
AOTI Debugging
FreeDiagnose and resolve AOTInductor errors effectively.
Free · Opens the source repo
What AOTI Debugging does
The AOTI Debugging skill is designed for developers and data scientists working with the AOTInductor in PyTorch. It provides a structured approach to identifying and resolving common errors that arise during the compilation and execution of models. Users can quickly navigate through various error patterns, ensuring they can troubleshoot issues such as segmentation faults, device mismatches, and runtime errors efficiently. The skill guides users through a series of checks and solutions tailored to the specific error encountered, streamlining the debugging process.
One of the key features of this skill is its focus on device and shape matching, which is crucial for avoiding runtime errors. The skill emphasizes the importance of ensuring that the compile and load devices are identical, as well as confirming that input shapes align with those used during model compilation. By following the provided guidelines, users can prevent a wide range of errors and ensure smoother model execution.
Additionally, the skill includes detailed instructions for handling more complex issues, such as CUDA illegal memory access errors. It suggests using specific debugging flags to gain clearer insights into the problems at hand, allowing users to pinpoint the source of the error effectively. The inclusion of logging and tracing tools further aids in diagnosing issues, making this skill a comprehensive resource for troubleshooting AOTInductor-related challenges.
Overall, the AOTI Debugging skill is an essential tool for anyone working with PyTorch's AOTInductor, providing the necessary guidance to overcome common pitfalls and enhance the debugging experience.
When to use it
Use this skill whenever you encounter AOTI errors, such as segmentation faults or device mismatch errors, during model compilation or execution.
When not to use it
This skill is not suitable for debugging issues unrelated to AOTInductor, such as general PyTorch errors or problems with other libraries.
What you can build with it
Debugging Segfaults
When experiencing segmentation faults during model execution, use the skill to check device and shape matching before proceeding to specific error guides.
Resolving Device Mismatches
If you encounter a device mismatch error, follow the skill's guidelines to ensure that your compile and load devices are the same.
Handling CUDA Errors
Utilize the skill's recommendations for debugging CUDA illegal memory access errors by setting the appropriate runtime flags for clearer diagnostics.
How to install AOTI Debugging
View source1. Install with the skills CLI
npx skills add pytorch/pytorch/aoti-debug --agent claude-code2. Or install it manually
Download the skill folder and drop it into ~/.claude/skills/ for all projects, or .claude/skills/ to scope it to one repo. Restart Claude Code so it picks up the new skill.
Anthropic's agentic coding CLI, and the reference implementation of Agent Skills. Drop a skill folder into ~/.claude/skills and Claude Code loads it automatically whenever a task matches the skill's description. Claude Code docs
Inside SKILL.md
Written by pytorchAOTI 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:
- Compile device == Load device: The model must be loaded on the same device type it was compiled on
- Input devices match: Runtime inputs must be on the same device as the compiled model
- 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
Common Error Patterns
1. Device Mismatch Segfault
Symptom: Segfault, exception, or crash during aot_load() or model execution.
Example error messages:
The specified pointer resides on host memory and is not registered with any CUDA device- Crash during constant loading in AOTInductorModelBase
Expected out tensor to have device cuda:0, but got cpu instead
Cause: Compile and load device types don't match (see "First Step" above).
Solution: Ensure compile and load use the same device type. If compiled on CPU, load on CPU. If compiled on CUDA, load on CUDA.
2. Input Device Mismatch at Runtime
Symptom: RuntimeError during model execution.
Cause: Input device doesn't match compile device (see "First Step" above).
Better Debugging: Run with AOTI_RUNTIME_CHECK_INPUTS=1 for clearer errors. This flag validates all input properties including device type, dtype, sizes, and strides:
AOTI_RUNTIME_CHECK_INPUTS=1 python your_script.py
This produces actionable error messages like:
Error: input_handles[0]: unmatched device type, expected: 0(cpu), but got: 1(cuda)
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=1checks if inputs satisfy the same guards used during compilationTORCHINDUCTOR_NAN_ASSERTS=1adds 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=1disables PyTorch's Caching Allocator, which allocates bigger buffers than needed immediately. This is usually why CUDA IMA errors are non-deterministic.CUDA_LAUNCH_BLOCKING=1forces 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
1to 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
| Variable | When | Purpose |
|---|---|---|
AOTI_RUNTIME_CHECK_INPUTS=1 | Compile time | Validate inputs match compilation guards |
TORCHINDUCTOR_NAN_ASSERTS=1 | Compile time | Check for NaN before/after kernels |
PYTORCH_NO_CUDA_MEMORY_CACHING=1 | Runtime | Make IMA errors deterministic |
CUDA_LAUNCH_BLOCKING=1 | Runtime | Force synchronous kernel launches |
AOT_INDUCTOR_DEBUG_INTERMEDIATE_VALUE_PRINTER=3 | Compile time | Print kernels at runtime |
TORCH_LOGS="+inductor,output_code" | Runtime | See PT2 internal logs |
TORCH_SHOW_CPP_STACKTRACES=1 | Runtime | Show C++ stack traces |
Frequently asked questions about AOTI Debugging
Similar skills
Agent Host Debug Logs
Analyze Agent Host debug logs for deeper insights.
Code OSS Dev - Launch + Debug
Launch and debug Code OSS with isolated profiles.
Phoenix CLI
Debug LLM applications with structured analysis tools.
Power Automate Debugging
Diagnose and fix Power Automate flow errors effectively.
Arize Trace
Inspect and export traces for LLM applications.
Runtime Behavior Probe
Investigate real runtime behavior with precision.
