For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo struct
EpiLoadPipelineStorage
struct EpiLoadPipelineStorage[num_stages: Int]
Storage for epilogue load pipeline (source C loading).
For EpilogueLoad warp β Epilogue warps synchronization. The epilogue load warp loads source tensor C into SMEM, and the epilogue warps consume it for residual operations.
Producer: EpilogueLoad warp (1 warp, 32 threads) Consumer: Epilogue warps (4 warps, 128 threads)
Parametersβ
- βnum_stages (
Int): Number of epilogue load pipeline stages (typically 2).
Fieldsβ
- βbarriers (
BarrierPair[num_stages]):
Implemented traitsβ
AnyType,
ImplicitlyDestructible
comptime membersβ
BarrierArrayβ
comptime BarrierArray = SMemArray[SharedMemBarrier, (num_stages * 2)]
Methodsβ
create_pipelineβ
create_pipeline(ref[AddressSpace._value] self) -> ProducerConsumerPipeline[num_stages]
Create runtime pipeline from this storage.
Returns:
ProducerConsumerPipeline[num_stages]
barrier_ptrβ
barrier_ptr(ref[AddressSpace._value] self) -> MbarPtr
Escape hatch: Get raw barrier pointer.
Returns:
MbarPtr
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!