U Kc",@sddlZddlmZddlmZmZmZmZddlm Z ddl Z ddl Z ddl Z ddl Z ddlZ ddlZ ddlZ ddlZ ddlmZddlmZddlmZeddd Zedd d ZGd d d e jjZGddde jjZddZddddZ GdddeZ!dS)N) nullcontext)AnyCallableDictSequence)warn)NvfuserPrimOperatorSupport)torch_function_passthrough)get_isolated_graphmodulec*Csttjftjtjjftjjtjjjftjtjjftjtjjftjtjjfg}tjjtjj tjj tjj tjj tjj tjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjjtjj tjj tjj!tjj!tjj"tjj"tjj#tjj#tjj$tjj$tjj%tjj%tjj&tj'j(tjj)tj'j)i}|D]0\}}|j*D]}|j+,|||j+,|<q^qPt-tjD].}|tjj*krtjj+,||t.tj|<q|S)z} Mapping of torch API functions to torch._refs functions. E.g. torch_to_refs_map()[torch.add] == torch._refs.add )/torchZ_refsnnZ functionalspecialZfftZlinalgZTensor __invert__Z bitwise_not__xor__Z bitwise_xor__and__Z bitwise_and__or__Z bitwise_or__eq__eq__rsub__Zrsub __rtruediv__Zrtruediv __floordiv__Z floor_divide __rfloordiv__Z rfloordiv__pow__pow__rpow__ZrpowZ new_emptyZnew_fullZ new_zerosZnew_onesZfill_Zzero_toZ sum_to_sizeZcopy__primsZcopy_toresize__all____dict__getdirgetattr)modulesrZ mod_torchZmod_refssr&8/tmp/pip-unpacked-wheel-gikjz4vx/torch/_prims/context.pytorch_to_refs_mapsr        r(cCsddtjjDS)zJ Set of all prim functions, e.g., torch._prims.add in all_prims() cSsh|]}tjj|qSr&)r rrr ).0r%r&r&r' Lszall_prims..)r rrr&r&r&r' all_primsGsr+c@s8eZdZdZddddZd eeeeeddd Z dS) NvfuserPrimsModea` Switches the interpretation of torch.ops.prims.* functions to use nvFuser's prims in torch.ops.nvprims.* >>> # xdoctest: +SKIP("undefined vars") >>> with NvfuserPrimsMode(): ... torch.ops.prims.add(x, y) # calls torch.ops.nvprims.add(x, y) By default, this context manager will fall back on the torch.ops.prims* if the nvprim does not exist. It's possible to skip certain prims by passing their names to the skip_ops argument. skip_ops is expected to be a sequence of strings, e.g., ["prims.add.default"] In order to check the expected name of a prim, one can use the `torch.overrides.resolve_name`. >>> # xdoctest: +SKIP("undefined vars") >>> with NvfuserPrimsMode(skips_ops=("prims.add.default")): ... torch.ops.prims.add.default(x, y) # does not call torch.ops.nvprims.add.default(x, y) r&skip_opscCs ||_dSNr-selfr.r&r&r'__init__dszNvfuserPrimsMode.__init__N orig_functypesargskwargscCs|dkr i}tj||jkr(|||St|tjjsDt|tjjrt| dd}t| dd}|dkrt tj j |d}|dk r|||S|||S)N.rZprims) r overrides resolve_namer. isinstance_ops OpOverloadOpOverloadPacketstrsplitr"opsnvprims)r1r4r5r6r7 namespacenameZnvfuncr&r&r'__torch_function__gs  z#NvfuserPrimsMode.__torch_function__)r&N) __name__ __module__ __qualname____doc__r2rrrrrFr&r&r&r'r,Osr,c@s>eZdZdZdddefddZd eeeee d d d Z dS) TorchRefsModeaB Switches the interpretation of torch.* functions and Tensor methods to use PrimTorch refs in torch._refs. (Direct calls to _refs are unaffected.) >>> # xdoctest: +SKIP >>> with TorchRefsMode(): ... torch.add(x, y) # calls torch._refs.add(x, y) By default, this context manager will fall back on the torch.* if the ref does not exist; set strict=True to error if this occurs. If the ref exists we still would like to fall back on the torch.