Comparing INT4 and NVFP4 Palettes on Real Gradient Tensors
Four-bit training quantizes every number to one of 16 values. NVFP4's menu is {0, ±0.5, ±1, ±1.5, ±2, ±3, ±4, ±6} , with one scale factor per block…
Tech news from the best sources
Four-bit training quantizes every number to one of 16 values. NVFP4's menu is {0, ±0.5, ±1, ±1.5, ±2, ±3, ±4, ±6} , with one scale factor per block…
Cloud TPU v6e-1 ( ct6e-standard-1t , one v6e chip, 32 GB HBM), Compute Engine flex-start, europe-west4-a. All timings below measured 2026-08-19 unle…
В этом месяце Alibaba выпустила две модели, и та, о которой все писали, оказалась не той, что мы ждали. Qwen3.8-Max — это API с 2,4 триллионами пара…
Quantized Google's MedGemma-1.5-4B (a medical vision-language model) to INT4 (W4A16) via llm-compressor 's GPTQModifier, for self-hosted deployment.…
Originally published at kunalganglani.com — read it there for inline code, hero image, and live links. LLM Quantization Levels Compared: Q4_K_M vs Q…
Развернул Gemma 4 31B на одной 4090 (48 ГБ) — и проверил нужен ли «честный» Q8, и переживает ли tool-calling 4-бита. Q8 не дал ничего (+0.007 — шум)…
Quantization formats compared: GGUF vs GPTQ vs AWQ vs NF4 You just finished fine-tuning a 7B parameter model. The raw FP16 weights are 14 GB. Your t…
requirements hugging face account https://huggingface.co/ Setup llama.cpp git clone https://github.com/ggml-org/llama.cpp.git cmake -S llama.cpp -B…
Part 3 of the quantization series. Yesterday I tested whether Part 1's drift-inversion intervention generalizes beyond granite. I wrote down a falsi…