text
stringlengths
0
1.73k
source
stringlengths
35
119
category
stringclasses
2 values
Returns: the dictionary used to save all logger stats Return type: target_dict class torch.ao.ns._numeric_suite.Logger Base class for stats logging forward(x) class torch.ao.ns._numeric_suite.ShadowLogger Class used in Shadow module to record the outputs of the original and shadow modules. forward(x, y) class torch.ao.ns._numeric_suite.OutputLogger Class used to log the outputs of the module forward(x) class torch.ao.ns._numeric_suite.Shadow(q_module, float_module, logger_cls) Shadow module attaches the float module to its matching quantized module as the shadow. Then it uses Logger module to process the outputs of both modules. Parameters: * q_module -- module quantized from float_module that we want to shadow * **float_module** -- float module used to shadow q_module * **logger_cls** -- type of logger used to process the outputs of q_module and float_module. ShadowLogger or custom loggers
https://pytorch.org/docs/stable/torch.ao.ns._numeric_suite.html
pytorch docs
can be used. forward(*x) Return type: *Tensor* add(x, y) Return type: *Tensor* add_scalar(x, y) Return type: *Tensor* mul(x, y) Return type: *Tensor* mul_scalar(x, y) Return type: *Tensor* cat(x, dim=0) Return type: *Tensor* add_relu(x, y) Return type: *Tensor* torch.ao.ns._numeric_suite.prepare_model_with_stubs(float_module, q_module, module_swap_list, logger_cls) Prepare the model by attaching the float module to its matching quantized module as the shadow if the float module type is in module_swap_list. Example usage: prepare_model_with_stubs(float_model, q_model, module_swap_list, Logger) q_model(data) ob_dict = get_logger_dict(q_model) Parameters: * float_module (Module) -- float module used to generate the q_module * **q_module** (*Module*) -- module quantized from float_module
https://pytorch.org/docs/stable/torch.ao.ns._numeric_suite.html
pytorch docs
module_swap_list (Set[type]) -- list of float module types to attach the shadow logger_cls (Callable) -- type of logger to be used in shadow module to process the outputs of quantized module and its float shadow module torch.ao.ns._numeric_suite.compare_model_stub(float_model, q_model, module_swap_list, *data, logger_cls=) Compare quantized module in a model with its floating point counterpart, feeding both of them the same input. Return a dict with key corresponding to module names and each entry being a dictionary with two keys 'float' and 'quantized', containing the output tensors of quantized and its matching float shadow module. This dict can be used to compare and compute the module level quantization error. This function first call prepare_model_with_stubs() to swap the quantized module that we want to compare with the Shadow module,
https://pytorch.org/docs/stable/torch.ao.ns._numeric_suite.html
pytorch docs
which takes quantized module, corresponding float module and logger as input, and creates a forward path inside to make the float module to shadow quantized module sharing the same input. The logger can be customizable, default logger is ShadowLogger and it will save the outputs of the quantized module and float module that can be used to compute the module level quantization error. Example usage: module_swap_list = [torchvision.models.quantization.resnet.QuantizableBasicBlock] ob_dict = compare_model_stub(float_model,qmodel,module_swap_list, data) for key in ob_dict: print(key, compute_error(ob_dict[key]['float'], ob_dict[key]['quantized'].dequantize())) Parameters: * float_model (Module) -- float model used to generate the q_model * **q_model** (*Module*) -- model quantized from float_model * **module_swap_list** (*Set**[**type**]*) -- list of float module types at which shadow modules will be attached.
https://pytorch.org/docs/stable/torch.ao.ns._numeric_suite.html
pytorch docs
data -- input data used to run the prepared q_model logger_cls -- type of logger to be used in shadow module to process the outputs of quantized module and its float shadow module Return type: Dict[str, Dict] torch.ao.ns._numeric_suite.get_matching_activations(float_module, q_module) Find the matching activation between float and quantized modules. Parameters: * float_module (Module) -- float module used to generate the q_module * **q_module** (*Module*) -- module quantized from float_module Returns: dict with key corresponding to quantized module names and each entry being a dictionary with two keys 'float' and 'quantized', containing the matching float and quantized activations Return type: act_dict torch.ao.ns._numeric_suite.prepare_model_outputs(float_module, q_module, logger_cls=, allow_list=None)
https://pytorch.org/docs/stable/torch.ao.ns._numeric_suite.html
pytorch docs
Prepare the model by attaching the logger to both float module and quantized module if they are in the allow_list. Parameters: * float_module (Module) -- float module used to generate the q_module * **q_module** (*Module*) -- module quantized from float_module * **logger_cls** -- type of logger to be attached to float_module and q_module * **allow_list** -- list of module types to attach logger torch.ao.ns._numeric_suite.compare_model_outputs(float_model, q_model, *data, logger_cls=, allow_list=None) Compare output activations between float and quantized models at corresponding locations for the same input. Return a dict with key corresponding to quantized module names and each entry being a dictionary with two keys 'float' and 'quantized', containing the activations of quantized model and float model at matching locations. This dict can be used to compare and compute the
https://pytorch.org/docs/stable/torch.ao.ns._numeric_suite.html
pytorch docs
propagation quantization error. Example usage: act_compare_dict = compare_model_outputs(float_model, qmodel, data) for key in act_compare_dict: print( key, compute_error( act_compare_dict[key]['float'], act_compare_dict[key]['quantized'].dequantize() ) ) Parameters: * float_model (Module) -- float model used to generate the q_model * **q_model** (*Module*) -- model quantized from float_model * **data** -- input data used to run the prepared float_model and q_model * **logger_cls** -- type of logger to be attached to float_module and q_module * **allow_list** -- list of module types to attach logger Returns: dict with key corresponding to quantized module names and each entry being a dictionary with two keys 'float' and 'quantized', containing the matching float and quantized activations Return type:
https://pytorch.org/docs/stable/torch.ao.ns._numeric_suite.html
pytorch docs
Return type: act_compare_dict
https://pytorch.org/docs/stable/torch.ao.ns._numeric_suite.html
pytorch docs
torch.utils.model_zoo Moved to torch.hub. torch.utils.model_zoo.load_url(url, model_dir=None, map_location=None, progress=True, check_hash=False, file_name=None) Loads the Torch serialized object at the given URL. If downloaded file is a zip file, it will be automatically decompressed. If the object is already present in model_dir, it's deserialized and returned. The default value of "model_dir" is "/checkpoints" where "hub_dir" is the directory returned by "get_dir()". Parameters: * url (str) -- URL of the object to download * **model_dir** (*str**, **optional*) -- directory in which to save the object * **map_location** (*optional*) -- a function or a dict specifying how to remap storage locations (see torch.load) * **progress** (*bool**, **optional*) -- whether or not to display a progress bar to stderr. Default: True
https://pytorch.org/docs/stable/model_zoo.html
pytorch docs
check_hash (bool, optional) -- If True, the filename part of the URL should follow the naming convention "filename-.ext" where "" is the first eight or more digits of the SHA256 hash of the contents of the file. The hash is used to ensure unique names and to verify the contents of the file. Default: False file_name (str, optional) -- name for the downloaded file. Filename from "url" will be used if not set. Return type: Dict[str, Any] -[ Example ]- state_dict = torch.hub.load_state_dict_from_url('https://s3.amazonaws.com/pytorch/models/resnet18-5c106cde.pth')
https://pytorch.org/docs/stable/model_zoo.html
pytorch docs
Warning: There are known non-determinism issues for RNN functions on some versions of cuDNN and CUDA. You can enforce deterministic behavior by setting the following environment variables:On CUDA 10.1, set environment variable "CUDA_LAUNCH_BLOCKING=1". This may affect performance.On CUDA 10.2 or later, set environment variable (note the leading colon symbol) "CUBLAS_WORKSPACE_CONFIG=:16:8" or "CUBLAS_WORKSPACE_CONFIG=:4096:2".See the cuDNN 8 Release Notes for more information.
