Nemo Mbridge Perf Cuda Graphs
NVIDIAOfficialPublished by NVIDIA
- Skill
- Tool integration
- Software development
- Apache-2.0
- Medium risk
AutoGineer summary
Written by AutoGineer from the skill’s own files, not by the author
This skill provides guidance on implementing CUDA graphs within the Megatron Bridge framework to minimize host-driver overhead. It covers two primary implementations: 'local' for full-iteration capture and 'transformer_engine' for scoped capture of specific modules like attention, MLP, and MoE. The documentation details configuration requirements, mandatory RNG tracker settings, performance harness usage, and critical constraints such as static tensor shapes and memory overhead considerations.
What it does
- Enables CUDA graph capture for reduced host-driver overhead.
- Supports local full-iteration graphs and Transformer Engine scoped graphs.
- Provides configuration templates for dense and MoE models.
- Includes a performance harness CLI for testing graph implementations.
- Enforces mandatory constraints like TE RNG tracker requirements.
- Offers verification steps via unit and functional tests.
When to use it
- When aiming to reduce host-driver overhead in training workloads.
- When troubleshooting performance regressions related to CUDA graph configurations.
- When optimizing dense models or dropless MoE architectures.
- When transitioning from eager execution to graphed callables.
Advantages
- Reduces CPU-side overhead by capturing and replaying GPU operations.
- Provides flexible scoping options to target specific performance bottlenecks.
- Integrates directly with Megatron Bridge configuration and training loops.
- Includes built-in validation for complex constraints like recompute and RNG tracking.
Usage examples
- “How do I enable TE-scoped CUDA graphs for a dense model?”
- Set cfg.model.cuda_graph_impl to 'transformer_engine', define the scope (e.g., ['attn', 'mlp']), and ensure use_te_rng_tracker is enabled.
- “What are the requirements for using full-iteration CUDA graphs?”
- Use 'local' implementation, set scope to ['full_iteration'], and disable check_for_nan_in_loss in the configuration.
Install
Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/nemo-mbridge-perf-cuda-graphs/downloadUnzip it into the cross-client skills folder of your project:
mkdir -p .agents/skills && unzip nemo-mbridge-perf-cuda-graphs.zip -d .agents/skills
Documentation by NVIDIA
CUDA Graphs
Stable documentation: @docs/training/cuda-graphs.md Card: @skills/nemo-mbridge-perf-cuda-graphs/card.yaml
<!-- NVSkills CI refresh: 2026-06-15. No instruction changes. -->
What It Is
CUDA graphs capture GPU operations once and replay them with minimal host-driver overhead. Bridge supports two implementations:
cuda_graph_impl | Mechanism | Scope support |
|---|---|---|
"local" | MCore FullCudaGraphWrapper wrapping entire fwd+bwd | full_iteration |
"transformer_engine" | TE make_graphed_callables() per layer | attn, mlp, moe, moe_router, moe_preprocess, mamba |
Quick Decision
Start with TE-scoped graphs for most training workloads, then verify replay timing against eager on the same dispatcher, layout, and container:
dense models:
attn, then optionallymlpdropless MoE:
attn moe_router moe_preprocessVLMs: the same dropless-MoE scope, but only after the real-data path is stable
Use local + full_iteration only when you specifically want full-iteration
capture and can satisfy the tighter constraints.
