[cuda backend][gemma4_31b] TQ4 SDPA: no-spill prefill kernel + analytic causal #20512

Merged

Gasoonjia merged 7 commits into

mainfrom

gemma4_31b-tq4-prefill-decode-tuned

Jun 25, 2026

Commits on Jun 24, 2026

[gemma4_31b][cuda] Export Gemma4-31B @128k under 32 GB
Gasoonjia
committed

Commits on Jun 25, 2026

Fix TurboQuant KV zeroed by low-mem export (993cff58ae): _is_kv_buffer only frees genuinely all-zero kv_cache.* buffers (count_nonzero==0); preserves TQ4 centroids/boundaries/rotation/rotation_T
Gasoonjia
committed
[executorch][cuda] gemma4_31b: fuse gate/up MLP projections (default-on)
Gasoonjia
committed
[gemma4_31b][cuda] length-aware bf16 global attention + head_dim-agnostic prefill autotune
Gasoonjia
committed
gemma4_31b TQ4 SDPA: no-spill prefill + analytic causal + autotune retune
Gasoonjia
committed
gemma4_31b TQ4 SDPA: add 5090-feasible autotune configs + comment updates
Gasoonjia
committed
Merge branch 'main' into gemma4_31b-tq4-prefill-decode-tuned
Gasoonjia
authored