brandonbeiler/InternVL3_5-14B-FP8-Dynamic

🔥 InternVL3_5-14B-FP8-Dynamic 🔥

This is a fp8 dynamic (w8a8) version of OpenGVLab/InternVL3_5-14B, optimized for high-performance inference with vLLM. The model utilizes fp8 dynamic (w8a8) for optimal performance and deployment.

Just Run It (vLLM serve)

You can serve the model using vLLM's OpenAI-compatible API server.

vllm serve brandonbeiler/InternVL3_5-14B-FP8-Dynamic \
    --quantization compressed-tensors \
    --served-model-name internvl3_5-14b \
    --reasoning-parser qwen3 \
    --trust-remote-code \
    --max-model-len 32768 \
    --tensor-parallel-size 1 # Adjust based on your GPU setup

Notes

• —32k max context length
• —reasoning parser ready to go, requires system prompt to run in thinking mode
• —still investigating tool calling

🚀 Key Features

• —FP8 Dynamic Quantization: No calibration required, ready to use immediately
• —Vision-Language Optimized: Specialized quantization recipe that preserves visual understanding
• —vLLM Ready: Seamless integration with vLLM for production deployment
• —Memory Efficient: ~50% memory reduction compared to FP16 original
• —Performance Boost: Significant faster inference on H100/L40S GPUs

📊 Model Details

• —Original Model: OpenGVLab/InternVL3_5-14B
• —Source Model: OpenGVLab/InternVL3_5-14B
• —Quantized Model: InternVL3_5-14B-FP8-Dynamic
• —Quantization Method: FP8 Dynamic (W8A8)
• —Quantization Library: LLM Compressor v0.7.1
• —Quantized by: brandonbeiler

🏗️ Technical Specifications

Hardware Requirements

• —Inference: ? VRAM (+ VRAM for context)
• —Supported GPUs: H100, L40S, A100 (80GB), RTX 4090 (2x for tensor parallelism)
• —GPU Architecture: Latest NVIDIA GPUs (Ada Lovelace, Hopper and later) and latest AMD GPUs. Recommended for NVIDIA GPUs with compute capability >=9.0 (Hopper and Blackwell)

Quantization Details

• —Weights: FP8 E4M3 with dynamic per-tensor scales
• —Activations: FP8 E4M3 with dynamic per-tensor scales
• —Preserved Components: Vision tower, embeddings, mlp1

🔬 Package Versions

This model was created using:

llmcompressor==0.7.1
compressed-tensors==0.10.2
transformers==4.55.0
torch==2.7.1
vllm==0.10.1.1

Quantized with ❤️ using LLM Compressor for the open-source community