eaddario/imatrix-calibration
Viewer • Updated • 299 • 9.13k • 75
Ed Addario PRO
eaddario
AI & ML interests
Finding ways to optimize LLMs' inference performance in resource-constrained environments (e.g. commodity hardware, desktops, laptops, mobiles, edge devices, etc.)
Recent Activity
posted an update 3 days ago
Experimental JQK and global target bits‑per‑weight quantization of Nanbeige/Nanbeige4.3-33B.
Unlike standard llama.cpp quantization that relies on fixed type heuristics (e.g., Q4_K_M), the Target BPW approach automatically optimizes per-tensor precision where it matters the most, and produces high quality models that meet a precise global size target.
Key Advantages:
- Better model quality: using JQK types can improve KLD between 10% to 50% compared to standard QK types, at the same bpw.
- VRAM Maximization: Can generate high quality models sized exactly to fit hardware constraints (e.g., fitting the model into exactly 24GB VRAM).
- Data-Driven Precision: Quantization mix is determined by actual weight error sensitivity rather than hardcoded rules, often yielding better PPL/KLD size trade-offs.
Information about JQK types: https://huggingface.co/eaddario/JQ2_K-JQ6_K-Test-Models
Full benchmarks (PPL, KLD, ARC, GPQA, MMLU, etc.) and methodology in the model's cards.
JQK Quantization: https://huggingface.co/eaddario/Nanbeige4.3-33B-JQK
Global target BPW: https://huggingface.co/eaddario/Nanbeige4.3-33B-GGUF updated a model 3 days ago
eaddario/Nanbeige4.3-33B-GGUF updated a model 3 days ago
eaddario/Nanbeige4.3-33B-JQKOrganizations