https://pytorch.org/docs/stable/cudnn_rnn_determinism.html
pytorch docs
PyTorch documentation PyTorch is an optimized tensor library for deep learning using GPUs and CPUs. Features described in this documentation are classified by release status: Stable: These features will be maintained long-term and there should generally be no major performance limitations or gaps in documentation. We also expect to maintain backwards compatibility (although breaking changes can happen and notice will be given one release ahead of time). Beta: These features are tagged as Beta because the API may change based on user feedback, because the performance needs to improve, or because coverage across operators is not yet complete. For Beta features, we are committing to seeing the feature through to the Stable classification. We are not, however, committing to backwards compatibility. Prototype: These features are typically not available as part of binary distributions like PyPI or Conda, except sometimes behind
https://pytorch.org/docs/stable/index.html
pytorch docs
run-time flags, and are at an early stage for feedback and testing. Community ^^^^^^^^^ PyTorch Governance | Build + CI PyTorch Contribution Guide PyTorch Design Philosophy PyTorch Governance | Mechanics PyTorch Governance | Maintainers Developer Notes ^^^^^^^^^^^^^^^ CUDA Automatic Mixed Precision examples Autograd mechanics Broadcasting semantics CPU threading and TorchScript inference CUDA semantics Distributed Data Parallel Extending PyTorch Extending torch.func with autograd.Function Frequently Asked Questions Gradcheck mechanics HIP (ROCm) semantics Features for large-scale deployments Modules MPS backend Multiprocessing best practices Numerical accuracy Reproducibility Serialization semantics Windows FAQ Language Bindings ^^^^^^^^^^^^^^^^^ C++ Javadoc torch::deploy Python API ^^^^^^^^^^ torch Tensors Generators Random sampling Serialization Parallelism
https://pytorch.org/docs/stable/index.html
pytorch docs
Serialization Parallelism Locally disabling gradient computation Math operations Utilities Symbolic Numbers Optimizations Operator Tags Engine Configuration torch.nn Parameter UninitializedParameter UninitializedBuffer Containers Convolution Layers Pooling layers Padding Layers Non-linear Activations (weighted sum, nonlinearity) Non-linear Activations (other) Normalization Layers Recurrent Layers Transformer Layers Linear Layers Dropout Layers Sparse Layers Distance Functions Loss Functions Vision Layers Shuffle Layers DataParallel Layers (multi-GPU, distributed) Utilities Quantized Functions Lazy Modules Initialization torch.nn.functional Convolution functions Pooling functions Non-linear activation functions Linear functions Dropout functions Sparse functions Distance functions Loss functions
https://pytorch.org/docs/stable/index.html
pytorch docs
Distance functions Loss functions Vision functions DataParallel functions (multi-GPU, distributed) torch.Tensor Data types Initializing and basic operations Tensor class reference Tensor Attributes torch.dtype torch.device torch.layout torch.memory_format Tensor Views torch.amp Autocasting Gradient Scaling Autocast Op Reference torch.autograd torch.autograd.backward torch.autograd.grad Forward-mode Automatic Differentiation Functional higher level API Locally disabling gradient computation Default gradient layouts In-place operations on Tensors Variable (deprecated) Tensor autograd functions Function Context method mixins Numerical gradient checking Profiler Anomaly detection Autograd graph torch.library torch.cuda StreamContext torch.cuda.can_device_access_peer torch.cuda.current_blas_handle torch.cuda.current_device
https://pytorch.org/docs/stable/index.html
pytorch docs
torch.cuda.current_device torch.cuda.current_stream torch.cuda.default_stream device torch.cuda.device_count device_of torch.cuda.get_arch_list torch.cuda.get_device_capability torch.cuda.get_device_name torch.cuda.get_device_properties torch.cuda.get_gencode_flags torch.cuda.get_sync_debug_mode torch.cuda.init torch.cuda.ipc_collect torch.cuda.is_available torch.cuda.is_initialized torch.cuda.memory_usage torch.cuda.set_device torch.cuda.set_stream torch.cuda.set_sync_debug_mode torch.cuda.stream torch.cuda.synchronize torch.cuda.utilization torch.cuda.OutOfMemoryError Random Number Generator Communication collectives Streams and events Graphs (beta) Memory management NVIDIA Tools Extension (NVTX) Jiterator (beta) Stream Sanitizer (prototype) torch.backends torch.backends.cuda torch.backends.cudnn torch.backends.mps
https://pytorch.org/docs/stable/index.html
pytorch docs
torch.backends.cudnn torch.backends.mps torch.backends.mkl torch.backends.mkldnn torch.backends.openmp torch.backends.opt_einsum torch.backends.xeon torch.distributed Backends Basics Initialization Post-Initialization Distributed Key-Value Store Groups Point-to-point communication Synchronous and asynchronous collective operations Collective functions Profiling Collective Communication Multi-GPU collective functions Third-party backends Launch utility Spawn utility Debugging "torch.distributed" applications Logging torch.distributed.algorithms.join torch.distributed.elastic Get Started Documentation torch.distributed.fsdp torch.distributed.optim torch.distributed.tensor.parallel torch.distributed.checkpoint torch.distributions Score function Pathwise derivative Distribution ExponentialFamily Bernoulli Beta Binomial
https://pytorch.org/docs/stable/index.html
pytorch docs
Bernoulli Beta Binomial Categorical Cauchy Chi2 ContinuousBernoulli Dirichlet Exponential FisherSnedecor Gamma Geometric Gumbel HalfCauchy HalfNormal Independent Kumaraswamy LKJCholesky Laplace LogNormal LowRankMultivariateNormal MixtureSameFamily Multinomial MultivariateNormal NegativeBinomial Normal OneHotCategorical Pareto Poisson RelaxedBernoulli LogitRelaxedBernoulli RelaxedOneHotCategorical StudentT TransformedDistribution Uniform VonMises Weibull Wishart KL Divergence Transforms Constraints Constraint Registry torch._dynamo torch.fft Fast Fourier Transforms Helper Functions torch.func What are composable function transforms? Why composable function transforms? Read More torch.futures torch.fx Overview Writing Transformations Debugging
https://pytorch.org/docs/stable/index.html
pytorch docs
Writing Transformations Debugging Limitations of Symbolic Tracing API Reference torch.hub Publishing models Loading models from Hub torch.jit TorchScript Language Reference Creating TorchScript Code Mixing Tracing and Scripting TorchScript Language Built-in Functions and Modules Debugging Frequently Asked Questions Known Issues Appendix torch.linalg Matrix Properties Decompositions Solvers Inverses Matrix Functions Matrix Products Tensor Operations Misc Experimental Functions torch.monitor API Reference torch.signal torch.signal.windows torch.special Functions torch.overrides Functions torch.package Tutorials How do I... Explanation API Reference torch.profiler Overview API Reference Intel Instrumentation and Tracing Technology APIs torch.nn.init torch.onnx Example: AlexNet from PyTorch to ONNX
https://pytorch.org/docs/stable/index.html
pytorch docs
Example: AlexNet from PyTorch to ONNX Tracing vs Scripting Avoiding Pitfalls Limitations Adding support for operators Frequently Asked Questions Contributing / developing Functions Classes torch.onnx diagnostics Overview Diagnostic Rules API Reference torch.optim How to use an optimizer Base class Algorithms How to adjust learning rate Stochastic Weight Averaging Complex Numbers Creating Complex Tensors Transition from the old representation Accessing real and imag Angle and abs Linear Algebra Serialization Autograd DDP Communication Hooks How to Use a Communication Hook? What Does a Communication Hook Operate On? Default Communication Hooks PowerSGD Communication Hook Debugging Communication Hooks Checkpointing of Communication Hooks Acknowledgements Pipeline Parallelism Model Parallelism using multiple GPUs Pipelined Execution
https://pytorch.org/docs/stable/index.html
pytorch docs
Pipelined Execution Pipe APIs in PyTorch Tutorials Acknowledgements Quantization Introduction to Quantization Quantization API Summary Quantization Stack Quantization Support Matrix Quantization API Reference Quantization Backend Configuration Quantization Accuracy Debugging Quantization Customizations Best Practices Frequently Asked Questions Common Errors Distributed RPC Framework Basics RPC RRef RemoteModule Distributed Autograd Framework Distributed Optimizer Design Notes Tutorials torch.random torch.masked Introduction Supported Operators torch.nested Introduction Construction size unbind Nested tensor constructor and conversion functions Supported operations torch.sparse Why and when to use sparsity Functionality overview Operator overview Sparse COO tensors Sparse Compressed Tensors Supported operations
https://pytorch.org/docs/stable/index.html
pytorch docs
Supported operations torch.Storage torch.testing torch.utils.benchmark torch.utils.bottleneck torch.utils.checkpoint torch.utils.cpp_extension torch.utils.data Dataset Types Data Loading Order and "Sampler" Loading Batched and Non-Batched Data Single- and Multi-process Data Loading Memory Pinning torch.utils.jit torch.utils.dlpack torch.utils.mobile_optimizer torch.utils.model_zoo torch.utils.tensorboard Type Info torch.finfo torch.iinfo Named Tensors Creating named tensors Named dimensions Name propagation semantics Explicit alignment by names Manipulating dimensions Autograd support Currently supported operations and subsystems Named tensor API reference Named Tensors operator coverage Keeps input names Removes dimensions Unifies names from inputs Permutes dimensions Contracts away dims Factory functions out function and in-place variants
https://pytorch.org/docs/stable/index.html
pytorch docs
out function and in-place variants torch.config Libraries ^^^^^^^^^ torchaudio TorchData TorchRec TorchServe torchtext torchvision PyTorch on XLA Devices Indices and tables * Index Module Index
https://pytorch.org/docs/stable/index.html
pytorch docs
Events Module contains events processing mechanisms that are integrated with the standard python logging. Example of usage: from torch.distributed.elastic import events event = events.Event(name="test_event", source=events.EventSource.WORKER, metadata={...}) events.get_logging_handler(destination="console").info(event) API Methods torch.distributed.elastic.events.record(event, destination='null') torch.distributed.elastic.events.get_logging_handler(destination='null') Return type: Handler Event Objects class torch.distributed.elastic.events.api.Event(name, source, timestamp=0, metadata=) The class represents the generic event that occurs during the torchelastic job execution. The event can be any kind of meaningful action. Parameters: * name (str) -- event name. * **source** (*EventSource*) -- the event producer, e.g. agent or worker
https://pytorch.org/docs/stable/elastic/events.html
pytorch docs
or worker * **timestamp** (*int*) -- timestamp in milliseconds when event occured. * **metadata** (*Dict**[**str**, **Optional**[**Union**[**str**, **int**, **float**, **bool**]**]**]*) -- additional data that is associated with the event. class torch.distributed.elastic.events.api.EventSource(value) Known identifiers of the event producers. torch.distributed.elastic.events.api.EventMetadataValue alias of "Optional"["Union"["str", "int", "float", "bool"]]
https://pytorch.org/docs/stable/elastic/events.html
pytorch docs
Metrics Metrics API Overview: The metrics API in torchelastic is used to publish telemetry metrics. It is designed to be used by torchelastic's internal modules to publish metrics for the end user with the goal of increasing visibility and helping with debugging. However you may use the same API in your jobs to publish metrics to the same metrics "sink". A "metric" can be thought of as timeseries data and is uniquely identified by the string-valued tuple "(metric_group, metric_name)". torchelastic makes no assumptions about what a "metric_group" is and what relationship it has with "metric_name". It is totally up to the user to use these two fields to uniquely identify a metric. Note: The metric group "torchelastic" is reserved by torchelastic for platform level metrics that it produces. For instance torchelastic may output the latency (in milliseconds) of a re-rendezvous operation from the agent as "(torchelastic, agent.rendezvous.duration.ms)"
https://pytorch.org/docs/stable/elastic/metrics.html
