[Dev] Refactor CUDA graph API: decompose cuda_graph_scope into full_iteration impl, inference scope, and per-layer capture modules

docs/discussions/deepseek-v3-gb200-optimization/deepseek-v3-gb200-reproduce-guide.md

@@ -263,7 +263,7 @@ This is a very important step on GB200.
 --recompute-granularity selective \
 --recompute-modules moe_act mlp \
 --cuda-graph-impl transformer_engine \
---cuda-graph-scope attn moe_router moe_preprocess \
+--cuda-graph-modules attn moe_router moe_preprocess \
 --te-rng-tracker \
 --pipeline-model-parallel-layout "Et|(tt|)*30L" \
 --moe-router-force-load-balancing \
@@ -304,7 +304,7 @@ The following arguments indicate key optimizations.
 
 ```bash
 --cuda-graph-impl transformer_engine \
---cuda-graph-scope attn moe_router moe_preprocess \
+--cuda-graph-modules attn moe_router moe_preprocess \
 --te-rng-tracker \
 ```
 

docs/user-guide/features/cuda_graph.md

@@ -0,0 +1,214 @@
+<!---
+   Copyright (c) 2022-2026, NVIDIA CORPORATION. All rights reserved.
+   NVIDIA CORPORATION and its licensors retain all intellectual property
+   and proprietary rights in and to this software, related documentation
+   and any modifications thereto. Any use, reproduction, disclosure or
+   distribution of this software or related documentation without an express
+   license agreement from NVIDIA CORPORATION is strictly prohibited.
+-->
+
+# CUDA Graph
+
+CUDA Graphs reduce kernel-launch overhead by recording GPU operations once and replaying the recording on subsequent iterations. Megatron-LM provides three CUDA graph implementations controlled by `--cuda-graph-impl`.
+
+For implementation background and design details, see NVIDIA's
+[Transformer Engine and Megatron-LM CUDA Graph Support](https://docs.nvidia.com/dl-cuda-graph/torch-cuda-graph/te-megatron-cuda-graphs.html).
+That article is a useful conceptual reference, but some examples there still use older flags such as
+`--enable-cuda-graph` or `--cuda-graph-scope full_iteration`; in this repository, prefer
+`--cuda-graph-impl local|transformer_engine|full_iteration` as documented below.
+
+## Overview
+
+CUDA graph behavior is set by three orthogonal flags:
+
+| Flag | Values | Purpose |
+|---|---|---|
+| `--cuda-graph-impl` | `none` / `local` / `transformer_engine` / `full_iteration` | Which capture backend or strategy to use |
+| `--cuda-graph-modules` | `attn` / `mlp` / `moe` / `moe_router` / `moe_preprocess` / `mamba` | Per-layer **training** capture coverage; multi-valued and only meaningful for `local` and `transformer_engine` |
+| `--inference-cuda-graph-scope` | `none` / `layer` / `block` | Granularity of CUDA graphs during **inference**; only `local` supports non-`none` values |
+
+Supported combinations:
+
+| `--cuda-graph-impl` | Backend | Training capture | Inference capture |
+|---|---|---|---|
+| `none` | — | off | off |
+| `local` | MCore `CudaGraphManager` | per-layer, controlled by `--cuda-graph-modules` | `layer` (default) or `block`, controlled by `--inference-cuda-graph-scope` |
+| `transformer_engine` | TE `make_graphed_callables()` | per-layer, controlled by `--cuda-graph-modules` | not supported (`none` only) |
+| `full_iteration` | MCore `FullCudaGraphWrapper` | one graph per training iteration; `--cuda-graph-modules` must be empty | not supported (`none` only) |
+
+---
+
+## CUDA Graph — Local Implementation (`--cuda-graph-impl local`)
+
+Uses MCore's built-in `CudaGraphManager`. During training, this is a per-layer mode:
+leaving `--cuda-graph-modules` unset captures the whole Transformer layer, while specifying
+modules restricts capture to selected sub-regions. During inference, `local` can instead attach
+graphs at either the layer boundary or the enclosing block boundary, as controlled by
+`--inference-cuda-graph-scope`.
+
+Operationally, this path is tightly integrated into MCore training and inference:
+
+- graphable modules create and own their `CudaGraphManager` instances automatically
+- the existing training schedules drive warmup/capture/replay automatically
+- users select the mode through config flags only; there is no separate helper API to
+  wire into a custom training loop or a separate need to handle static input buffers
+
+### Usage
+
+```bash
+--cuda-graph-impl local
+```
+
+### `--cuda-graph-modules` options
+
+| Module | What is captured |
+|---|---|
+| *(empty / not set)* | Entire Transformer layer (default) |
+| `attn` | `TransformerLayer._forward_attention()` |
+| `mlp` | `TransformerLayer._forward_mlp()` for dense layers |
+| `moe` | `TransformerLayer._forward_mlp()` for MoE layers (drop-and-pad only) |
+| `moe_router` | MoE router + shared experts (if not EP-comm-overlapped) |
+| `moe_preprocess` | `MoELayer.preprocess()` — must be paired with `moe_router` |
+| `mamba` | Mamba SSM layer |
+
+**Example — MoE model, capture attention and router:**
+```bash
+--cuda-graph-impl local \
+# Optionally restrict captured modules (default: capture whole layer, but not working for MoE dynamic shapes)
+--cuda-graph-modules attn moe_router moe_preprocess
+```
+
+---
+
+## CUDA Graph — Transformer Engine Implementation (`--cuda-graph-impl transformer_engine`)
+
+Uses Transformer Engine's `make_graphed_callables()` path. In Megatron-LM's CLI, this has the
+same training granularity as `local`: leaving `--cuda-graph-modules` unset captures the whole
