Skip to content

VRAM reference

Qwen2.5 72B VRAM requirements

Fine-tuning Qwen2.5 72B takes about 51 GB of GPU memory with QLoRA, 155 GB with LoRA on 16-bit weights and 1,169 GB for full fine-tuning, at 2,048 tokens per sequence. Running it takes 148 GB in 16-bit or 44 GB in 4-bit with a 8K-token context.

By Will Schulz · Architecture and prices checked 15 November 2026 · Prices as of November 2026

Fine-tuning memory

At 2,048 tokens per sequence, batch 1, LoRA rank 16, gradient checkpointing on
MethodWeightsTraining stateActivationsTotalSmallest single GPU80 GB GPUs needed
QLoRA (4-bit base)41 GB3.4 GB6.0 GB51 GBA100 80GB (80 GB)1
LoRA (16-bit base)145 GB3.4 GB6.0 GB155 GBB200 (180 GB)3
Full fine-tuning145 GB1,018 GB6.0 GB1,169 GBNone17

On a 24 GB card such as the RTX 4090: QLoRA: no, it needs about 51 GB at 2K tokens, more than the 24 GB available.

Inference memory

One sequence, fp16 KV cache
PrecisionWeightsKV cache at 8KTotalAt 128K contextSmallest single GPU
16-bit (BF16/FP16)145 GB2.7 GB148 GB188 GBB200 (180 GB)
8-bit73 GB2.7 GB75 GB116 GBH200 (141 GB)
4-bit41 GB2.7 GB44 GB84 GBA100 80GB (80 GB)

Which GPUs fit Qwen2.5 72B

One GPU unless a count is shown. Follow a GPU's name for its maximum sequence length, batch size, LoRA rank and rental cost.

GPUQLoRALoRAFull fine-tuningRun, fp16Run, int8Run, int4
RTX 3090 24 GB3×8×55×7×4×3×
RTX 4090 24 GB3×8×55×7×4×3×
RTX 5090 32 GB2×6×41×6×3×2×
L40S 48 GB2×4×28×4×2×2×
A100 40GB 40 GB2×5×33×5×3×2×
A100 80GB 80 GBYes3×17×3×2×Yes
H100 SXM 80 GBYes3×17×3×2×Yes
H200 141 GBYes2×10×2×YesYes
B200 180 GBYesYes8×YesYesYes

What's specific to Qwen2.5 72B

  • Slightly larger than Llama 70B (72.7B parameters), and its license differs from the Apache 2.0 used for the smaller Qwen2.5 models.
  • 16-bit inference needs about 145 GB, so two 80 GB GPUs or one B200.
ArchitectureQwen2.5 72B
Parameters72.7B
Layers80
Hidden size8,192
MLP width29,568
Attention heads / KV heads64 / 8 (head dim 128)
Vocabulary152,064 tokens
Context128K tokens (32K native)
LicenseQwen License
ReleasedSeptember 2024

Try other settings

Pre-filled for Qwen2.5 72B. Change the sequence length, batch, LoRA rank, context or GPU.

JobMemoryRTX 4090 (24 GB)Smallest single GPU
QLoRA (4-bit base)2K tokens × batch 1, rank 1651 GB3 GPUsA100 80GB (80 GB)
LoRA (16-bit base)2K tokens × batch 1, rank 16155 GB8 GPUsB200 (180 GB)
Full fine-tuning2K tokens × batch 11,169 GB55 GPUsNone: needs several
Inference, 16-bit (BF16/FP16)8K-token context × 1 sequence148 GB7 GPUsB200 (180 GB)
Inference, 8-bit8K-token context × 1 sequence75 GB4 GPUsH200 (141 GB)
Inference, 4-bit8K-token context × 1 sequence44 GB3 GPUsA100 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.

Frequently asked questions

Can I fine-tune Qwen2.5 72B on a 24 GB GPU?
On an RTX 3090 or 4090 (24 GB): QLoRA: no, it needs about 51 GB at 2K tokens, more than the 24 GB available. LoRA: no, it needs about 155 GB at 2K tokens, more than the 24 GB available. Estimates use gradient checkpointing and leave 10% of memory free.
How much VRAM do you need to run Qwen2.5 72B?
About 148 GB in 16-bit and 44 GB in 4-bit, at 8K tokens of context. The smallest single GPU here for 4-bit is the A100 80GB (80 GB); for 16-bit, the B200 (180 GB).
How many GPUs does full fine-tuning Qwen2.5 72B take?
About 1,169 GB with mixed-precision Adam at 2K tokens, so at least 17 × 80 GB H100s with the training state sharded (FSDP or ZeRO-3). LoRA needs 155 GB and QLoRA 51 GB.
How long a context can Qwen2.5 72B handle on an 80 GB GPU?
Its KV cache costs 43 GB at the full 128K tokens (fp16 cache). On one 80 GB GPU: 16-bit weights don't fit, 64K tokens with 4-bit weights.

Terms used on this page: LoRA, QLoRA, KV cache, gradient checkpointing, quantization. For any model size, use the fine-tuning VRAM calculator; for the method in depth, the Academy lesson Memory first: what hardware can train what.

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]Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
  5. [5]Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  6. [6]Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
  7. [7]Korthikanti et al., 2022. Reducing Activation Recomputation in Large Transformer Models (arXiv:2205.05198)
  8. [8]Chen et al., 2016. Training Deep Nets with Sublinear Memory Cost (arXiv:1604.06174)

Architecture read from config.json and parameter count from the Hugging Face safetensors metadata on 15 November 2026.

Find out if you need to fine-tune at all.

Kojen's prompt tuning and model search run for real on your own test cases, on hosted models. Measure the cheaper levers before you rent a GPU.

Start training free