hxSrSSKJr SSKrSSKJr SSKJr SSKJ r J r J r J r SSK Jr S\ RS\R SS4S \R"S \R$S \ S \S \ S\S\\S\S\ 4SjjrS\S\S\\4Sjr\R.R0"SS\R2R455rg)z9 Utilities for scaling high precision tensors to float8. )OptionalN)ScalingGranularity)tensor_already_casted_to_fp8)Float8TrainingTensor GemmInputRoleLinearMMConfighp_tensor_and_scale_to_float8)tensor_to_scaleF hp_tensor float8_dtypelinear_mm_config reduce_amaxgemm_input_rolescaling_granularity axiswise_dimround_scales_to_power_of_2returnc D[UUUUUUU5n [UU UUUU5$)a Given a high precision tensor `hp_tensor`, scales `hp_tensor` dynamically and returns a `Float8TrainingTensor` of the result. Args: hp_tensor: the tensor to convert float8_dtype: the float8 dtype to use linear_mm_config: Defines the configuration for the scaled_mm for the 3 fwd/bwd gemms of linear reduce_amax: whether to reduce the max(abs(hp_tensor)) value across distributed ranks gemm_input_role: Defines the role of this tensor (input, weight or grad_output) in the 3 fwd/bwd gemms of linear scaling_granularity: Defines the scaling granularity axiswise_dim: if axiswise granularity is used, defines the dim to scale across round_scales_to_power_of_2: if true, round scaling factor down to the nearest power of 2. )r r ) r r r rr device_meshrrrscales ]/home/james-whalen/.local/lib/python3.13/site-packages/torchao/float8/float8_scaling_utils.pyhp_tensor_to_float8_dynamicrsD6 " E )   c.U[RLaU$g)a Convenience function which takes in an axiswise dim which is only relevant for axiswise scaing, and a scaling type. The output is pass-through if scaling type is axiswise, and None otherwise. This is done to keep the logic from choosing the axiswise dim out of the scaling function. N)rAXISWISE)rrs rget_maybe_axiswise_dimrKs0999 rcX\rSrSrSr\S\S\R4Sj5r \S5r Sr g) NoopFwToFloat8BwDynamicZzF Forward: no-op Backward: convert to float8_e5m2 with dynamic scaling r target_dtypecX lX0lU$N)r r )ctxtensorr r s rforwardNoopFwToFloat8BwDynamic.forwardas 0' rc[U5(aUSS4$[XR5n[UUURUR[ R 5nUSS4$r")rr r r r r GRAD_OUTPUT)r#gradY gradY_scale fp8_tensors rbackward NoopFwToFloat8BwDynamic.backwardlsd ' . .$$ $%e-=-=> 2        % %  4%%rN) __name__ __module__ __qualname____firstlineno____doc__ staticmethodrtorchdtyper%r,__static_attributes__r.rrrrZsG )kk  & &rr)r3typingrr5torchao.float8.configr torchao.float8.distributed_utilsr%torchao.float8.float8_training_tensorrrrr torchao.float8.float8_utilsr INPUT TENSORWISETensorr6boolintrr_dynamoallow_in_graphautogradFunctionrr.rrrFs 4I %2%8%8.@.K.K"&',+||++++%+ + # +,+3-+!%++\  + c] &enn55&&r