graph: re-enable sdpa training bwd#4814

Open

ElaineBao wants to merge 23 commits intomainfrom

yixin/sdpa_ukernel_train

Commits on Mar 19, 2026

common: sdpa: adds backwards sdpa primitives

syurkevi
authored and
TaoLv
committed
common: sdpa: prepares internal primitive for bwd testing

syurkevi
authored and
TaoLv
committed
gtests: internals: adds sdpa training tests

syurkevi
authored and
TaoLv
committed
xe: sdpa: prepares config structs for bwd pass

syurkevi
authored and
TaoLv
committed
gpu: gemm: jit: allow LLR gemms

syurkevi
authored and
TaoLv
committed
gpu: gemm: jit: enables ukernel with multiple kernel sources

syurkevi
authored and
TaoLv
committed
xe: sdpa: splits fwd/bwd gpu training primitives

syurkevi
authored and
TaoLv
committed
xe: sdpa: split gemm setup for backwards pass

syurkevi
authored and
TaoLv
committed
xe: sdpa: calculates forward logusmexp to ws

syurkevi
authored and
TaoLv
committed
xe: sdpa: updates tile_ops packed load/stores

syurkevi
authored and
TaoLv
committed
xe: sdpa: adds bwd kernel implementation

syurkevi
authored and
TaoLv
committed
xe: sdpa: adds create_sdpa_pd for backwards pass

syurkevi
authored and
TaoLv
committed
xe: sdpa: rename fwd/bwd primitives

syurkevi
authored and
TaoLv
committed
gtests: internals: separate sdpa from internals

syurkevi
authored and
TaoLv
committed
common: sdpa: move prop_kind param before attrs

syurkevi
authored and
TaoLv
committed
xe: sdpa: enable transpose_k for training

syurkevi
authored and
TaoLv
committed
common: sdpa: refactors pd accessors, misc cleanup

syurkevi
authored and
TaoLv
committed
graph: backend: dnnl: move softmax decomposition to a transform pass

ElaineBao
authored and
TaoLv
committed
graph: backend: dnnl: enable sdpa microkernel for training fwd

ElaineBao
authored and
TaoLv
committed
graph: backend: dnnl: enable sdpa microkernel for training bwd

ElaineBao
authored and
TaoLv
committed
tests, examples: update sdpa training bwd cases

ElaineBao
authored and
TaoLv
committed
tests: benchdnn: graph: remove legacy sdpa bwd cases

ElaineBao
authored and
TaoLv
committed
graph: backend: dnnl: fix code format

ElaineBao
authored and
TaoLv
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

graph: re-enable sdpa training bwd#4814

graph: re-enable sdpa training bwd#4814
ElaineBao wants to merge 23 commits intomainfrom
yixin/sdpa_ukernel_train

Commits on Mar 19, 2026

common: sdpa: adds backwards sdpa primitives

common: sdpa: prepares internal primitive for bwd testing

gtests: internals: adds sdpa training tests

xe: sdpa: prepares config structs for bwd pass

gpu: gemm: jit: allow LLR gemms

gpu: gemm: jit: enables ukernel with multiple kernel sources

xe: sdpa: splits fwd/bwd gpu training primitives

xe: sdpa: split gemm setup for backwards pass

xe: sdpa: calculates forward logusmexp to ws

xe: sdpa: updates tile_ops packed load/stores

xe: sdpa: adds bwd kernel implementation

xe: sdpa: adds create_sdpa_pd for backwards pass

xe: sdpa: rename fwd/bwd primitives

gtests: internals: separate sdpa from internals

common: sdpa: move prop_kind param before attrs

xe: sdpa: enable transpose_k for training

common: sdpa: refactors pd accessors, misc cleanup

graph: backend: dnnl: move softmax decomposition to a transform pass

graph: backend: dnnl: enable sdpa microkernel for training fwd

graph: backend: dnnl: enable sdpa microkernel for training bwd

tests, examples: update sdpa training bwd cases

tests: benchdnn: graph: remove legacy sdpa bwd cases

graph: backend: dnnl: fix code format