For recompute-heavy workloads:
TE-scoped graphs pair naturally with selective recompute
full recompute usually pushes you toward
localfull-iteration graphs or away from graphs entirely
Related docs:
@docs/training/cuda-graphs.md
@docs/training/activation-recomputation.md
Enablement
Local full-iteration graph
cfg.model.cuda_graph_impl = "local"
cfg.model.cuda_graph_scope = ["full_iteration"]
cfg.model.cuda_graph_warmup_steps = 3
cfg.model.use_te_rng_tracker = True
cfg.rng.te_rng_tracker = True
cfg.rerun_state_machine.check_for_nan_in_loss = False
cfg.ddp.check_for_nan_in_grad = FalseTE scoped graph (dense model)
cfg.model.cuda_graph_impl = "transformer_engine"
cfg.model.cuda_graph_scope = ["attn"] # or ["attn", "mlp"]
cfg.model.cuda_graph_warmup_steps = 3
cfg.model.use_te_rng_tracker = True
cfg.rng.te_rng_tracker = TrueTE scoped graph (MoE model)
cfg.model.cuda_graph_impl = "transformer_engine"
cfg.model.cuda_graph_scope = ["attn", "moe_router", "moe_preprocess"]
cfg.model.cuda_graph_warmup_steps = 3
cfg.model.use_te_rng_tracker = True
cfg.rng.te_rng_tracker = TruePerformance harness CLI
uv run python scripts/performance/run_script.py \
-m qwen \
-mr qwen3_30b_a3b \
--task pretrain \
-g h100 \
-c bf16 \
-ng 16 \
--cuda_graph_impl transformer_engine \
--cuda_graph_scope attn,moe_router,moe_preprocess \
...Valid CLI values live in scripts/performance/argument_parser.py:
VALID_CUDA_GRAPH_IMPLS:["none", "local", "transformer_engine"]VALID_CUDA_GRAPH_SCOPES:["full_iteration", "attn", "mlp", "moe", "moe_router", "moe_preprocess", "mamba"]
The performance harness uses a comma-separated --cuda_graph_scope value and
auto-enables model.use_te_rng_tracker plus rng.te_rng_tracker when
--cuda_graph_impl is not none.
Required constraints
use_te_rng_tracker = True(enforced ingpt_provider.py)full_iterationscope only withcuda_graph_impl = "local"full_iterationscope requirescheck_for_nan_in_loss = FalseDo not combine
moescope andmoe_routerscopeTensor shapes must be static (fixed seq_length, fixed micro_batch_size)
MoE token-dropless routing limits graphable scope to dense modules
With
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, setNCCL_GRAPH_REGISTER=0(MCore enforces for local impl on arch < sm_100; TE impl asserts unconditionally)CPU offloading is incompatible with CUDA graphs
moe_preprocessscope requiresmoe_routerscope to also be set
Practical bring-up order
Stabilize the eager run first.
Fix sequence length and micro-batch size.
Enable the narrowest useful graph scope.
Confirm replay is active and memory is still acceptable.
Compare eager against graph replay iterations after warmup and capture; do not include the capture step in steady-state timing.
Only then widen scope or combine with overlap features.
Code Anchors
Bridge config and validation
# CUDA graph scope validation: check_for_nan_in_loss must be disabled with full_iteration graph
if self.model.cuda_graph_impl == "local" and CudaGraphScope.full_iteration in self.model.cuda_graph_scope:
assert not self.rerun_state_machine.check_for_nan_in_loss, (
"check_for_nan_in_loss must be disabled when using full_iteration CUDA graph. "
"Set rerun_state_machine.check_for_nan_in_loss=False."
)
if self.model.cuda_graph_impl == "none":
self.model.cuda_graph_scope = []TE RNG tracker requirement
if self.cuda_graph_impl != "none":
assert getattr(self, "use_te_rng_tracker", False), (
"Transformer engine's RNG tracker is required for cudagraphs, it can be "
"enabled with use_te_rng_tracker=True'."Graph creation and capture in training loop
# Capture CUDA Graphs.
cuda_graph_helper = None
if model_config.cuda_graph_impl == "transformer_engine":
cuda_graph_helper = TECudaGraphHelper(...)
# ...
if config.model.cuda_graph_impl == "local" and CudaGraphScope.full_iteration in config.model.cuda_graph_scope:
forward_backward_func = FullCudaGraphWrapper(
forward_backward_func, cuda_graph_warmup_steps=config.model.cuda_graph_warmup_steps