pytorch docs
agent.rendezvous.duration.ms)" A sensible way to use metric groups is to map them to a stage or module in your job. You may also encode certain high level properties the job such as the region or stage (dev vs prod). Publish Metrics: Using torchelastic's metrics API is similar to using python's logging framework. You first have to configure a metrics handler before trying to add metric data. The example below measures the latency for the "calculate()" function. import time import torch.distributed.elastic.metrics as metrics # makes all metrics other than the one from "my_module" to go /dev/null metrics.configure(metrics.NullMetricsHandler()) metrics.configure(metrics.ConsoleMetricsHandler(), "my_module") def my_method(): start = time.time() calculate() end = time.time() metrics.put_metric("calculate_latency", int(end-start), "my_module") You may also use the torch.distributed.elastic.metrics.prof` decorator to conveniently and succinctly profile functions
https://pytorch.org/docs/stable/elastic/metrics.html
pytorch docs
to conveniently and succinctly profile functions # -- in module examples.foobar -- import torch.distributed.elastic.metrics as metrics metrics.configure(metrics.ConsoleMetricsHandler(), "foobar") metrics.configure(metrics.ConsoleMetricsHandler(), "Bar") @metrics.prof def foo(): pass class Bar(): @metrics.prof def baz(): pass "@metrics.prof" will publish the following metrics .success - 1 if the function finished successfully .failure - 1 if the function threw an exception .duration.ms - function duration in milliseconds Configuring Metrics Handler: torch.distributed.elastic.metrics.MetricHandler is responsible for emitting the added metric values to a particular destination. Metric groups can be configured with different metric handlers. By default torchelastic emits all metrics to "/dev/null". By adding
https://pytorch.org/docs/stable/elastic/metrics.html
pytorch docs
the following configuration metrics, "torchelastic" and "my_app" metric groups will be printed out to console. import torch.distributed.elastic.metrics as metrics metrics.configure(metrics.ConsoleMetricHandler(), group = "torchelastic") metrics.configure(metrics.ConsoleMetricHandler(), group = "my_app") Writing a Custom Metric Handler: If you want your metrics to be emitted to a custom location, implement the torch.distributed.elastic.metrics.MetricHandler interface and configure your job to use your custom metric handler. Below is a toy example that prints the metrics to "stdout" import torch.distributed.elastic.metrics as metrics class StdoutMetricHandler(metrics.MetricHandler): def emit(self, metric_data): ts = metric_data.timestamp group = metric_data.group_name name = metric_data.name value = metric_data.value print(f"[{ts}][{group}]: {name}={value}") metrics.configure(StdoutMetricHandler(), group="my_app")
https://pytorch.org/docs/stable/elastic/metrics.html
pytorch docs
Now all metrics in the group "my_app" will be printed to stdout as: [1574213883.4182858][my_app]: my_metric= [1574213940.5237644][my_app]: my_metric= Metric Handlers Below are the metric handlers that come included with torchelastic. class torch.distributed.elastic.metrics.api.MetricHandler class torch.distributed.elastic.metrics.api.ConsoleMetricHandler class torch.distributed.elastic.metrics.api.NullMetricHandler Methods torch.distributed.elastic.metrics.configure(handler, group=None) torch.distributed.elastic.metrics.prof(fn=None, group='torchelastic') @profile decorator publishes duration.ms, count, success, failure metrics for the function that it decorates. The metric name defaults to the qualified name ("class_name.def_name") of the function. If the function does not belong to a class, it uses the leaf module name instead. Usage @metrics.prof def x(): pass @metrics.prof(group="agent")
https://pytorch.org/docs/stable/elastic/metrics.html
pytorch docs
pass @metrics.prof(group="agent") def y(): pass torch.distributed.elastic.metrics.put_metric(metric_name, metric_value, metric_group='torchelastic') Publishes a metric data point. Usage put_metric("metric_name", 1) put_metric("metric_name", 1, "metric_group_name")
https://pytorch.org/docs/stable/elastic/metrics.html
pytorch docs
Quickstart To launch a fault-tolerant job, run the following on all nodes. torchrun --nnodes=NUM_NODES --nproc_per_node=TRAINERS_PER_NODE --max_restarts=NUM_ALLOWED_FAILURES --rdzv_id=JOB_ID --rdzv_backend=c10d --rdzv_endpoint=HOST_NODE_ADDR YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...) To launch an elastic job, run the following on at least "MIN_SIZE" nodes and at most "MAX_SIZE" nodes. torchrun --nnodes=MIN_SIZE:MAX_SIZE --nproc_per_node=TRAINERS_PER_NODE --max_restarts=NUM_ALLOWED_FAILURES_OR_MEMBERSHIP_CHANGES --rdzv_id=JOB_ID --rdzv_backend=c10d --rdzv_endpoint=HOST_NODE_ADDR YOUR_TRAINING_SCRIPT.py (--arg1 ... train script args...) Note: TorchElastic models failures as membership changes. When a node fails, this is treated as a "scale down" event. When the failed node is replaced by the scheduler, it is a "scale up" event. Hence for
https://pytorch.org/docs/stable/elastic/quickstart.html
pytorch docs
both fault tolerant and elastic jobs, "--max_restarts" is used to control the total number of restarts before giving up, regardless of whether the restart was caused due to a failure or a scaling event. "HOST_NODE_ADDR", in form [:] (e.g. node1.example.com:29400), specifies the node and the port on which the C10d rendezvous backend should be instantiated and hosted. It can be any node in your training cluster, but ideally you should pick a node that has a high bandwidth. Note: If no port number is specified "HOST_NODE_ADDR" defaults to 29400. Note: The "--standalone" option can be passed to launch a single node job with a sidecar rendezvous backend. You don’t have to pass "-- rdzv_id", "--rdzv_endpoint", and "--rdzv_backend" when the "-- standalone" option is used. Note: Learn more about writing your distributed training script here. If "torchrun" does not meet your requirements you may use our APIs directly for more powerful customization. Start by taking a look at
https://pytorch.org/docs/stable/elastic/quickstart.html
pytorch docs
the elastic agent API.
https://pytorch.org/docs/stable/elastic/quickstart.html
pytorch docs
Examples Please refer to the elastic/examples README.
https://pytorch.org/docs/stable/elastic/examples.html
pytorch docs
Customization This section describes how to customize TorchElastic to fit your needs. Launcher The launcher program that ships with TorchElastic should be sufficient for most use-cases (see torchrun (Elastic Launch)). You can implement a custom launcher by programmatically creating an agent and passing it specs for your workers as shown below. # my_launcher.py if name == "main": args = parse_args(sys.argv[1:]) rdzv_handler = RendezvousHandler(...) spec = WorkerSpec( local_world_size=args.nproc_per_node, fn=trainer_entrypoint_fn, args=(trainer_entrypoint_fn args.fn_args,...), rdzv_handler=rdzv_handler, max_restarts=args.max_restarts, monitor_interval=args.monitor_interval, ) agent = LocalElasticAgent(spec, start_method="spawn") try: run_result = agent.run() if run_result.is_failed(): print(f"worker 0 failed with: run_result.failures[0]") else:
https://pytorch.org/docs/stable/elastic/customization.html
pytorch docs
else: print(f"worker 0 return value is: run_result.return_values[0]") except Exception ex: # handle exception Rendezvous Handler To implement your own rendezvous, extend "torch.distributed.elastic.rendezvous.RendezvousHandler" and implement its methods. Warning: Rendezvous handlers are tricky to implement. Before you begin make sure you completely understand the properties of rendezvous. Please refer to Rendezvous for more information. Once implemented you can pass your custom rendezvous handler to the worker spec when creating the agent. spec = WorkerSpec( rdzv_handler=MyRendezvousHandler(params), ... ) elastic_agent = LocalElasticAgent(spec, start_method=start_method) elastic_agent.run(spec.role) Metric Handler TorchElastic emits platform level metrics (see Metrics). By default metrics are emitted to /dev/null so you will not see them. To have
https://pytorch.org/docs/stable/elastic/customization.html
pytorch docs
the metrics pushed to a metric handling service in your infrastructure, implement a torch.distributed.elastic.metrics.MetricHandler and configure it in your custom launcher. # my_launcher.py import torch.distributed.elastic.metrics as metrics class MyMetricHandler(metrics.MetricHandler): def emit(self, metric_data: metrics.MetricData): # push metric_data to your metric sink def main(): metrics.configure(MyMetricHandler()) spec = WorkerSpec(...) agent = LocalElasticAgent(spec) agent.run() Events Handler TorchElastic supports events recording (see Events). The events module defines API that allows you to record events and implement custom EventHandler. EventHandler is used for publishing events produced during torchelastic execution to different sources, e.g. AWS CloudWatch. By default it uses torch.distributed.elastic.events.NullEventHandler that ignores events. To configure custom events handler you need to implement
https://pytorch.org/docs/stable/elastic/customization.html
pytorch docs
torch.distributed.elastic.events.EventHandler interface and configure it in your custom launcher. # my_launcher.py import torch.distributed.elastic.events as events class MyEventHandler(events.EventHandler): def record(self, event: events.Event): # process event def main(): events.configure(MyEventHandler()) spec = WorkerSpec(...) agent = LocalElasticAgent(spec) agent.run()
https://pytorch.org/docs/stable/elastic/customization.html
pytorch docs
Multiprocessing Library that launches and manages "n" copies of worker subprocesses either specified by a function or a binary. For functions, it uses "torch.multiprocessing" (and therefore python "multiprocessing") to spawn/fork worker processes. For binaries it uses python "subprocessing.Popen" to create worker processes. Usage 1: Launching two trainers as a function from torch.distributed.elastic.multiprocessing import Std, start_processes def trainer(a, b, c): pass # train # runs two trainers # LOCAL_RANK=0 trainer(1,2,3) # LOCAL_RANK=1 trainer(4,5,6) ctx = start_processes( name="trainer", entrypoint=trainer, args={0: (1,2,3), 1: (4,5,6)}, envs={0: {"LOCAL_RANK": 0}, 1: {"LOCAL_RANK": 1}}, log_dir="/tmp/foobar", redirects=Std.ALL, # write all worker stdout/stderr to a log file tee={0: Std.ERR}, # tee only local rank 0's stderr to console )
https://pytorch.org/docs/stable/elastic/multiprocessing.html
pytorch docs
) # waits for all copies of trainer to finish ctx.wait() Usage 2: Launching 2 echo workers as a binary # same as invoking # echo hello # echo world > stdout.log ctx = start_processes( name="echo" entrypoint="echo", log_dir="/tmp/foobar", args={0: "hello", 1: "world"}, redirects={1: Std.OUT}, ) Just like "torch.multiprocessing", the return value of the function "start_processes()" is a process context ("api.PContext"). If a function was launched, a "api.MultiprocessContext" is returned and if a binary was launched a "api.SubprocessContext" is returned. Both are specific implementations of the parent "api.PContext" class. Starting Multiple Workers torch.distributed.elastic.multiprocessing.start_processes(name, entrypoint, args, envs, log_dir, start_method='spawn', redirects=Std.NONE, tee=Std.NONE) Starts "n" copies of "entrypoint" processes with the provided
https://pytorch.org/docs/stable/elastic/multiprocessing.html
pytorch docs
options. "entrypoint" is either a "Callable" (function) or a "str" (binary). The number of copies is determined by the number of entries for "args" and "envs" arguments, which need to have the same key set. "args" and "env" parameters are the arguments and environment variables to pass down to the entrypoint mapped by the replica index (local rank). All local ranks must be accounted for. That is, the keyset should be "{0,1,...,(nprocs-1)}". Note: When the "entrypoint" is a binary ("str"), "args" can only be strings. If any other type is given, then it is casted to a string representation (e.g. "str(arg1)"). Furthermore, a binary failure will only write an "error.json" error file if the main function is annotated with "torch.distributed.elastic.multiprocessing.errors.record". For function launches, this is done by default and there is no need to manually annotate with the "@record" annotation.
