torch._logging.set_logs — PyTorch 2.7 documentation (original) (raw)

torch._logging.set_logs(*, all=None, dynamo=None, aot=None, autograd=None, dynamic=None, inductor=None, distributed=None, c10d=None, ddp=None, fsdp=None, dtensor=None, onnx=None, bytecode=False, aot_graphs=False, aot_joint_graph=False, ddp_graphs=False, graph=False, graph_code=False, graph_breaks=False, graph_sizes=False, guards=False, recompiles=False, recompiles_verbose=False, trace_source=False, trace_call=False, trace_bytecode=False, output_code=False, kernel_code=False, schedule=False, perf_hints=False, pre_grad_graphs=False, post_grad_graphs=False, ir_pre_fusion=False, ir_post_fusion=False, onnx_diagnostics=False, fusion=False, overlap=False, export=None, modules=None, cudagraphs=False, sym_node=False, compiled_autograd=False, compiled_autograd_verbose=False, cudagraph_static_inputs=False, benchmarking=False, autotuning=False, graph_region_expansion=False)[source][source]¶

Sets the log level for individual components and toggles individual log artifact types.

Warning

This feature is a prototype and may have compatibility breaking changes in the future.

Note

The TORCH_LOGS environment variable has complete precedence over this function, so if it was set, this function does nothing.

A component is a set of related features in PyTorch. All of the log messages emitted from a given component have their own log levels. If the log level of a particular message has priority greater than or equal to its component’s log level setting, it is emitted. Otherwise, it is suppressed. This allows you to, for instance, silence large groups of log messages that are not relevant to you and increase verbosity of logs for components that are relevant. The expected log level values, ordered from highest to lowest priority, are:

• logging.CRITICAL
• logging.ERROR
• logging.WARNING
• logging.INFO
• logging.DEBUG
• logging.NOTSET

See documentation for the Python logging module for more information on log levels: https://docs.python.org/3/library/logging.html#logging-levels

An artifact is a particular type of log message. Each artifact is assigned to a parent component. A component can emit many different kinds of artifacts. In general, an artifact is emitted if either its corresponding setting in the argument list below is turned on or if its parent component is set to a log level less than or equal to the log level of the artifact.

Keyword Arguments

• all (Optional[int]) – The default log level for all components. Default: logging.WARN
• dynamo (Optional[int]) – The log level for the TorchDynamo component. Default: logging.WARN
• aot (Optional[int]) – The log level for the AOTAutograd component. Default: logging.WARN
• autograd (Optional[int]) – The log level for autograd. Default: logging.WARN
• inductor (Optional[int]) – The log level for the TorchInductor component. Default: logging.WARN
• dynamic (Optional[int]) – The log level for dynamic shapes. Default: logging.WARN
• distributed (Optional[int]) – Whether to log c10d communication operations and other debug info from PyTorch Distributed components. Default: logging.WARN
• c10d (Optional[int]) – Whether to log c10d communication operations related debug info in PyTorch Distributed components. Default: logging.WARN
• ddp (Optional[int]) – Whether to log debug info related to DistributedDataParallel``(DDP) from PyTorch Distributed components. Default: ``logging.WARN
• fsdp (Optional[int]) – Whether to log debug info related to FullyShardedDataParallel``(FSDP) in PyTorch Distributed components. Default: ``logging.WARN
• dtensor (Optional[int]) – Whether to log debug info related to DTensor``(DTensor) in PyTorch Distributed components. Default: ``logging.WARN
• onnx (Optional[int]) – The log level for the ONNX exporter component. Default: logging.WARN
• bytecode (bool) – Whether to emit the original and generated bytecode from TorchDynamo. Default: False
• aot_graphs (bool) – Whether to emit the graphs generated by AOTAutograd. Default: False
• aot_joint_graph (bool) – Whether to emit the joint forward-backward graph generated by AOTAutograd. Default: False
• ddp_graphs (bool) – Whether to emit graphs generated by DDPOptimizer. Default: False
• graph (bool) – Whether to emit the graph captured by TorchDynamo in tabular format. Default: False
• graph_code (bool) – Whether to emit the python source of the graph captured by TorchDynamo. Default: False
• graph_breaks (bool) – Whether to emit the graph breaks encountered by TorchDynamo. Default: False
• graph_sizes (bool) – Whether to emit tensor sizes of the graph captured by TorchDynamo. Default: False
• guards (bool) – Whether to emit the guards generated by TorchDynamo for each compiled function. Default: False
• recompiles (bool) – Whether to emit a guard failure reason and message every time TorchDynamo recompiles a function. Default: False
• recompiles_verbose (bool) – Whether to emit all guard failure reasons when TorchDynamo recompiles a function, even those that are not actually run. Default: False
• trace_source (bool) – Whether to emit when TorchDynamo begins tracing a new line. Default: False
• trace_call (bool) – Whether to emit detailed line location when TorchDynamo creates an FX node corresponding to function call. Python 3.11+ only. Default: False
• trace_bytecode (bool) – Whether to emit bytecode instructions and traced stack state as TorchDynamo traces bytecode. Default: False
• output_code (bool) – Whether to emit the TorchInductor output code on a per-graph basis. Default: False
• kernel_code (bool) – Whether to emit the TorchInductor output code on a per-kernel bases. Default: False
• schedule (bool) – Whether to emit the TorchInductor schedule. Default: False
• perf_hints (bool) – Whether to emit the TorchInductor perf hints. Default: False
• pre_grad_graphs (bool) – Whether to emit the graphs before inductor grad passes. Default: False
• post_grad_graphs (bool) – Whether to emit the graphs generated by after post grad passes. Default: False
• ir_pre_fusion (bool) – Whether to emit the graphs before inductor fusion passes. Default: False
• ir_post_fusion (bool) – Whether to emit the graphs after inductor fusion passes. Default: False
• onnx_diagnostics (bool) – Whether to emit the ONNX exporter diagnostics in logging. Default: False
• fusion (bool) – Whether to emit detailed Inductor fusion decisions. Default: False
• overlap (bool) – Whether to emit detailed Inductor compute/comm overlap decisions. Default: False
• sym_node (bool) – Whether to emit debug info for various SymNode opterations. Default: False
• export (Optional[int]) – The log level for export. Default: logging.WARN
• benchmarking (bool) – Whether to emit detailed Inductor benchmarking information. Default: False
• modules (dict) – This argument provides an alternate way to specify the above log component and artifact settings, in the format of a keyword args dictionary given as a single argument. There are two cases where this is useful (1) if a new log component or artifact has been registered but a keyword argument for it has not been added to this function and (2) if the log level for an unregistered module needs to be set. This can be done by providing the fully-qualified module name as the key, with the log level as the value. Default: None
• cudagraph_static_inputs (bool) – Whether to emit debug info for cudagraph static input detection. Default: False
• autotuning (bool) – Autotuning choice logs, such as kernel source, perf, and tuning parameters. Default: False
• graph_region_expansion (bool) – Whether to emit the detailed steps of the duplicate graph region tracker expansion algorithm. Default: False

Example:

import logging

The following changes the "dynamo" component to emit DEBUG-level

logs, and to emit "graph_code" artifacts.

torch._logging.set_logs(dynamo=logging.DEBUG, graph_code=True)

The following enables the logs for a different module

torch._logging.set_logs(modules={"unregistered.module.name": logging.DEBUG})