Hugging Face's logo

Datasets:
odyn-network
/
benchmark-dataset-finetune

Tasks:
Other
Modalities:
Tabular
Text
Formats:
csv
Size:
< 1K
ArXiv:
Tags:
fine-tuning
vram
lora
qlora
benchmark
gpu
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
description
string
expected_peak_vram_gb
float64
validation_status
string
math_engine_min_vram_gb
float64
math_engine_min_vram_v2
float64
math_engine_tier_gb
int64
math_engine_tier_gb_v2
int64
vram_vs_expected_pct
float64
vram_vs_expected_percentage
float64
tier_vs_expected_pct
float64
tier_vs_expected_percentage
float64
breakdown_weights_gb
float64
breakdown_activations_gb
float64
breakdown_optimizer_gb
float64
breakdown_gradients_gb
float64
breakdown_overhead_gb
float64
input_param_b
float64
input_context_length
int64
input_batch_size
int64
input_gradient_accumulation_steps
int64
input_lora_rank
int64
input_precision
string
gradient_checkpointing
bool
source
string
source_url
string
Llama 3.1 8B LoRA bf16 ctx=4096 bs=1 with GC
27
estimated
21.8
27
24
40
-19.3
0
-11.1
0
14.9
5.6
0.08
0.04
2.5
8
4,096
1
1
16
bf16
true
Estimated: 65% GC activation reduction applied to ctx=4096 estimate
https://arxiv.org/pdf/2308.03303
Llama 2 7B LoRA bf16 ctx=8192 bs=1 with GC+FA2
25.7
confirmed
19.3
26.1
24
40
-24.9
1.6
-6.6
0
13.04
9.21
0.63
0.32
2.5
7
8,192
1
1
16
bf16
true
LongLoRA arXiv:2309.12307 Table 12
https://arxiv.org/pdf/2309.12307v1
Llama 2 7B LoRA bf16 ctx=16384 bs=1 with GC+FA2
34.7
confirmed
19.3
32.6
24
40
-44.4
-6.1
-30.8
0
13.04
18.21
0.63
0.32
2.5
7
16,384
1
1
16
bf16
true
LongLoRA arXiv:2309.12307 Table 12
https://arxiv.org/pdf/2309.12307v1
Llama 2 7B LoRA bf16 ctx=32768 bs=1 with GC+FA2
46.5
confirmed
19.3
45.7
24
48
-58.5
-1.7
-48.4
0
13.04
30.01
0.63
0.32
2.5
7
32,768
1
1
16
bf16
true
LongLoRA arXiv:2309.12307 Table 12
https://arxiv.org/pdf/2309.12307v1
Llama 3 8B QLoRA ctx=2048 bs=2 with GC
23
confirmed
10.9
17.5
12
24
-52.6
-23.9
-47.8
0
3.73
3.4
0.08
0.04
1.5
8
2,048
2
2
32
nf4
true
Medical LoRA arXiv:2408.10715
https://arxiv.org/pdf/2408.10715
Mistral 7B LoRA bf16 ctx=4096 bs=1 with GC
28
estimated
19.3
22.6
24
24
-31.1
-19.3
-14.3
-40
13.04
12
0.25
0.12
2.5
7
4,096
1
1
16
bf16
true
Extrapolated: LoRA-FA activation scaling + GC reduction
https://arxiv.org/pdf/2308.03303
Qwen 2.5 14B LoRA bf16 ctx=2048 bs=1 with GC
40
estimated
36.8
41
40
48
-8
2.5
0
20
26.08
8
0.5
0.25
1.5
14
2,048
1
1
16
bf16
true
Rule-of-thumb estimate (exxactcorp.com)
https://www.exxactcorp.com/blog/deep-learning/ai-fine-tuning-with-lora
Qwen 2.5 14B LoRA bf16 ctx=4096 bs=1 with GC
55
estimated
36.8
45.3
40
48
-33.1
-17.6
-27.3
-40
26.08
22
0.5
0.25
2.5
14
4,096
1
1
16
bf16
true
Extrapolated: ctx=2048 estimate scaled to ctx=4096
https://arxiv.org/pdf/2308.03303
Llama 3.2 3B LoRA bf16 ctx=8192 bs=1 with GC
18
estimated
9.3
20.2
10
24
-48.3
12.2
-44.4
0
5.59
10.5
0.09
0.04
1.5
3
8,192
1
1
16
bf16
true
Extrapolated from LoRA-FA scaling
https://arxiv.org/pdf/2308.03303
Llama 3.1 8B LoRA r=64 bf16 ctx=2048 bs=1 with GC
22
unverified
25.6
26.3
40
40
16.4
19.5
81.8
66.7
14.9
1.96
0.33
0.16
2.5
8
2,048
1
1
64
bf16
true
arXiv:2406.02290 + GC
https://arxiv.org/html/2406.02290v2

Fine-Tuning VRAM Benchmark Dataset

Benchmark dataset for evaluating the accuracy of the Odyn Smart Digester VRAM Math Engine for fine-tuning workloads.

Compares the V1 (initial) and V2 (updated) engine estimates against expected peak VRAM values sourced from published research papers and hardware measurements.

Dataset Details

  • 10 workload rows — all with gradient checkpointing enabled
  • Methods covered — LoRA (bf16) and QLoRA (NF4)
  • Models — Llama 2 7B, Llama 3/3.1/3.2 8B/3B, Mistral 7B, Qwen 2.5 14B
  • Context lengths — 2048 to 32768 tokens

Accuracy Summary

Engine Mean VRAM Accuracy
V1 (initial) 66.3%
V2 (updated) 89.6%

Columns

Column Description
description Workload identifier
expected_peak_vram_gb Ground truth VRAM from paper or hardware measurement
validation_status confirmed (hardware measured), estimated (extrapolated), or unverified
math_engine_min_vram_gb V1 engine recommended minimum VRAM (GB)
math_engine_min_vram_v2 V2 engine recommended minimum VRAM (GB)
math_engine_tier_gb V1 GPU tier selected
math_engine_tier_gb_v2 V2 GPU tier selected
vram_vs_expected_pct V1 error % vs expected
vram_vs_expected_percentage V2 error % vs expected
tier_vs_expected_pct V1 tier error % vs expected tier
tier_vs_expected_percentage V2 tier error % vs expected tier
breakdown_* Per-component VRAM breakdown (weights, activations, optimizer, gradients, overhead)
input_* Input parameters used (param count, context length, batch size, LoRA rank, precision)
gradient_checkpointing Always true in this dataset
source Description of the expected value source
source_url Link to source paper or reference

V2 Engine Improvements

The V2 math engine includes four calibrated fixes over V1:

  1. GC activation coefficient — changed from 2 to 4, calibrated from LongLoRA Table 12 hardware measurements
  2. ZeRO-2 distributed training — added num_gpus parameter; optimizer and gradient states sharded ÷ n_gpus with communication buffer
  3. QLoRA activation memory — replaced dequant-only formula with Korthikanti O(s) formula + dequantisation buffer
  4. QLoRA optimizer — corrected from 12 bytes/param (fp32 AdamW) to 4 bytes/param (8-bit paged AdamW)

Calibration Sources

Downloads last month
25
Number of rows:
10
Total file size:
8.32 kB

Papers for odyn-network/benchmark-dataset-finetune