torch.compile

torch._inductor The default compiler in torch.compile, builds loops schedules and codegens triton or C++ kernels. FX (Functional Transformations) FX is a toolkit for developers to use to transform nn.Module instances. FX consists of three main components: a symbolic tracer, an intermediate representation , and Python code generation . torch._dynamo PyTorch’s compiler frontend which can output subgraphs when programs aren’t traceable.

Topic	Replies	Views	Activity
About the torch.compile category torch.compile	0	1098	January 9, 2023
A node type in export IR graph FX (Functional Transformations)	1	6	November 28, 2024
Scaled_dot_product_attention higher head num cost much more memory torch.compile	1	10	November 28, 2024
CUDA memory allocation for result tensor torch.compile	0	7	November 26, 2024
Compile and vmap in custom op with quantile torch.compile	0	14	November 25, 2024
Compiling vmapped custom op torch.compile	5	19	November 25, 2024
Closures are being gc'd and causing failures to compile torch.compile	1	16	November 24, 2024
Why does the inductor reduction Triton Codegen use the Welford algorithm instead of the Naive? torch.compile	1	18	November 20, 2024
Image_process.postprocess slow after torch.compile torch.compile	0	17	November 18, 2024
Compiling a method other than forward torch.compile	2	21	November 19, 2024
Error module torchvision in CUDA 11.4 torch.compile	2	16	November 19, 2024
Increased memory footprint with custom kernel and all reduce torch.compile	2	22	November 18, 2024
The forward graphs captured by torch.export and aot_export_module are different torch._dynamo	2	31	November 17, 2024
Dynamic slicing torch.export torch._dynamo	2	23	November 16, 2024
Discrepancies Between Compiled and Non-Compiled Models with Convolutional Layers in PyTorch torch.compile	1	38	November 16, 2024
Any chance to preserve some ops while decomposing PT2E model? FX (Functional Transformations)	1	13	November 16, 2024
Multiple compiled versions of the same model torch.compile	2	33	November 16, 2024
Torch.compile - what is the best scope of compilation? torch.compile	7	2235	November 16, 2024
Dynamo Trace with Parameter Lifting torch._dynamo	1	22	November 16, 2024
AOTInductor autograd support torch.compile	1	13	November 16, 2024
Is it possible to ignore part of the code for torch compile torch.compile	4	31	November 16, 2024
Inconsistent Results with torch.compile on Identical Environments and GPUs torch.compile	3	54	November 8, 2024
Torch compile with forward-mode automatic differentiation torch.compile	2	15	November 6, 2024
Profiling torch.compile CUDA code torch.compile	3	29	November 5, 2024
Handling LSTM states as model's inputs/outputs using fx.symbolic_trace FX (Functional Transformations)	0	9	November 5, 2024
How to limit torch.compile to CPU only? torch._dynamo	3	78	November 1, 2024
Does max-autotune is useful for A5000? torch.compile	0	11	October 31, 2024
What caused matrix inputs to no longer be transposed in PyTorch 2.5? torch._inductor	0	14	October 29, 2024
Torch.fx.symbolic_trace with multiple GPUs FX (Functional Transformations)	1	21	October 29, 2024
CUDA 12.6 and torch.cuda.is_available return false torch.compile	7	533	October 28, 2024