+Transformer layer, while specifying modules restricts capture to selected sub-regions. The main difference from
+`local` is the backend implementation and feature compatibility. Unlike `local`, this path does
+not support inference CUDA graphs.
+
+Compared to `local`, this path exposes a more general and self-contained API via TE's
+`make_graphed_callables()`, giving users greater flexibility and control over how CUDA graphs are
+wired into custom training loops. The trade-off is that it requires more manual setup:
+
+- the training loop must instantiate `TECudaGraphHelper`
+- the training loop must call helper methods such as `create_cudagraphs()` and
+  `cuda_graph_set_manual_hooks()` at the correct points
+
+Megatron-LM's stock training loop already wires these calls in `megatron/training/training.py`,
+but custom training scripts must do the same work themselves.
+
+### Usage
+
+```bash
+--cuda-graph-impl transformer_engine \
+--cuda-graph-modules attn moe_router moe_preprocess
+```
+
+The same training `--cuda-graph-modules` options apply as for `local`, and the default is likewise
+whole-layer training capture when the flag is omitted.
+
+---
+
+## Full-Iteration Training CUDA Graph (`--cuda-graph-impl full_iteration`)
+
+Captures the entire training iteration (excluding optimizer) as a single CUDA graph. The same
+wrapper is also used for training-loop validation/eval in forward-only mode. This provides the
+largest training/validation latency reduction.
+
+This implementation does not create inference CUDA graphs. For inference, use
+`--cuda-graph-impl local --inference-cuda-graph-scope layer|block`.
+
+### Requirements
+
+- `--no-check-for-nan-in-loss-and-grad` is required: NaN checks involve CPU-GPU synchronization
+  which cannot run inside a CUDA graph.
+- `--cuda-graph-modules` must be omitted (or left empty): per-module selection has no meaning
+  when the entire iteration is captured as a single graph.
+
+### Example
+
+```bash
+--cuda-graph-impl full_iteration \
+--no-check-for-nan-in-loss-and-grad
+```
+
+---
+
+## Common Configuration Examples
+
+### Dense Model Training
+
+All three implementations work for dense models:
+
+```bash
+# Per-layer (local)
+--cuda-graph-impl local
+# equivalent: --cuda-graph-impl local --cuda-graph-modules attn mlp
+
+# Per-layer (TE)
+--cuda-graph-impl transformer_engine
+# equivalent: --cuda-graph-impl transformer_engine --cuda-graph-modules attn mlp
+
+# Full-iteration
+--cuda-graph-impl full_iteration \
+--no-check-for-nan-in-loss-and-grad
+```
+
+### MoE Model Training
+
+MoE expert dispatch involves dynamic shapes and cannot be captured. `--cuda-graph-modules` is used
+to capture only the static parts (attention, router, preprocess) while leaving expert compute in
+eager mode. Example using `transformer_engine` (`local` works the same way):
+
+```bash
+--cuda-graph-impl transformer_engine \
+--cuda-graph-modules attn moe_router moe_preprocess
+```
+
+With [paged stash](paged_stash.md), expert dispatch shapes become static (pre-sized via
+`--moe-expert-rank-capacity-factor`), which allows full-iteration CUDA graphs to be used on MoE
+models as well:
+
+```bash
+--cuda-graph-impl full_iteration \
+--no-check-for-nan-in-loss-and-grad \
+--moe-flex-dispatcher-backend hybridep \
+--use-transformer-engine-op-fuser \
+--moe-expert-rank-capacity-factor <float> \
+--moe-paged-stash
+```
+
+---
+
+## Additional Notes
+
+- `--cuda-graph-warmup-steps` (default: 3) controls how many warmup steps run before CUDA graph
+  capture. Setting it to 0 is not recommended: some operations rely on the first few iterations
+  for lazy initialization or autotuning, and capturing too early may produce incorrect or
+  suboptimal graphs.
+- Inference CUDA graphs (serving or RL rollout) currently require
+  `--cuda-graph-impl local`. Use `--inference-cuda-graph-scope layer|block` with
+  `local`; all other implementations must set `--inference-cuda-graph-scope none`,
+  meaning inference runs in eager mode.
+- Background reference: [Transformer Engine and Megatron-LM CUDA Graph Support](https://docs.nvidia.com/dl-cuda-graph/torch-cuda-graph/te-megatron-cuda-graphs.html),
+  which also covers PyTorch CUDA Graph best practices and lessons learned.
+
+---
+
+## Migration Guide
+
+Legacy configurations (including `--enable-cuda-graph`, `--external-cuda-graph`, the renamed
+`--cuda-graph-scope` flag (now `--cuda-graph-modules`), and deprecated module values such as
+`full_iteration` and `full_iteration_inference`) are still accepted and automatically migrated
+at runtime, but we encourage updating your configs to the new forms:
+
+| Old command | New command |
+|---|---|
+| `--enable-cuda-graph` | `--cuda-graph-impl local` |
+| `--external-cuda-graph` | `--cuda-graph-impl transformer_engine` |
+| `--cuda-graph-scope <modules>` | `--cuda-graph-modules <modules>` |
+| `--cuda-graph-impl local --cuda-graph-scope full_iteration` | `--cuda-graph-impl full_iteration` |
+| `--cuda-graph-impl local --cuda-graph-scope full_iteration_inference` | `--cuda-graph-impl local --inference-cuda-graph-scope block` |
+| `--cuda-graph-impl local --cuda-graph-scope attn moe_router moe_preprocess full_iteration_inference` | `--cuda-graph-impl local --cuda-graph-modules attn moe_router moe_preprocess --inference-cuda-graph-scope block` |

