TunableOp#
Created On: Jun 03, 2024 | Last Updated: Mar 07, 2025
Overview#
This module exposes a TunableOp interface.
Some operations, such as GEMMs, could be implemented using more than one library or more than one technique. For example, a GEMM could be implemented for CUDA or ROCm using either the blas or blasLt libraries. Further, ROCm’s rocblas and hipblaslt libraries allow the user to query for all possible algorithms and then choose one. How does one know which implementation is the fastest and should be chosen? That’s what TunableOp provides.
Enabling TunableOp and Tuning Separately#
The TunableOp feature is enabled separately from enabling the tuning phase itself. Enabling TunableOp means that PyTorch will replace any standard operators with their Tunable implementations. Any call to a TunableOp first checks whether it has already been tuned for the given operator inputs. If so, it will immediately call the tuned operation; no further tuning will take place even when the tuning setting is enabled. Instead if no tuning result is found, and tuning is enabled, the TunableOp will benchmark every registered implementation of that operator for the given set of inputs and select the fastest.
File Input and Output#
The first time any TunableOp is invoked, the internal database of tuned operations will be prepared by attempting to read the results from the given file. The default filename is ‘tunableop_results.csv’. To support tuning when multiple GPUs are used across multiple processes, the GPU device ordinal is automatically inserted into the filename to avoid multiple processes overwriting the same file.
If tuning is enabled and new tunings are discovered during the course of your workload, it will also write out to this same filename with all tunings, both the ones it read in at startup as well as the new ones found at runtime. This can be used, for example, to build up a tunings file across many workloads by reusing the same file. The output file is automatically created when the application terminates. This behavior can be controlled by the C++ and Python APIs but not the environment variables.
Assuming you specified a filename, you’ll end up with a CSV file with contents like so:
Validator,PT_VERSION,2.2.0
Validator,ROCM_VERSION,6.0.0.0-12969-1544e39
Validator,HIPBLASLT_VERSION,0.6.0-a9c5cc7
Validator,ROCBLAS_VERSION,4.0.0-72e57364-dirty
GemmTunableOp_float_NT,nt_25088_4096_64,Gemm_Hipblaslt_1219,1.262
GemmTunableOp_float_NT,nt_4096_4096_64,Gemm_Rocblas_1216,0.033
Note the “Validator” lines. If you change a library version, or ROCm version, or PyTorch version, TunableOp will detect this and reject the tunings file because the prior tunings are likely affected by other software changes.
The remaining lines are the tuned solutions for each TunableOp encountered during your execution. Each line consists of 4 comma-separated fields: operator name, operator parameters, solution name, and average execution time. The execution time is an optional field. The CSV file can be edited, but with caution. For example, the solution name (field 3) can be changed to “Default” and it will fall back to the original PyTorch untuned implementation. Or, in the case of ROCm’s hipBLAS or hipBLASLt libraries, if you know the specific solution index you can override the solution that TunableOp selected by replacing the value. The operator name and parameters (fields 1 and 2) are internally named and should not be modified. In the case of GemmTunableOp, field 1 indicates the datatype and whether the inputs are transposed (T) or not (N) and field 2 indicates the M, N, K input shapes.
There is an option to enable verbose output but it is only recommended for debugging purposes. This will produce a lot of diagnostic messages but may be useful to see if TunableOp is being used at all. Otherwise, TunableOp is completely silent, besides file output, unless there is a warning or error during its use. The verbose option is only available by setting the environment variable PYTORCH_TUNABLEOP_VEROBSE=1.
A Note on Tuning Behavior, Warmup, and Cache Effects#
Tuning an operator consists of iterating through the list or registered implementations and profiling each one. The profile is established by running a single implementation in a loop multiple times and taking the average execution time. There is also an optional warmup phase prior to tuning that can help with reaching stable power states by the hardware. During tuning of a workload the various hardware caches will more likely produce hits than when not tuning. There are options for flushing the instruction cache and rotate the input tensors which might help produce a more faithful profile of the tuned operator as if the operator were run within a larger workload instead of in a tight, repetitive loop.
By default, each possible solution for a given operator will be run for either 100 iterations or as many iterations that can be run within 30ms, whichever is smaller, and its average execution will be calculated. The fastest solution among all that were successfully profiled will be chosen. A profile might fail if the given solution doesn’t achieve the same accuracy as the default implementation or if the solution returns an error code.
Current Tunable Operators#
TunableGemm for ROCm#
Currently only a TunableGemm for ROCm is implemented. Note that CUDA builds of PyTorch will function correctly when using TunableOp but the only solution available to CUDA builds is the ‘Default’ implementation i.e. the original cuBLAS default, now called through TunableOp. Any call to at::cuda::blas::gemm() or ::bgemm() will be routed through TunableOp when enabled. Calling gemm() for a given set of input arguments (transa, transb, m, n, k) will attempt to use the fastest available implementation across both rocblas and hipblaslt.
Offline Tuning#
Motivation#
There are several use cases for offline tuning.
One use case involves a workload with a high-memory utilization, where regular tuning might lead to running out of memory.
Another use case is for compute-intensive workloads. In such cases, it is more resource-efficient to collect the GEMMs for the workload once and then tune repeatedly with different tuning parameters or libraries.