https://pytorch.org/docs/stable/elastic/multiprocessing.html
pytorch docs
"redirects" and "tee" are bitmasks specifying which std stream(s) to redirect to a log file in the "log_dir". Valid mask values are defined in "Std". To redirect/tee only certain local ranks, pass "redirects" as a map with the key as the local rank to specify the redirect behavior for. Any missing local ranks will default to "Std.NONE". "tee" acts like the unix "tee" command in that it redirects + prints to console. To avoid worker stdout/stderr from printing to console, use the "redirects" parameter. For each process, the "log_dir" will contain: "{local_rank}/error.json": if the process failed, a file with the error info "{local_rank}/stdout.json": if "redirect & STDOUT == STDOUT" "{local_rank}/stderr.json": if "redirect & STDERR == STDERR" Note: It is expected that the "log_dir" exists, is empty, and is a directory. Example: log_dir = "/tmp/test" # ok; two copies of foo: foo("bar0"), foo("bar1")
https://pytorch.org/docs/stable/elastic/multiprocessing.html
pytorch docs
start_processes( name="trainer", entrypoint=foo, args:{0:("bar0",), 1:("bar1",), envs:{0:{}, 1:{}}, log_dir=log_dir ) # invalid; envs missing for local rank 1 start_processes( name="trainer", entrypoint=foo, args:{0:("bar0",), 1:("bar1",), envs:{0:{}}, log_dir=log_dir ) # ok; two copies of /usr/bin/touch: touch file1, touch file2 start_processes( name="trainer", entrypoint="/usr/bin/touch", args:{0:("file1",), 1:("file2",), envs:{0:{}, 1:{}}, log_dir=log_dir ) # caution; arguments casted to string, runs: # echo "1" "2" "3" and echo "[1, 2, 3]" start_processes( name="trainer", entrypoint="/usr/bin/echo", args:{0:(1,2,3), 1:([1,2,3],), envs:{0:{}, 1:{}}, log_dir=log_dir ) Parameters: * name (str) -- a human readable short name that describes
https://pytorch.org/docs/stable/elastic/multiprocessing.html
pytorch docs
what the processes are (used as header when tee'ing stdout/stderr outputs) * **entrypoint** (*Union**[**Callable**, **str**]*) -- either a "Callable" (function) or "cmd" (binary) * **args** (*Dict**[**int**, **Tuple**]*) -- arguments to each replica * **envs** (*Dict**[**int**, **Dict**[**str**, **str**]**]*) -- env vars to each replica * **log_dir** (*str*) -- directory used to write log files * **start_method** (*str*) -- multiprocessing start method (spawn, fork, forkserver) ignored for binaries * **redirects** (*Union**[**Std**, **Dict**[**int**, **Std**]**]*) -- which std streams to redirect to a log file * **tee** (*Union**[**Std**, **Dict**[**int**, **Std**]**]*) -- which std streams to redirect + print to console Return type: PContext Process Context
https://pytorch.org/docs/stable/elastic/multiprocessing.html
pytorch docs
Process Context class torch.distributed.elastic.multiprocessing.api.PContext(name, entrypoint, args, envs, stdouts, stderrs, tee_stdouts, tee_stderrs, error_files) The base class that standardizes operations over a set of processes that are launched via different mechanisms. The name "PContext" is intentional to disambiguate with "torch.multiprocessing.ProcessContext". Warning: stdouts and stderrs should ALWAYS be a superset of tee_stdouts and tee_stderrs (respectively) this is b/c tee is implemented as a redirect + tail -f <stdout/stderr.log> class torch.distributed.elastic.multiprocessing.api.MultiprocessContext(name, entrypoint, args, envs, stdouts, stderrs, tee_stdouts, tee_stderrs, error_files, start_method) "PContext" holding worker processes invoked as a function. class torch.distributed.elastic.multiprocessing.api.SubprocessContext(name, entrypoint, args, envs, stdouts, stderrs, tee_stdouts, tee_stderrs, error_files)
https://pytorch.org/docs/stable/elastic/multiprocessing.html
pytorch docs
"PContext" holding worker processes invoked as a binary. class torch.distributed.elastic.multiprocessing.api.RunProcsResult(return_values=, failures=, stdouts=, stderrs=) Results of a completed run of processes started with "start_processes()". Returned by "PContext". Note the following: All fields are mapped by local rank "return_values" - only populated for functions (not the binaries). "stdouts" - path to stdout.log (empty string if no redirect) "stderrs" - path to stderr.log (empty string if no redirect)
https://pytorch.org/docs/stable/elastic/multiprocessing.html
pytorch docs
Elastic Agent Server The elastic agent is the control plane of torchelastic. It is a process that launches and manages underlying worker processes. The agent is responsible for: Working with distributed torch: the workers are started with all the necessary information to successfully and trivially call "torch.distributed.init_process_group()". Fault tolerance: monitors workers and upon detecting worker failures or unhealthiness, tears down all workers and restarts everyone. Elasticity: Reacts to membership changes and restarts workers with the new members. The simplest agents are deployed per node and works with local processes. A more advanced agent can launch and manage workers remotely. Agents can be completely decentralized, making decisions based on the workers it manages. Or can be coordinated, communicating to other agents (that manage workers in the same job) to make a collective decision.
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
collective decision. Below is a diagram of an agent that manages a local group of workers. [image] Concepts This section describes the high-level classes and concepts that are relevant to understanding the role of the "agent" in torchelastic. class torch.distributed.elastic.agent.server.ElasticAgent Agent process responsible for managing one or more worker processes. The worker processes are assumed to be regular distributed PyTorch scripts. When the worker process is created by the agent, the agent provides the necessary information for the worker processes to properly initialize a torch process group. The exact deployment topology and ratio of agent-to-worker is dependent on the specific implementation of the agent and the user's job placement preferences. For instance, to run a distributed training job on GPU with 8 trainers (one per GPU) one can: Use 8 x single GPU instances, place an agent per instance, managing 1 worker per agent.
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
managing 1 worker per agent. Use 4 x double GPU instances, place an agent per instance, managing 2 workers per agent. Use 2 x quad GPU instances, place an agent per instance, managing 4 workers per agent. Use 1 x 8 GPU instance, place an agent per instance, managing 8 workers per agent. Usage group_result = agent.run() if group_result.is_failed(): # workers failed failure = group_result.failures[0] log.exception(f"worker 0 failed with exit code : {failure.exit_code}") else: return group_result.return_values[0] # return rank 0's results abstract get_worker_group(role='default') Returns: The "WorkerGroup" for the given "role". Note that the worker group is a mutable object and hence in a multi- threaded/process environment it may change state. Implementors are encouraged (but not required) to return a defensive read-only copy. Return type:
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
Return type: WorkerGroup abstract run(role='default') Runs the agent, retrying the worker group on failures up to "max_restarts". Returns: The result of the execution, containing the return values or failure details for each worker mapped by the worker's global rank. Raises: **Exception - any other failures NOT related to worker process** -- Return type: *RunResult* class torch.distributed.elastic.agent.server.WorkerSpec(role, local_world_size, rdzv_handler, fn=None, entrypoint=None, args=(), max_restarts=3, monitor_interval=30.0, master_port=None, master_addr=None, local_addr=None, redirects=Std.NONE, tee=Std.NONE) Contains blueprint information about a particular type of worker. For a given role, there must only exist a single worker spec. Worker spec is expected to be homogenous across all nodes (machine), that is each node runs the same number of workers for a particular spec.
