IFM Lab releases xLLM training library: hot-swappable tokenizers and 1k-line Jinja chat templates
HildeKuehne · x · 2026-09-29
IFM Lab open-sourced xLLM, a training library built around a flexible online data pipeline for fine-tuning on chat/instruction data: tokenizers and data mixtures stay changeable during training, so a new tokenizer or chat template is just a config change, not a reprocessing run. Features include tokenization straight from JSONL, configurable mixtures with buffered shuffle, bestfit packing that keeps documents intact, and throughput unaffected by overlapping CPU prep with GPU training. The team says flexibility mattered in practice — they kept finding bugs in chat templates (their final Jinja is 1k lines) and only fast tokenizer/template swaps let them fix things mid-training.
More from Infra
- Training-free page-sparse decode in llama.cpp: 1.72× faster 128k decoding on an RTX 5090 — topppggg · 2026-11-16
- llama.cpp offloads CI caches and runners to Hugging Face to survive 35 PRs/day — ggerganov · 2026-11-16
- India's Sarvam launches Token Factory, hosting top open models on local infra — HarveenChadha · 2026-11-16
- ZGX Fury stress-tested: 351 tok/s on Qwen3.8-Flash-Next, offline agents and local RAG demos — Sam Witteveen · 2026-11-16
- Gemini 3.5 Flash now runs on air-gapped Google Distributed Cloud, offline up to 12 months — shashib · 2026-11-16
- Gemma.c: full Gemma 1/2/3 multimodal inference in a single C file, ~72 tok/s on CPU at INT8 — This-Purchase-3325 · 2026-11-16