Training Platform Core: minimal MLP end-to-end training on Siracusa#182

Open

runwangdl wants to merge 30 commits into

pulp-platform:develfrom

runwangdl:training-platform-core

Commits on Apr 10, 2026

training-platform core: minimal MLP end-to-end training on Siracusa
runwangdl
committed
training-platform core: apply pre-commit formatting (yapf/isort/clang-format)
runwangdl
committed
training-platform core: canonicalise SoftmaxCrossEntropyLoss to 2 outputs
runwangdl
committed
training-platform core: collapse InPlaceAccumulatorV2 template to single binding
runwangdl
committed
training-platform core: generalise SGD alias comment to L2 or L3
runwangdl
committed
training-platform core: trim InPlaceAccumulatorV2 tile constraint comments
runwangdl
committed
training-platform core: drop leftover egress-target comment
runwangdl
committed
training-platform core: drop stray ReluGradTileConstraint
runwangdl
committed
training-platform core: delete stray SoftmaxCrossEntropyLossDualOutputTileConstraint
runwangdl
committed
training-platform core: simplify MiniMalloc alias-skip block
runwangdl
committed
training-platform core: restore per-layer { } block in generateInferenceCode
runwangdl
committed
training-platform core: restore upstream SoftmaxCrossEntropy kernel test fixtures
runwangdl
committed
training-platform core: drop legacy 1-output Softmax/CrossEntropy kernel test
runwangdl
committed
training-platform core: propagate loss verification result + drop dead code
runwangdl
committed
training-platform core: label Step B in run_optimizer_step
runwangdl
committed
training-platform core: drop _augment_path PATH manipulation
runwangdl
committed
training-platform core: extract training codegen helpers to trainingUtils.py
runwangdl
committed
training-platform core: collapse duplicate tiled/non-tiled training branches in generate_network
runwangdl
committed
training-platform core: extract training codegen argparse builders to trainingUtils.py
runwangdl
committed
training-platform core: lift execution.py training subprocess helpers to trainingUtils
runwangdl
committed
training-platform core: lift _resolve_optimizer_dir to trainingUtils
runwangdl
committed
training-platform core: decouple execution.py from training pipeline
runwangdl
committed
training-platform core: extract training codegen helpers to codeGenerateTraining.py
runwangdl
committed
training-platform core: drop redundant top-level deeployTrainingRunner.py
runwangdl
committed
training-platform core: drop non-training helper wrappers from trainingUtils
runwangdl
committed
training-platform core: drop unused loop var and populate zero-sized input defaults
runwangdl
committed
training-platform core: apply pre-commit yapf + autoflake autofixes
runwangdl
committed
training-platform core: drop CCT2_FT2 from siracusa tiled L3 CI list
runwangdl
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Training Platform Core: minimal MLP end-to-end training on Siracusa#182

Training Platform Core: minimal MLP end-to-end training on Siracusa#182
runwangdl wants to merge 30 commits into
pulp-platform:develfrom
runwangdl:training-platform-core

Commits on Apr 10, 2026

training-platform core: minimal MLP end-to-end training on Siracusa

training-platform core: apply pre-commit formatting (yapf/isort/clang-format)

training-platform core: canonicalise SoftmaxCrossEntropyLoss to 2 outputs

training-platform core: collapse InPlaceAccumulatorV2 template to single binding

training-platform core: generalise SGD alias comment to L2 or L3

training-platform core: trim InPlaceAccumulatorV2 tile constraint comments

training-platform core: drop leftover egress-target comment

training-platform core: drop stray ReluGradTileConstraint

training-platform core: delete stray SoftmaxCrossEntropyLossDualOutputTileConstraint

training-platform core: simplify MiniMalloc alias-skip block

training-platform core: restore per-layer { } block in generateInferenceCode

training-platform core: restore upstream SoftmaxCrossEntropy kernel test fixtures

training-platform core: drop legacy 1-output Softmax/CrossEntropy kernel test

training-platform core: propagate loss verification result + drop dead code

training-platform core: label Step B in run_optimizer_step

training-platform core: drop _augment_path PATH manipulation

training-platform core: extract training codegen helpers to trainingUtils.py

training-platform core: collapse duplicate tiled/non-tiled training branches in generate_network

training-platform core: extract training codegen argparse builders to trainingUtils.py

training-platform core: lift execution.py training subprocess helpers to trainingUtils

training-platform core: lift _resolve_optimizer_dir to trainingUtils

training-platform core: decouple execution.py from training pipeline

training-platform core: extract training codegen helpers to codeGenerateTraining.py

training-platform core: drop redundant top-level deeployTrainingRunner.py

training-platform core: drop non-training helper wrappers from trainingUtils

training-platform core: drop unused loop var and populate zero-sized input defaults

training-platform core: apply pre-commit yapf + autoflake autofixes

training-platform core: drop CCT2_FT2 from siracusa tiled L3 CI list

Commits on Apr 13, 2026

Add simpleMLP training to CI

fix linting