ATOM Compilation & CUDA Graphs Guide

Quick Reference

Concept

Key Class / Enum

Import

Compilation Levels

CompilationLevel

from atom.config import CompilationLevel

Compilation Config

CompilationConfig

from atom.config import CompilationConfig

CUDA Graph Modes

CUDAGraphMode

from atom.config import CUDAGraphMode

CUDA Graph Wrapper

CUDAGraphWrapper

from atom.utils.cuda_graph import CUDAGraphWrapper

Forward Context

ForwardContext

from atom.utils.forward_context import ForwardContext

Compiler Backend

VllmBackend

from atom.utils.backends import VllmBackend

Compiler Manager

CompilerManager

from atom.utils.backends import CompilerManager

Compiler Interface

CompilerInterface

from atom.utils.compiler_inferface import CompilerInterface

Inductor Adaptor

InductorAdaptor

from atom.utils.compiler_inferface import InductorAdaptor

Piecewise Backend

PiecewiseBackend

from atom.utils.cuda_piecewise_backend import PiecewiseBackend

Compile Decorator

@support_torch_compile

from atom.utils.decorators import support_torch_compile

Custom Op Registration

direct_register_custom_op

from atom.utils.custom_register import direct_register_custom_op

Compilation Levels at a Glance

Level

Name

Behavior

0

NO_COMPILATION

Pure eager execution, no torch.compile

1

DYNAMO_AS_IS

torch.compile with backend="eager"

2

DYNAMO_ONCE

torch.compile with Inductor

3

PIECEWISE

Piecewise compilation with CUDA graph capture (production default)

CUDA Graph Modes at a Glance

Mode

Value

Behavior

NONE

0

No graph capture

PIECEWISE

1

Per-subgraph capture (default for level 3)

FULL

2

Whole-model capture

FULL_DECODE_ONLY

(FULL, NONE)

Full for decode, none for mixed batches

FULL_AND_PIECEWISE

(FULL, PIECEWISE)

Full for decode, piecewise for prefill

1. Compilation Levels

ATOM provides four compilation levels via the CompilationLevel class in atom/config.py. The level is set through CompilationConfig.level and controls how torch.compile is applied to the model.

Level 0 – NO_COMPILATION

No torch.compile is applied. The model runs in pure eager mode. This is the simplest mode and is useful for debugging or when using models that are incompatible with torch.compile.

When level=0, the @support_torch_compile decorator sets self.do_not_compile = True and the model’s __call__ method bypasses compilation entirely, calling self.forward() directly.

Level 1 – DYNAMO_AS_IS

Uses torch.compile with backend="eager" and fullgraph=True. This runs Dynamo’s bytecode analysis and graph capture but does not apply any compiler optimizations. It is useful as a quick check to verify that a model is compatible with Dynamo’s tracing.

Like level 0, DYNAMO_AS_IS causes the decorator to set self.do_not_compile = True, since the model runner (rather than the decorator) handles the compilation at this level.

Level 2 – DYNAMO_ONCE

Uses torch.compile with the Inductor backend. The model graph is traced by Dynamo and compiled once through Inductor for optimized GPU kernel generation. The @support_torch_compile decorator’s custom dispatcher is activated when compilation_level >= DYNAMO_ONCE, allowing compiled bytecode to be dispatched directly after the first compilation without repeated guard evaluation.

Level 3 – PIECEWISE (Production Default)

The most advanced level. When Config.__post_init__ detects level == PIECEWISE, it:

Calls CompilationConfig.set_splitting_ops_for_v1() to configure the splitting operations (default: ["aiter.unified_attention_with_output", "aiter.mla_attention"]).
Calls Config._set_cudagraph_sizes() to compute the graph batch sizes.
Sets cudagraph_mode = CUDAGraphMode.PIECEWISE.
Calls CompilationConfig.init_with_cudagraph_sizes() to finalize compile sizes.

The VllmBackend is then used as the torch.compile backend. It splits the model graph into subgraphs at the splitting operations and compiles each subgraph independently via PiecewiseBackend.

