name: aoti-debug description: Debug AOTInductor (AOTI) errors and crashes. Use when encountering AOTI segfaults, device mismatch errors, constant loading failures, or runtime errors from aot_compile, aot_load, aoti_compile_and_package, or aoti_load_package.

AOTI Debugging Guide

This skill helps diagnose and fix common AOTInductor issues.

Error Pattern Routing

Check the error message and route to the appropriate sub-guide:

Triton Index Out of Bounds

If the error matches this pattern:

Assertion `index out of bounds: 0 <= tmpN < ksM` failed

→ Follow the guide in triton-index-out-of-bounds.md

All Other Errors

Continue with the sections below.

First Step: Always Check Device and Shape Matching

For ANY AOTI error (segfault, exception, crash, wrong output), ALWAYS check these first:

1. Compile device == Load device: The model must be loaded on the same device type it was compiled on
2. Input devices match: Runtime inputs must be on the same device as the compiled model
3. Input shapes match: Runtime input shapes must match the shapes used during compilation (or satisfy dynamic shape constraints)

# During compilation - note the device and shapes
model = MyModel().eval()           # What device? CPU or .cuda()?
inp = torch.randn(2, 10)           # What device? What shape?
compiled_so = torch._inductor.aot_compile(model, (inp,))

# During loading - device type MUST match compilation
loaded = torch._export.aot_load(compiled_so, "???")  # Must match model/input device above

# During inference - device and shapes MUST match
out = loaded(inp.to("???"))  # Must match compile device, shape must match

If any of these don't match, you will get errors ranging from segfaults to exceptions to wrong outputs.

Key Constraint: Device Type Matching

AOTI requires compile and load to use the same device type.

• If you compile on CUDA, you must load on CUDA (device index can differ)
• If you compile on CPU, you must load on CPU
• Cross-device loading (e.g., compile on GPU, load on CPU) is NOT supported

Debugging CUDA Illegal Memory Access (IMA) Errors

If you encounter CUDA illegal memory access errors, follow this systematic approach:

Step 1: Sanity Checks

Before diving deep, try these debugging flags:

AOTI_RUNTIME_CHECK_INPUTS=1
TORCHINDUCTOR_NAN_ASSERTS=1

These flags take effect at compilation time (at codegen time):

• AOTI_RUNTIME_CHECK_INPUTS=1 checks if inputs satisfy the same guards used during compilation
• TORCHINDUCTOR_NAN_ASSERTS=1 adds codegen before and after each kernel to check for NaN

Step 2: Pinpoint the CUDA IMA

CUDA IMA errors can be non-deterministic. Use these flags to trigger the error deterministically:

PYTORCH_NO_CUDA_MEMORY_CACHING=1
CUDA_LAUNCH_BLOCKING=1

These flags take effect at runtime:

• PYTORCH_NO_CUDA_MEMORY_CACHING=1 disables PyTorch's Caching Allocator, which allocates bigger buffers than needed immediately. This is usually why CUDA IMA errors are non-deterministic.
• CUDA_LAUNCH_BLOCKING=1 forces kernels to launch one at a time. Without this, you get "CUDA kernel errors might be asynchronously reported" warnings since kernels launch asynchronously.

Step 3: Identify Problematic Kernels with Intermediate Value Debugger

Use the AOTI Intermediate Value Debugger to pinpoint the problematic kernel:

AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=3

This prints kernels one by one at runtime. Together with previous flags, this shows which kernel was launched right before the error.

To inspect inputs to a specific kernel:

AOT_INDUCTOR_FILTERED_KERNELS_TO_PRINT="triton_poi_fused_add_ge_logical_and_logical_or_lt_231,_add_position_embeddings_kernel_5" AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=2

If inputs to the kernel are unexpected, inspect the kernel that produces the bad input.

Additional Debugging Tools

Logging and Tracing

• tlparse / TORCH_TRACE: Provides complete output codes and records guards used
• TORCH_LOGS: Use TORCH_LOGS="+inductor,output_code" to see more PT2 internal logs
• TORCH_SHOW_CPP_STACKTRACES: Set to 1 to see more stack traces

Common Sources of Issues

• Dynamic shapes: Historically a source of many IMAs. Pay special attention when debugging dynamic shape scenarios.
• Custom ops: Especially when implemented in C++ with dynamic shapes. The meta function may need to be Symint'ified.

API Notes

Deprecated API

torch._export.aot_compile()  # Deprecated
torch._export.aot_load()     # Deprecated

Current API

torch._inductor.aoti_compile_and_package()
torch._inductor.aoti_load_package()

The new API stores device metadata in the package, so aoti_load_package() automatically uses the correct device type. You can only change the device index (e.g., cuda:0 vs cuda:1), not the device type.

Environment Variables Summary