Skip to content

TritonKernelCall: CUDA graph compatibility #3000

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
tdophung wants to merge 2 commits into
base: main
Choose a base branch
from
+20 −6
Open

TritonKernelCall: CUDA graph compatibility #3000

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c1bb105
TritonKernelCall: CUDA graph compatibility
tdophung May 15, 2026
931ff20
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] May 15, 2026
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
23 changes: 17 additions & 6 deletions transformer_engine/jax/triton_extensions/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,14 +44,17 @@
from packaging import version

from jax import core
from jaxlib.mlir import ir
import jax
import jax.numpy as jnp

from ..version_utils import (
TRITON_AUTOTUNED_INPUT_OUTPUT_ALIAS_MIN_JAX_VERSION,
TRITON_EXTENSION_MIN_JAX_VERSION,
TRITON_EXTENSION_CUDA_GRAPH_MIN_JAX_VERSION,
is_triton_autotuned_alias_safe,
is_triton_extension_supported,
jax_version_meet_requirement,
)


Expand Down Expand Up @@ -581,18 +584,26 @@ def lowering(ctx, x, *, block_size):

serialized_metadata = b""
call_proto = kernel_call.to_proto(actual_kernel_fn.__name__, serialized_metadata)
compressed_call_proto = zlib.compress(call_proto)

if input_output_aliases:
ffi_operand_output_aliases = input_output_aliases
else:
ffi_operand_output_aliases = None

# Use JAX FFI lowering with compressed protobuf
rule = jax.ffi.ffi_lowering(
"triton_kernel_call", # Custom call target registered in gpu_triton.py
api_version=2,
backend_config=zlib.compress(call_proto),
operand_output_aliases=ffi_operand_output_aliases,
)
if jax_version_meet_requirement(TRITON_EXTENSION_CUDA_GRAPH_MIN_JAX_VERSION):
rule = jax.ffi.ffi_lowering(
"triton_kernel_call_ffi", # Custom call target registered in gpu_triton.py
backend_config={"kernel_call_proto": ir.StringAttr.get(compressed_call_proto)},
operand_output_aliases=ffi_operand_output_aliases,
)
else:
rule = jax.ffi.ffi_lowering(
"triton_kernel_call", # Custom call target registered in gpu_triton.py
api_version=2,
backend_config=compressed_call_proto,
operand_output_aliases=ffi_operand_output_aliases,
)

return rule(ctx, *array_args)
3 changes: 3 additions & 0 deletions transformer_engine/jax/version_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,9 @@ def jax_version_meet_requirement(version: str):
# Minimum JAX version required for Triton kernel dispatch (jaxlib < 0.8.0 segfaults).
TRITON_EXTENSION_MIN_JAX_VERSION = "0.8.0"

# Minimum JAX version for non-legacy Triton kernel FFI (supporting CUDA graph capture)
TRITON_EXTENSION_CUDA_GRAPH_MIN_JAX_VERSION = "0.10.1.dev0"
Comment on lines +28 to +29
Copy link
Copy Markdown
Contributor

@greptile-apps greptile-apps Bot May 15, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Using "0.10.1.dev0" as the floor accepts every 0.10.1.devYYYYMMDD nightly build (because any real date integer > 0), including early builds that may predate the CUDA-graph support. The existing pattern in this file pins to the exact nightly that first shipped the fix (e.g. "0.9.2.dev20260317"). If the feature was not present from the very first 0.10.1 dev build, this floor will silently enable the new code path on builds that don't have the matching C++ registration, likely producing a KeyError or undefined-symbol crash at lowering time.

Suggested change
# Minimum JAX version for non-legacy Triton kernel FFI (supporting CUDA graph capture)
TRITON_EXTENSION_CUDA_GRAPH_MIN_JAX_VERSION = "0.10.1.dev0"
# Minimum JAX version for non-legacy Triton kernel FFI (supporting CUDA graph capture).
# Pin to the exact nightly that first shipped the triton_kernel_call_ffi registration
# (analogous to _TRITON_AUTOTUNED_ALIAS_NIGHTLY_FLOOR above).
# TODO: replace dev0 with the first nightly date that contains the feature, e.g.
# "0.10.1.dev20YYMMDD", and add a separate stable floor if 0.10.1 stable ships the fix.
TRITON_EXTENSION_CUDA_GRAPH_MIN_JAX_VERSION = "0.10.1.dev0"


# Nightly and stable floors for safe input_output_aliases in TritonAutotunedKernelCall.
# jaxlib/gpu/triton_kernels.cc had a bug in the autotuning save/restore loop:
# it iterated over all declared aliases unconditionally, but input_copies only
Expand Down
Loading