2. CUDA Graph Modes

The CUDAGraphMode enum in atom/config.py controls how CUDA graphs are captured and replayed. CUDA graphs record a sequence of GPU operations and replay them with minimal CPU overhead, which is critical for low-latency decode steps.

NONE (value: 0)

No CUDA graph capture or replay. Every forward pass launches kernels individually. This mode is used during profiling, warmup, or when CUDA graphs are not supported.

PIECEWISE (value: 1)

The default mode for level 3 compilation. CUDA graphs are captured per subgraph (one for each piecewise-compiled region). Attention operations, which are split out by splitting_ops, run outside CUDA graphs because they may need dynamic metadata that changes between steps.

The CUDAGraphWrapper class wraps each subgraph with runtime_mode=CUDAGraphMode.PIECEWISE for capture and replay.

FULL (value: 2)

The entire model forward pass is captured as a single CUDA graph. This is suitable for small models or workloads with small, uniform batch sizes. Not all attention backends support full CUDA graph capture.

FULL_DECODE_ONLY (value: (FULL, NONE))

A tuple mode that applies different strategies to different batch types:

Decode batches: Captured with full CUDA graphs.
Mixed prefill-decode batches: Run without CUDA graphs.

This is useful for prefill/decode disaggregated (P/D) setups where decode latency matters more than prefill performance.

FULL_AND_PIECEWISE (value: (FULL, PIECEWISE))

A tuple mode combining both strategies:

Decode batches: Captured with full CUDA graphs.
Prefill and mixed batches: Captured with piecewise CUDA graphs.

This is described in the code as “the most performant mode for most models.”

Helper Methods

The CUDAGraphMode enum provides several helper methods for runtime dispatch:

Method	Returns	Purpose
`decode_mode()`	`CUDAGraphMode`	Returns the mode to use for decode batches. For tuple modes, returns the first element.
`mixed_mode()`	`CUDAGraphMode`	Returns the mode to use for mixed batches. For tuple modes, returns the second element.
`separate_routine()`	`bool`	Returns `True` if the mode is a tuple (different strategies for decode vs. mixed).
`has_full_cudagraphs()`	`bool`	Returns `True` if any part of the mode uses `FULL` capture.
`requires_piecewise_compilation()`	`bool`	Returns `True` if either decode or mixed mode uses `PIECEWISE`.
`max_cudagraph_mode()`	`CUDAGraphMode`	Returns the highest-valued mode across both decode and mixed modes.

3. CUDA Graph Capture

CUDA graph capture is handled by ModelRunner.capture_cudagraph() in atom/model_engine/model_runner.py. This method is called at startup (under @torch.inference_mode()) to pre-capture graphs for a set of batch sizes.

Capture Flow

capture_cudagraph()
  |
  +-- Determine graph_bs list
  |     |-- If cudagraph_capture_sizes is set: use directly
  |     |-- If cuda_graph_sizes has 1 value N: [1, 2, 4, 8, 16, 32, ..., N]
  |     +-- If cuda_graph_sizes has >1 values: use the provided list
  |
  +-- Sort graph_bs in descending order (largest batch first)
  |
  +-- Assert max batch size <= max_num_seqs
  |
  +-- Initialize graph storage: self.graphs = dict()
  |
  +-- For each batch size bs (with progress bar on rank 0):
  |     |
  |     +-- Compute max_q_len (= mtp_k + 1 if MTP drafter, else 1)
  |     +-- Compute num_tokens = bs * max_q_len
  |     +-- Prepare cu_seqlens_q, positions
  |     +-- Build attn_metadata and context via attn_metadata_builder
  |     +-- Handle DP padding via get_dp_padding()
  |     +-- Set forward context (set_forward_context)
  |     +-- Warmup run: model(input_ids[:num_tokens], positions[:num_tokens])
  |     +-- Capture: torch.cuda.graph(graph, self.graph_pool, stream=gc.stream)
  |     +-- Share graph_pool across captures (set on first capture)
  |     +-- Store: self.graphs[(bs, max_q_len)] = graph
  |     +-- torch.cuda.synchronize()
  |
  +-- Sort graph_bs back to ascending order
  +-- Return (elapsed_time, graph_bs)

Graph Keying

Each captured graph is stored in a dictionary keyed by a (graph_bs, max_q_len) tuple:

self.graphs: dict[tuple[int, int], torch.cuda.CUDAGraph] = dict()

graph_bs: The padded batch size used during capture.
max_q_len: The maximum query length per sequence. For standard decode, this is 1. For MTP (Multi-Token Prediction) speculative decoding, this is mtp_k + 1.