docs/user-guide/features/fine_grained_activation_offloading.md

@@ -42,13 +42,13 @@ Supported offloading modules are `"attn_norm"`, `"core_attn"`, `"attn_proj"`, `"
 ```bash
 # Optional: cap inflight D2H offloads per offload group to N (omit or None in most setups).
 # Required as a non-None non-negative integer when fine-grained activation offloading is used with
-# local full-iteration CUDA graphs (full_iteration in cuda_graph_scope); see prose below.
+# full-iteration CUDA graphs (--cuda-graph-impl full_iteration); see prose below.
 --fine_grained_offloading_max_inflight_offloads <N>
 ```
 
 TransformerConfig.fine_grained_offloading_max_inflight_offloads caps, per offload group (for example `moe_act`, `qkv_linear`), how many D2H copies may be in flight before a main-stream wait_event. 0 waits after each offload; larger values allow more overlap; None skips these joins.
 
-With full-iteration CUDA graphs (local graph impl, full_iteration in cuda_graph_scope) and fine-grained activation offloading enabled, set it to a non-None integer: that path does not rely on record_stream, so explicit joins are required.
+With full-iteration CUDA graphs (`--cuda-graph-impl full_iteration`) and fine-grained activation offloading enabled, set it to a non-None integer: that path does not rely on record_stream, so explicit joins are required.
 
