Module layers

Structs§

AvgPool2d
CausalMaskConfig: Configuration for CausalMasker::make_causal_mask.
CausalMasker
Conv3dConfig
Conv3dNoBias
DeepSeekV2RopeConfig
DeepSeekV2RotaryEmbedding
DiaRotaryEmbedding
F32RmsNorm
FloatInfo
Gemma3RopeScalingConfig
Gemma3RotaryEmbedding
Gemma3nRopeScalingConfig
Gemma3nRotaryEmbedding
GemmaRmsNorm: Gemma-style RmsNorm that adds +1.0 to the weight during initialization.
GptOssRotaryEmbedding: GPT-OSS style rotary embedding with YARN scaling support. Uses chunked/GPT-NeoX style rotation and applies attention scaling.
Llama3RopeConfig
Llama3RotaryEmbedding: RoPE for Llama3
MatMul: Device/configurable intelligent matrix multiplication
Mlp
Phi4MMRopeScalingConfig
Phi4MMRotaryEmbedding
PhiRopeConfig
PhiRotaryEmbedding: RoPE supporting LongRope
QLinear
QRmsNorm
Qwen2VLRotaryEmbedding
Qwen2_5VLRotaryEmbedding
Qwen3VLRotaryEmbedding: Qwen3 VL uses interleaved MRoPE (not chunked like Qwen2 VL). Frequencies are arranged as THW THW THW… TTTT pattern. See apply_interleaved_mrope in modeling_qwen3_vl.py.
ReflectionPad2d: Applies 2D reflection padding to a tensor of shape (N, C, H, W).
RmsNorm
RotaryEmbedding
ScaledEmbedding
Sdpa
SmolLm3RopeConfig
SmolLm3RotaryEmbedding: RoPE for SmolLm3