Skip to content

Cherry-pick ROCm fixes from PR #740 to rocm-jaxlib-v0.8.2 #786

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
srinivamd wants to merge 1 commit into
base: rocm-jaxlib-v0.8.2
Choose a base branch
from
Open

Cherry-pick ROCm fixes from PR #740 to rocm-jaxlib-v0.8.2 #786

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 24 additions & 1 deletion jax/_src/cudnn/scaled_matmul_stablehlo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@
from jax._src import core
from jax._src import dispatch
from jax._src import dtypes
from jax._src import lax as lax_internal
from jax._src import numpy as jnp
from jax._src import tree_util
from jax._src.custom_derivatives import custom_vjp
Expand Down Expand Up @@ -104,6 +105,23 @@ def _scaled_matmul_gpu_lowering(
return [out.result]


def _scaled_matmul_rocm_lowering(
ctx, a, b, a_scales, b_scales, preferred_element_type
):
def _scaled_dot_lowering_impl(lhs, rhs, lhs_scales, rhs_scales):
return lax_internal.scaled_dot(
lhs,
rhs,
lhs_scale=lhs_scales,
rhs_scale=rhs_scales,
dimension_numbers=(((2,), (2,)), ((0,), (0,))),
preferred_element_type=preferred_element_type,
)
return mlir.lower_fun(_scaled_dot_lowering_impl, multiple_results=False)(
ctx, a, b, a_scales, b_scales
)


def _scaled_matmul_abstract(a, b, a_scale, b_scale, *, preferred_element_type):
batch, non_contracting_lhs, contracting_lhs = a.shape
_, non_contracting_rhs, _ = b.shape
Expand All @@ -122,9 +140,14 @@ def _scaled_matmul_abstract(a, b, a_scale, b_scale, *, preferred_element_type):
_scaled_matmul_gpu_lowering,
platform="cuda",
)

# Keep CUDA lowering on the existing `__op$block_scaled_dot` custom call.
# ROCm is registered separately because AMD's backend does not support that
# custom call; instead, lowering through `lax.scaled_dot` lets XLA emit
# `kScaledDot`, which ROCm can then fuse via Triton or hipBLASLt when possible.
mlir.register_lowering(
_scaled_matmul_p,
_scaled_matmul_gpu_lowering,
_scaled_matmul_rocm_lowering,
platform="rocm",
)

Expand Down
4 changes: 4 additions & 0 deletions tests/ann_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,6 +119,10 @@ def test_autodiff(self, shape, dtype, k, is_max_k):
)
def test_pmap(self, qy_shape, db_shape, dtype, k, recall):
num_devices = jax.device_count()
# TODO(araganes): Re-enable once upstream HloShardingV3 lands (JAX 0.9.2+).
# New pmap's SPMD tiling can't convert back to 1D pmap mesh on ROCm.
if jtu.is_device_rocm():
self.skipTest("IndivisibleError: SPMD tiling incompatible with 1D pmap mesh on ROCm")
rng = jtu.rand_default(self.rng())
qy = rng(qy_shape, dtype)
db = rng(db_shape, dtype)
Expand Down