Datasets:

odyn-network
/

benchmark-finetune-lora-v1

description stringlengths 38 100	expected_peak_vram_gb float64 3 74	validation_status stringclasses 3 values	math_engine_peak_vram_gb float64 2.55 77	math_engine_tier_gb int64 8 80	vram_vs_expected_pct float64 -58.9 59.9	tier_vs_expected_pct float64 -30.4 233	breakdown_weights_gb float64 1.51 65.4	breakdown_activations_gb float64 0.17 56	breakdown_optimizer_gb float64 0.06 5.13	breakdown_gradients_gb float64 0.01 0.85	breakdown_temp_buffers_gb float64 0.18 23.5	breakdown_overhead_gb float64 0.25 2	measurement_scope stringclasses 2 values	input_param_b float64 0.8 35	input_context_length int64 256 65.5k	input_batch_size int64 1 8	input_gradient_accumulation_steps int64 1 8	input_lora_rank float64 16 128 ⌀	input_precision stringclasses 3 values	input_num_gpus int64 1 8	input_parallelism stringclasses 3 values	tolerance_pct int64 10 15	gradient_checkpointing bool 1 class	source stringlengths 21 80	source_url stringlengths 32 97
Llama 2 7B plain LoRA per-GPU peak ctx=8192, bs=1/GPU, grad_accum=8, FA2, ZeRO-2, 8×A100	25.7	confirmed	24.95	40	-2.9	55.6	13.12	4	0.06	0.01	5.87	2	per_gpu_distributed	7	8,192	1	8	null	bf16	8	ddp_zero2	15	true	LongLoRA arXiv:2309.12307 Table 12 (plain LoRA row)	https://arxiv.org/html/2309.12307v2
Llama 2 7B plain LoRA per-GPU peak ctx=16384, bs=1/GPU, grad_accum=8, FA2, ZeRO-2, 8×A100	34.7	confirmed	40.82	48	17.6	38.3	13.12	14	0.06	0.01	11.74	2	per_gpu_distributed	7	16,384	1	8	null	bf16	8	ddp_zero2	15	true	LongLoRA arXiv:2309.12307 Table 12 (plain LoRA row)	https://arxiv.org/html/2309.12307v2
Llama 2 7B plain LoRA per-GPU peak ctx=32768, bs=1/GPU, grad_accum=8, FA2, ZeRO-2, 8×A100	46.5	confirmed	66.56	80	43.1	72	13.12	28	0.06	0.01	23.48	2	per_gpu_distributed	7	32,768	1	8	null	bf16	8	ddp_zero2	15	true	LongLoRA arXiv:2309.12307 Table 12 (plain LoRA row)	https://arxiv.org/html/2309.12307v2
Llama 2 7B plain LoRA per-GPU peak ctx=65536, bs=1/GPU, grad_accum=8, FA2, ZeRO-2, 8×A100	71.1	confirmed	76.95	80	8.2	12.5	13.12	56	0.06	0.01	5.87	2	per_gpu_distributed	7	65,536	1	8	null	bf16	8	ddp_zero2	15	true	LongLoRA arXiv:2309.12307 Table 12 (plain LoRA row)	https://arxiv.org/html/2309.12307v2
Llama 2 7B LongLoRA per-GPU peak ctx=8192, bs=1/GPU, grad_accum=8, FA2, ZeRO-2, 8×A100	25.6	confirmed	24.95	40	-2.5	56.2	13.12	4	0.06	0.01	5.87	2	per_gpu_distributed	7	8,192	1	8	null	bf16	8	ddp_zero2	15	true	LongLoRA arXiv:2309.12307 Table 12 (LongLoRA row)	https://arxiv.org/html/2309.12307v2
Llama 2 7B LongLoRA per-GPU peak ctx=16384, bs=1/GPU, grad_accum=8, FA2, ZeRO-2, 8×A100	34.6	confirmed	40.82	48	18	38.7	13.12	14	0.06	0.01	11.74	2	per_gpu_distributed	7	16,384	1	8	null	bf16	8	ddp_zero2	15	true	LongLoRA arXiv:2309.12307 Table 12 (LongLoRA row)	https://arxiv.org/html/2309.12307v2
Llama 2 7B LongLoRA per-GPU peak ctx=32768, bs=1/GPU, grad_accum=8, FA2, ZeRO-2, 8×A100	46.4	confirmed	66.56	80	43.5	72.4	13.12	28	0.06	0.01	23.48	2	per_gpu_distributed	7	32,768	1	8	null	bf16	8	ddp_zero2	15	true	LongLoRA arXiv:2309.12307 Table 12 (LongLoRA row)	https://arxiv.org/html/2309.12307v2
Llama 2 7B LongLoRA per-GPU peak ctx=65536, bs=1/GPU, grad_accum=8, FA2, ZeRO-2, 8×A100	69.8	confirmed	76.95	80	10.2	14.6	13.12	56	0.06	0.01	5.87	2	per_gpu_distributed	7	65,536	1	8	null	bf16	8	ddp_zero2	15	true	LongLoRA arXiv:2309.12307 Table 12 (LongLoRA row)	https://arxiv.org/html/2309.12307v2
