Skip to content

VRAM reference

Qwen2.5 72B on an H200

Yes, you can fine-tune Qwen2.5 72B on one H200 (141 GB) with QLoRA. 4-bit inference: yes, 41 GB of weights, room for 128K tokens of context or 64 concurrent 4K-token sequences.

By Will Schulz · Checked 15 November 2026 · Prices as of November 2026

Verdict by method

Fine-tuningMemory at 2K tokensOne H200?Longest sequence (batch 1)Largest batch at 2KLargest LoRA rank
QLoRA (4-bit base)51 GBYes16K tokens13256
LoRA (16-bit base)155 GBNo———
Full fine-tuning1,169 GBNo——n/a
InferenceMemory at 8KOne H200?Longest contextConcurrent 4K sequences
16-bit (BF16/FP16)148 GBNo——
8-bit75 GBYes128K tokens40
4-bit44 GBYes128K tokens64
  • QLoRA: yes, about 51 GB at 2K tokens, batch 1, rank 16; up to 16K-token sequences at batch 1 or batch 13 at 2K, LoRA rank up to 256.
  • LoRA: no, it needs about 155 GB at 2K tokens, more than the 141 GB available.
  • Full fine-tuning: no, it needs about 1,169 GB at 2K tokens, more than the 141 GB available.

Where the memory goes

At 2,048 tokens per sequence, batch 1, LoRA rank 16, gradient checkpointing on
MethodWeightsTraining stateActivationsTotalOn one H20080 GB GPUs needed
QLoRA (4-bit base)41 GB3.4 GB6.0 GB51 GBFits1
LoRA (16-bit base)145 GB3.4 GB6.0 GB155 GBDoesn't fit3
Full fine-tuning145 GB1,018 GB6.0 GB1,169 GBDoesn't fit17
One sequence, fp16 KV cache
PrecisionWeightsKV cache at 8KTotalAt 128K contextOn one H200
16-bit (BF16/FP16)145 GB2.7 GB148 GB188 GBDoesn't fit
8-bit73 GB2.7 GB75 GB116 GBFits
4-bit41 GB2.7 GB44 GB84 GBFits

The cheapest way to run it

Renting an H200 costs $5.29 an hour on Runpod Secure Cloud ($3.59 on Runpod Community Cloud). QLoRA on it processes about 1,061 tokens a second by our estimate, so 10 million training tokens take 2.6 hours and cost about $13.85.

Job (default settings)MemoryCheapest rental that fitsPer hour
QLoRA fine-tuning51 GBA100 80GB$2.19
LoRA fine-tuning155 GB4 × L40S$4.36
4-bit inference44 GBA100 80GB$2.19
16-bit inference148 GB4 × L40S$4.36

On-demand rates seen on 15 November 2026: Runpod Secure Cloud, Lambda for the A100 40GB. Consumer cards count as single GPUs only; data-centre GPUs up to eight per node. Kojen's GPU training can launch the H200 on Runpod, Vast.ai or Lambda, or pick the cheapest GPU in stock for you.

Try other settings

JobMemoryH200 (141 GB)Smallest single GPU
QLoRA (4-bit base)2K tokens × batch 1, rank 1651 GBFitsA100 80GB (80 GB)
LoRA (16-bit base)2K tokens × batch 1, rank 16155 GB2 GPUsB200 (180 GB)
Full fine-tuning2K tokens × batch 11,169 GB10 GPUsNone: needs several
Inference, 16-bit (BF16/FP16)8K-token context × 1 sequence148 GB2 GPUsB200 (180 GB)
Inference, 8-bit8K-token context × 1 sequence75 GBFitsH200 (141 GB)
Inference, 4-bit8K-token context × 1 sequence44 GBFitsA100 80GB (80 GB)

A job fits when the estimate is within 90% of the card's memory. Training assumes gradient checkpointing and FlashAttention; multi-GPU counts assume the memory splits evenly (FSDP, ZeRO-3 or tensor parallelism).

How we estimate it

Every figure comes from the model's published architecture (layers, hidden size, attention and KV heads, vocabulary) rather than a rule of thumb, using the same functions as the calculator above. Sizes are in GB (10⁹ bytes).

  • Full fine-tuning: 16 bytes per parameter for mixed-precision Adam (bf16 weights and gradients, fp32 master weights and two moments), as in the ZeRO paper.
  • LoRA: frozen bf16 weights (2 bytes per parameter) plus 16 bytes per adapter parameter, rank 16 on every linear layer. For mixture-of-experts models, adapters go on attention only and the experts stay frozen.
  • QLoRA: NF4 weights with double quantisation (4.127 bits per parameter); embeddings and the output head stay in bf16, as bitsandbytes leaves them.
  • Activations: gradient checkpointing and FlashAttention. Each layer's bf16 input is kept, one layer is recomputed at a time, and the output logits take 8 bytes per vocabulary entry per token (fp32 values and gradients). Chunked cross-entropy, as in Unsloth or Liger kernels, cuts that last term.
  • Inference: 2, 1 or about 0.56 bytes per weight (4-bit with group-wise scales), plus an fp16 KV cache of 2 × layers × KV heads × head dim × 2 bytes per token; sliding-window layers only cache their window.
  • Fits means the estimate is within 90% of the card's listed memory; the rest covers the CUDA context, kernels and fragmentation. Multi-GPU counts assume memory splits evenly.

Speed estimates use H200's dense BF16 peak of 989 TFLOPS, about 4 FLOPs per active parameter per token for LoRA and 40% utilisation; see fine-tuning and training costs for the method.

Frequently asked questions

Can I fine-tune Qwen2.5 72B on an H200?
QLoRA: yes, about 51 GB at 2K tokens, batch 1, rank 16; up to 16K-token sequences at batch 1 or batch 13 at 2K, LoRA rank up to 256. LoRA: no, it needs about 155 GB at 2K tokens, more than the 141 GB available. Full fine-tuning: no, it needs about 1,169 GB at 2K tokens, more than the 141 GB available.
Can I run Qwen2.5 72B on an H200?
16-bit (BF16/FP16) inference: no, the weights alone are 145 GB. 8-bit inference: yes, 73 GB of weights, room for 128K tokens of context or 40 concurrent 4K-token sequences. 4-bit inference: yes, 41 GB of weights, room for 128K tokens of context or 64 concurrent 4K-token sequences.
What does fine-tuning Qwen2.5 72B on an H200 cost?
With QLoRA at our planning rate of 40% utilisation, about 2.6 hours for 10 million training tokens, around $13.85 at $5.29 an hour (Runpod Secure Cloud, November 2026).

Sources

  1. [1]Hugging Face. Qwen/Qwen2.5-72B-Instruct config.json
  2. [2]Alibaba Qwen. Qwen/Qwen2.5-72B-Instruct model card
  3. [3]Alibaba Qwen. Qwen2.5 Technical Report (arXiv:2412.15115)
  4. [4]NVIDIA. H200 specifications
  5. [5]Runpod. GPU pricing (updated 12 November 2026)
  6. [6]Lambda. GPU cloud pricing
  7. [7]Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
  8. [8]Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  9. [9]Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
  10. [10]Korthikanti et al., 2022. Reducing Activation Recomputation in Large Transformer Models (arXiv:2205.05198)

Measure before you rent the GPU.

Prompt tuning and model search in Kojen run on hosted models against your own test cases, so you know whether fine-tuning Qwen2.5 72B is worth the GPU time.

Start training free