Default Capture Sizes

When cuda_graph_sizes has a single value (e.g., [512], the default), the capture sizes follow this pattern:

[1, 2, 4, 8] + [i for i in range(16, cuda_graph_sizes[0] + 1, 16)]
# Example with default 512:
# [1, 2, 4, 8, 16, 32, 48, 64, ..., 496, 512]

Graph Replay in run_model()

During inference, ModelRunner.run_model() decides whether to use eager execution or graph replay:

def run_model(self, input_ids):
    forward_context = get_forward_context()
    context = forward_context.context
    bs = context.batch_size
    is_prefill = context.is_prefill
    positions = context.positions

    if is_prefill or self.enforce_eager or bs > self.graph_bs[-1]:
        # Eager path: prefills, enforce_eager mode, or oversized batches
        hidden_states = self.model(input_ids, positions)
    else:
        # Graph replay path: decode batches within captured range
        graph_bs = context.graph_bs
        max_q_len = forward_context.attn_metadata.max_seqlen_q
        graph_key = (graph_bs, max_q_len)
        self.graphs[graph_key].replay()
        num_tokens = context.batch_size * max_q_len
        hidden_states = self.forward_vars["outputs"][:num_tokens]

    return self.model.compute_logits(hidden_states), hidden_states

Key decisions:

Prefill: Always eager (variable sequence lengths make CUDA graphs impractical).
Decode with bs <= max captured size: Replay the pre-captured graph.
Decode with bs > max captured size: Fall back to eager execution.
enforce_eager=True: Always eager, regardless of batch size.

4. Piecewise Compilation

Piecewise compilation splits the model’s computation graph at specified operations and compiles each subgraph independently. This enables CUDA graph capture for the compilable parts while leaving incompatible operations (primarily attention) to run eagerly.

Splitting Operations

The splitting_ops field in CompilationConfig defines which operations split the graph. When set_splitting_ops_for_v1() is called (automatically at level 3), the default splitting ops are:

["aiter.unified_attention_with_output", "aiter.mla_attention"]

These attention operations are split out because:

They require dynamic metadata (sequence lengths, block tables) that changes per step.
Some attention backends are not compatible with CUDA graph capture.
Attention kernels are already highly optimized, so Inductor compilation provides minimal additional benefit.

Compilation Pipeline

The VllmBackend.__call__ method orchestrates the piecewise compilation:

Graph splitting: split_graph() divides the traced model graph at the splitting operations into a sequence of SplitItem objects, each containing a subgraph.
Submodule identification: Subgraphs that are not splitting operations are identified as candidates for compilation.
Dynamic-shape compilation: PiecewiseCompileInterpreter runs the split graph with fake inputs and compiles each non-splitting subgraph via CompilerManager.compile() for a general (dynamic) shape.
Backend creation: For each compiled subgraph, a PiecewiseBackend instance is created. It holds:
- compiled_graph_for_general_shape: The Inductor-compiled graph for dynamic shapes.
- concrete_size_entries: A dictionary mapping specific runtime shapes to ConcreteSizeEntry objects for shape-specialized compilation.
Runtime dispatch: When PiecewiseBackend.__call__ is invoked:
- On the first run, it uses the general-shape compiled graph.
- For subsequent runs, if the runtime shape is in compile_sizes, it lazily compiles a shape-specialized version via CompilerManager.compile() and caches it.
- For shapes not in compile_sizes, it falls back to the general-shape compiled graph.

Cache Management

The CompilerManager caches compiled graphs using a key of (runtime_shape, graph_index, backend_name). The cache is stored in a Python file (vllm_compile_cache.py) at the local cache directory (~/.cache/atom/torch_compile_cache/<hash>/rank_<i>/<prefix>/).

On subsequent runs with the same model and configuration, compiled graphs are loaded from the cache, bypassing Inductor compilation entirely.

