Qwen2.5 72B Instruct (FP16) — 47.4 GBon NVIDIA L40S

FP16 NVIDIA L40S

Overview

Qwen2.5 72B Instruct is a 72.71B parameter dense language model by Qwen, with code, multilingual, tool-calls capabilities. It supports a context window of up to 32,768 tokens.

Qwen2.5 72B Instruct is a 72.71-billion-parameter dense transformer from Alibaba's Qwen team, fine-tuned for instruction following, code generation, and multilingual tasks. It competes with other leading 70B instruct models while supporting 14 languages including English, Chinese, Arabic, and Japanese. The model provides native tool calling and structured output capabilities. With a 32K context window and grouped-query attention, it quantizes efficiently for self-hosted inference on high-end consumer or server-class GPU configurations.

At FP16 quantization (full-precision quality tier), the model weighs 135.84 GB. This exceeds the 48 GB of VRAM on NVIDIA L40S. Inference is still possible via CPU offload or memory-mapped loading from disk, but expect significantly reduced performance.

The NVIDIA L40S is a datacenter GPU with 48 GB of GDDR6 VRAM and 864 GB/s memory bandwidth. It delivers 362 FP16 TFLOPS with Ada Lovelace architecture. A versatile GPU for AI inference, training, and graphics workloads. Handles quantized models up to 40B parameters comfortably.

Hardware Requirements

Model size	135.84 GB
VRAM available	48 GB
VRAM used	47.4 GB
Min RAM required	90 GB
GPU layers	27 / 80
Context size	1,013
Backend	cuda13
Flash attention	Yes

Performance Notes

Reduced performance

53 of 80 layers run on CPU. Your system needs at least 90 GB RAM. If system RAM is insufficient, data will be read from disk on demand, with significant performance impact.

Shorter conversations

Context size has been reduced due to memory constraints. This may affect the model's ability to handle long conversations.

Deploy

Server Local

Prerequisites

Ensure your GPU nodes are prepared with the NVIDIA container toolkit:

ansible-playbook prositronic.infra.nvidia_container_toolkit

Command

helmfile --state-values-file <(curl -s https://www.prositronic.eu/values/qwen2-5-72b-instruct/fp16/nvidia-l40s.yaml) apply

Generated values.yaml

/values/qwen2-5-72b-instruct/fp16/nvidia-l40s.yaml

Loading values…

Install llama.cpp

Install llama.cpp from the official build instructions for your platform and backend.

Download Model

curl -L -o qwen2-5-72b-instruct.gguf "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct-GGUF/resolve/main/qwen2.5-72b-instruct-fp16-00001-of-00042.gguf"

Start Server

llama-server \
  -m qwen2-5-72b-instruct.gguf \
  --n-gpu-layers 27 \
  --ctx-size 1013 \
  --flash-attn

Verify

curl http://localhost:8080/health

Explore

• Configure a deployment
• Browse models
• Browse hardware

Other quantizations

• Qwen2.5 72B Instruct (Q8_0)
• Qwen2.5 72B Instruct (Q5_K_M)
• Qwen2.5 72B Instruct (Q4_K_M)

Other hardware

• Qwen2.5 72B Instruct on CPU Only
• Qwen2.5 72B Instruct on NVIDIA A100 80GB
• Qwen2.5 72B Instruct on NVIDIA H100

Frequently Asked Questions

How much VRAM does Qwen2.5 72B Instruct (FP16) need?

The FP16 quantization of Qwen2.5 72B Instruct requires 135.84 GB. 27 of 80 layers fit in the 48 GB of VRAM on NVIDIA L40S; remaining layers are offloaded to CPU.

Can I run Qwen2.5 72B Instruct on NVIDIA L40S?

Yes, with reduced performance. NVIDIA L40S can run Qwen2.5 72B Instruct (FP16), but only 27 of 80 layers fit in VRAM. The rest are offloaded to CPU.

What is quantization?

Quantization reduces a model's numerical precision from its original floating-point format to a more compact representation. This shrinks the file size and VRAM footprint, making it possible to run large models on consumer hardware. The trade-off is a small reduction in output quality. FP16 compresses Qwen2.5 72B Instruct from its original size down to 135.84 GB.

What quantization should I choose for Qwen2.5 72B Instruct?

FP16 is a full-precision quantization. Higher-quality quants (Q8, Q6) preserve more model accuracy but need more VRAM. Lower quants (Q4, Q3, Q2) reduce VRAM usage at the cost of some quality. Choose based on your available hardware and quality requirements.

What is flash attention and why is it enabled?

Flash attention is a memory-efficient algorithm that speeds up the attention mechanism in transformer models. It reduces VRAM usage during inference by avoiding the materialisation of the full attention matrix. For Qwen2.5 72B Instruct on NVIDIA L40S, flash attention is enabled to maximise context length and throughput within the available 48 GB of VRAM.

Why are some layers offloaded to CPU?

NVIDIA L40S has 48 GB of VRAM, but Qwen2.5 72B Instruct (FP16) requires approximately 135.84 GB. Only 27 of 80 layers fit in VRAM; the remaining layers run on CPU, which is slower but still functional.

How do I run Qwen2.5 72B Instruct (FP16) with Ollama?

Run ollama run qwen2.5:72b-instruct-fp16 to start Qwen2.5 72B Instruct (FP16). Ollama handles downloading the model weights automatically on first run.

Last updated: March 5, 2026