perf-torch-cuda-graphs

CUDA Graphs for PyTorch

CUDA Graphs capture a sequence of GPU operations once and replay them with minimal CPU overhead. This skill guides applying CUDA Graphs to PyTorch training and inference workloads using native PyTorch APIs, Transformer Engine, and Megatron-LM.

When to Use

Reach for this skill when you encounter:

• Triggers: User wants to optimize with CUDA Graphs, reduce kernel launch overhead, or speed up training/inference loops
• Symptoms: Low GPU utilization (<80%), many small kernel launches (<50 us each), CPU-bound training, high kernel launch latency visible in Nsight Systems profiles
• Keywords: "CUDA graph", "torch.cuda.graph", "make_graphed_callables", "reduce-overhead", "graph capture", "graph replay", "kernel launch overhead", "CudaGraphManager", "FullCudaGraphWrapper", "full-iteration graph", "stream capture"

Do NOT use this skill for:

• General PyTorch performance tuning unrelated to kernel launch overhead
• CUDA kernel development or custom CUDA C++ code
• Host-device sync elimination only (use perf-torch-sync-free skill instead)
• Nsight Systems profiling (use perf-nsight-systems skill)
• TensorFlow/JAX graph compilation (different APIs entirely)

Requirements

Dependency	Version	Notes
PyTorch	>= 1.10	torch.cuda.graph() available
CUDA	>= 11.0	Graph update APIs
GPU	NVIDIA (any)	Required for CUDA
Nsight Systems	any	Optional, for profiling
APEX	any	Optional, for capturable optimizers
Transformer Engine	>= 2.2	Optional, for FP8-aware graphing
Megatron-LM	core >= 0.14.0	Optional, for CudaGraphManager / FullCudaGraphWrapper

API Selection Guide

Choose the API based on your framework and performance needs.

Situation	API	Workflow
Quick experiment, unknown graph boundaries	torch.compile(mode="reduce-overhead")	Workflow 2
Training, need autograd, no FP8/PP	torch.cuda.make_graphed_callables()	Workflow 3
Any PyTorch model, FP8 or PP support	TE make_graphed_callables	Workflow 4
Megatron-LM, per-layer, automatic	MCore CudaGraphManager	Workflow 5
Maximum perf, full-iteration capture	MCore FullCudaGraphWrapper	Workflow 6
Full manual control, custom pipelines	torch.cuda.graph()	Workflow 7

Decision flowchart:

1. Using Megatron-LM with FP8/PP?
   • Yes, want maximum perf with static workload --> Workflow 6 (FullCudaGraphWrapper)
   • Yes, want per-layer automatic graphing --> Workflow 5 (CudaGraphManager)
   • Yes, want manual control over what gets graphed --> Workflow 4 (TE make_graphed_callables)
2. Using Transformer Engine without Megatron?
   • Yes, need FP8 or PP --> Workflow 4 (TE make_graphed_callables)
3. General PyTorch?
   • Want zero effort, okay with fragmented graphs --> Workflow 2 (torch.compile)
   • Want autograd support, training loop --> Workflow 3 (PyTorch make_graphed_callables)
   • Want full manual control --> Workflow 7 (torch.cuda.graph)

Strategy: Start with the highest-level API available for your framework. Move to lower-level APIs only if you need more control, hit limitations, or do not achieve the expected performance improvement.

Workflows

Workflow 1: Profile and Decide Whether Graphs Help

Goal: Determine if CUDA Graphs will benefit your workload before investing effort.

1. Profile with Nsight Systems:

   nsys profile --cuda-graph-trace=graph python train.py
   

2. Check GPU utilization -- if already >95%, graphs won't help much.
3. Look for gaps between kernel launches (CPU overhead) and many small kernels (<50 us each). These are the targets for graphing.
4. Annotate regions of interest to correlate idle GPU time with code:

   with torch.cuda.nvtx.range("forward"):
       output = model(input)
   

5. Estimate benefit: count kernels per iteration. Workloads with hundreds of small kernels and <80% GPU utilization are strong candidates.

Expected result: Identified bottleneck regions with low GPU occupancy between kernels. Proceed to the appropriate workflow from the API Selection Guide.

Workflow 2: torch.compile(mode="reduce-overhead")

Goal: Automatic CUDA Graph capture with zero manual effort.

When to use: Quick experiment, unknown graph boundaries, already using torch.compile.

