Skip to content

Scaled dot fixes: 4D tensor handling#996

Open
pemeliya wants to merge 3 commits into
rocm-jaxlib-v0.10.2from
pemeliya/v0.10.2_scaled_dot_fixes
Open

Scaled dot fixes: 4D tensor handling#996
pemeliya wants to merge 3 commits into
rocm-jaxlib-v0.10.2from
pemeliya/v0.10.2_scaled_dot_fixes

Conversation

@pemeliya

@pemeliya pemeliya commented Jun 25, 2026

Copy link
Copy Markdown
Collaborator

Cherry-picked from: #940

Also, fixing compilation under rocm 7.13:

external/com_google_absl/absl/base/casts.h:183:23: note: while substituting prior template arguments into non-type template parameter [with Dest = _Float16, Source = uint16_t]
  183 | inline constexpr Dest bit_cast(const Source& source) {
      |                       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  184 |   return __builtin_bit_cast(Dest, source);
      |   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  185 | }
      | ~
xla/stream_executor/rocm/buffer_debug_float_check_kernel_rocm.cu.cc:45:5: note: while substituting deduced template arguments into function template 'bit_cast' [with Dest = _Float16, Source = uint16_t, $2 = (no value)]
   45 |     absl::bit_cast<_Float16>(uint16_t{0x7C00});
      |     ^

pemeliya added 2 commits June 25, 2026 04:37
[ROCm] Enable scaled dot for ROCm.

fixed test after adding triton backend

refactor: BlasLt interface reform and plan split

- Refactor BlasLt::Get() to take StreamExecutor* and return StatusOr<BlasLt*>
- Remove static convenience methods (GetMatmulPlan, GetGroupedMatmulPlan)
- Remove Stream* parameter from GetAlgorithms
- Split ROCm MatmulPlan into RegularMatmulPlan and GroupedMatmulPlan
- Remove redundant ExecuteOnStream overloads from base class
- Rename internal members (parent_ -> executor_, blas_lt_ -> handle_)
- Update all callers (autotuners, SYCL backend)

clang format

clang tidy

clang tidy

clang tidy

clang tidy

fixes after rebase

fixing deps

clang format

fix

fixing cuda test

fixing cuda

inlined NYI GroupedMatmulPlan for SyCL

addressing comments

Revert "refactor: BlasLt interface reform and plan split"

This reverts commit 690e8f6.

cherry-picking gpu_blas_lt refactoring changes from upstream

Revert "cherry-picking gpu_blas_lt refactoring changes from upstream"

This reverts commit 6aa1289.

updated text

added batch size check back

small refactoring and more robust way on scaled-dot rewriting
@pemeliya pemeliya requested review from ScXfjiang and i-chaochen June 25, 2026 05:11
@i-chaochen

Copy link
Copy Markdown
Collaborator

We also have

0.10.0: https://github.com/ROCm/xla/tree/rocm-jaxlib-v0.10.0
0.10.1: https://github.com/ROCm/xla/tree/rocm-jaxlib-v0.10.1

as well.

So it's best to upstream it first and let JAX team to decide which release branches they will pick up

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants