Mojo package
amd
AMD CDNA GPU attention kernels for GFX942/GFX950 architectures.
Includes MHA prefill/decode, MLA, matrix-multiply-accumulate primitives, shared-memory buffers, and softmax helpers. RDNA kernels are in amd_rdna/.
Modules
-
attention: -
buffers: -
kv_buffer: KV cache buffer for structured MHA kernels (TileTensor hot path). -
mha_gfx942: -
mha_gfx950: -
mha_structured: MHA prefill kernel for gfx950 with structured scheduling. -
mla: -
mma: -
softmax: -
utils:
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!