Steps:

1. Decorate the training step with @torch.compile(mode="reduce-overhead"):

   @torch.compile(mode="reduce-overhead")
   def train_step(model, x, target, criterion):
       output = model(x)
       loss = criterion(output, target)
       loss.backward()
       return loss
   

2. Run the training loop normally -- graphs are captured automatically.
3. Profile with Nsight Systems to see captured graphs:

   nsys profile --cuda-graph-trace=graph python train.py
   

4. If you see too many small graphs (graph fragmentation), check for graph breaks: .item(), print(), data-dependent control flow. Fix these or escalate to Workflow 3+.

Trade-offs:

• Zero effort, but may create fragmented small graphs.
• Limited control over what gets graphed.
• Graph fragmentation limits performance gains compared to manual approaches.

Workflow 3: torch.cuda.make_graphed_callables()

Goal: Training with autograd support. Separate forward/backward graphs.

When to use: Training with custom loops, non-FP8, need autograd.

Steps:

1. Prepare sample inputs matching training batch shape:

   sample_input = torch.randn(batch_size, seq_len, hidden_size, device="cuda")
   

2. Create the graphed model:

   graphed_model = torch.cuda.make_graphed_callables(
       model, (sample_input,), num_warmup_iters=3
   )
   

3. Use graphed_model as a drop-in replacement in the training loop:

   for data, target in dataloader:
       optimizer.zero_grad()
       output = graphed_model(data)
       loss = criterion(output, target)
       loss.backward()
       optimizer.step()
   

4. If using AMP, set cache_enabled=False:

   for data, target in dataloader:
       optimizer.zero_grad()
       with torch.amp.autocast("cuda", cache_enabled=False):
           output = graphed_model(data)
           loss = criterion(output, target)
       loss.backward()
       optimizer.step()
   

5. If using DDP, construct DDP on a side stream and use 11 warmup iters:

   os.environ["TORCH_NCCL_ASYNC_ERROR_HANDLING"] = "0"
   s = torch.cuda.Stream()
   with torch.cuda.stream(s):
       model = DistributedDataParallel(model)
   torch.cuda.current_stream().wait_stream(s)
   
   graphed_model = torch.cuda.make_graphed_callables(
       model, (sample_input,), num_warmup_iters=11
   )
   

Limitations:

• No double backward (higher-order gradients).
• No module hooks during capture.
• Module structure is frozen after graphing (no add/remove parameters).
• Argument signature must match sample_args exactly.

Workflow 4: TE make_graphed_callables

Goal: Per-callable graphing with FP8 support and pipeline parallelism.

When to use: FP8 training, PP with manual scheduling, non-Megatron models needing FP8, or any PyTorch model that needs FP8-aware CUDA Graphs.

Steps:

1. Import and configure:

   from transformer_engine.pytorch.graph import make_graphed_callables
   from transformer_engine.pytorch.fp8 import fp8_autocast
   

2. Prepare sample inputs (one per callable per microbatch per chunk):

   sample_args = tuple(
       (torch.randn(batch_size, seq_len, hidden_size, device="cuda"),)
       for _ in range(num_callables * num_microbatches)
   )
   

3. Define pipeline schedule if using PP (1-indexed chunk IDs, positive=fwd, negative=bwd):

   # Example: 2 chunks, 3 microbatches
   layer_order = [1, 2, 1, 2, 1, 2, -2, -1, -2, -1, -2, -1]
   

4. Wrap layers in CUDA Graphs:

   graphed_layers = make_graphed_callables(
       tuple(layers),
       sample_args=sample_args,
       fp8_enabled=True,
       fp8_recipe=fp8_recipe,
       fp8_weight_caching=True,
       _order=layer_order,  # None for no PP
   )
   

5. Training loop -- wrap with fp8_autocast during replay:

   with fp8_autocast(enabled=True, fp8_recipe=fp8_recipe):
       for layer in graphed_layers[start:end]:
           x = layer(x, is_first_microbatch=(mb_idx == 0))
   # FP8 scaling auto-updated on fp8_autocast exit
   optimizer.step()
   

Key points:

• AOT capture: Graphs captured before the training loop when you call make_graphed_callables().
• Replay order must match _order: The training loop must execute graphs in the same interleaved order as specified during capture.
• fp8_autocast required during replay: Without it, FP8 state is not properly configured.
• Weight caching: fp8_weight_caching=True caches FP8 weight quantization across microbatches; pass is_first_microbatch kwarg to control when weights are requantized.

For full API details, see references/api-te-megatron.md.

Workflow 5: MCore CudaGraphManager (Per-Layer)

Goal: Automatic per-layer graphing for Megatron-LM training.

When to use: Megatron-LM training, especially with PP > 1. Default choice for Megatron users.

Steps:

1. Enable via CLI flags (no code changes needed):

   python pretrain_gpt.py \
       --enable-cuda-graph \
       --cuda-graph-num-warmup-steps 3
   

2. Or enable via Python config:

   config = TransformerConfig(
       enable_cuda_graph=True,
       cuda_graph_num_warmup_steps=3,
   )
   

3. Training loop is unchanged -- graphs are captured automatically after warmup iterations.

Key points:

• Megatron layers only: Works with TransformerLayer and MambaLayer.
• JIT capture: Records execution order during warmup, captures graphs after warmup completes, then replays on subsequent iterations.
• Automatic FP8 handling: Uses fp8_autocast(..., _graph=True) to skip per-layer amax reduction; reduction happens once after all backward graphs.
• Automatic PP support: Handles microbatch interleaving automatically.
• Memory savings: Set cuda_graph_share_io_buffers=True to share I/O buffers between layers (requires no operations between layers).
• Memory pool strategy: Default uses separate pools per microbatch for graph reuse. Set cuda_graph_use_single_mempool=True for shared pool (higher graph count but may reduce fragmentation).

Workflow 6: MCore FullCudaGraphWrapper (Full-Iteration)

Goal: Maximum performance. Captures forward+backward for all microbatches as a single graph.

When to use: Maximum performance priority, static workloads, Megatron-LM training.

Steps:

1. Enable via CLI flags:

   python pretrain_gpt.py \
       --enable-cuda-graph \
       --cuda-graph-scope full_iteration \
       --cuda-graph-warmup-steps 1 \
       --te-rng-tracker \
       --no-check-for-nan-in-loss-and-grad
   

2. Ensure all forward+backward code is capturable (no .item(), no NaN check, no dynamic control flow).
3. Optimizer remains in eager mode by default (outside the graph). Can be included inside the graph for maximum performance.

Key points:

• Only 2 graphs total: One for training, one for validation.
• --te-rng-tracker required: Standard RNG uses CPU scalars that cannot be captured; TE RNG uses device tensors compatible with graphs.
• --no-check-for-nan-in-loss-and-grad mandatory: NaN checking uses .item() which requires CPU-GPU sync, forbidden during capture.
• StaticBufferLoader: Pre-allocates input buffers for all microbatches during warmup.
• Optimizer in/out of graph: Inside = maximum performance (all optimizer kernels captured). Outside = more flexible (can change optimizer/LR without recapture).
• JIT capture: Graph captured during training at iteration warmup_steps + 1.

Workflow 7: torch.cuda.graph() (Manual)

Goal: Full control over capture and replay. Custom pipelines, full-iteration capture without Megatron.

When to use: Need fine-grained control, non-Megatron full-iteration capture, custom pipelines.

Inference pattern:

1. Pre-allocate static input/output tensors:

   static_input = torch.randn(batch_size, *shape, device="cuda")
   

2. Warmup on a side stream (3 iterations, 11 for DDP):

   s = torch.cuda.Stream()
   with torch.cuda.stream(s):
       for _ in range(3):
           _ = model(static_input)
   torch.cuda.current_stream().wait_stream(s)
   

3. Capture the graph:

   g = torch.cuda.CUDAGraph()
   with torch.cuda.graph(g):
       static_output = model(static_input)
   

4. Replay loop -- update inputs via .copy_(), clone outputs:

   for data in loader:
       static_input.copy_(data)
       g.replay()
       result = static_output.clone()
   

Full training pattern (fwd+bwd+optimizer in one graph):

model = MyModel().cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
criterion = torch.nn.CrossEntropyLoss()

static_input = torch.randn(batch_size, *shape, device="cuda")
static_target = torch.randint(0, num_classes, (batch_size,), device="cuda")

# Warmup
s = torch.cuda.Stream()
with torch.cuda.stream(s):
    for _ in range(3):
        optimizer.zero_grad()
        with torch.amp.autocast("cuda", cache_enabled=False):
            out = model(static_input)
            loss = criterion(out, static_target)
        loss.backward()
torch.cuda.current_stream().wait_stream(s)

# Capture
g = torch.cuda.CUDAGraph()
with torch.cuda.graph(g):
    optimizer.zero_grad()
    with torch.amp.autocast("cuda", cache_enabled=False):
        static_output = model(static_input)
        static_loss = criterion(static_output, static_target)
    static_loss.backward()