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
particular spec. Parameters: * role (str) -- user-defined role for the workers with this spec * **local_world_size** (*int*) -- number local workers to run * **fn** (*Optional**[**Callable**]*) -- (deprecated use entrypoint instead) * **entrypoint** (*Optional**[**Union**[**Callable**, **str**]**]*) -- worker function or command * **args** (*Tuple*) -- arguments to pass to "entrypoint" * **rdzv_handler** (*RendezvousHandler*) -- handles rdzv for this set of workers * **max_restarts** (*int*) -- number of max retries for the workers * **monitor_interval** (*float*) -- monitor status of workers every "n" seconds * **master_port** (*Optional**[**int**]*) -- fixed port to run the c10d store on rank 0 if not specified then will chose a random free port * **master_addr** (*Optional**[**str**]*) -- fixed master_addr
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
to run the c10d store on rank 0 if not specified then will chose hostname on agent rank 0 * **redirects** (*Union**[**Std**, **Dict**[**int**, **Std**]**]*) -- redirect std streams to a file, selectively redirect for a particular local rank by passing a map * **tee** (*Union**[**Std**, **Dict**[**int**, **Std**]**]*) -- tees the specified std stream(s) to console + file, selectively tee for a particular local rank by passing a map, takes precedence over "redirects" settings. get_entrypoint_name() If the entrypoint is a function (e.g. "Callable") returns its "__qualname__", else if the entrypoint is a binary (e.g. "str"), returns the binary name. class torch.distributed.elastic.agent.server.WorkerState(value) State of the "WorkerGroup". Workers in a worker group change state as a unit. If a single worker in a worker group fails the entire set is considered failed:
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
set is considered failed: UNKNOWN - agent lost track of worker group state, unrecoverable INIT - worker group object created not yet started HEALTHY - workers running and healthy UNHEALTHY - workers running and unhealthy STOPPED - workers stopped (interrupted) by the agent SUCCEEDED - workers finished running (exit 0) FAILED - workers failed to successfully finish (exit !0) A worker group starts from an initial "INIT" state, then progresses to "HEALTHY" or "UNHEALTHY" states, and finally reaches a terminal "SUCCEEDED" or "FAILED" state. Worker groups can be interrupted and temporarily put into "STOPPED" state by the agent. Workers in "STOPPED" state are scheduled to be restarted in the near future by the agent. Some examples of workers being put into "STOPPED" state are: Worker group failure|unhealthy observed Membership change detected When actions (start, stop, rdzv, retry, etc) on worker group fails
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
and results in the action being partially applied to the worker group the state will be "UNKNOWN". Typically this happens on uncaught/unhandled exceptions during state change events on the agent. The agent is not expected to recover worker groups in "UNKNOWN" state and is better off self terminating and allowing the job manager to retry the node. static is_running(state) Returns: True if the worker state represents workers still running (e.g. that the process exists but not necessarily healthy). Return type: bool class torch.distributed.elastic.agent.server.Worker(local_rank, global_rank=- 1, role_rank=- 1, world_size=- 1, role_world_size=- 1) Represents a worker instance. Contrast this with "WorkerSpec" that represents the specifications of a worker. A "Worker" is created from a "WorkerSpec". A "Worker" is to a "WorkerSpec" as an object is to a class. The "id" of the worker is interpreted by the specific
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
implementation of "ElasticAgent". For a local agent, it could be the "pid (int)" of the worker, for a remote agent it could be encoded as "host:port (string)". Parameters: * id (Any) -- uniquely identifies a worker (interpreted by the agent) * **local_rank** (*int*) -- local rank of the worker * **global_rank** (*int*) -- global rank of the worker * **role_rank** (*int*) -- rank of the worker across all workers that have the same role * **world_size** (*int*) -- number of workers (globally) * **role_world_size** (*int*) -- number of workers that have the same role class torch.distributed.elastic.agent.server.WorkerGroup(spec) Represents the set of "Worker" instances for the given "WorkerSpec" managed by "ElasticAgent". Whether the worker group contains cross instance workers or not depends on the implementation of the agent. Implementations
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
Implementations Below are the agent implementations provided by torchelastic. class torch.distributed.elastic.agent.server.local_elastic_agent.LocalElasticAgent(spec, start_method='spawn', exit_barrier_timeout=300, log_dir=None) An implementation of "torchelastic.agent.server.ElasticAgent" that handles host-local workers. This agent is deployed per host and is configured to spawn "n" workers. When using GPUs, "n" maps to the number of GPUs available on the host. The local agent does not communicate to other local agents deployed on other hosts, even if the workers may communicate inter-host. The worker id is interpreted to be a local process. The agent starts and stops all worker processes as a single unit. The worker function and argument passed to the worker function must be python multiprocessing compatible. To pass multiprocessing data structures to the workers you may create the data structure in the
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
same multiprocessing context as the specified "start_method" and pass it as a function argument. The "exit_barrier_timeout" specifies the amount of time (in seconds) to wait for other agents to finish. This acts as a safety net to handle cases where workers finish at different times, to prevent agents from viewing workers that finished early as a scale- down event. It is strongly advised that the user code deal with ensuring that workers are terminated in a synchronous manner rather than relying on the exit_barrier_timeout. A named pipe based watchdog can be enabled in "LocalElasticAgent" if an environment variable "TORCHELASTIC_ENABLE_FILE_TIMER" with value 1 has been defined in the "LocalElasticAgent" process. Optionally, another environment variable "TORCHELASTIC_TIMER_FILE" can be set with a unique file name for the named pipe. If the environment variable "TORCHELASTIC_TIMER_FILE" is not set, "LocalElasticAgent" will
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
internally create a unique file name and set it to the environment variable "TORCHELASTIC_TIMER_FILE", and this environment variable will be propagated to the worker processes to allow them to connect to the same named pipe that "LocalElasticAgent" uses. Example launching function def trainer(args) -> str: return "do train" def main(): start_method="spawn" shared_queue= multiprocessing.get_context(start_method).Queue() spec = WorkerSpec( role="trainer", local_world_size=nproc_per_process, entrypoint=trainer, args=("foobar",), ...<OTHER_PARAMS...>) agent = LocalElasticAgent(spec, start_method) results = agent.run() if results.is_failed(): print("trainer failed") else: print(f"rank 0 return value: {results.return_values[0]}")
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
prints -> rank 0 return value: do train Example launching binary def main(): spec = WorkerSpec( role="trainer", local_world_size=nproc_per_process, entrypoint="/usr/local/bin/trainer", args=("--trainer_args", "foobar"), ...<OTHER_PARAMS...>) agent = LocalElasticAgent(spec) results = agent.run() if not results.is_failed(): print("binary launches do not have return values") Extending the Agent To extend the agent you can implement "`ElasticAgent" directly, however we recommend you extend "SimpleElasticAgent" instead, which provides most of the scaffolding and leaves you with a few specific abstract methods to implement. class torch.distributed.elastic.agent.server.SimpleElasticAgent(spec, exit_barrier_timeout=300) An "ElasticAgent" that manages workers ("WorkerGroup") for a single
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
"WorkerSpec" (e.g. one particular type of worker role). _assign_worker_ranks(store, group_rank, group_world_size, spec) Determines proper ranks for worker processes. The rank assignment is done according to the following algorithm: 1. Each agent writes its configuration(group_rank, group_world_size , num_workers) to the common store. 2. Each agent retrieves configuration for all agents and performs two level sort using role and rank. 3. Determine the global rank: the global rank of the workers for the current agent is the offset of the infos array up to group_rank of the agent. The offset is computed as a sum of local_world_size of all agents that have rank less than the group_rank. The workers would have the ranks: [offset, offset+local_world_size) 4. Determine the role rank: The role rank is determined using the algorithms in the point 3 with the exception that the
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
offset is done from the first agent that has the same role as current one and has the minimum group rank. Return type: *List*[*Worker*] _exit_barrier() Wait for "exit_barrier_timeout" seconds for all agents to finish executing their local workers (either successfully or not). This acts as a safety guard against user scripts that terminate at different times. This barrier keeps the agent process alive until all workers finish. _initialize_workers(worker_group) Starts a fresh set of workers for the worker_group. Essentially a rendezvous followed by a start_workers. The caller should first call "_stop_workers()" to stop running workers prior to calling this method. Optimistically sets the state of the worker group that just started as "HEALTHY" and delegates the actual monitoring of state to "_monitor_workers()" method abstract _monitor_workers(worker_group)
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
abstract _monitor_workers(worker_group) Checks on the workers for the "worker_group" and returns the new state of the worker group. Return type: *RunResult* _rendezvous(worker_group) Runs rendezvous for the workers specified by worker spec. Assigns workers a new global rank and world size. Updates the rendezvous store for the worker group. _restart_workers(worker_group) Restarts (stops, rendezvous, starts) all local workers in the group. abstract _shutdown(death_sig=Signals.SIGTERM) Cleans up any resources that were allocated during the agent's work. Parameters: **death_sig** (*Signals*) -- Signal to send to the child process, SIGTERM is default abstract _start_workers(worker_group) Starts "worker_group.spec.local_world_size" number of workers according to worker spec for the worker group . Returns a map of "local_rank" to worker "id". Return type:
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
Return type: Dict[int, Any] abstract _stop_workers(worker_group) Stops all workers in the given worker group. Implementors must deal with workers in all states defined by "WorkerState". That is, it must gracefully handle stopping non-existent workers, unhealthy (stuck) workers, etc. class torch.distributed.elastic.agent.server.api.RunResult(state, return_values=, failures=) Results returned by the worker executions. Run results follow an "all-or-nothing" policy where the run is successful if and only if ALL local workers managed by this agent complete successfully. If the result is successful (e.g. "is_failed() = False") then the "return_values" field contains the outputs (return values) of the workers managed by THIS agent mapped by their GLOBAL ranks. That is "result.return_values[0]" is the return value of global rank 0. Note: "return_values" are only meaningful for when the worker
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
entrypoint is a function. Workers specified as a binary entrypoint do not canonically have a return value and the "return_values" field is meaningless and may be empty. If "is_failed()" returns "True" then the "failures" field contains the failure information, again, mapped by the GLOBAL rank of the worker that failed. The keys in "return_values" and "failures" are mutually exclusive, that is, a worker's final state can only be one of: succeeded, failed. Workers intentionally terminated by the agent according to the agent's restart policy, are not represented in either "return_values" nor "failures". Watchdog in the Agent A named pipe based watchdog can be enabled in "LocalElasticAgent" if an environment variable "TORCHELASTIC_ENABLE_FILE_TIMER" with value 1 has been defined in the "LocalElasticAgent" process. Optionally, another environment variable "TORCHELASTIC_TIMER_FILE" can be set
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
with a unique file name for the named pipe. If the environment variable "TORCHELASTIC_TIMER_FILE" is not set, "LocalElasticAgent" will internally create a unique file name and set it to the environment variable "TORCHELASTIC_TIMER_FILE", and this environment variable will be propagated to the worker processes to allow them to connect to the same named pipe that "LocalElasticAgent" uses.
