We used AWQ to quantize a model with the same architecture as LLaMA2. After quantization, the VRAM usage during loading was only 6567M, but the VRAM usage reached 32223M when generating up to 500 tokens during inference. Is this characteristic inherent to AWQ, or is there an error in our implementation? Looking forward to your response.
We used AWQ to quantize a model with the same architecture as LLaMA2. After quantization, the VRAM usage during loading was only 6567M, but the VRAM usage reached 32223M when generating up to 500 tokens during inference. Is this characteristic inherent to AWQ, or is there an error in our implementation? Looking forward to your response.