# Replay loop
for data, target in loader:
    static_input.copy_(data)
    static_target.copy_(target)
    g.replay()
    optimizer.step()

DDP setup:

os.environ["TORCH_NCCL_ASYNC_ERROR_HANDLING"] = "0"

s = torch.cuda.Stream()
with torch.cuda.stream(s):
    model = DistributedDataParallel(model)

# 11 warmup iterations for DDP
with torch.cuda.stream(s):
    for _ in range(11):
        out = model(static_input)
        out.sum().backward()
torch.cuda.current_stream().wait_stream(s)

# Capture on the same side stream
with torch.cuda.graph(g):
    static_output = model(static_input)

Memory pool sharing for multiple graphs:

g1 = torch.cuda.CUDAGraph()
with torch.cuda.graph(g1):
    out1 = model_a(static_in_a)

# Second graph shares first graph's memory pool
g2 = torch.cuda.CUDAGraph()
with torch.cuda.graph(g2, pool=g1.pool()):
    out2 = model_b(static_in_b)

Custom RNG registration:

gen = torch.cuda.default_generators[0]
g = torch.cuda.CUDAGraph()
g.register_generator_state(gen)
with torch.cuda.graph(g):
    out = model(static_input)  # RNG state properly captured

Navigating Between Workflows

• torch.compile gives insufficient speedup --> escalate to make_graphed_callables (Workflow 3) for larger, fewer graphs.
• make_graphed_callables can't handle FP8/PP --> TE make_graphed_callables (Workflow 4).
• Need Megatron per-layer automatic --> CudaGraphManager (Workflow 5).
• Want maximum perf --> FullCudaGraphWrapper (Workflow 6) or manual full-iteration capture (Workflow 7).
• Something too hard to graph --> partial capture (graph what you can, leave the rest in eager mode).
• User wants best absolute perf --> skip directly to Workflow 6 (Megatron) or Workflow 7 (manual).
• Start small, expand progressively: Begin with one module/layer. Verify correctness. Then expand to more layers, full forward pass, add backward, and eventually full iteration with optimizer.

Making Code Graph-Compatible

These principles apply to all workflows. Code inside the captured region must satisfy three constraints.

Principle 1: GPU-Only

Only GPU operations are captured. CPU-side code (Python logic, I/O, logging) executes during capture but is eliminated during replay.

Violations:

• File I/O: data = torch.load("file.pt") won't reload on replay
• CPU preprocessing: tokens = tokenizer.encode(text) won't re-tokenize
• Logging: print(f"Step {i}") won't print during replay
• CPU RNG: random.randint(0, 10) won't regenerate
• CPU bookkeeping: buffer.append(tensor) won't populate during replay

Fix: Move all CPU-side operations outside the graphed region.

Principle 2: Sync-Free

No CPU-GPU synchronization inside the graph. The CPU queues work continuously without waiting for GPU results.

Violations:

• .item() to get scalar values
• .cpu() to move tensors for inspection
• torch.cuda.synchronize() or stream.synchronize()
• print(tensor) (implicitly syncs)

Fix: Invoke the perf-torch-sync-free skill for systematic detection and elimination of sync points. Use torch.cuda.set_sync_debug_mode("warn") to find hidden syncs.

Principle 3: Static

All operations, control flow, memory addresses, and shapes must be fixed across all replays.

Violations and fixes:

Dynamic aspect	Fix
if loss > threshold:	torch.where(condition, a, b)
input = new_tensor (address changes)	Pre-allocate + .copy_()
Python scalars (lr, temperature)	GPU tensor + .fill_()
Variable batch size / sequence length	Padding or bucketing
MoE / dynamic routing	Partial graphing

For detailed patterns, see references/patterns-dynamic.md.

Compatibility Checklist

Verify every item before attempting capture:

• No .item(), .cpu(), .numpy(), print(tensor) inside graph
• No torch.cuda.synchronize() or stream.synchronize()
• No if tensor_value: -- use torch.where() instead
• All inputs pre-allocated, updated via .copy_()
• All shapes fixed (use padding or bucketing for variable sizes)
• Python scalars --> GPU tensors with .fill_()
• Output tensors .clone()d before next replay
• cache_enabled=False with torch.amp.autocast
• Custom RNG generators registered with graph.register_generator_state()
• Use graphsafe_get_state() / graphsafe_set_state() for RNG
• Warmup completed (3 standard, 11 for DDP)
• DDP: TORCH_NCCL_ASYNC_ERROR_HANDLING=0, construct on side stream
• DDP: NCCL >= 2.9.6 for full graph capture
• Libraries/extensions use torch.cuda.current_stream(), not default stream
• No pinned memory allocation during capture (triggers hidden event query)
• activation_checkpointing: preserve_rng_state=False
• Global tensors used in graph kept alive (not deleted/reassigned)
• No torch.compile functions inside manual capture without prior warmup
• Gradient clipping uses sync-free clip_grad_norm_ (PyTorch >= 1.13)

For the complete checklist with references, see references/patterns-compatibility.md.

