Community
Hobbyist runs DeepSeek-V4.1-Flash on eight V100 GPUs at over 200 tok/s for code
Using a 2-bit GGUF quant with custom kernels and pipeline parallelism across eight 2018 V100s, the author reports 215 tok/s on code after a 1K prompt and 524K context per request. Figures are the author's own measurements with greedy decoding and no prefix cache.