vit-base-breast-cancer-phi-lattice

Second-Generation (v2) Release Available: See vit-base-breast-cancer-phi-lattice-v2 for the updated phi-lattice v2 build (19.51 dB 4-bit SQNR vs 18.60 dB in v1, 0.00028 logit MSE vs 0.00047 in v1, plus standalone low_vram_4bit and triton_4bit loaders in 46.3 MiB VRAM). This repository is preserved as the first-generation (v1) baseline for historical comparison and reproducibility.

Experimental build of the 85.8M-parameter medical Vision Transformer (Falah/vit-base-breast-cancer) quantized and pruned using the phi-lattice post-training optimization pass. Uploaded for testing and evaluation.

Included Formats

1. vit_phi_packed_4bit.pt (43.3 MB) — Bit-packed 4-bit uint8 linear projections (7.57x smaller than the 327.4 MB FP32 checkpoint).
2. model.safetensors (163.7 MB) — Standard float16 HuggingFace transformers checkpoint for direct loading via AutoModelForImageClassification.

Benchmark: Standard Uniform vs phi-lattice on vit-base-breast-cancer

Usage (transformers)

from transformers import AutoImageProcessor, AutoModelForImageClassification

repo_id = "ununununununununununununununununununun/vit-base-breast-cancer-phi-lattice"
processor = AutoImageProcessor.from_pretrained(repo_id)
model = AutoModelForImageClassification.from_pretrained(repo_id)