Kernel Tuning

PMetal automatically selects kernel parameters based on your device tier and GPU family.

Matrix Tile Size (GEMM, LoRA forward)

Tier	Apple7–9	Apple10 (M5+, NAX)
Base	32×32×32	64×32×32
Pro	64×32×32	64×64×32
Max	64×64×32	128×64×32
Ultra	64×64×32	128×64×32

FlashAttention Block Sizes

Block size selection per head dimension:

Head Dim	Base	Pro	Max	Ultra
64	64×32	64×32	64×64	64×64
80	64×32	64×32	64×64	64×64
96	64×32	64×32	64×64	64×64
128	32×32	32×32	64×64	64×64
256	32×16	32×16	32×32	32×32

Fused Kernel Threadgroup Sizes

Fused RMSNorm + LoRA

Tier	Threadgroup Size
Base	128
Pro	128
Max	256
Ultra	256

Fused SwiGLU

Tier	Threadgroup Size
Base	256
Pro	256
Max	512
Ultra	512

Batch Size Multiplier

Tier	Multiplier
Base	1×
Pro	2×
Max	4×
Ultra	8×

Metal GPU Optimizations

Kernel	Description
FlashAttention	O(n) memory attention with fused softmax, tier-aware block sizes
Fused GDN	Gated Delta Network recurrence kernel — single-pass state update
Fused LoRA	Combined forward pass for adapter layers (~2× speedup)
Fused Cross-Entropy	Unsloth-style chunked loss computation
Fused Linear Cross-Entropy	Skips logits materialization entirely
Fused RoPE	Rotary position embeddings in-kernel
Fused SwiGLU	Fused gate + activation with tier-tuned threadgroups
Fused RMSNorm + LoRA	Combined normalization and adapter projection
Fused Sampler	JIT-compiled token sampling
Fused MLP	Combined gate/up/down projections
Async Scheduler	Double/triple-buffered GPU command scheduling

See Also

• Apple Silicon Support — Hardware detection matrix
• pmetal bench — Benchmark on your hardware