)TE graph capture after warmup
# Capture CUDA Graphs after warmup.
if (
model_config.cuda_graph_impl == "transformer_engine"
and cuda_graph_helper is not None
and not cuda_graph_helper.graphs_created()
and global_state.train_state.step - start_iteration == model_config.cuda_graph_warmup_steps
):
if model_config.cuda_graph_warmup_steps > 0 and should_toggle_forward_pre_hook:
disable_forward_pre_hook(model, param_sync=False)
cuda_graph_helper.create_cudagraphs()
if model_config.cuda_graph_warmup_steps > 0 and should_toggle_forward_pre_hook:
enable_forward_pre_hook(model)
cuda_graph_helper.cuda_graph_set_manual_hooks()RNG initialization
_set_random_seed(
rng_config.seed,
rng_config.data_parallel_random_init,
rng_config.te_rng_tracker,
rng_config.inference_rng_tracker,
use_cudagraphable_rng=(model_config.cuda_graph_impl != "none"),
pg_collection=pg_collection,
)Delayed wgrad + CUDA graph interaction
cuda_graph_scope = getattr(model_cfg, "cuda_graph_scope", []) or []
# ... scope parsing ...
if wgrad_in_graph_scope:
assert is_te_min_version("2.12.0"), ...
assert model_cfg.gradient_accumulation_fusion, ...
if attn_scope_enabled:
assert not model_cfg.add_bias_linear and not model_cfg.add_qkv_bias, ...Perf harness override helper
def _set_cuda_graph_overrides(
recipe, cuda_graph_impl=None, cuda_graph_scope=None
):
# Sets impl, scope, and auto-enables te_rng_trackerGraph cleanup
def _delete_cuda_graphs(cuda_graph_helper):
# Deletes FullCudaGraphWrapper and TE graph objects to free NCCL buffersMCore classes (in 3rdparty/Megatron-LM)
CudaGraphManager:megatron/core/transformer/cuda_graphs.pyTECudaGraphHelper:megatron/core/transformer/cuda_graphs.pyFullCudaGraphWrapper:megatron/core/full_cuda_graph.pyCudaGraphScopeenum:megatron/core/transformer/enums.py
Positive recipe anchors
src/megatron/bridge/perf_recipes/deepseek/gb300/deepseek_v3.pysrc/megatron/bridge/perf_recipes/qwen/gb300/qwen3_moe.pysrc/megatron/bridge/perf_recipes/gpt_oss/gb300/gpt_oss.py
Tests
| File | Coverage |
|---|---|
tests/unit_tests/training/test_config.py | full_iteration NaN-check constraint |
tests/unit_tests/training/test_comm_overlap.py | delay_wgrad + CUDA graph interaction |
tests/unit_tests/models/test_gpt_full_te_layer_autocast_spec.py | TE autocast with CUDA graphs |
tests/functional_tests/test_groups/recipes/test_llama_recipes_pretrain_cuda_graphs.py | End-to-end local and TE graph smoke tests |
tests/unit_tests/recipes/kimi/test_kimi_k2.py | TE + CUDA graph recipe config |
tests/unit_tests/recipes/gpt/test_gpt3_175b.py | TE + CUDA graph recipe config |
tests/unit_tests/recipes/qwen_vl/test_qwen25_vl_recipes.py | VLM CUDA graph settings |
Pitfalls
TE RNG tracker is mandatory: Setting
cuda_graph_implwithoutuse_te_rng_tracker=Trueandrng.te_rng_tracker=Truewill assert in the provider.full_iterationrequires NaN checks disabled: The entire fwd+bwd is captured, so loss-NaN checking cannot inspect intermediate values.MoE scope restrictions:
moescope andmoe_routerscope are mutually exclusive. Token-dropless MoE can only graphmoe_routerandmoe_preprocess, not the full expert dispatch.Memory overhead: CUDA graphs pin all intermediate buffers for the graph's lifetime (no memory reuse). TE scoped graphs add a few GB; full-iteration graphs can increase peak memory by 1.5–2×.
PP > 1compounds overhead since each stage holds its own graph.Delayed wgrad interaction: When
delay_wgrad_compute=Trueand attention or MoE router is incuda_graph_scope, additional constraints apply: TE >= 2.12.0,gradient_accumulation_fusion=True, and no attention bias.Variable-length sequences break graphs: Sequence lengths must be constant across steps. Use padded packed sequences if packing is needed.
