Module gqa 

Grouped Query Attention kernel.

Matches gqa-kernel-v1.yaml. KV head broadcasting: kv_head = query_head / (num_heads / num_kv_heads)

Each function provides one of three backends:

• fn gqa_scalar(...) – Pure Rust scalar reference (ground truth)
• unsafe fn gqa_avx2(...) – AVX2 SIMD implementation
• fn gqa_ptx() -> &'static str – PTX assembly source string

Functions

gqa_avx2^⚠

AVX2 Grouped Query Attention – delegates to scalar.

gqa_ptx

PTX assembly for Grouped Query Attention.

gqa_scalar

Grouped Query Attention (scalar reference).