5. Forward Context & Stateless Dispatch

The ForwardContext dataclass in atom/utils/forward_context.py provides a module-level global mechanism for passing metadata to layers during the forward pass. This is critical for CUDA graphs because captured graphs cannot accept new arguments – all dynamic metadata must be accessible through a side channel.

ForwardContext Fields

Field	Type	Purpose
`no_compile_layers`	`dict[int, Any]`	Layers that should skip compilation (from `static_forward_context`)
`attn_metadata`	`AttentionMetaData` or `dict`	Attention-specific metadata (sequence lengths, block tables, etc.)
`kv_cache_data`	`dict[str, KVCacheTensor]`	KV cache tensors for each layer
`context`	`Context`	Basic forward pass context (positions, is_prefill, batch_size, graph_bs)
`dp_metadata`	`DPMetadata`	Data-parallel metadata (token counts across DP ranks)
`spec_decode_metadata`	`SpecDecodeMetadata`	Speculative decoding metadata (draft tokens, logits indices)

Lifecycle

The forward context follows a set-use-reset lifecycle:

Set: Before each forward pass, set_forward_context() is called with attention metadata, the ATOM config, a Context object, and optional DP/speculative decoding metadata.
Access: During the forward pass, any layer can call get_forward_context() to retrieve the current metadata without needing it passed as a function argument. This is used by both eager execution and CUDA graph replay paths.
Reset: After the forward pass, reset_forward_context() replaces the global context with an empty ForwardContext().

Context Dataclass

The Context object carries the most frequently accessed per-step state:

@dataclass
class Context:
    positions: torch.Tensor    # Token position IDs
    is_prefill: bool = False   # Whether this is a prefill step
    batch_size: int = 0        # Number of sequences in the batch
    graph_bs: int = 0          # Padded batch size for graph lookup
    is_draft: bool = False     # Whether this is a draft model forward

The graph_bs field is particularly important for CUDA graph dispatch: it holds the padded batch size that maps to a pre-captured graph key.

Integration with CUDA Graphs

For ModelRunner’s direct CUDA graph path (non-piecewise), the forward context is set before run_model() via set_forward_context(), and run_model() reads context.graph_bs and attn_metadata.max_seqlen_q to look up the correct pre-captured graph.

For the piecewise path, CUDAGraphWrapper (in atom/utils/cuda_graph.py) expects batch_descriptor and cudagraph_runtime_mode fields on the forward context to decide whether to capture, replay, or run eagerly:

forward_context = get_forward_context()
batch_descriptor = forward_context.batch_descriptor
cudagraph_runtime_mode = forward_context.cudagraph_runtime_mode

Note: The piecewise CUDAGraphWrapper integration is under development. The batch_descriptor and cudagraph_runtime_mode fields are expected by CUDAGraphWrapper.__call__() but are not currently defined on the ForwardContext dataclass. The per-subgraph wrapping in backends.py is also currently commented out. The direct CUDA graph path in ModelRunner is the active production path.

6. Compiler Backend

CompilerManager

CompilerManager in atom/utils/backends.py manages the full compilation lifecycle:

Initialization: Creates a CompilerInterface via make_compiler(). Uses InductorStandaloneAdaptor for PyTorch 2.8+ or InductorAdaptor for earlier versions.
Caching: Maintains a dictionary mapping (runtime_shape, graph_index, backend_name) to compiler-specific handles. Caches are serialized to a Python file using pprint.
Compile-or-load: On each call to compile(), first attempts load() from the cache. On miss, delegates to the compiler and stores the result.

CompilerInterface

CompilerInterface in atom/utils/compiler_inferface.py (note the typo in the filename) defines the abstract interface that all compiler backends must implement:

Method	Purpose
`initialize_cache(cache_dir, disable_cache, prefix)`	Set up cache directories for the compiler
`compute_hash(vllm_config)`	Generate a hash of compiler-specific state for cache invalidation
`compile(graph, example_inputs, compiler_config, runtime_shape, key)`	Compile a graph, returning `(compiled_callable, handle)`
`load(handle, graph, example_inputs, graph_index, runtime_shape)`	Load a previously compiled graph from the handle

InductorAdaptor

The default compiler for PyTorch < 2.8. Uses torch._inductor.compile_fx.compile_fx and monkey-patches several internal functions to:

Extract the compilation hash for caching.
Provide a dummy shape environment (AlwaysHitShapeEnv) so Inductor cache lookups succeed outside of Dynamo’s tracing context.
Force caching of graphs that Inductor would normally refuse to cache.

