+ "ݳjRt^RIHt^RIt^RIHt^RIHt^RIH t H t H t H t ^RI HtR] PR]P RR3RR lltR R lt]P&P(!R R ]P*P,44tR#)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_scaleFcV^8dQhR\PR\PR\R\R\ R\ R\\,R\R \/ #)  hp_tensor float8_dtypelinear_mm_config reduce_amaxgemm_input_rolescaling_granularity axiswise_dimround_scales_to_power_of_2return) torchTensordtyperboolrrrintr)formats"d/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/float8/float8_scaling_utils.py __annotate__rsn++||++++%+ + # +,+3-+!%++c D\VVVVVVV4p \VV VVVV4#)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 rrr device_meshrrrscales &&&&&&&&& rhp_tensor_to_float8_dynamicr!sD6 " E )   rcRV^8dQhR\R\R\\,/#)r rrr)rrr)rs"rrrKs*   + c] rc2V\PJdV#R#)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_dimr%Ks0999 rcPa]tRt^ZtoRt]V3RlRl4t]R4tRtVt R#)NoopFwToFloat8BwDynamiczF Forward: no-op Backward: convert to float8_e5m2 with dynamic scaling c:<V^8dQhRS[RS[P/#)r r target_dtype)rrr)r __classdict__s"rr$NoopFwToFloat8BwDynamic.__annotate__bs%)kk rcW nW0nV#N)rr))ctxtensorrr)s&&&&rforwardNoopFwToFloat8BwDynamic.forwardas 0' rc\V4'dVRR3#\WP4p\VVVPVP\ P 4pVRR3#r-)rr r)rrr GRAD_OUTPUT)r.gradY gradY_scale fp8_tensors&& rbackward NoopFwToFloat8BwDynamic.backwardlsd ' . .$$ $%e-=-=> 2        % %  4%%rN) __name__ __module__ __qualname____firstlineno____doc__ staticmethodr0r7__static_attributes____classdictcell__)r*s@rr'r'Zs2  & &rr')r>typingrrtorchao.float8.configr torchao.float8.distributed_utilsr%torchao.float8.float8_training_tensorrrrrtorchao.float8.float8_utilsr INPUT TENSORWISEr!r%_dynamoallow_in_graphautogradFunctionr'r9rrrMs 4I %2%8%8.@.K.K"&',+\ &enn55&&r