* CI: add manual aiter prebuilt upload flow

Add a workflow_dispatch GHA to build/upload aiter prebuilts using a chosen image and GPU arch list, reusing the shell script.

Make ci/aiter_upload.sh handle build + package + upload with optional env-based upload, respecting GPU_ARCHS input and defaulting to gfx942;gfx950.

Strip upload/packaging logic out of the CMake helper so normal builds only download/use prebuilts.

* Addressed reviews

Move aiter upload helper to .github/scripts, add copyright header, and use a temp gitconfig for safe.directory/commit lookup

set CK_TILE_FLOAT_TO_BFLOAT16_DEFAULT, added functionality to verify remote SHA after upload

Trim workflow diagnostics/cleanup, use --rm container, pass GPU_ARCHS input directly

* Adressed comments

* Copyright update

* Run core with all GPUs

* Change n_parallel_jobs number to use all GPUs

* Update MNIST dataset loading to use 'ylecun/mnist' instead of 'mnist' for both training and testing datasets.

* Update dataset loading in encoder tests to use 'nyu-mll/glue' for both training and validation datasets.

---------

Co-authored-by: sugovind <sugovind@amd.com>

* Fixed setting safe.directory for aiter upload

* Added EOF to docker exec

* Fixed dubious ownership

* Added container cleanup

…kages (ROCm#433)

* Clean up testing of outdated behavior

* Minimize cumulative diff from upstream

* [CI] Skipped test_gpt_full_activation_recompute tests for gfx950

* [CI] Skipped unsupported test_basic_linear_quantized tests on gfx950

* [CI] Fixed test_numerics, test_norms, test_fused_optimizer failures for gfx950 ci enablement

* [CI] Disabled gfx950 support until FP8 GEMM layout coverage is verified with hipblaslt

* [CI] [gfx950] Disable cudaGraph for gemmm and grouped-gemm

* Addressed reviews

* [CI] Add MI355 nodes to github actions workflow

* [CI] Update docker image

* [CI] add MI355 runner matrix and keep matrix legs independent

* Skip unstable Gemm tests on gfx950

* Addressed reviews

* Guard gfx950 TN skip by ROCm version and adjust MXFP8 Dq test size

* Removed ROCM7.2 guards

* Reverted ROCM7.2 guards

* Corrected Normalization scale_inv padding removal

* Updated cast behavior, added safeguards around MXFP8 GEMM

* Add partial progress

* Improved guards on LayerNormMLP tests

* Remove swizzle in JAX GEMM primitive

* Added unique factors for sharding scales, added xfail to test

* Removed old code

* Added bias parameterization

* Refactored test guard, added bias guard in hipblasgemm

* PR comments adressed

* Minor typo

* Updated test per PR comments

* Improve test to cover padded scale_inv for mxfp8 gemm

* Address PR review comments

* Added arch specific guard for FP8 GEMM config

* Reformat inline comment

* Minor code reformat

* Remove debug statement, reformat code

* Formatting

* Reverted unnecessary shardy changes

---------

Co-authored-by: Veera Rajasekhar Reddy Gopu <veerarajasekharreddy.gopu@amd.com>