Llama 2 7B ZeRO-3+LoRA per-GPU peak on 8×V100, seq=256, eff batch=8, rank=64	4.74	confirmed	2.76	8	-41.7	111	1.74	1	0.23	0.04	1.47	0.25	per_gpu_distributed	7	256	8	1	64	bf16	8	ddp_zero3	15	true	Singh et al. arXiv:2406.02290 Table 2	https://arxiv.org/html/2406.02290v2
Llama 2 13B ZeRO-3+LoRA per-GPU peak on 8×V100, seq=256, eff batch=8, rank=64	6.22	confirmed	2.55	8	-58.9	60.8	3.67	1.56	0.32	0.05	0.37	0.25	per_gpu_distributed	13	256	8	1	64	bf16	8	ddp_zero3	15	true	Singh et al. arXiv:2406.02290 Table 2	https://arxiv.org/html/2406.02290v2
Gemma-2B LoRA peak on A100 40GB, PubMed CPT, token_batch=512, rank=128, GC+FA+Unsloth	7.91	confirmed	7.59	8	-4	26.4	3.91	0.17	1.13	0.19	0.18	2	single_gpu	2	1,024	1	1	128	bf16	1	none	10	true	LlamaFactory arXiv:2403.13372 Table 4 (LoRA peak memory GB)	https://arxiv.org/html/2403.13372v4
Llama2-7B LoRA peak on A100 40GB, PubMed CPT, token_batch=512, rank=128, GC+FA+Unsloth	16.32	confirmed	21.27	24	30.3	47.1	13.66	0.5	3.75	0.62	0.73	2	single_gpu	7	1,024	1	1	128	bf16	1	none	10	true	LlamaFactory arXiv:2403.13372 Table 4 (LoRA peak memory GB)	https://arxiv.org/html/2403.13372v4
Llama2-13B LoRA peak on A100 40GB, PubMed CPT, token_batch=512, rank=128, GC+FA+Unsloth	30.09	confirmed	34.01	40	13	32.9	25.07	0.78	5.13	0.85	0.18	2	single_gpu	13	1,024	1	1	128	bf16	1	none	10	true	LlamaFactory arXiv:2403.13372 Table 4 (LoRA peak memory GB)	https://arxiv.org/html/2403.13372v4
Meta-Llama-3-8B LoRA FP16 r=16 batch=1 (GigaGPU measured VRAM table)	22	confirmed	19.99	24	-9.1	9.1	14.98	1	0.47	0.08	1.47	2	single_gpu	8	2,048	1	1	16	fp16	1	none	10	true	GigaGPU Llama 3 8B LoRA guide VRAM table	https://gigagpu.com/fine-tune-llama-3-8b-lora-gpu-guide/
Meta-Llama-3-8B LoRA FP16 r=16 batch=4 (GigaGPU measured VRAM table)	28	confirmed	27.4	40	-2.1	42.9	14.98	4	0.47	0.08	5.87	2	single_gpu	8	2,048	4	1	16	fp16	1	none	10	true	GigaGPU Llama 3 8B LoRA guide VRAM table	https://gigagpu.com/fine-tune-llama-3-8b-lora-gpu-guide/
Meta-Llama-3-8B LoRA FP16 r=64 batch=1 (GigaGPU measured VRAM table)	24	confirmed	21.87	24	-8.9	0	15.21	1	1.88	0.31	1.47	2	single_gpu	8	2,048	1	1	64	fp16	1	none	10	true	GigaGPU Llama 3 8B LoRA guide VRAM table	https://gigagpu.com/fine-tune-llama-3-8b-lora-gpu-guide/
Llama 3 8B LoRA r=64 peak on RTX 5090 32GB, Alpaca 52K, bf16, r=16, bs=4, grad_accum=4, 3 epochs	18.4	confirmed	29.27	40	59.1	30.4	15.21	1	1.88	0.31	1.47	2	single_gpu	8	2,048	4	4	64	bf16	1	none	10	true	GigaGPU best-GPU benchmarks LoRA rank 64 table	https://gigagpu.com/best-gpu-for-fine-tuning-llms/
Llama 3 8B LoRA r=64 peak on RTX 5090 24GB, Alpaca 52K, bf16, r=16, bs=4, grad_accum=4, 3 epochs	18.3	confirmed	29.27	40	59.9	31.1	15.21	1	1.88	0.31	1.47	2	single_gpu	8	2,048	4	4	64	bf16	1	none	10	true	GigaGPU best-GPU benchmarks LoRA rank 64 table	https://gigagpu.com/best-gpu-for-fine-tuning-llms/
Llama 3 8B LoRA r=64 peak on RTX 3090 24GB, Alpaca 52K, bf16, r=16, bs=4, grad_accum=4, 3 epochs	18.5	confirmed	29.27	40	58.2	29.7	15.21	1	1.88	0.31	1.47	2	single_gpu	8	2,048	4	4	64	bf16	1	none	10	true	GigaGPU best-GPU benchmarks LoRA rank 64 table	https://gigagpu.com/best-gpu-for-fine-tuning-llms/
Llama 3 8B LoRA r=64 peak on RTX 6000 Pro 48GB, Alpaca 52K, bf16, r=16, bs=4, grad_accum=4, 3 epochs	18.3	confirmed	29.27	40	59.9	31.1	15.21	1	1.88	0.31	1.47	2	single_gpu	8	2,048	4	4	64	bf16	1	none	10	true	GigaGPU best-GPU benchmarks LoRA rank 64 table	https://gigagpu.com/best-gpu-for-fine-tuning-llms/
