GGUF-based LoRA training fits Qwen3.5-35B-A3B into 16GB VRAM
woct0rdho · reddit · 2026-07-21
A GitHub project shows how to train Qwen3.5-35B-A3B LoRA models in 16 GB VRAM using GGUF as the base model format. The author claims GGUF is becoming a better base than bitsandbytes for low-VRAM training because it supports newer model types and quantization schemes, and reports that with APEX quant the model can fit at 13.3 GiB with fused dequant-matmul/MoE kernels. On Strix Halo, the setup reportedly reaches 6.5 s/it with batch size 1, 2048 token chunks, and LoRA rank 4, and the repo also spun out torch-ggml-ops for PyTorch bindings.
More from Infra
- Nearly half of Claude revenue now flows through cloud marketplace commitments, up from 11% — BenBajarin · 2026-10-26
- Anthropic ARR goes vertical after price cuts: tokens up 4x, $100B+ ARR back in play — GavinSBaker · 2026-10-26
- Alex Banks argues power caps GPU supply, real risk is demand cooling first — thealexbanks · 2026-10-26
- Analyst: Ant has the most fungible AI workload of any model provider today — BenBajarin · 2026-10-26
- Anthropic S-1: training infra edge cut per-task cost 40% and doubled revenue growth — BenBajarin · 2026-10-26
- Rewriting scaling law study: optimal synthetic data share rises from 10% to 32% with compute — heghbalz · 2026-10-26