5 comments

  • ilc1 hour ago
    Watch the video carefully. DFlash2&#x27;s tool call fails on python syntax.<p>Usually models in this class nail things like that 1 shot, which the other side did.<p>I don&#x27;t know the cause. It may be nothing. But I&#x27;d like to see the model doing something where its path is a bit more constrained, to help out rule out such oddities.
    • zackangelo18 minutes ago
      DFlash is lossless so this would be a bug in the implementation if it is indeed a regression against the target model.
      • liuliu3 minutes ago
        Only if you do greedy sampling. With probabilisitic sampling (categorical sampling), you will end up with different trajectory just “mathematically equivalent”.
  • hypfer3 hours ago
    Amazing tech<p>&gt; An agent writes in an afternoon what a chatbot writes in a month<p>But can you just.. not.<p>Your tech is so good, it speaks for itself. Don&#x27;t ruin that.
  • adefa3 hours ago
    I&#x27;m getting around 27 tokens per second decode using vLLM + Qwen 3.8 27b nvfp4 + DFlash 2 on the DGX Spark.
  • sarjann3 hours ago
    Great news, has made low memory bandwidth model usage so much nicer.
  • verdverm4 hours ago
    vllm PR for DFlash2: <a href="https:&#x2F;&#x2F;github.com&#x2F;vllm-project&#x2F;vllm&#x2F;pull&#x2F;52816" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;vllm-project&#x2F;vllm&#x2F;pull&#x2F;52816</a>
    • cogman103 hours ago
      llama.cpp PR <a href="https:&#x2F;&#x2F;github.com&#x2F;ggml-org&#x2F;llama.cpp&#x2F;pull&#x2F;27342" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;ggml-org&#x2F;llama.cpp&#x2F;pull&#x2F;27342</a>