Phi-3 Mini 3.8B LoRA r=16 measured requirement (GigaGPU LoRA vs full FT table)	8	confirmed	11.28	16	41	50	7.12	0.44	0.23	0.04	1.47	2	single_gpu	3.8	2,048	1	1	16	bf16	1	none	10	true	GigaGPU best-GPU benchmarks VRAM requirements table	https://gigagpu.com/best-gpu-for-fine-tuning-llms/
Phi-3 Mini 3.8B LoRA r=64 measured requirement (GigaGPU LoRA vs full FT table)	10	confirmed	12.19	16	21.9	60	7.23	0.44	0.9	0.15	1.47	2	single_gpu	3.8	2,048	1	1	64	bf16	1	none	10	true	GigaGPU best-GPU benchmarks VRAM requirements table	https://gigagpu.com/best-gpu-for-fine-tuning-llms/
Mistral 7B v0.3 LoRA r=64 measured requirement (GigaGPU LoRA vs full FT table)	17	confirmed	20.01	24	17.7	41.2	13.35	1	1.88	0.31	1.47	2	single_gpu	7	2,048	1	1	64	bf16	1	none	10	true	GigaGPU best-GPU benchmarks VRAM requirements table	https://gigagpu.com/best-gpu-for-fine-tuning-llms/
Qwen 2.5 14B LoRA r=64 measured requirement (GigaGPU LoRA vs full FT table)	30	confirmed	34.78	40	15.9	33.3	26.5	1.56	2.55	0.43	1.74	2	single_gpu	14	2,048	1	1	64	bf16	1	none	10	true	GigaGPU best-GPU benchmarks VRAM requirements table	https://gigagpu.com/best-gpu-for-fine-tuning-llms/
Llama 3.1 8B LoRA bf16 peak via Unsloth, seq=2048 (Clore framework comparison)	18	confirmed	27.4	40	52.2	122.2	14.98	4	0.47	0.08	5.87	2	single_gpu	8	2,048	4	1	16	bf16	1	none	10	true	Clore.ai fine-tuning tools comparison VRAM table (LoRA bf16 row)	https://docs.clore.ai/guides/comparisons/finetuning-comparison
Llama 3.1 8B LoRA bf16 peak via Axolotl, seq=2048 (Clore framework comparison)	24	confirmed	27.4	40	14.2	66.7	14.98	4	0.47	0.08	5.87	2	single_gpu	8	2,048	4	1	16	bf16	1	none	10	true	Clore.ai fine-tuning tools comparison VRAM table (LoRA bf16 row)	https://docs.clore.ai/guides/comparisons/finetuning-comparison
Llama 3.1 8B LoRA bf16 peak via LLaMA-Factory, seq=2048 (Clore framework comparison)	25	confirmed	27.4	40	9.6	60	14.98	4	0.47	0.08	5.87	2	single_gpu	8	2,048	4	1	16	bf16	1	none	10	true	Clore.ai fine-tuning tools comparison VRAM table (LoRA bf16 row)	https://docs.clore.ai/guides/comparisons/finetuning-comparison
Llama 3.1 8B LoRA bf16 peak via TRL, seq=2048 (Clore framework comparison)	26	confirmed	27.4	40	5.4	53.8	14.98	4	0.47	0.08	5.87	2	single_gpu	8	2,048	4	1	16	bf16	1	none	10	true	Clore.ai fine-tuning tools comparison VRAM table (LoRA bf16 row)	https://docs.clore.ai/guides/comparisons/finetuning-comparison
Llama 3.1 8B LoRA 16-bit peak via Unsloth, bs=4, A100 80GB (Clore speed benchmark)	22	confirmed	27.4	40	24.5	81.8	14.98	4	0.47	0.08	5.87	2	single_gpu	8	2,048	4	1	16	bf16	1	none	10	true	Clore.ai fine-tuning tools comparison speed benchmark (Unsloth full 16-bit LoRA)	https://docs.clore.ai/guides/comparisons/finetuning-comparison
Llama 3 8B bf16 LoRA community peak 16 GB (Unsloth blog, rank=32, Unsloth stack)	16	confirmed	23.09	24	44.3	50	15.06	1	0.94	0.16	1.47	2	single_gpu	8	2,048	2	4	32	bf16	1	none	10	true	Unsloth blog llama3 community bf16 LoRA measurement	https://www.unsloth.ai/blog/llama3
Qwen3.5-0.8B bf16 LoRA measured VRAM (Unsloth fine-tuning guide)	3	confirmed	4.39	8	46.3	233.3	1.51	0.34	0.14	0.02	0.37	2	single_gpu	0.8	2,048	1	1	16	bf16	1	none	10	true	Unsloth Qwen3.5 fine-tuning guide bf16 LoRA VRAM table	https://www.unsloth.ai/docs/models/qwen3.5/fine-tune
