nemo-mbridge-perf-cuda-graphs

CUDA Graphs

Stable documentation: @docs/training/cuda-graphs.md Card: @skills/nemo-mbridge-perf-cuda-graphs/card.yaml

What It Is

CUDA graphs capture GPU operations once and replay them with minimal host-driver overhead. Bridge supports two implementations:

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 optionally mlp
• dropless MoE: attn moe_router moe_preprocess
• VLMs: 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 local full-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 =

TE scoped graph (dense model)

TE scoped graph (MoE model)

Performance harness CLI

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 in gpt_provider.py)
• full_iteration scope only with cuda_graph_impl = "local"
• full_iteration scope requires check_for_nan_in_loss = False
• Do not combine moe scope and moe_router scope
• Tensor 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, set NCCL_GRAPH_REGISTER=0 (MCore enforces for local impl on arch < sm_100; TE impl asserts unconditionally)
• CPU offloading is incompatible with CUDA graphs
• moe_preprocess scope requires moe_router scope to also be set

Practical bring-up order

1. Stabilize the eager run first.
2. Fix sequence length and micro-batch size.
3. Enable the narrowest useful graph scope.
4. Confirm replay is active and memory is still acceptable.
5. Compare eager against graph replay iterations after warmup and capture; do not include the capture step in steady-state timing.
6. 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_tracker

Graph cleanup

def _delete_cuda_graphs(cuda_graph_helper):
    # Deletes FullCudaGraphWrapper and TE graph objects to free NCCL buffers

MCore classes (in 3rdparty/Megatron-LM)

• CudaGraphManager: megatron/core/transformer/cuda_graphs.py
• TECudaGraphHelper: megatron/core/transformer/cuda_graphs.py
• FullCudaGraphWrapper: megatron/core/full_cuda_graph.py
• CudaGraphScope enum: megatron/core/transformer/enums.py

Positive recipe anchors

• src/megatron/bridge/perf_recipes/deepseek/gb300/deepseek_v3.py
• src/megatron/bridge/perf_recipes/qwen/gb300/qwen3_moe.py
• src/megatron/bridge/perf_recipes/gpt_oss/gb300/gpt_oss.py

Tests

Pitfalls

1. TE RNG tracker is mandatory: Setting cuda_graph_impl without use_te_rng_tracker=True and rng.te_rng_tracker=True will assert in the provider.

2. full_iteration requires NaN checks disabled: The entire fwd+bwd is captured, so loss-NaN checking cannot inspect intermediate values.

3. MoE scope restrictions: moe scope and moe_router scope are mutually exclusive. Token-dropless MoE can only graph moe_router and moe_preprocess, not the full expert dispatch.

4. 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 > 1 compounds overhead since each stage holds its own graph.

5. Delayed wgrad interaction: When delay_wgrad_compute=True and attention or MoE router is in cuda_graph_scope, additional constraints apply: TE >= 2.12.0, gradient_accumulation_fusion=True, and no attention bias.

6. Variable-length sequences break graphs: Sequence lengths must be constant across steps. Use padded packed sequences if packing is needed.

7. 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.

8. Older GPU architectures: On GPUs with compute capability < 10.0 (pre-Blackwell), set NCCL_GRAPH_REGISTER=0 when using PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True. Enforced in MCore CudaGraphManager (cuda_graphs.py:1428) and TECudaGraphHelper (cuda_graphs.py:1697). The TE impl asserts unconditionally regardless of arch.

9. CPU offloading incompatible: CUDA graphs cannot be used with CPU offloading. Enforced in MCore transformer_config.py:1907.

10. MoE recompute + moe_router scope: MoE recompute is not supported with moe_router CUDA graph scope when using cuda_graph_impl = "transformer_engine". Enforced in MCore transformer_config.py:1977.

11. Layer-level recompute requires full_iteration scope: Using recompute_granularity="full" with recompute_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_V1 uses cuda_graph_impl= "transformer_engine", cuda_graph_scope="mlp"). Fix: use submodule recompute (recompute_granularity="selective" + recompute_modules), disable CUDA graphs, or switch to local + full_iteration. Enforced in MCore transformer_config.py:2001-2005. See also @skills/nemo-mbridge-perf-activation-recompute/SKILL.md.

12. 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.

13. 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_preprocess graphs successfully (48 graphable layers, about 6.9 s capture time on rank 0), but replay iterations 5-8 averaged 42.00 s versus 41.36 s for 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" -q

Functional smoke test (requires GPU)

uv run python -m pytest \
  tests/functional_tests/test_groups/recipes/test_llama_recipes_pretrain_cuda_graphs.py -q

Success criteria

• Unit tests pass, covering config validation for both local and transformer_engine implementations.
• Functional test completes training steps with both CUDA graph implementations.
• No NCCL errors or illegal memory access in logs.