U d@sddlmZddlZddlmZddlmZmZddlm Z ddl Zej r^ddl mZmZdde ed d d Zed d dZdS))CallableN) GraphModule) TensorMetagetnvFuserDtype) PrimContext)FusionFusionDefinitionatenexecutor)ctxr c Osb|dkr ti|j}|j||S|dkrLtjs@s zexecute..zQReceived unexpected value for 'executor': {0}. Allowed values are: aten, nvfuser.N) rgraphZforwardrcuda is_available RuntimeErrorlenAssertionError placeholder_rootappendrrrrZ define_tensorsizeZstriderZdtypeZ add_inputZnodesoptargetZ impl_nvfuserargsZ add_outputexecutetupleformat ValueError) r r r#kwargsZgmZgraph_fdZfusionr Znv_argsrxoutmsgrrrr$sP                r$fncsddfdd }|S)a Returns a function that, when called, will trace its torch operations to prims and then execute those prims on the requested trace executor (possibly lowering them to that trace executor first). Only supports the torch operations defined in _torch_to_reference_map in context.py and operations with positional args. All args must be tensors and the function must return a single tensor. In the near future all these restrictions will be lifted. Example usage: def foo(a, b): return torch.add(a, b) traced_foo = make_traced(foo) a = torch.randn((1, 2, 3, 4, 5), device='cuda') b = torch.randn((1, 2, 3, 4, 5), device='cuda') result = traced_foo(a, b, executor='nvfuser') Executor may be either 'aten' or 'nvfuser'. r r c s~tjtX}g}|D]6}t|tjr<||t|q|||q|}| |W5QRXt |f|d|iS)Nr ) rZ overridesZpush_torch_function_moderrrrrroutputr$)r r#r Z placeholdersrresultr,rr_tracedcs zmake_traced.._tracedr)r-r0rr,r make_tracedIsr1)typingrrZtorch.fxrZtorch._prims.utilsrrZtorch._prims.contextrZtorch.overridesrrZtorch._C._nvfuserrrstrr$r1rrrrs    ;