Skip to content

cp: CUDA_DEVICE_MAX_CONNECTIONS update for Blackwell (1287) into r0.2.0 #1293

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鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ananthsub merged 1 commit into from
Nov 11, 2025
Merged

cp: CUDA_DEVICE_MAX_CONNECTIONS update for Blackwell (1287) into r0.2.0 #1293

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
18 changes: 5 additions & 13 deletions scripts/performance/perf_plugins.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,7 +194,6 @@ class PerfEnvPlugin(Plugin):
cp_size: int = 1
pp_size: int = 1
script_args_converter_fn: Optional[Callable[[PerfEnvPluginScriptArgs], List[str]]] = None
num_gpus: int = 8
moe_a2a_overlap: bool = False
model_name: str
model_size: str
Expand All @@ -207,26 +206,21 @@ def _set_num_cuda_device_max_connections(
self,
task: Union["run.Partial", "run.Script"],
executor: "run.Executor",
num_gpus: int,
tp_size: int,
cp_size: int,
pp_size: int,
moe_a2a_overlap: bool,
enable_deepep: bool,
gpu_sm100_or_newer: bool,
):
dp_size = num_gpus // (tp_size * cp_size * pp_size)

cuda_device_max_connections = 8
if enable_deepep:
cuda_device_max_connections = 32
if gpu_sm100_or_newer:
if (tp_size > 1 or cp_size > 1) and (dp_size > 1 or pp_size > 1):
"""
We need extra connections to avoid serialization of streams, so we use max connections of 32 instead
of the default device connection of 8.
"""
cuda_device_max_connections = 32
"""
We need extra connections to avoid serialization of streams, so we use max connections of 32 instead
of the default device connection of 8.
"""
cuda_device_max_connections = 32
else:
# Hopper or earlier generation GPUs
if (tp_size > 1 or cp_size > 1) and not moe_a2a_overlap:
Expand Down Expand Up @@ -370,10 +364,8 @@ def setup(self, task: Union["run.Partial", "run.Script"], executor: "run.Executo
self._set_num_cuda_device_max_connections(
task,
executor,
self.num_gpus,
tp_size,
cp_size,
pp_size,
moe_a2a_overlap=moe_a2a_overlap,
enable_deepep=enable_deepep,
gpu_sm100_or_newer=self.gpu in ["b200", "gb200", "gb300"],
Expand Down
3 changes: 0 additions & 3 deletions scripts/performance/setup_experiment.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,7 +50,6 @@ def main(
compute_dtype: str,
fp8_recipe: str,
gpu: str,
num_gpus: int,
hf_token: str,
custom_mounts: List[str],
detach: bool,
Expand Down Expand Up @@ -87,7 +86,6 @@ def main(
plugins.append(
PerfEnvPlugin(
enable_vboost=enable_vboost,
num_gpus=num_gpus,
moe_a2a_overlap=moe_a2a_overlap,
tp_size=tp_size,
pp_size=pp_size,
Expand Down Expand Up @@ -150,7 +148,6 @@ def main(
compute_dtype=args.compute_dtype,
fp8_recipe=args.fp8_recipe,
gpu=args.gpu,
num_gpus=args.num_gpus,
hf_token=args.hf_token,
custom_mounts=args.custom_mounts,
detach=args.detach,
Expand Down