Output Formats

Success indicators:

• g.replay() completes without errors
• Outputs match eager mode within tolerance (torch.allclose)
• Nsight Systems profile shows single graph launch replacing many kernels
• GPU utilization increases, training/inference latency decreases

Key metrics:

Metric	How to Check
Correctness	torch.allclose(eager, graphed, rtol=1e-5)
Speedup	Wall-clock time comparison
GPU utilization	nvidia-smi or Nsight Systems timeline
Memory overhead	torch.cuda.memory_summary()

Error Handling

Error	Cause	Fix
StreamCaptureUnsupported (900)	Sync op during capture (.item(), .cpu())	Move sync outside graph
StreamCaptureInvalidated (901)	Background thread (e.g., pin_memory)	capture_error_mode="thread_local"
StreamCaptureUnjoined (904)	Side stream didn't rejoin capture stream	capture_stream.wait_stream(side_stream)
StreamCaptureImplicit (906)	AccumulateGrad on default stream	Warmup on side stream before capture
Illegal memory access	Input tensor freed/reassigned	Keep persistent ref, use .copy_()
Wrong numerical results	Dynamic behavior frozen at capture	See references/patterns-compatibility.md
OOM with multiple graphs	Pools can't share memory	pool=g1.pool() for sequential graphs
No speedup	Already GPU-bound or wrong capture scope	Profile with nsys first (Workflow 1)
FP8 scaling corruption	TE without fp8_autocast during replay	Wrap with fp8_autocast(enabled=True)
PP replay order mismatch	Wrong execution order during replay	Match _order / capture sequence exactly
FullCudaGraphWrapper capture fail	NaN check or sync enabled	--no-check-for-nan-in-loss-and-grad
RNG failure with FullCudaGraphWrapper	Standard RNG not capturable	--te-rng-tracker
DDP capture failure	Async error handling watchdog	TORCH_NCCL_ASYNC_ERROR_HANDLING=0
DDP AccumulateGrad on default stream	DDP constructed on default stream	Construct DDP in side stream context
Autocast cache invalidation	Cached cast tensors freed on exit	cache_enabled=False

For detailed troubleshooting, see references/troubleshooting.md.

Finding More Information

Use this 3-tier lookup hierarchy -- start at Tier 1 and escalate only when needed.

Tier 1: This File (SKILL.md)

You are reading it now. The workflows, compatibility checklist, and error table above cover the most common tasks. Search this file first before going deeper.

Tier 2: references/ Directory

The references/ directory beside this file contains distilled reference material -- API details, patterns, and troubleshooting pages.

How to search:

1. Grep for your keyword across references/ -- headers are designed to be grep-friendly.
2. Read only the file that grep points you to. Do not read every file.

Available references:

• references/api-pytorch.md -- PyTorch CUDA Graph APIs (torch.cuda.graph, make_graphed_callables, torch.compile reduce-overhead)
• references/api-te-megatron.md -- TE make_graphed_callables, CudaGraphManager, FullCudaGraphWrapper implementations
• references/patterns-compatibility.md -- GPU-only, sync-free, and static principles with full checklist
• references/patterns-dynamic.md -- Dynamic control flow, tensors, scalars, shapes: workarounds and patterns
• references/troubleshooting.md -- Capture failures, numerical errors, memory issues, performance issues

Tier 3: Original Documentation

If Tiers 1-2 do not answer the question, consult the original sources:

• NVIDIA guide: https://docs.nvidia.com/dl-cuda-graph/latest/index.html
• PyTorch docs: https://docs.pytorch.org/docs/stable/notes/cuda.html (CUDA Graphs section)
• TE docs: https://docs.nvidia.com/deeplearning/transformer-engine/user-guide/index.html
• Megatron Core docs: https://docs.nvidia.com/megatron-core/developer-guide/latest/index.html

Return to Tier 2 afterward and consider whether the answer should be distilled into the references directory for next time.