Tools for AI developers

7 tools

  • Sizes the cache for models using grouped query attention, MLA, hybrid linear attention, or sparse attention.

    Open
  • Estimates how long REAP expert pruning takes for a mixture of experts model, whether one expert block fits your GPU, and how much smaller the pruned model gets.

    Open
  • Estimates the cost of training a DSpark speculative-decoding drafter against a custom target model: the size of the target hidden state cache, the length of the run, and whether it fits in VRAM.

    Open
  • Sizes the KV cache for a model and then finds the smallest GPU configuration that serves it in the weight format the checkpoint publishes, from the resident weights, the cache, an activation buffer, and a framework reserve.

    Open
  • Finds the session size at which summarising it starts to pay off, and the widest summary cap that still beats carrying the session.

    Open
  • Turns a token count into bytes, KB, MB, GB, TB, and PB using the model’s own tokenizer, read from HuggingFace or ModelScope, and sizes the same tokens as text, JSONL, Parquet, or Arrow.

    Open
  • Ranks every model on the clef-evals Decision Model Leaderboard by the smallest GPU in the hardware list on this site that holds it, at a workload you set.

    Open