Triton bypassLDS changes#695

Merged

jataylo merged 32 commits intorelease/internal/3.1.xfrom

plognjen/release/internal/3.1.x

Jan 9, 2025

Commits on Jan 5, 2025

Add blocked to dot shortcut
oplavsic
committed
pack tensors in vectors instead of structures

binarman
authored and
oplavsic
committed
fix

binarman
authored and
oplavsic
committed
add moe bypass option

binarman
authored and
oplavsic
committed
initial commit

Ognjen Plavsic
authored and
oplavsic
committed
fix

Ognjen Plavsic
authored and
oplavsic
committed
fix

Ognjen Plavsic
authored and
oplavsic
committed
add missing configurations and add more checks in passes

binarman
authored and
oplavsic
committed
adjust global load layout for vllm swizzling format

binarman
authored and
oplavsic
committed
Remove debug print

vgokhale
authored and
oplavsic
committed
make load width dependable on data type

binarman
authored and
oplavsic
committed
fix int 8 logic

binarman
authored and
oplavsic
committed
generalize load analysis: return last load in dependant laod chain instead of 2

binarman
authored and
oplavsic
committed
Add message for assert failure

zhanglx13
authored and
oplavsic
committed
add k=512/1024 cases

binarman
authored and
oplavsic
committed
[BACKEND] Add memory space to memdesc type. (#4027)

ThomasRaoux
authored and
oplavsic
committed
[BACKEND] Fix memory side effects of `tt.dot` (#4033)

Jokeren
authored and
oplavsic
committed
remove streamPipelinev2
oplavsic
committed
[TEST] NFC: Drop irrelevant NVIDIA specific attributes (#4384)

antiagainst
authored and
oplavsic
committed
[Pipeliner] NFC: Expose Pipeliner infrastructure for use by other target backends (#4155)

sjw36
authored and
oplavsic
committed
[BACKEND] Fix regression in pipeliner pre-checks. (#4196)

ThomasRaoux
authored and
oplavsic
committed
[Backend][AMD] Introduce stream pipeliner v2 (#4148)

authored and
oplavsic
committed
[AMD] Prefetch loads and independent local_stores (#4429)

sjw36
authored and
oplavsic
committed
[Pipeliner] Implement dynamic loop peeling

SJW
authored and
oplavsic
committed
* disabled for num_stages > 2

SJW
authored and
oplavsic
committed
* guard each stage of ramp-down in epilogue

SJW
authored and
oplavsic
committed
* pipeline reg buffers

SJW
authored and
oplavsic
committed
[AMD] Fixed bug with tritonamdgpu-reorder-instructions

SJW
authored and
oplavsic
committed
* only move ops early

SJW
authored and
oplavsic
committed
Fix in streamPipelinerV2
oplavsic
committed

Commits on Jan 6, 2025

Fix lit tests
oplavsic
committed
[Backend][AMD] Add temporary environment variable for pipeliner v2 (#4430)

authored and
oplavsic
committed