Module thunk 

Thunks — pre-compiled kernel dispatch with zero per-call overhead.

At compile time, the graph is lowered into a flat Vec<Thunk> where each thunk holds pre-computed arena offsets, dimensions, and kernel type. At runtime, the executor just iterates thunks and calls kernels directly. No match dispatch, no HashMap lookup, no dimension computation.

Structs

ThunkSchedule

Compiled thunk schedule — the runtime hot path. Nop thunks are filtered out at compile time for zero iteration overhead.

Enums

Thunk

A pre-compiled kernel call with all args resolved to arena offsets.

Functions

compile_thunks

Compile graph into thunk schedule.

dequant_matmul_nvfp4

execute_axial_rope2d_f32^⚠

Host axial 2-D RoPE for Metal (and other) fallbacks on unified memory.

execute_compiled

Execute a thunk schedule on a raw arena buffer. Fastest executor: call pre-compiled closures sequentially. Zero match dispatch — each closure is a direct kernel call.

execute_conv_transpose2d_nchw_f32^⚠

Host fallback for NCHW ConvTranspose2d.

execute_cumsum_backward_f32^⚠

execute_dequant_grouped_matmul_gguf_f32^⚠

Host-fallback entry for GGUF Op::DequantGroupedMatMul (MoE expert stack).

execute_dequant_matmul_fp8_f32^⚠

Host-fallback entry for FP8 Op::DequantMatMul (Metal unified memory).

execute_dequant_matmul_gguf_f32^⚠

Host-fallback entry for GGUF Op::DequantMatMul (Metal unified memory).

execute_dequant_matmul_int4_f32^⚠

Host-fallback entry for Int4 Op::DequantMatMul (Metal unified memory).

execute_dequant_matmul_nvfp4_f32^⚠

Host-fallback entry for NVFP4 Op::DequantMatMul (Metal unified memory).

execute_fft1d^⚠

Dtype-dispatching host entry for Op::Fft (shared by GPU host fallbacks).

execute_fft1d_c64^⚠

C64 interleaved layout: each complex element is [re: f32, im: f32].

execute_fft1d_f32^⚠

f32 mirror of execute_fft1d_f64. Same public-host-fallback role.

execute_fft1d_f64^⚠

Execute a batched 1D FFT in the f64 2N-real-block layout. Each “row” is 2N f64 elements: first N real, then N imag. The outer rows are independent and processed sequentially.

execute_gated_delta_net_f16^⚠

Host-fallback entry for f16 Op::GatedDeltaNet tensors on Metal.

execute_gated_delta_net_f32^⚠

f32 counterpart of execute_fft1d_f64. Same 2N-real-block layout (first N real, second N imag per row), same unnormalized convention; only the element width differs. Twiddle factors are computed in f64 and cast to f32 to keep large-N error closer to the f64 path (the savings from f32 are in memory bandwidth, not in twiddle precision). Host-fallback entry for Op::GatedDeltaNet (Metal / unified memory). When state == 0, uses a zero-initialized scratch state per batch item.

execute_gather_backward_f32^⚠

execute_group_norm_nchw_f32^⚠

Host fallback for NCHW group norm (Metal unified-memory arena).

execute_layer_norm2d_nchw_f32^⚠

Host fallback for NCHW LayerNorm2d (SAM / candle semantics).

execute_resize_nearest_2x_f32^⚠

Host fallback for nearest 2× upsample on NCHW.

execute_rms_norm_backward_beta_f32^⚠

execute_rms_norm_backward_gamma_f32^⚠

execute_rms_norm_backward_input_f32^⚠

Host-fallback: Op::RmsNormBackwardInput (GPU unified-memory / D2H arenas).

execute_rope_backward_f32^⚠

execute_thunks

execute_thunks_active

Active-extent execution stub. The runtime calls this when it has an active-extent hint set. CPU doesn’t implement per-thunk active-extent scaling yet — return false so the caller falls back to the full execute_thunks path.