When runtime_shape is an integer (specific batch size), it enables max_autotune and coordinate_descent_tuning for Triton kernel parameter optimization.

InductorStandaloneAdaptor

The preferred compiler for PyTorch 2.8+. Uses torch._inductor.standalone_compile which provides a cleaner interface without the monkey-patching required by InductorAdaptor. Compiled artifacts are saved to disk in “unpacked” format and can be loaded directly.

VllmBackend

VllmBackend in atom/utils/backends.py serves as the torch.compile backend for level 3 (piecewise) compilation. When Dynamo calls it:

Computes a cache directory hash from config, traced files, and compiler state.
Splits the graph at splitting_ops using split_graph().
Runs PiecewiseCompileInterpreter to compile each non-splitting subgraph.
Saves the computation graph to computation_graph.py for debugging.
Returns the stitching graph module (split_gm) as the callable.

If cudagraph_copy_inputs is True, it wraps the callable to copy input tensors into static buffers before each call, ensuring CUDA graph input address stability.

@support_torch_compile Decorator

The @support_torch_compile decorator in atom/utils/decorators.py augments a model class to support torch.compile:

Class modification: Adds TorchCompileWrapperWithCustomDispatcher as a base class and overrides __init__ and __call__.
Dynamic shape marking: On the first compilation, it inspects the forward method signature, identifies torch.Tensor arguments, and calls torch._dynamo.mark_dynamic() to mark their batch dimensions as dynamic.
Custom dispatch: After the first compilation, if use_custom_dispatcher is True (levels >= 2), subsequent calls bypass Dynamo’s guard mechanism and dispatch directly to the compiled bytecode via dispatch_to_code(0).
Safety check: The bytecode hook checks for update in the compiled code’s co_names, raising an error if the model modifies nn.Module buffers during the forward pass (which would cause silent errors with CUDA graphs).

Custom Op Registration

direct_register_custom_op() in atom/utils/custom_register.py registers custom operators with PyTorch’s torch.library system:

direct_register_custom_op(
    op_name="my_op",
    op_func=my_kernel,
    mutates_args=["output"],
    fake_impl=my_fake_impl,
)

This registers the op under the "aiter" library namespace (e.g., aiter.my_op), making it visible to Dynamo’s tracing. The fake_impl is used during tracing to compute output shapes without executing the real kernel. The dispatch_key defaults to "CUDA" for GPU operations.

Registered custom ops can be used as splitting_ops in piecewise compilation (e.g., "aiter.unified_attention_with_output").

7. Configuration Options

All compilation-related configuration fields from CompilationConfig:

Field	Type	Default	Description
`level`	`int`	`0`	Compilation level (0-3). See Section 1.
`use_cudagraph`	`bool`	`True`	Whether CUDA graph capture is enabled.
`cudagraph_capture_sizes`	`Optional[list[int]]`	`None`	Explicit list of batch sizes to capture. Overrides `cuda_graph_sizes`.
`cuda_graph_sizes`	`list[int]`	`[512]`	Controls auto-generated capture sizes. 1 value = generate pattern; >1 values = use directly.
`cudagraph_mode`	`Optional[CUDAGraphMode]`	`None`	CUDA graph mode. Set to `PIECEWISE` automatically at level 3.
`splitting_ops`	`Optional[list[str]]`	`None`	Operations that split the graph for piecewise compilation. Auto-set at level 3.
`cudagraph_copy_inputs`	`bool`	`False`	Copy input tensors to static buffers for CUDA graph address stability. Only effective in `PIECEWISE` mode.
`use_inductor`	`bool`	`True`	Whether to use the Inductor compiler backend.
`compile_sizes`	`Optional[list[Union[int, str]]]`	`None`	Specific sizes to compile with Inductor. Supports `"cudagraph_capture_sizes"` string.
`inductor_compile_config`	`dict`	`{}`	Additional Inductor configuration (e.g., `max_autotune`).
`debug_dump_path`	`str`	`""`	Path to dump debug information (traced graphs, decompiled code).
`cache_dir`	`str`	`""`	Custom cache directory. Auto-generated if empty (`~/.cache/atom/torch_compile_cache/<hash>/`).

