(&self, on_disk_bytes: u64, dtype_bytes: u64)
| 94 | } |
| 95 | |
| 96 | fn estimate_layer_vram(&self, on_disk_bytes: u64, dtype_bytes: u64) -> u64 { |
| 97 | // FP8 is 1 byte per element on disk, expands to dtype_bytes in memory |
| 98 | on_disk_bytes * dtype_bytes |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | /// GPTQ 4-bit quantization — wraps `gptq::load_gptq_var_builder`. |