Expand description
WGSL compute shaders for GPU operations
Modules§
- backward
- WGSL backward (gradient) shaders for training
Constants§
- CAUSAL_
ATTENTION_ SHADER - Causal multi-head attention (WGSL) — scaled dot-product with GQA
- COLUMN_
GATHER_ SHADER - Column gather shader — extracts columns from a wide matrix into a chunk.
- COLUMN_
SCATTER_ SHADER - Column scatter shader — copies chunk columns into a wider row-major matrix.
- LORA_
ADDMM_ SHADER - Fused LoRA addmm: output += (input @ A) @ B * scale
- MATMUL_
SHADER - Matrix multiplication compute shader (WGSL) — tiled shared memory
- TILED_
GEMM_ SHADER - CUTLASS-style tiled GEMM compute shader (WGSL) — 64×64 output tiles
- TRANSPOSE_
SHADER - Scaled transpose: B[j,i] = scale * A[i,j] Contract: wgsl-transpose-v1