Skip to content

Pass input_output_alias to TritonAutotunedKernelCall #2814

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

Merged
tdophung merged 3 commits into from
Apr 2, 2026
Merged

Pass input_output_alias to TritonAutotunedKernelCall #2814

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
918df1e
Pass input_output_alias to TritonAutotunedKernelCall
nvjax Mar 31, 2026
7ed5f95
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Mar 31, 2026
f4e7f52
Add jax version guard for the input_output_aliasing fix
tdophung Mar 31, 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
27 changes: 13 additions & 14 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 @@ -474,23 +474,22 @@ def lowering(ctx, x, *, block_size):

kernel_calls.append((config_call, str(config)))

# IMPORTANT: We pass an empty tuple for input_output_aliases_with_sizes.
#
# Background:
# 1. jax.ffi.ffi_lowering(operand_output_aliases=...) is a HINT to XLA that an
# output can reuse an input's buffer. XLA may or may not honor this.
# 2. TritonAutotunedKernelCall's input_output_aliases_with_sizes triggers
# save/restore logic during autotuning (see jaxlib/gpu/triton_kernels.cc:630-701).
#
# The problem: The save phase (triton_kernels.cc:632) only saves if buffers[input_idx] == buffers[output_idx],
# but the restore phase (triton_kernels.cc:697-700) unconditionally iterates over all aliases and tries
# to access input_copies[input_idx]. If XLA didn't actually alias the buffers, input_copies[input_idx] doesn't exist, creating an empty vector whose .data() returns nullptr, causing CUDA_ERROR_INVALID_VALUE during the restore memcpy.
#
# WAR: Don't pass aliases to TritonAutotunedKernelCall.
input_output_aliases_with_sizes = ()
Comment thread
jberchtold-nvidia marked this conversation as resolved.
if input_output_aliases:
num_inputs = len(ctx.avals_in)
aliases = []
for input_idx, output_idx in input_output_aliases.items():
aval = ctx.avals_in[input_idx]
size_bytes = aval.size * jnp.dtype(aval.dtype).itemsize
# AutotunedKernelCall expects buffer indices (inputs + outputs).
buffer_output_idx = num_inputs + output_idx
aliases.append((input_idx, buffer_output_idx, size_bytes))
input_output_aliases_with_sizes = tuple(aliases)

kernel_call = gpu_triton.TritonAutotunedKernelCall(
f"{actual_kernel_fn.__name__}_autotuned",
kernel_calls,
(), # Empty to avoid buggy save/restore in jaxlib/gpu/triton_kernels.cc
input_output_aliases_with_sizes,
)

else:
Expand Down
Loading