https://pytorch.org/docs/stable/elastic/agent.html
pytorch docs
Expiration Timers Expiration timers are set up on the same process as the agent and used from your script to deal with stuck workers. When you go into a code- block that has the potential to get stuck you can acquire an expiration timer, which instructs the timer server to kill the process if it does not release the timer by the self-imposed expiration deadline. Usage: import torchelastic.timer as timer import torchelastic.agent.server as agent def main(): start_method = "spawn" message_queue = mp.get_context(start_method).Queue() server = timer.LocalTimerServer(message, max_interval=0.01) server.start() # non-blocking spec = WorkerSpec( fn=trainer_func, args=(message_queue,), ...<OTHER_PARAMS...>) agent = agent.LocalElasticAgent(spec, start_method) agent.run() def trainer_func(message_queue): timer.configure(timer.LocalTimerClient(message_queue))
https://pytorch.org/docs/stable/elastic/timer.html
pytorch docs
with timer.expires(after=60): # 60 second expiry # do some work In the example above if "trainer_func" takes more than 60 seconds to complete, then the worker process is killed and the agent retries the worker group. Client Methods torch.distributed.elastic.timer.configure(timer_client) Configures a timer client. Must be called before using "expires". torch.distributed.elastic.timer.expires(after, scope=None, client=None) Acquires a countdown timer that expires in "after" seconds from now, unless the code-block that it wraps is finished within the timeframe. When the timer expires, this worker is eligible to be reaped. The exact meaning of "reaped" depends on the client implementation. In most cases, reaping means to terminate the worker process. Note that the worker is NOT guaranteed to be reaped at exactly "time.now() + after", but rather the worker is "eligible" for being reaped and the "TimerServer" that the client
https://pytorch.org/docs/stable/elastic/timer.html
pytorch docs
talks to will ultimately make the decision when and how to reap the workers with expired timers. Usage: torch.distributed.elastic.timer.configure(LocalTimerClient()) with expires(after=10): torch.distributed.all_reduce(...) Server/Client Implementations Below are the timer server and client pairs that are provided by torchelastic. Note: Timer server and clients always have to be implemented and used in pairs since there is a messaging protocol between the server and client. Below is a pair of timer server and client that is implemented based on a "multiprocess.Queue". class torch.distributed.elastic.timer.LocalTimerServer(mp_queue, max_interval=60, daemon=True) Server that works with "LocalTimerClient". Clients are expected to be subprocesses to the parent process that is running this server. Each host in the job is expected to start its own timer server locally and each server instance manages timers for local workers
https://pytorch.org/docs/stable/elastic/timer.html
pytorch docs
(running on processes on the same host). class torch.distributed.elastic.timer.LocalTimerClient(mp_queue) Client side of "LocalTimerServer". This client is meant to be used on the same host that the "LocalTimerServer" is running on and uses pid to uniquely identify a worker. This is particularly useful in situations where one spawns a subprocess (trainer) per GPU on a host with multiple GPU devices. Below is another pair of timer server and client that is implemented based on a named pipe. class torch.distributed.elastic.timer.FileTimerServer(file_path, max_interval=10, daemon=True, log_event=None) Server that works with "FileTimerClient". Clients are expected to be running on the same host as the process that is running this server. Each host in the job is expected to start its own timer server locally and each server instance manages timers for local workers (running on processes on the same host). Parameters:
https://pytorch.org/docs/stable/elastic/timer.html
pytorch docs
Parameters: * file_path (str) -- str, the path of a FIFO special file to be created. * **max_interval** (*float*) -- float, max interval in seconds for each watchdog loop. * **daemon** (*bool*) -- bool, running the watchdog thread in daemon mode or not. A daemon thread will not block a process to stop. * **log_event** (*Callable**[**[**str**, **Optional**[**FileTimerRequest**]**]**, **None**]*) -- Callable[[Dict[str, str]], None], an optional callback for logging the events in JSON format. class torch.distributed.elastic.timer.FileTimerClient(file_path, signal=Signals.SIGKILL) Client side of "FileTimerServer". This client is meant to be used on the same host that the "FileTimerServer" is running on and uses pid to uniquely identify a worker. This client uses a named_pipe to send timer requests to the "FileTimerServer". This client is a producer while the "FileTimerServer" is a consumer. Multiple
https://pytorch.org/docs/stable/elastic/timer.html
pytorch docs
clients can work with the same "FileTimerServer". Parameters: * file_path (str) -- str, the path of a FIFO special file. "FileTimerServer" must have created it by calling os.mkfifo(). * **signal** -- signal, the signal to use to kill the process. Using a negative or zero signal will not kill the process. Writing a custom timer server/client To write your own timer server and client extend the "torch.distributed.elastic.timer.TimerServer" for the server and "torch.distributed.elastic.timer.TimerClient" for the client. The "TimerRequest" object is used to pass messages between the server and client. class torch.distributed.elastic.timer.TimerRequest(worker_id, scope_id, expiration_time) Data object representing a countdown timer acquisition and release that is used between the "TimerClient" and "TimerServer". A negative "expiration_time" should be interpreted as a "release" request. Note:
https://pytorch.org/docs/stable/elastic/timer.html
pytorch docs
request. Note: the type of "worker_id" is implementation specific. It is whatever the TimerServer and TimerClient implementations have on to uniquely identify a worker. class torch.distributed.elastic.timer.TimerServer(request_queue, max_interval, daemon=True) Entity that monitors active timers and expires them in a timely fashion. This server is responsible for reaping workers that have expired timers. abstract clear_timers(worker_ids) Clears all timers for the given "worker_ids". abstract get_expired_timers(deadline) Returns all expired timers for each worker_id. An expired timer is a timer for which the expiration_time is less than or equal to the provided deadline. Return type: *Dict*[str, *List*[*TimerRequest*]] abstract register_timers(timer_requests) Processes the incoming timer requests and registers them with the server. The timer request can either be a acquire-timer or
https://pytorch.org/docs/stable/elastic/timer.html
pytorch docs
release-timer request. Timer requests with a negative expiration_time should be interpreted as a release-timer request. class torch.distributed.elastic.timer.TimerClient Client library to acquire and release countdown timers by communicating with the TimerServer. abstract acquire(scope_id, expiration_time) Acquires a timer for the worker that holds this client object given the scope_id and expiration_time. Typically registers the timer with the TimerServer. abstract release(scope_id) Releases the timer for the "scope_id" on the worker this client represents. After this method is called, the countdown timer on the scope is no longer in effect.
https://pytorch.org/docs/stable/elastic/timer.html
pytorch docs
Train script If your train script works with "torch.distributed.launch" it will continue working with "torchrun" with these differences: No need to manually pass "RANK", "WORLD_SIZE", "MASTER_ADDR", and "MASTER_PORT". "rdzv_backend" and "rdzv_endpoint" can be provided. For most users this will be set to "c10d" (see rendezvous). The default "rdzv_backend" creates a non-elastic rendezvous where "rdzv_endpoint" holds the master address. Make sure you have a "load_checkpoint(path)" and "save_checkpoint(path)" logic in your script. When any number of workers fail we restart all the workers with the same program arguments so you will lose progress up to the most recent checkpoint (see elastic launch). "use_env" flag has been removed. If you were parsing local rank by parsing the "--local_rank" option, you need to get the local rank from the environment variable "LOCAL_RANK" (e.g. "int(os.environ["LOCAL_RANK"])").
https://pytorch.org/docs/stable/elastic/train_script.html
pytorch docs
"int(os.environ["LOCAL_RANK"])"). Below is an expository example of a training script that checkpoints on each epoch, hence the worst-case progress lost on failure is one full epoch worth of training. def main(): args = parse_args(sys.argv[1:]) state = load_checkpoint(args.checkpoint_path) initialize(state) # torch.distributed.run ensures that this will work # by exporting all the env vars needed to initialize the process group torch.distributed.init_process_group(backend=args.backend) for i in range(state.epoch, state.total_num_epochs) for batch in iter(state.dataset) train(batch, state.model) state.epoch += 1 save_checkpoint(state) For concrete examples of torchelastic-compliant train scripts, visit our examples page.
https://pytorch.org/docs/stable/elastic/train_script.html
pytorch docs
TorchElastic Kubernetes Please refer to our GitHub's Kubernetes README for more information on Elastic Job Controller and custom resource definition.