Related fields on Config:

Field	Type	Default	Description
`enforce_eager`	`bool`	`False`	Force eager execution, skip all compilation and CUDA graphs.
`graph_bs`	`Optional[list[int]]`	`None`	Final list of batch sizes for CUDA graph capture (computed from `CompilationConfig`).
`compilation_config`	`CompilationConfig`	`CompilationConfig()`	The compilation configuration dataclass.

8. Decision Tree

Use this decision tree to select the right compilation level and CUDA graph mode for your workload:

Is the model supported by torch.compile?
|
+-- No --> Level 0 (NO_COMPILATION)
|          enforce_eager=True
|
+-- Yes
    |
    +-- Debugging / profiling?
    |   |
    |   +-- Yes --> Level 0 (NO_COMPILATION)
    |
    +-- Quick compatibility check?
    |   |
    |   +-- Yes --> Level 1 (DYNAMO_AS_IS)
    |
    +-- Want Inductor optimization without CUDA graphs?
    |   |
    |   +-- Yes --> Level 2 (DYNAMO_ONCE)
    |
    +-- Production deployment
        |
        +-- Level 3 (PIECEWISE) [recommended]
            |
            +-- Standard serving --> cudagraph_mode=PIECEWISE (default)
            |
            +-- Small model / uniform batches --> cudagraph_mode=FULL
            |
            +-- P/D disaggregated (decode instance) --> cudagraph_mode=FULL_DECODE_ONLY
            |
            +-- Maximum performance --> cudagraph_mode=FULL_AND_PIECEWISE

Common Configurations

Default production setup (level 3, piecewise CUDA graphs):

CompilationConfig(level=3)
# Automatically sets:
#   splitting_ops = ["aiter.unified_attention_with_output", "aiter.mla_attention"]
#   cudagraph_mode = CUDAGraphMode.PIECEWISE
#   cuda_graph_sizes = [512]
#   graph_bs = [1, 2, 4, 8, 16, 32, ..., 512]

Custom capture sizes:

CompilationConfig(level=3, cudagraph_capture_sizes=[1, 2, 4, 8])

Debugging with full eager execution:

Config(model="...", enforce_eager=True)
# or
CompilationConfig(level=0)

Inductor with debug dump:

CompilationConfig(level=3, debug_dump_path="/tmp/atom_debug")
# Dumps traced graphs and decompiled code to /tmp/atom_debug/rank_0/

Source Files

File	Description
`atom/config.py`	`CompilationLevel`, `CompilationConfig`, `CUDAGraphMode`, `Config.__post_init__` (compilation setup)
`atom/utils/cuda_graph.py`	`CUDAGraphEntry`, `CUDAGraphOptions`, `CUDAGraphWrapper`, `BatchDescriptor`
`atom/utils/backends.py`	`CompilerManager`, `VllmBackend`, `SplitItem`, `split_graph()`, `PiecewiseCompileInterpreter`
`atom/utils/forward_context.py`	`ForwardContext`, `Context`, `AttentionMetaData`, `DPMetadata`, `set_forward_context()`, `get_forward_context()`
`atom/utils/compiler_inferface.py`	`CompilerInterface`, `InductorAdaptor`, `InductorStandaloneAdaptor`, `AlwaysHitShapeEnv`
`atom/utils/cuda_piecewise_backend.py`	`PiecewiseBackend`, `ConcreteSizeEntry`
`atom/utils/decorators.py`	`@support_torch_compile`, `TorchCompileWrapperWithCustomDispatcher`, `start_monitoring_torch_compile`
`atom/utils/custom_register.py`	`direct_register_custom_op()`, `aiter_lib` (Library instance)
`atom/model_engine/model_runner.py`	`ModelRunner.capture_cudagraph()`, `ModelRunner.run_model()`