[DRAFT] Add Adam Support for Generic and Siracusa Platforms - #169

Closed

Storiann wants to merge 22 commits into

pulp-platform:develfrom

Commits on May 2, 2026

Merge training-platform-core: minimal SimpleMLP training on Siracusa
runwangdl
committed
22db7f2
ci: disable auto-trigger on non-training workflows (#1)
runwangdl
authored
e4bfb22
ci: add siracusa-train-kernel CI job + train_kernel pytest marker (#2)
runwangdl
authored
24147d5
feat(kernel): add AveragePool (forward + Grad) on Siracusa/PULPOpen (#3)
runwangdl
authored
1e64ae3
feat(kernel): add GlobalAveragePool family + AveragePoolGrad fixture (#5)
runwangdl
authored
ea24efb
feat(kernel): add ConvGrad family (X/W/B + DW + PW) on Siracusa/PULPOpen (#6)
runwangdl
authored
0520c40
feat(kernel): add MSELoss (forward + Grad) on Siracusa/PULPOpen (#7)
runwangdl
authored
80292b2
feat(kernel): add MaxPoolGrad + ReluGrad + LayerNormGrad backward kernels (#8)
runwangdl
authored
4cfd6d1
feat(kernel): add BatchNorm family (Internal + Grad + 4 split helpers) on Siracusa/PULPOpen (#9)
runwangdl
authored
1d2674a

Commits on May 3, 2026

feat(e2e): add 4 MLPerf-Tiny end-to-end training models (#10)
runwangdl
authored
fedc4df

Commits on May 4, 2026

Commits on May 6, 2026

Commits on May 8, 2026

fix(tiling): MiniMalloc cost variable + SGD rank-2 DMA constraint (#18)
runwangdl
authored
950b1fd

Commits on May 10, 2026

feat(cct): update CCT train/optimizer fixtures to img_size=32, embedding_dim=128, n_conv_layers=2 (#23)

authored
42f87d2

Commits on May 12, 2026

fix(codegen): eliminate dead optimizer L3 arena allocation (#25)

authored
d88b620

Commits on May 14, 2026

perf(ConvGrad): GradWStrategy framework + ConvGradX/W speedups (#28)
runwangdl
authored
022f045

Commits on Jul 17, 2026

Commits on Jul 21, 2026

fix(optimizer): read graph input dtype from ONNX instead of forcing float32
Storiann
committed
a0e1e53