8192 × 1024
172	Output	Output	1 × 8192 × 1024

What the verifier says

infoAt 28 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 ai-sage/Giga-Embeddings-instruct-480M-0826: 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 ai-sage/Giga-Embeddings-instruct-480M-0826 --plan --share

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