 ## Compatible With Fine-Grained Recomputation
 

docs/user-guide/features/index.md

@@ -14,6 +14,7 @@ Guides for Megatron Core training features.
 ```{toctree}
 :maxdepth: 2
 
+cuda_graph
 fine_grained_activation_offloading
 moe
 context_parallel

examples/rl/model_configs/nemotron6_3b_moe.sh

@@ -65,7 +65,6 @@ MODEL_OPTIONS="\
   --inference-dynamic-batching-num-cuda-graphs 2 \
   --decode-only-cuda-graphs \
   --cuda-graph-impl local \
-  --cuda-graph-scope full \
   --use-checkpoint-args \
   --enable-experimental \
   --cross-entropy-loss-fusion \

megatron/core/inference/engines/dynamic_engine.py

@@ -45,7 +45,7 @@
 from megatron.core.inference.utils import Counter, await_process_call
 from megatron.core.process_groups_config import ProcessGroupCollection
 from megatron.core.transformer.cuda_graphs import delete_cuda_graphs
-from megatron.core.transformer.enums import CudaGraphScope
+from megatron.core.transformer.enums import InferenceCudaGraphScope
 from megatron.core.transformer.moe.router_replay import RouterReplay, RouterReplayAction
 from megatron.core.utils import (
     deprecate_args,
@@ -224,7 +224,8 @@ def __init__(self, controller: TextGenerationController, context: DynamicInferen
         self.disable_ep_consensus = inference_config.disable_ep_consensus
         self.ep_consensus_interval = inference_config.ep_consensus_interval
         self.cuda_graph_impl = model_config.cuda_graph_impl
-        self.cuda_graph_scope = model_config.cuda_graph_scope
+        self.inference_cuda_graph_scope = model_config.inference_cuda_graph_scope
+        self.cuda_graph_modules = model_config.cuda_graph_modules
         # Initialize engine.
         self.reset()
 
@@ -331,17 +332,11 @@ def create_cuda_graphs(self, reset_context: bool = True):
             reset_context (bool): Whether to reset the context after building cuda graphs.
         """
 
-        if self.cuda_graph_impl != "local":
+        if self.inference_cuda_graph_scope == InferenceCudaGraphScope.none:
             return
 
-        if (
-            CudaGraphScope.full_iteration in self.cuda_graph_scope
-            and CudaGraphScope.full_iteration_inference not in self.cuda_graph_scope
-        ):
-            warnings.warn(
-                "\n\n*** WARNING: 'full_iteration' CUDA graph scope used during inference! "
-                "This will not create inference CUDA graphs. Use '--cuda-graph-scope=full_iteration_inference' instead. ***\n"
-            )
+        if self.cuda_graph_impl != "local":
+            return
 
         context = self.context
         controller = self.controller

megatron/core/inference/text_generation_controllers/text_generation_controller.py

@@ -36,7 +36,6 @@
     gather_from_sequence_parallel_region,
     scatter_to_sequence_parallel_region,
 )
-from megatron.core.transformer.enums import CudaGraphScope
 from megatron.core.transformer.moe.moe_layer import BaseMoELayer
 from megatron.core.transformer.moe.router_replay import RouterReplay, RouterReplayAction
 from megatron.core.transformer.utils import set_model_to_sequence_parallel
@@ -1924,10 +1923,7 @@ def generate_all_output_tokens_static_batch(
         )
 
         # Check whether CUDA graphs are enabled
-        enable_cuda_graph = (
-            model_config.cuda_graph_impl == "local"
-            and CudaGraphScope.full_iteration not in model_config.cuda_graph_scope
-        )
+        enable_cuda_graph = model_config.cuda_graph_impl == "local"
 
         # Pad batch tokens if necessary
         batch_size = len(active_requests)

megatron/core/models/common/language_module/language_module.py

@@ -22,7 +22,7 @@
     is_vp_last_stage,
 )
 from megatron.core.process_groups_config import ProcessGroupCollection
-from megatron.core.transformer.enums import AttnBackend, CudaGraphScope
+from megatron.core.transformer.enums import AttnBackend
 from megatron.core.transformer.module import MegatronModule
 from megatron.core.transformer.multi_token_prediction import tie_word_embeddings_state_dict
 from megatron.core.transformer.transformer_config import TransformerConfig
@@ -159,8 +159,8 @@ def compute_language_model_loss(self, labels: Tensor, logits: Tensor) -> Tensor:
                     labels = torch.as_strided(labels, labels.size(), (labels.size()[1], 1))
                     # Use is_cg_capturable=True for full iteration CUDA graphs to avoid torch.equal checks
                     is_cg_capturable = (
-                        hasattr(self.config, 'cuda_graph_scope')
-                        and CudaGraphScope.full_iteration in self.config.cuda_graph_scope
+                        hasattr(self.config, 'cuda_graph_impl')
+                        and self.config.cuda_graph_impl == "full_iteration"
                     )
                     if is_cg_capturable and not is_te_min_version("2.7.0"):
                         from megatron.core.utils import get_te_version
@@ -169,7 +169,7 @@ def compute_language_model_loss(self, labels: Tensor, logits: Tensor) -> Tensor:
                         raise AssertionError(
                             f"CUDA graph compatible cross entropy requires TransformerEngine >= 2.7.0, "
                             f"but found version {current_version}. Please upgrade TransformerEngine "
-                            f"or set cuda_graph_scope to a value other than 'full_iteration'."
+                            f"or set cuda_graph_impl to a value other than 'full_iteration'."
                         )
 
                     loss = te_parallel_cross_entropy(

megatron/core/models/common/model_chunk_schedule_plan.py

@@ -14,7 +14,7 @@
     get_comm_stream,
     get_comp_stream,
 )
-from megatron.core.transformer.enums import CudaGraphScope
+from megatron.core.transformer.enums import CudaGraphModule
 from megatron.core.utils import nvtx_range_pop, nvtx_range_push
 
 
@@ -175,7 +175,7 @@ def create_node(stream, module, name):
             self.mtp_post_process = NoopScheduleNode()
 
         # mlp and combine may receive dgrad from attn, which is managed by cuda graph.
-        if CudaGraphScope.attn in self.config.cuda_graph_scope:
+        if CudaGraphModule.attn in self.config.cuda_graph_modules:
             self.mlp.manual_grads_release = False
             self.moe_combine.manual_grads_release = False