Offload Final Linear Layer#48

Open

kurtis-b-1 wants to merge 21 commits intoamd:develfrom

kurtis-b-1:offload-final-linear-layer-2

Commits on Dec 5, 2025

Use output directory instead of /tmp directory for building
kurtis-b-1
committed
Can run GEMM with a wide B matrix
kurtis-b-1
committed
Moved creating buffers for buffer lists B/C inside golden reference so that the major order logic is there instead
kurtis-b-1
committed
Simplifying some logic in the operator
kurtis-b-1
committed
Further simplification of the gemm operator
kurtis-b-1
committed
Formatting
kurtis-b-1
committed
Adjust comments for C tile streaming through shim DMA based on the separate_c_tiles parameter
kurtis-b-1
committed
Remove using the separated C tile runtime streams in test
kurtis-b-1
committed
Can offload the last linear layer, but TTFT goes way up--reason seems to be that the actual forward operation for last linear layer is ~4.5s, likely due to reading the output buffers and converting…
kurtis-b-1
committed

Commits on Dec 6, 2025

Commits on Dec 9, 2025

Commits on Dec 11, 2025

Fix comments in numpy/torch conversion utils
kurtis-b-1
committed