Skip to content

[https://nvbugs/5647400] [fix] Enlarged the AllReduce workspace size to 64MB. Added AllReduce strategy to AD config. #9145

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
MrGeva merged 16 commits into from
Nov 25, 2025
Merged

[https://nvbugs/5647400] [fix] Enlarged the AllReduce workspace size to 64MB. Added AllReduce strategy to AD config. #9145

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
66952c7
fixed auto
MrGeva Nov 10, 2025
1554ced
strategy configurable , added test
MrGeva Nov 10, 2025
9dbd383
test to tp2 instead of tp4, fixed llm_root fixture
MrGeva Nov 10, 2025
c8ea3ca
moved allreduce_strategy to BaseLlmArgs
MrGeva Nov 10, 2025
f77c3ba
moved to be sharding config param
MrGeva Nov 11, 2025
c926392
fixed param in pt be
MrGeva Nov 11, 2025
d01cca9
fixed strategy in fused ops, and make it mandatory param
MrGeva Nov 11, 2025
3334637
allreduce_strategy to be mandatory. ShardingConfig to inject strategy…
MrGeva Nov 13, 2025
34aa1a8
fixed CR, fixed failing tests, added test
MrGeva Nov 16, 2025
6786c40
Merge branch 'main' into egeva/pass_as_param
MrGeva Nov 18, 2025
7e42197
Merge branch 'main' into egeva/pass_as_param
MrGeva Nov 19, 2025
58a9864
fixed hf issue in test
MrGeva Nov 23, 2025
d582e92
fixed hf model, fixed missing strategy param in rms
MrGeva Nov 23, 2025
db96d5b
Merge branch 'main' into egeva/pass_as_param
MrGeva Nov 23, 2025
301b977
fixed model path in test
MrGeva Nov 23, 2025
92b0944
fixed model_path arg in test
MrGeva Nov 24, 2025
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
Prev Previous commit
Next Next commit
moved allreduce_strategy to BaseLlmArgs 
Signed-off-by: Eran Geva <19514940+MrGeva@users.noreply.github.com>
  • Loading branch information
@MrGeva
MrGeva committed Nov 10, 2025
commit c8ea3ca85863b4fba17c0026bad0335d5a9ce22f
19 changes: 0 additions & 19 deletions tensorrt_llm/_torch/auto_deploy/llm_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,25 +123,6 @@ class AutoDeployConfig(DynamicYamlMixInForSettings, BaseSettings):

device: str = Field(default="cuda", description="The device to use for the model.", frozen=True)

allreduce_strategy: Literal[
"AUTO",
"NCCL",
"ONESHOT",
"TWOSHOT",
"MIN_LATENCY",
"LOWPRECISION",
"UB",
"MNNVL",
"NCCL_SYMMETRIC",
] = Field(
default="AUTO",
description="AllReduce strategy for distributed inference. Options: AUTO (automatic selection), "
"NCCL (NCCL-based), ONESHOT (single-phase fusion kernel), TWOSHOT (two-phase fusion kernel), "
"MIN_LATENCY (minimum latency heuristic), LOWPRECISION (low precision allreduce), "
"UB (unified buffer), MNNVL (multi-node NVLINK), NCCL_SYMMETRIC (NCCL symmetric). "
"AUTO is recommended for most use cases.",
)

# TODO: see if we can just remove this field and use kv_cache_config.dtype instead?
kv_cache_dtype: str = Field(
default="auto",
Expand Down
18 changes: 12 additions & 6 deletions tensorrt_llm/llmapi/llm_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1577,6 +1577,18 @@ class BaseLlmArgs(StrictBaseModel):
default=None,
description="The expert parallel size for MoE models's expert weights.")

allreduce_strategy: Optional[Literal[
'AUTO', 'NCCL', 'UB', 'MINLATENCY', 'ONESHOT', 'TWOSHOT',
'LOWPRECISION', 'MNNVL', 'NCCL_SYMMETRIC']] = Field(
default='AUTO',
description=
"AllReduce strategy for distributed inference. Options: AUTO (automatic selection), "
"NCCL (NCCL-based), ONESHOT (single-phase fusion kernel), TWOSHOT (two-phase fusion kernel), "
"MIN_LATENCY (minimum latency heuristic), LOWPRECISION (low precision allreduce), "
"UB (unified buffer), MNNVL (multi-node NVLINK), NCCL_SYMMETRIC (NCCL symmetric). "
"AUTO is recommended for most use cases.",
status="beta")

enable_attention_dp: bool = Field(
default=False,
description="Enable attention data parallel.",
Expand Down Expand Up @@ -2531,12 +2543,6 @@ class TorchLlmArgs(BaseLlmArgs):
status="prototype",
)

allreduce_strategy: Optional[Literal[
'AUTO', 'NCCL', 'UB', 'MINLATENCY', 'ONESHOT', 'TWOSHOT',
'LOWPRECISION', 'MNNVL',
'NCCL_SYMMETRIC']] = Field(default='AUTO',
description="Allreduce strategy to use.",
status="beta")
checkpoint_loader: Optional[object] = Field(
default=None,
description=
Expand Down