Workflow#
There are basically two steps:
1) Set the environment variables to collect the untuned GEMM and this will generate tunableop_untuned0.csv:
PYTORCH_TUNABLEOP_ENABLED=1
PYTORCH_TUNABLEOP_TUNING=0
PYTORCH_TUNABLEOP_RECORD_UNTUNED=1
...
Run a Python script that reads the
tunableop_untuned0.csvand generates thetunableop_results0.csv, like this:
import torch.cuda.tunable as tunable
import os
os.putenv('PYTORCH_TUNABLEOP_ENABLED', '1')
os.putenv('PYTORCH_TUNABLEOP_TUNING', '1')
os.putenv('PYTORCH_TUNABLEOP_RECORD_UNTUNED', '0')
tunable.tune_gemm_in_file("tunableop_untuned0.csv")
It is also possible to take multiple untuned files and distribute the GEMMs for tuning to multiple GPUs
within a single node. In the first step, the GEMMs are first gathered and duplicate GEMMs are eliminated.
Next, the GEMMs are distributed to different GPUs for tuning. After all GEMMs are tuned, the results from
all the GPUs are then gathered into a single file whose base filename has _full0 appended to it
(for example tunableop_results_full0.csv). Finally, this new file, containing the gathered results, will be
duplicated N times, once for each GPU as convenience to the user will run the workload with the tuned
configuration on N GPUs.
if __name__ == "__main__":
num_gpus = 8 # number of GPUs that will be used during the tuning process
tunable.mgpu_tune_gemm_in_file("tunableop_untuned?.csv", num_gpus)
Note that the usage of the mgpu_tune_gemm_in_file API is different from its single GPU counterpart
(tune_gemm_in_file). The body of the Python script that calls the API must be wrapped in main() as shown
due to the use of concurrent futures module. The argument to mgpu_tune_gemm_in_file must contain a wild card
expression (? or *) to generate the list of untuned files containing the GEMMs to be processed. The num_gpus
must between 1 and the total number of GPUs available.
Tuning Context#
The behavior of TunableOp is currently manipulated through environment variables, the C++ interface of at::cuda::tunable::getTuningContext(), or the torch.cuda.tunable python interfaces. The environment variables take precedence over any setting you manipulate using the C++ or Python APIs.
Environment Variable Interface#
Environment variables are cached the first time they are read. You cannot use the environment variable interface programmatically since the settings become fixed. Use the C++ or Python APIs instead.
API Reference#
- torch.cuda.tunable.enable(val=True)[source]#
This is the big on/off switch for all TunableOp implementations.
- torch.cuda.tunable.is_enabled()[source]#
Returns whether the TunableOp feature is enabled.
- Return type
- torch.cuda.tunable.tuning_enable(val=True)[source]#
Enable tuning of TunableOp implementations.
When enabled, if a tuned entry isn’t found, run the tuning step and record the entry.
- torch.cuda.tunable.tuning_is_enabled()[source]#
Returns whether TunableOp implementations can be tuned.
- Return type
- torch.cuda.tunable.record_untuned_enable(val=True)[source]#
Enable recording untuned of TunableOp perations for offline tuning.
When enabled, if a tuned entry isn’t found, write it to the untuned file.
- torch.cuda.tunable.record_untuned_is_enabled()[source]#
Returns whether TunableOp operations are recorded for offline tuning.
- Return type
- torch.cuda.tunable.set_max_tuning_duration(duration)[source]#
Set max time in milliseconds to spend tuning a given solution.
If both max tuning duration and iterations are set, the smaller of the two will be honored. At minimum 1 tuning iteration will always be run.
- torch.cuda.tunable.get_max_tuning_duration()[source]#
Get max time to spend tuning a given solution.
- Return type
- torch.cuda.tunable.set_max_tuning_iterations(iterations)[source]#
Set max number of iterations to spend tuning a given solution.
If both max tuning duration and iterations are set, the smaller of the two will be honored. At minimum 1 tuning iteration will always be run.
- torch.cuda.tunable.get_max_tuning_iterations()[source]#
Get max iterations to spend tuning a given solution.
- Return type
- torch.cuda.tunable.set_filename(filename, insert_device_ordinal=False)[source]#
Set the filename to use for input/output of tuning results.
If
insert_device_ordinalisTruethen the current device ordinal will be added to the given filename automatically. This can be used in a 1-process-per-gpu cenario to ensure all processes write to a separate file.
- torch.cuda.tunable.write_file_on_exit(val)[source]#
During Tuning Context destruction, write file to disk.
This is useful as a final flush of your results to disk if your application terminates as result of normal operation or an error. Manual flushing of your results can be achieved by manually calling
write_file().
- torch.cuda.tunable.write_file(filename=None)[source]#
Write results to a CSV file.
If
filenameis not given,get_filename()is called.- Return type
- torch.cuda.tunable.read_file(filename=None)[source]#
Read results from a TunableOp CSV file.
If
filenameis not given,get_filename()is called.- Return type
- torch.cuda.tunable.mgpu_tune_gemm_in_file(filename_pattern, num_gpus)[source]#
Process one or more files and distribute work over one or more GPUs.