Copy page View as Markdown

Guide

• Loading: how a checkpoint becomes a StandardizedTransformer.
• Names: the module names every family answers to.
• Residual stream: the stream leaving a block and what each sublayer adds to it.
• Attention: the tensors inside the attention module, from the queries to the head outputs.
• The whole model: the values on the root, from the input ids to the logits.
• Layouts: the axis layout of every value.
• Availability: which values a checkpoint has.
• Mixture of experts: the router and the experts on every MoE family.
• Gated DeltaNet: the linear-attention mixer on Qwen3-Next, Qwen3.5, OLMo-Hybrid and Kimi-Linear.
• Mamba: the selective-scan mixer on Mamba, Falcon-Mamba and Jamba.
• Mamba-2: the state-space mixer on Mamba-2, Nemotron-H, Bamba and Falcon-H1.
• Vision-language: the image side of an image-text-to-text checkpoint.
• Methods: steer, skip_layers, project_on_vocab and get_topk_closest_tokens.
• Generation: the values under model.generate.
• Helpers: activations and next-token distributions over many prompts.
• Remote: a StandardizedTransformer trace on NDIF.