torch.compile (original) (raw)

• model (Callable) – Module/function to optimize
• fullgraph (bool) – If False (default), torch.compile attempts to discover compileable regions in the function that it will optimize. If True, then we require that the entire function be capturable into a single graph. If this is not possible (that is, if there are graph breaks), then this will raise an error.
• dynamic (bool or None) – Use dynamic shape tracing. When this is True, we will up-front attempt to generate a kernel that is as dynamic as possible to avoid recompilations when sizes change. This may not always work as some operations/optimizations will force specialization; use TORCH_LOGS=dynamic to debug overspecialization. When this is False, we will NEVER generate dynamic kernels, we will always specialize. By default (None), we automatically detect if dynamism has occurred and compile a more dynamic kernel upon recompile.
• backend (str or Callable) –
  backend to be used
  • ”inductor” is the default backend, which is a good balance between performance and overhead
  • Non experimental in-tree backends can be seen with torch._dynamo.list_backends()
  • Experimental or debug in-tree backends can be seen with torch._dynamo.list_backends(None)
  • To register an out-of-tree custom backend:https://pytorch.org/docs/main/torch.compiler_custom_backends.html#registering-custom-backends
• mode (str) –
  Can be either “default”, “reduce-overhead”, “max-autotune” or “max-autotune-no-cudagraphs”
  • ”default” is the default mode, which is a good balance between performance and overhead
  • ”reduce-overhead” is a mode that reduces the overhead of python with CUDA graphs, useful for small batches. Reduction of overhead can come at the cost of more memory usage, as we will cache the workspace memory required for the invocation so that we do not have to reallocate it on subsequent runs. Reduction of overhead is not guaranteed to work; today, we only reduce overhead for CUDA only graphs which do not mutate inputs. There are other circumstances where CUDA graphs are not applicable; use TORCH_LOG=perf_hints to debug.
  • ”max-autotune” is a mode that leverages Triton or template based matrix multiplications on supported devices and Triton based convolutions on GPU. It enables CUDA graphs by default on GPU.
  • ”max-autotune-no-cudagraphs” is a mode similar to “max-autotune” but without CUDA graphs
  • To see the exact configs that each mode sets you can call torch._inductor.list_mode_options()
• options (dict) –
  A dictionary of options to pass to the backend. Some notable ones to try out are
  • epilogue_fusion which fuses pointwise ops into templates. Requires max_autotune to also be set
  • max_autotune which will profile to pick the best matmul configuration
  • fallback_random which is useful when debugging accuracy issues
  • shape_padding which pads matrix shapes to better align loads on GPUs especially for tensor cores
  • triton.cudagraphs which will reduce the overhead of python with CUDA graphs
  • trace.enabled which is the most useful debugging flag to turn on
  • trace.graph_diagram which will show you a picture of your graph after fusion
  • For inductor you can see the full list of configs that it supports by calling torch._inductor.list_options()
• disable (bool) – Turn torch.compile() into a no-op for testing