Expand description
Opt-in FlashInfer GDN prefill via dlopen(libatlasgdn.so) — behind ATLAS_GDN_FLASHINFER=1.
Bridges Atlas’s native packed-QKV + interleaved gate/beta buffers to the AOT-exported
FlashInfer chunked gated-delta-rule scan (tensor-core, ~11× the scalar FLA chunk_delta_h
at the Holo shape — see 3rdparty_patches/gdn_aot/STATUS.md). The C-ABI shim
(atlas_gdn_prefill_packed) takes Atlas’s exact native pointers: it deinterleaves
gate/beta in-shim and reads q/k/v straight out of the packed buffer via conv_dim
strides (no copy). Atlas’s gate is already linear α (the kernel does the logf),
so there is NO gate-space conversion.
dlopen (not link-time) keeps this fully opt-in: the binary builds and runs without the
library; it is only loaded when the flag is set. ATLAS_GDN_LIB overrides the path.
Functions§
- available
- True when
ATLAS_GDN_FLASHINFER=1AND the library + symbols loaded successfully. - flashinfer_
gdn_ prefill - Run one prefill GDN scan through the FlashInfer kernel on Atlas’s native buffers.