* sometimes, this behavior can be customized by passing a function to should_fallback_fn. FcGsdS)NFr&)_r&r&r'zTorchRefsMode.cCs||_||_||_dSr/strictshould_fallback_fnprims_mode_cls)r1rPrQrRr&r&r'r2szTorchRefsMode.__init__r&Nr3c Cs|dkr i}|tks|tkrD||||W5QRSQRXt}||d}|dkr|t|tjjr|tj j |d}|dk r| |||||r|||S||||W5QRSQRX|j rt dtj||||S)Nzno _refs support for )r r+rRr(r r<r r=r>_decompZdecomposition_tablerQrP RuntimeErrorr:r;)r1r4r5r6r7mappingfuncr&r&r'rFs&   z TorchRefsMode.__torch_function__)r&N) rGrHrIrJrr2rrrrrFr&r&r&r'rKs rKcCs|jdkot|jdddk S)N call_functionZ impl_nvfuser)opr"target)noder&r&r'_is_node_supported_nvfusers r[r&r-c stj||krdS|dzt|||}WnLtk rv}z.td|jdt|WYW5QRdSd}~XYnXW5QRXtt dd|j j }t fdd|D} | S) a This function traces the `func` under `torch_function_mode` and checks if any of the traced nodes are not supported by nvFuser. If so, we should fallback to the original function. `skip_ops` argument is expected to be a list of strings of function names that would match with `torch.overrides.resolve_name`. Args: torch_function_mode: The torch_function_mode context manager. orig_func: The original function, its name will be used to check if it should be skipped. func: The function to be traced. args: The args to be passed to the function. kwargs: The kwargs to be passed to the function. Keyword args: skip_ops: A list of ops to skip when checking if the function is supported. Tz2get_isolated_graphmodule failed on decomposition: z with error message: NcSs |jdkS)NrW)rX)nr&r&r'rMrNz._is_func_unsupported_nvfuser..c3s|]}d| VqdSr/)Zis_node_supported)r)rZZ supported_opsr&r' sz/_is_func_unsupported_nvfuser..) r r:r;r ExceptionrrGr@rfiltergraphZnodesany) Ztorch_function_moder4rVr6r7r.ZgmeZcall_function_nodesZany_unsupportedr&r]r'_is_func_unsupported_nvfusers,, rdcsXeZdZddfdd ZddZddZd d Zdeeee e d fd d Z Z S)TorchRefsNvfuserCapabilityModer&r-cs*tjdtjt|dtjt|dddS)NFr-rO)superr2 functoolspartialrdr,r0 __class__r&r'r2s z'TorchRefsNvfuserCapabilityMode.__init__cCs8dtj|kp6t|tjjs,t|tjjo6dt|kS)Nztorch.var_meanz aten.var_mean)r r:r;r<r=r>r?r@r1rVr&r&r' _is_var_means   z+TorchRefsNvfuserCapabilityMode._is_var_meancCs.dtj|kp,|tjjjjkp,|tjjjkS)Nztorch.native_batch_norm)r r:r;rBatennative_batch_normdefaultrkr&r&r'_is_native_batch_norm s z4TorchRefsNvfuserCapabilityMode._is_native_batch_normcCs2dtj|kp,|tjjjkp,|tjjjjk}|S)Nztorch.rand_like)r r:r;rBrm rand_likero)r1rVresultr&r&r' _is_rand_likesz,TorchRefsNvfuserCapabilityMode._is_rand_likeNr3cs~|dkr i}||r&tjjj||S||r@tjjj||S||rlt|dkr^t dtjjj |St ||||S)Nrzrand_like has ignored kwars!) rlr rBrCZvar_meanrprnrslenrrqrfrF)r1r4r5r6r7rir&r'rFs    z1TorchRefsNvfuserCapabilityMode.__torch_function__)r&N) rGrHrIr2rlrprsrrrrrF __classcell__r&r&rir'res   re)"rg contextlibrtypingrrrrwarningsrr Z torch._decompZ torch._primsZ torch._refsZtorch._refs.nnZtorch._refs.nn.functionalZtorch._refs.specialZtorch.overridesZtorch._prims.nvfuser_executorrZtorch._prims_commonr Z"torch.fx.experimental.proxy_tensorr lru_cacher(r+r:ZTorchFunctionModer,rKr[rdrer&r&r&r's0      0 3A /