https://pytorch.org/docs/stable/elastic/kubernetes.html
pytorch docs
Automatic Mixed Precision package - torch.amp "torch.amp" provides convenience methods for mixed precision, where some operations use the "torch.float32" ("float") datatype and other operations use lower precision floating point datatype ("lower_precision_fp"): "torch.float16" ("half") or "torch.bfloat16". Some ops, like linear layers and convolutions, are much faster in "lower_precision_fp". Other ops, like reductions, often require the dynamic range of "float32". Mixed precision tries to match each op to its appropriate datatype. Ordinarily, "automatic mixed precision training" with datatype of "torch.float16" uses "torch.autocast" and "torch.cuda.amp.GradScaler" together, as shown in the CUDA Automatic Mixed Precision examples and CUDA Automatic Mixed Precision recipe. However, "torch.autocast" and "torch.cuda.amp.GradScaler" are modular, and may be used separately if desired. As shown in the CPU example section of "torch.autocast",
https://pytorch.org/docs/stable/amp.html
pytorch docs
"automatic mixed precision training/inference" on CPU with datatype of "torch.bfloat16" only uses "torch.autocast". For CUDA and CPU, APIs are also provided separately: "torch.autocast("cuda", args...)" is equivalent to "torch.cuda.amp.autocast(args...)". "torch.autocast("cpu", args...)" is equivalent to "torch.cpu.amp.autocast(args...)". For CPU, only lower precision floating point datatype of "torch.bfloat16" is supported for now. Autocasting Gradient Scaling Autocast Op Reference Op Eligibility CUDA Op-Specific Behavior CUDA Ops that can autocast to "float16" CUDA Ops that can autocast to "float32" CUDA Ops that promote to the widest input type Prefer "binary_cross_entropy_with_logits" over "binary_cross_entropy" CPU Op-Specific Behavior CPU Ops that can autocast to "bfloat16" CPU Ops that can autocast to "float32" CPU Ops that promote to the widest input type Autocasting
https://pytorch.org/docs/stable/amp.html
pytorch docs
Autocasting class torch.autocast(device_type, dtype=None, enabled=True, cache_enabled=None) Instances of "autocast" serve as context managers or decorators that allow regions of your script to run in mixed precision. In these regions, ops run in an op-specific dtype chosen by autocast to improve performance while maintaining accuracy. See the Autocast Op Reference for details. When entering an autocast-enabled region, Tensors may be any type. You should not call "half()" or "bfloat16()" on your model(s) or inputs when using autocasting. "autocast" should wrap only the forward pass(es) of your network, including the loss computation(s). Backward passes under autocast are not recommended. Backward ops run in the same type that autocast used for corresponding forward ops. Example for CUDA Devices: # Creates model and optimizer in default precision model = Net().cuda() optimizer = optim.SGD(model.parameters(), ...)
https://pytorch.org/docs/stable/amp.html
pytorch docs
for input, target in data: optimizer.zero_grad() # Enables autocasting for the forward pass (model + loss) with autocast(): output = model(input) loss = loss_fn(output, target) # Exits the context manager before backward() loss.backward() optimizer.step() See the CUDA Automatic Mixed Precision examples for usage (along with gradient scaling) in more complex scenarios (e.g., gradient penalty, multiple models/losses, custom autograd functions). "autocast" can also be used as a decorator, e.g., on the "forward" method of your model: class AutocastModel(nn.Module): ... @autocast() def forward(self, input): ... Floating-point Tensors produced in an autocast-enabled region may be "float16". After returning to an autocast-disabled region, using them with floating-point Tensors of different dtypes may cause type
https://pytorch.org/docs/stable/amp.html
pytorch docs
mismatch errors. If so, cast the Tensor(s) produced in the autocast region back to "float32" (or other dtype if desired). If a Tensor from the autocast region is already "float32", the cast is a no-op, and incurs no additional overhead. CUDA Example: # Creates some tensors in default dtype (here assumed to be float32) a_float32 = torch.rand((8, 8), device="cuda") b_float32 = torch.rand((8, 8), device="cuda") c_float32 = torch.rand((8, 8), device="cuda") d_float32 = torch.rand((8, 8), device="cuda") with autocast(): # torch.mm is on autocast's list of ops that should run in float16. # Inputs are float32, but the op runs in float16 and produces float16 output. # No manual casts are required. e_float16 = torch.mm(a_float32, b_float32) # Also handles mixed input types f_float16 = torch.mm(d_float32, e_float16) # After exiting autocast, calls f_float16.float() to use with d_float32
https://pytorch.org/docs/stable/amp.html
pytorch docs
g_float32 = torch.mm(d_float32, f_float16.float()) CPU Training Example: # Creates model and optimizer in default precision model = Net() optimizer = optim.SGD(model.parameters(), ...) for epoch in epochs: for input, target in data: optimizer.zero_grad() # Runs the forward pass with autocasting. with torch.autocast(device_type="cpu", dtype=torch.bfloat16): output = model(input) loss = loss_fn(output, target) loss.backward() optimizer.step() CPU Inference Example: # Creates model in default precision model = Net().eval() with torch.autocast(device_type="cpu", dtype=torch.bfloat16): for input in data: # Runs the forward pass with autocasting. output = model(input) CPU Inference Example with Jit Trace: class TestModel(nn.Module): def __init__(self, input_size, num_classes):
https://pytorch.org/docs/stable/amp.html
pytorch docs
super(TestModel, self).init() self.fc1 = nn.Linear(input_size, num_classes) def forward(self, x): return self.fc1(x) input_size = 2 num_classes = 2 model = TestModel(input_size, num_classes).eval() # For now, we suggest to disable the Jit Autocast Pass, # As the issue: https://github.com/pytorch/pytorch/issues/75956 torch._C._jit_set_autocast_mode(False) with torch.cpu.amp.autocast(cache_enabled=False): model = torch.jit.trace(model, torch.randn(1, input_size)) model = torch.jit.freeze(model) # Models Run for _ in range(3): model(torch.randn(1, input_size)) Type mismatch errors in an autocast-enabled region are a bug; if this is what you observe, please file an issue. "autocast(enabled=False)" subregions can be nested in autocast- enabled regions. Locally disabling autocast can be useful, for
https://pytorch.org/docs/stable/amp.html
pytorch docs
example, if you want to force a subregion to run in a particular "dtype". Disabling autocast gives you explicit control over the execution type. In the subregion, inputs from the surrounding region should be cast to "dtype" before use: # Creates some tensors in default dtype (here assumed to be float32) a_float32 = torch.rand((8, 8), device="cuda") b_float32 = torch.rand((8, 8), device="cuda") c_float32 = torch.rand((8, 8), device="cuda") d_float32 = torch.rand((8, 8), device="cuda") with autocast(): e_float16 = torch.mm(a_float32, b_float32) with autocast(enabled=False): # Calls e_float16.float() to ensure float32 execution # (necessary because e_float16 was created in an autocasted region) f_float32 = torch.mm(c_float32, e_float16.float()) # No manual casts are required when re-entering the autocast-enabled region.
https://pytorch.org/docs/stable/amp.html
pytorch docs
torch.mm again runs in float16 and produces float16 output, regardless of input types. g_float16 = torch.mm(d_float32, f_float32) The autocast state is thread-local. If you want it enabled in a new thread, the context manager or decorator must be invoked in that thread. This affects "torch.nn.DataParallel" and "torch.nn.parallel.DistributedDataParallel" when used with more than one GPU per process (see Working with Multiple GPUs). Parameters: * device_type (str, required) -- Whether to use 'cuda' or 'cpu' device * **enabled** (*bool**, **optional*) -- Whether autocasting should be enabled in the region. Default: "True" * **dtype** (*torch_dtype**, **optional*) -- Whether to use torch.float16 or torch.bfloat16. * **cache_enabled** (*bool**, **optional*) -- Whether the weight cache inside autocast should be enabled. Default: "True"
https://pytorch.org/docs/stable/amp.html
pytorch docs
class torch.cuda.amp.autocast(enabled=True, dtype=torch.float16, cache_enabled=True) See "torch.autocast". "torch.cuda.amp.autocast(args...)" is equivalent to "torch.autocast("cuda", args...)" torch.cuda.amp.custom_fwd(fwd=None, *, cast_inputs=None) Helper decorator for "forward" methods of custom autograd functions (subclasses of "torch.autograd.Function"). See the example page for more detail. Parameters: cast_inputs ("torch.dtype" or None, optional, default=None) -- If not "None", when "forward" runs in an autocast-enabled region, casts incoming floating-point CUDA Tensors to the target dtype (non-floating-point Tensors are not affected), then executes "forward" with autocast disabled. If "None", "forward"'s internal ops execute with the current autocast state. Note: If the decorated "forward" is called outside an autocast-enabled region, "custom_fwd" is a no-op and "cast_inputs" has no effect.