Graph cleanup is required: CUDA graph objects hold NCCL buffer references. Bridge handles this in
_delete_cuda_graphs()at the end of training, but early exits must call it explicitly.Older GPU architectures: On GPUs with compute capability < 10.0 (pre-Blackwell), set
NCCL_GRAPH_REGISTER=0when usingPYTORCH_CUDA_ALLOC_CONF=expandable_segments:True. Enforced in MCoreCudaGraphManager(cuda_graphs.py:1428) andTECudaGraphHelper(cuda_graphs.py:1697). The TE impl asserts unconditionally regardless of arch.CPU offloading incompatible: CUDA graphs cannot be used with CPU offloading. Enforced in MCore
transformer_config.py:1907.MoE recompute + moe_router scope: MoE recompute is not supported with
moe_routerCUDA graph scope when usingcuda_graph_impl = "transformer_engine". Enforced in MCoretransformer_config.py:1977.Layer-level recompute requires
full_iterationscope: Usingrecompute_granularity="full"withrecompute_num_layers(recompute N whole transformer layers) is incompatible with TE-scoped graphs. MCore calls this "full" granularity even though you're selecting how many layers — the name refers to recomputing the full layer, not full model. Any TE-scoped scope (attn,mlp,moe_router, etc.) will assert:AssertionError: full recompute is only supported with full iteration CUDA graph.This commonly hits FP8 configs that default to TE-scoped graphs (e.g.LLAMA3_70B_SFT_CONFIG_H100_FP8_CS_V1usescuda_graph_impl= "transformer_engine",cuda_graph_scope="mlp"). Fix: use submodule recompute (recompute_granularity="selective"+recompute_modules), disable CUDA graphs, or switch tolocal+full_iteration. Enforced in MCoretransformer_config.py:2001-2005. See also @skills/nemo-mbridge-perf-activation-recompute/SKILL.md.Benchmark numbers are workload-specific: graph wins are usually real when host overhead is visible, but the exact gain depends on batch shape, PP depth, recompute, dispatcher backend, and whether the eager baseline was already optimized.
A successful capture is not a speedup guarantee: On 2026-05-18, Qwen3 30B A3B H100 BF16 pretrain with the all-to-all dispatcher captured TE-scoped
attn,moe_router,moe_preprocessgraphs successfully (48graphable layers, about6.9 scapture time on rank 0), but replay iterations 5-8 averaged42.00 sversus41.36 sfor eager. Treat scoped graphs as a bring-up candidate and validate on the target stack.
Verification
Unit tests
uv run python -m pytest \
tests/unit_tests/training/test_config.py -k "cuda_graph" \
tests/unit_tests/training/test_comm_overlap.py -k "cuda_graph" \
tests/unit_tests/models/test_gpt_full_te_layer_autocast_spec.py -k "cuda_graph" -qFunctional smoke test (requires GPU)
uv run python -m pytest \
tests/functional_tests/test_groups/recipes/test_llama_recipes_pretrain_cuda_graphs.py -qSuccess criteria
Unit tests pass, covering config validation for both
localandtransformer_engineimplementations.Functional test completes training steps with both CUDA graph implementations.
No NCCL errors or illegal memory access in logs.
Files
- ATTRIBUTION.mdAdded by AutoGineer857 B
- BENCHMARK.md3.9 KiB
- LICENSELicence file added by AutoGineer11.2 KiB
- SKILL.md13.8 KiB
- card.yaml13.5 KiB
- evals/evals.json1.4 KiB
- skill-card.md4.1 KiB
- skill.oms.sig4.7 KiB
LICENSE
Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [yyyy] [name of copyright owner]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
Requirements and compatibility
Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta
“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.
Needs: Needs Python
Security
What our scanner noticed· Medium risk
- Large base64-encoded blobs
A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.
Attribution and licence
- Published by
- NVIDIA
- Publisher
- NVIDIA (Official)
- Source
- NVIDIA/skills · skills/nemo-mbridge-perf-cuda-graphs · commit fce3cca
- Licence
- Apache License 2.0 (redistribution allowed)
- LICENSE (added by AutoGineer from LICENSE-APACHE)
- ATTRIBUTION.md (written by AutoGineer)
- Synced
AutoGineer is not affiliated with or endorsed by NVIDIA or NVIDIA.
Download ZIP
The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download nemo-mbridge-perf-cuda-graphs.zip