Qwen3.5-2B bf16 LoRA measured VRAM (Unsloth fine-tuning guide)	5	confirmed	6.62	8	32.4	100	3.75	0.34	0.14	0.02	0.37	2	single_gpu	2	2,048	1	1	16	bf16	1	none	10	true	Unsloth Qwen3.5 fine-tuning guide bf16 LoRA VRAM table	https://www.unsloth.ai/docs/models/qwen3.5/fine-tune
Qwen3.5-4B bf16 LoRA measured VRAM (Unsloth fine-tuning guide)	10	confirmed	11.66	16	16.6	20	7.49	0.44	0.23	0.04	1.47	2	single_gpu	4	2,048	1	1	16	bf16	1	none	10	true	Unsloth Qwen3.5 fine-tuning guide bf16 LoRA VRAM table	https://www.unsloth.ai/docs/models/qwen3.5/fine-tune
Qwen3.5-9B bf16 LoRA measured VRAM (Unsloth fine-tuning guide)	22	confirmed	21.86	24	-0.6	9.1	16.84	1	0.47	0.08	1.47	2	single_gpu	9	2,048	1	1	16	bf16	1	none	10	true	Unsloth Qwen3.5 fine-tuning guide bf16 LoRA VRAM table	https://www.unsloth.ai/docs/models/qwen3.5/fine-tune
Qwen3.5-27B bf16 LoRA measured VRAM (Unsloth fine-tuning guide)	56	confirmed	57.36	80	2.4	42.9	50.5	3.05	1.24	0.21	0.37	2	single_gpu	27	2,048	1	1	16	bf16	1	none	10	true	Unsloth Qwen3.5 fine-tuning guide bf16 LoRA VRAM table	https://www.unsloth.ai/docs/models/qwen3.5/fine-tune
Qwen3.5-35B-A3B bf16 LoRA measured VRAM (Unsloth fine-tuning guide)	74	confirmed	71.6	80	-3.2	8.1	65.37	2.62	1.07	0.18	0.37	2	single_gpu	35	2,048	1	1	16	bf16	1	none	10	true	Unsloth Qwen3.5 fine-tuning guide bf16 LoRA VRAM table	https://www.unsloth.ai/docs/models/qwen3.5/fine-tune
gpt-oss-20b bf16 LoRA measured VRAM requirement (Unsloth gpt-oss guide)	44	confirmed	43.41	48	-1.3	9.1	37.36	1.56	0.64	0.11	1.74	2	single_gpu	20	2,048	1	1	16	bf16	1	none	10	true	Unsloth gpt-oss fine-tuning guide BF16 LoRA requirements	https://unsloth.ai/docs/models/gpt-oss-how-to-run-and-fine-tune/tutorial-how-to-fine-tune-gpt-oss
Llama 3.1 8B LoRA bf16 ctx=4096 bs=1 with GC	27	estimated	22.46	24	-16.8	-11.1	14.98	2	0.47	0.08	2.94	2	single_gpu	8	4,096	1	1	16	bf16	1	none	10	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	25.53	40	-0.7	24.5	13.12	4	0.47	0.08	5.87	2	single_gpu	7	8,192	1	1	16	bf16	1	none	10	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	41.41	48	19.3	38.3	13.12	14	0.47	0.08	11.74	2	single_gpu	7	16,384	1	1	16	bf16	1	none	10	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	67.15	80	44.4	72	13.12	28	0.47	0.08	23.48	2	single_gpu	7	32,768	1	1	16	bf16	1	none	10	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	12.53	16	-45.5	-30.4	3.73	4.4	0.31	0.16	1.47	2	single_gpu	8	2,048	2	2	32	nf4	1	none	10	true	Medical LoRA arXiv:2408.10715	https://arxiv.org/pdf/2408.10715
Mistral 7B LoRA bf16 ctx=4096 bs=1 with GC	28	estimated	20.6	24	-26.4	-14.3	13.12	2	0.47	0.08	2.94	2	single_gpu	7	4,096	1	1	16	bf16	1	none	10	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	32.23	40	-19.4	0	26.18	1.56	0.64	0.11	1.74	2	single_gpu	14	2,048	1	1	16	bf16	1	none	10	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	35.53	40	-35.4	-27.3	26.18	3.12	0.64	0.11	3.48	2	single_gpu	14	4,096	1	1	16	bf16	1	none	10	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	15.51	16	-13.8	-11.1	5.63	1.75	0.23	0.04	5.87	2	single_gpu	3	8,192	1	1	16	bf16	1	none	10	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	21.87	24	-0.6	9.1	15.21	1	1.88	0.31	1.47	2	single_gpu	8	2,048	1	1	64	bf16	1	none	10	true	arXiv:2406.02290 + GC	https://arxiv.org/html/2406.02290v2

