Metal GPU + CoreML/ANE support for transformer nets (SiLU, GQA, learnable RoPE)#1205

Open

ChinChangYang wants to merge 13 commits into

lightvector:masterfrom

ChinChangYang:feature/metal-transformer-silu-b10c384h6

Commits on Jun 1, 2026

Add Metal GPU + CoreML/ANE transformer support for b10c384h6nbttflrs (v15)

ChinChangYang
and
claude
committed
Merge branch 'master' into feature/metal-transformer-silu-b10c384h6
ChinChangYang
committed
Fix Metal GPU crash on GQA transformer attention (NDArray INT_MAX)

ChinChangYang
and
claude
committed
Fix dropped SiLU activation in CoreML value/policy/meta heads

ChinChangYang
and
claude
committed
Implement GQA support in CoreML/ANE MIL attention builder

ChinChangYang
and
claude
committed
Fix CoreML/ANE FP16 transformer accuracy via precision tiers

ChinChangYang
and
claude
committed
Refactor: dedupe CoreML global-pooling FP32 wrap

ChinChangYang
and
claude
committed

Commits on Jun 22, 2026

Metal: fix ~1GB/game memory leak in MPSGraph apply()

ChinChangYang
and
claude
committed