https://pytorch.org/docs/stable/amp.html
pytorch docs
torch.cuda.amp.custom_bwd(bwd) Helper decorator for backward methods of custom autograd functions (subclasses of "torch.autograd.Function"). Ensures that "backward" executes with the same autocast state as "forward". See the example page for more detail. class torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16, cache_enabled=True) See "torch.autocast". "torch.cpu.amp.autocast(args...)" is equivalent to "torch.autocast("cpu", args...)" Gradient Scaling If the forward pass for a particular op has "float16" inputs, the backward pass for that op will produce "float16" gradients. Gradient values with small magnitudes may not be representable in "float16". These values will flush to zero ("underflow"), so the update for the corresponding parameters will be lost. To prevent underflow, "gradient scaling" multiplies the network's loss(es) by a scale factor and invokes a backward pass on the scaled loss(es). Gradients flowing backward through the network are then
https://pytorch.org/docs/stable/amp.html
pytorch docs
scaled by the same factor. In other words, gradient values have a larger magnitude, so they don't flush to zero. Each parameter's gradient (".grad" attribute) should be unscaled before the optimizer updates the parameters, so the scale factor does not interfere with the learning rate. class torch.cuda.amp.GradScaler(init_scale=65536.0, growth_factor=2.0, backoff_factor=0.5, growth_interval=2000, enabled=True) get_backoff_factor() Returns a Python float containing the scale backoff factor. get_growth_factor() Returns a Python float containing the scale growth factor. get_growth_interval() Returns a Python int containing the growth interval. get_scale() Returns a Python float containing the current scale, or 1.0 if scaling is disabled. Warning: "get_scale()" incurs a CPU-GPU sync. is_enabled() Returns a bool indicating whether this instance is enabled. load_state_dict(state_dict)
https://pytorch.org/docs/stable/amp.html
pytorch docs
load_state_dict(state_dict) Loads the scaler state. If this instance is disabled, "load_state_dict()" is a no-op. Parameters: **state_dict** (*dict*) -- scaler state. Should be an object returned from a call to "state_dict()". scale(outputs) Multiplies ('scales') a tensor or list of tensors by the scale factor. Returns scaled outputs. If this instance of "GradScaler" is not enabled, outputs are returned unmodified. Parameters: **outputs** (*Tensor** or **iterable of Tensors*) -- Outputs to scale. set_backoff_factor(new_factor) Parameters: **new_scale** (*float*) -- Value to use as the new scale backoff factor. set_growth_factor(new_factor) Parameters: **new_scale** (*float*) -- Value to use as the new scale growth factor. set_growth_interval(new_interval) Parameters: **new_interval** (*int*) -- Value to use as the new growth
https://pytorch.org/docs/stable/amp.html
pytorch docs
interval. state_dict() Returns the state of the scaler as a "dict". It contains five entries: * ""scale"" - a Python float containing the current scale * ""growth_factor"" - a Python float containing the current growth factor * ""backoff_factor"" - a Python float containing the current backoff factor * ""growth_interval"" - a Python int containing the current growth interval * ""_growth_tracker"" - a Python int containing the number of recent consecutive unskipped steps. If this instance is not enabled, returns an empty dict. Note: If you wish to checkpoint the scaler's state after a particular iteration, "state_dict()" should be called after "update()". step(optimizer, args, *kwargs) "step()" carries out the following two operations: 1. Internally invokes "unscale_(optimizer)" (unless "unscale_()" was explicitly called for "optimizer" earlier in the
https://pytorch.org/docs/stable/amp.html
pytorch docs
iteration). As part of the "unscale_()", gradients are checked for infs/NaNs. 2. If no inf/NaN gradients are found, invokes "optimizer.step()" using the unscaled gradients. Otherwise, "optimizer.step()" is skipped to avoid corrupting the params. "*args" and "**kwargs" are forwarded to "optimizer.step()". Returns the return value of "optimizer.step(*args, **kwargs)". Parameters: * **optimizer** (*torch.optim.Optimizer*) -- Optimizer that applies the gradients. * **args** -- Any arguments. * **kwargs** -- Any keyword arguments. Warning: Closure use is not currently supported. unscale_(optimizer) Divides ("unscales") the optimizer's gradient tensors by the scale factor. "unscale_()" is optional, serving cases where you need to modify or inspect gradients between the backward pass(es) and "step()". If "unscale_()" is not called explicitly, gradients will be
https://pytorch.org/docs/stable/amp.html
pytorch docs
unscaled automatically during "step()". Simple example, using "unscale_()" to enable clipping of unscaled gradients: ... scaler.scale(loss).backward() scaler.unscale_(optimizer) torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm) scaler.step(optimizer) scaler.update() Parameters: **optimizer** (*torch.optim.Optimizer*) -- Optimizer that owns the gradients to be unscaled. Note: "unscale_()" does not incur a CPU-GPU sync. Warning: "unscale_()" should only be called once per optimizer per "step()" call, and only after all gradients for that optimizer's assigned parameters have been accumulated. Calling "unscale_()" twice for a given optimizer between each "step()" triggers a RuntimeError. Warning: "unscale_()" may unscale sparse gradients out of place, replacing the ".grad" attribute. update(new_scale=None)
https://pytorch.org/docs/stable/amp.html
pytorch docs
update(new_scale=None) Updates the scale factor. If any optimizer steps were skipped the scale is multiplied by "backoff_factor" to reduce it. If "growth_interval" unskipped iterations occurred consecutively, the scale is multiplied by "growth_factor" to increase it. Passing "new_scale" sets the new scale value manually. ("new_scale" is not used directly, it's used to fill GradScaler's internal scale tensor. So if "new_scale" was a tensor, later in-place changes to that tensor will not further affect the scale GradScaler uses internally.) Parameters: **new_scale** (float or "torch.cuda.FloatTensor", optional, default=None) -- New scale factor. Warning: "update()" should only be called at the end of the iteration, after "scaler.step(optimizer)" has been invoked for all optimizers used this iteration. Autocast Op Reference Op Eligibility
https://pytorch.org/docs/stable/amp.html
pytorch docs
Op Eligibility Ops that run in "float64" or non-floating-point dtypes are not eligible, and will run in these types whether or not autocast is enabled. Only out-of-place ops and Tensor methods are eligible. In-place variants and calls that explicitly supply an "out=..." Tensor are allowed in autocast-enabled regions, but won't go through autocasting. For example, in an autocast-enabled region "a.addmm(b, c)" can autocast, but "a.addmm_(b, c)" and "a.addmm(b, c, out=d)" cannot. For best performance and stability, prefer out-of-place ops in autocast- enabled regions. Ops called with an explicit "dtype=..." argument are not eligible, and will produce output that respects the "dtype" argument. CUDA Op-Specific Behavior The following lists describe the behavior of eligible ops in autocast- enabled regions. These ops always go through autocasting whether they are invoked as part of a "torch.nn.Module", as a function, or as a
https://pytorch.org/docs/stable/amp.html
pytorch docs
"torch.Tensor" method. If functions are exposed in multiple namespaces, they go through autocasting regardless of the namespace. Ops not listed below do not go through autocasting. They run in the type defined by their inputs. However, autocasting may still change the type in which unlisted ops run if they're downstream from autocasted ops. If an op is unlisted, we assume it's numerically stable in "float16". If you believe an unlisted op is numerically unstable in "float16", please file an issue. CUDA Ops that can autocast to "float16" "__matmul__", "addbmm", "addmm", "addmv", "addr", "baddbmm", "bmm", "chain_matmul", "multi_dot", "conv1d", "conv2d", "conv3d", "conv_transpose1d", "conv_transpose2d", "conv_transpose3d", "GRUCell", "linear", "LSTMCell", "matmul", "mm", "mv", "prelu", "RNNCell" CUDA Ops that can autocast to "float32" "pow", "rdiv", "rpow", "rtruediv", "acos", "asin",
https://pytorch.org/docs/stable/amp.html
pytorch docs
"binary_cross_entropy_with_logits", "cosh", "cosine_embedding_loss", "cdist", "cosine_similarity", "cross_entropy", "cumprod", "cumsum", "dist", "erfinv", "exp", "expm1", "group_norm", "hinge_embedding_loss", "kl_div", "l1_loss", "layer_norm", "log", "log_softmax", "log10", "log1p", "log2", "margin_ranking_loss", "mse_loss", "multilabel_margin_loss", "multi_margin_loss", "nll_loss", "norm", "normalize", "pdist", "poisson_nll_loss", "pow", "prod", "reciprocal", "rsqrt", "sinh", "smooth_l1_loss", "soft_margin_loss", "softmax", "softmin", "softplus", "sum", "renorm", "tan", "triplet_margin_loss" CUDA Ops that promote to the widest input type ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ These ops don't require a particular dtype for stability, but take multiple inputs and require that the inputs' dtypes match. If all of the inputs are "float16", the op runs in "float16". If any of the inputs is "float32", autocast casts all inputs to "float32" and runs the op in "float32".
https://pytorch.org/docs/stable/amp.html
pytorch docs
the op in "float32". "addcdiv", "addcmul", "atan2", "bilinear", "cross", "dot", "grid_sample", "index_put", "scatter_add", "tensordot" Some ops not listed here (e.g., binary ops like "add") natively promote inputs without autocasting's intervention. If inputs are a mixture of "float16" and "float32", these ops run in "float32" and produce "float32" output, regardless of whether autocast is enabled. Prefer "binary_cross_entropy_with_logits" over "binary_cross_entropy" ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The backward passes of "torch.nn.functional.binary_cross_entropy()" (and "torch.nn.BCELoss", which wraps it) can produce gradients that aren't representable in "float16". In autocast-enabled regions, the forward input may be "float16", which means the backward gradient must be representable in "float16" (autocasting "float16" forward inputs to "float32" doesn't help, because that cast must be reversed in backward). Therefore, "binary_cross_entropy" and "BCELoss" raise an
https://pytorch.org/docs/stable/amp.html
pytorch docs
error in autocast-enabled regions. Many models use a sigmoid layer right before the binary cross entropy layer. In this case, combine the two layers using "torch.nn.functional.binary_cross_entropy_with_logits()" or "torch.nn.BCEWithLogitsLoss". "binary_cross_entropy_with_logits" and "BCEWithLogits" are safe to autocast. CPU Op-Specific Behavior The following lists describe the behavior of eligible ops in autocast- enabled regions. These ops always go through autocasting whether they are invoked as part of a "torch.nn.Module", as a function, or as a "torch.Tensor" method. If functions are exposed in multiple namespaces, they go through autocasting regardless of the namespace. Ops not listed below do not go through autocasting. They run in the type defined by their inputs. However, autocasting may still change the type in which unlisted ops run if they're downstream from autocasted ops. If an op is unlisted, we assume it's numerically stable in "bfloat16".
https://pytorch.org/docs/stable/amp.html
pytorch docs
If you believe an unlisted op is numerically unstable in "bfloat16", please file an issue. CPU Ops that can autocast to "bfloat16" ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ "conv1d", "conv2d", "conv3d", "bmm", "mm", "baddbmm", "addmm", "addbmm", "linear", "matmul", "_convolution" CPU Ops that can autocast to "float32" ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ "conv_transpose1d", "conv_transpose2d", "conv_transpose3d", "avg_pool3d", "binary_cross_entropy", "grid_sampler", "grid_sampler_2d", "_grid_sampler_2d_cpu_fallback", "grid_sampler_3d", "polar", "prod", "quantile", "nanquantile", "stft", "cdist", "trace", "view_as_complex", "cholesky", "cholesky_inverse", "cholesky_solve", "inverse", "lu_solve", "orgqr", "inverse", "ormqr", "pinverse", "max_pool3d", "max_unpool2d", "max_unpool3d", "adaptive_avg_pool3d", "reflection_pad1d", "reflection_pad2d", "replication_pad1d", "replication_pad2d", "replication_pad3d", "mse_loss", "ctc_loss", "kl_div", "multilabel_margin_loss", "fft_fft", "fft_ifft", "fft_fft2",
https://pytorch.org/docs/stable/amp.html
pytorch docs