[GDN] Eliminate GPU->CPU sync in prepare_chunk_indices during prefill#38361

Merged

vadiklyutiy merged 14 commits intovllm-project:mainfrom

arpera:artem/remove-extra-d2h-copy

Apr 3, 2026

Commits on Mar 31, 2026

[GDN] Eliminate GPU->CPU sync in prepare_chunk_indices during prefill
arpera
committed
Fix gemini-code issues: extract _insert helper in tensor_cache, add TODO for chunk_size constant
arpera
committed
Fix mypy: add type: ignore for dynamic register attribute
arpera
committed
Extract hardcoded chunk_size=64 into FLA_CHUNK_SIZE constant
arpera
committed
Fix: skip chunk_indices pre-registration on pure decode steps
arpera
committed
[GDN] Pre-compute chunk_indices/chunk_offsets in metadata builder
arpera
committed
Remove dead register() code and duplicate prefill block
arpera
committed
Fix Claude review comments: backend guard, kda chunk_size, BT simplification
arpera
committed
Remove use_flashinfer backend guard for chunk_indices pre-computation
arpera
committed
Merge branch 'main' into artem/remove-extra-d2h-copy
vadiklyutiy
authored

Commits on Apr 2, 2026

Commits on Apr 3, 2026