Odyn benchmark: LoRA fine-tuning peak VRAM (V1)

Curated benchmark rows for validating GPU memory estimators during LoRA fine-tuning. Each row pairs a published or measured expected peak VRAM with inputs to a math engine (model size, context length, batch, LoRA rank, precision, parallelism) plus optional VRAM breakdown and provenance.

This dataset is not Alpaca-style training JSONL. It is evaluation ground truth for placement / scheduler memory models (Odyn Smart Digester math engine).

Schema

Column	Type	Description
`description`	string	Human-readable scenario label
`expected_peak_vram_gb`	float	Reference peak VRAM (GB) from source
`validation_status`	string	`confirmed`, `estimated`, or `unverified`
`math_engine_peak_vram_gb`	float	Odyn math engine estimate (GB)
`math_engine_tier_gb`	float	Recommended GPU tier (GB)
`vram_vs_expected_pct`	float	`(math_engine - expected) / expected * 100`
`tier_vs_expected_pct`	float	Tier headroom vs expected
`breakdown_*_gb`	float	Weights, activations, optimizer, gradients, temp buffers, overhead
`measurement_scope`	string	e.g. `single_gpu`, `per_gpu_distributed`
`input_param_b`	float	Model size (billions of parameters)
`input_context_length`	int	Sequence / context length
`input_batch_size`	int	Per-step batch size
`input_gradient_accumulation_steps`	int	Gradient accumulation
`input_lora_rank`	int	LoRA rank (nullable)
`input_precision`	string	e.g. `bf16`, `fp16`, `nf4`
`input_num_gpus`	int	GPU count
`input_parallelism`	string	e.g. `none`, `ddp_zero2`, `ddp_zero3`
`tolerance_pct`	int	Acceptance band used in eval
`gradient_checkpointing`	bool	GC enabled
`source`	string	Citation / origin
`source_url`	string	Link to primary source

Sources

Rows cite LongLoRA, Singh et al. (ZeRO-3+LoRA), LlamaFactory, GigaGPU, Clore.ai, Unsloth, and other public VRAM tables. See source and source_url per row.

Usage

from datasets import load_dataset

ds = load_dataset("odyn-network/benchmark-finetune-lora-v1", split="train")
print(ds[0]["description"], ds[0]["expected_peak_vram_gb"])

Version

V1 — 48 scenarios (benchmark_finetune_lora_dataset_V1.csv)

Downloads last month: 47

Number of rows:

Total file size:

19.1 kB