DeepSeek V4.1 Flash 1017 on 2x DGX Spark: 57 tok/s text, 92 tok/s code with TensorFold (2.1-2.3x the vLLM kit)
英文标题：DeepSeek V4.1 Flash 1017 on 2x DGX Spark: 57 tok/s text, 92 tok/s code with TensorFold (2.1-2.3x the vLLM kit)
**TL;DR:** DeepSeek-V4.1-Flash-1017 (EXL3 2.9bpw, 1M context) on 2x DGX Spark, TP=2 over the CX7 cable. TensorFold is 2.1-2.3x faster than the vLLM kit on the same pair, same weights, same prompts, same day. vLLM kit TensorFold Decode, text 27.0 tok/s **56.8 tok/s** Decode, code 39.9 tok/s *
打开原文相关 2 · 机会 0
中文
**TL;DR:** DeepSeek-V4.1-Flash-1017 (EXL3 2.9bpw, 1M context) on 2x DGX Spark, TP=2 over the CX7 cable. TensorFold is 2.1-2.3x faster than the vLLM kit on the same pair, same weights, same prompts, same day. vLLM kit TensorFold Decode, text 27.0 tok/s **56.8 tok/s** Decode, code 39.9 tok/s *
中文全文翻译正在补齐，可先查看下方英文原文。
English / 原文
**TL;DR:** DeepSeek-V4.1-Flash-1017 (EXL3 2.9bpw, 1M context) on 2x DGX Spark, TP=2 over the CX7 cable. TensorFold is 2.1-2.3x faster than the vLLM kit on the same pair, same weights, same prompts, same day. vLLM kit TensorFold Decode, text 27.0 tok/s **56.8 tok/s** Decode, code 39.9 tok/s *