Skip to content

Enhancement: Add rocprofv2 trace support for AMD GPUs #817

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
shcho wants to merge 5 commits into
base: main
Choose a base branch
from
+49 −15
Open

Enhancement: Add rocprofv2 trace support for AMD GPUs #817

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
20ebb38
Enhancement: Add rocprofv2 trace support for AMD GPUs
May 13, 2026
9b1dd82
Use shlex.quote() for trace output paths in runner.py
May 19, 2026
a9485ef
Rename nsys_prefix to trace_prefix for tool-agnostic naming
May 19, 2026
8a22f3e
Use getattr for safe torch.version.cuda/hip access in pytorch_base.py
Jul 7, 2026
a351951
Add -- separator after rocprofv2 options in runner.py
Jul 7, 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
6 changes: 4 additions & 2 deletions superbench/benchmarks/model_benchmarks/pytorch_base.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -591,8 +591,10 @@ def _benchmark(self):
Run the benchmark then handle post-run model log save/compare.
Set SB_ENABLE_PYTORCH_PROFILER='1' to enable profiling.
"""
# Check if this is a Nvidia GPU
if not (torch.cuda.is_available() and torch.version.cuda is not None):
# Check if this is a Nvidia or AMD GPU
if not (torch.cuda.is_available() and (
getattr(torch.version, 'cuda', None) is not None or getattr(torch.version, 'hip', None) is not None
)):
ok = super()._benchmark()
self._post_run_model_log()
return ok
Expand Down
58 changes: 45 additions & 13 deletions superbench/runner/runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@
import sys
import json
import random
import shlex
import signal
from pathlib import Path
from pprint import pformat
Expand Down Expand Up @@ -135,12 +136,25 @@ def __get_mode_command(self, benchmark_name, mode, timeout=None):
enable_nsys = os.environ.get('SB_ENABLE_NSYS', '') == '1'
trace_dir = os.environ.get('SB_NSYS_TRACE_DIR', self._sb_output_dir)

# Enable rocprofv2 profiling based on environment variable
enable_rocprof = os.environ.get('SB_ENABLE_ROCPROF', '') == '1'
rocprof_trace_dir = os.environ.get('SB_ROCPROF_TRACE_DIR', self._sb_output_dir)
Comment thread
shcho marked this conversation as resolved.

mode_command = exec_command
if mode.name == 'local':
trace_command = (
f'nsys profile --output {trace_dir}/{benchmark_name}_{mode.proc_rank}_traces '
f'--backtrace none --sample none --force-overwrite true --cpuctxsw none --trace cuda,nvtx '
) if enable_nsys and mode.proc_rank == 0 else ''
trace_command = ''
if enable_nsys and mode.proc_rank == 0:
trace_output = shlex.quote(f'{trace_dir}/{benchmark_name}_{mode.proc_rank}_traces')
trace_command = (
f'nsys profile --output {trace_output} '
f'--backtrace none --sample none --force-overwrite true --cpuctxsw none --trace cuda,nvtx '
)
elif enable_rocprof and mode.proc_rank == 0:
trace_output = shlex.quote(f'{rocprof_trace_dir}/{benchmark_name}_{mode.proc_rank}_traces')
trace_command = (
f'rocprofv2 --hip-trace --kernel-trace --plugin json '
f'-d {trace_output} -- '
)
Comment thread
shcho marked this conversation as resolved.
Comment thread
shcho marked this conversation as resolved.
Comment thread
shcho marked this conversation as resolved.
# Build the command parts, only including trace if it's not empty
command_parts = []
prefix = mode.prefix.format(proc_rank=mode.proc_rank, proc_num=mode.proc_num)
Expand All @@ -159,23 +173,41 @@ def __get_mode_command(self, benchmark_name, mode, timeout=None):
'--nnodes=$NNODES --node_rank=$NODE_RANK --master_addr=$MASTER_ADDR --master_port=$MASTER_PORT '
)

nsys_prefix = (
f'nsys profile --output {trace_dir}/{benchmark_name}_traces '
f'--backtrace none --sample none --force-overwrite true --cpuctxsw none --trace cuda,nvtx '
) if enable_nsys else ''
trace_prefix = ''
if enable_nsys:
trace_output = shlex.quote(f'{trace_dir}/{benchmark_name}_traces')
trace_prefix = (
f'nsys profile --output {trace_output} '
f'--backtrace none --sample none --force-overwrite true --cpuctxsw none --trace cuda,nvtx '
)
elif enable_rocprof:
trace_output = shlex.quote(f'{rocprof_trace_dir}/{benchmark_name}_traces')
trace_prefix = (
f'rocprofv2 --hip-trace --kernel-trace --plugin json '
f'-d {trace_output} -- '
)
Comment thread
shcho marked this conversation as resolved.

mode_command = (
f'{nsys_prefix}'
f'{trace_prefix}'
f'torchrun'
f' --no_python --nproc_per_node={mode.proc_num} {torch_dist_params}{exec_command}'
f' superbench.benchmarks.{benchmark_name}.parameters.distributed_impl=ddp'
f' superbench.benchmarks.{benchmark_name}.parameters.distributed_backend=nccl'
)
elif mode.name == 'mpi':
trace_command = (
f'nsys profile --output {trace_dir}/{benchmark_name}_{mode.proc_rank}_traces '
f'--backtrace none --sample none --force-overwrite true --cpuctxsw none --trace cuda,nvtx '
) if enable_nsys else ''
trace_command = ''
if enable_nsys:
trace_output = shlex.quote(f'{trace_dir}/{benchmark_name}_{mode.proc_rank}_traces')
trace_command = (
f'nsys profile --output {trace_output} '
f'--backtrace none --sample none --force-overwrite true --cpuctxsw none --trace cuda,nvtx '
)
elif enable_rocprof:
trace_output = shlex.quote(f'{rocprof_trace_dir}/{benchmark_name}_{mode.proc_rank}_traces')
trace_command = (
f'rocprofv2 --hip-trace --kernel-trace --plugin json '
f'-d {trace_output} -- '
)
mode_command = (
'{trace} '
'mpirun ' # use default OpenMPI in image
Expand Down
Loading