Match Qwen head-chunked attention execution

This commit is contained in:
Georg Bauer committed 2026-10-04 16:45:19 +02:00
1 parent 7b894fbfee
commit d9026aa7bf
70 files changed
+4979 -10

No files matched your search

+4
View File
@@ -25,6 +25,10 @@ fn falsey(value: Option<&str>) -> bool {
})
}
pub(super) fn sdpa_head_chunk_enabled(get: impl Fn(&'static str) -> Option<String>) -> bool {
!falsey(get("MTPLX_QWEN4_VERIFY_SDPA_HEAD_CHUNK").as_deref())
}
pub(super) fn sorted_gather_enabled(get: impl Fn(&'static str) -> Option<String>) -> bool {
!falsey(get("MTPLX_MOE_SORTED_GATHER_KERNEL").as_deref())
&& !truthy(get("MTPLX_FORCE_GPU_FAMILY_FALLBACK").as_deref())