+ j-:Rt^RIt!RR4t!RR4tR#)aN Helper classes for working with low precision floating point types that align with the opencompute (OCP) microscaling (MX) specification. * MXFP4Tensor: 4-bit E2M1 floating point data * MXScaleTensor: 8-bit E8M0 floating point data Reference: https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf NcHa]tRt^ toR RltRtRtRtRtRt Rt Vt R#) MXFP4TensorNc W0nVeQ\V\P4'gQR4hVPVnVP V4VnR#Ve#\V\ 4'dTMV3VnR#\R4h)a< Tensor class for working with four bit E2M1 floating point data as defined by the opencompute microscaling specification. Parameters: - data: A torch tensor of float32 numbers to convert to fp4e2m1 microscaling format. - size: The size of the tensor to create. - device: The device on which to create the tensor. N%Parameter data must be a torch tensor.Either parameter data or size must be provided device isinstancetorchTensor _from_floatdatatuplesize ValueErrorselfr rrs&&&&R/home/luca/screen-checking/.venv/lib/python3.14/site-packages/triton/tools/mxfp.py__init__MXFP4Tensor.__init__sr  dELL11 Z3Z Z1++DK((.DI   *4 7 7dXDIMN Nc\P!^^VP\PVPR7p\P!^^VP\PVPR7p\P!^^VP\PVPR7pV^,V^,,V,P \P4VnV#)rdtyper)r randintruint8rtyper )rSEMs& rrandomMXFP4Tensor.random#s MM!QTYYekk$++ V MM!QTYYekk$++ V MM!QTYYekk$++ V1fa(1,225;;?  rc FV\P8XgQR4hVPpV^, ^,PV4pV^, ^,PV4pV^,PV4p\P!V4pV^8HV^8H,pV(pVP 4'dW8,p WH,p WX,p \P !RV 4p \P!V ^8HW^, 4p \P!V ^8HV R,RV R,,4pV \P !^V 4,V,pWV&WgV^8H,;;,R,uu&VP\P4#)zj Convert fp4e2m1 data to float32. Returns: - A torch tensor of type dtype representing the fp4e2m1 data. zCCurrently only float32 is supported for fp4e2m1 to float conversion??)r float32r r zeros_likeanypowwhere)rrr rrr valueis_zero non_zero_maskS_nzE_nzM_nzsignexponentmantissavalue_nzs&& rtoMXFP4Tensor.to+sV %l'll%yyai3  $ $U +ai3  $ $U + CZ  e $  #6a1f%      #D#D#D99R&D{{419d1H=H{{419dSj#s :JKHeii844x?H#+-  a !R'!zz%--((rc 4\P!V4P\P4p\P!V4pV^8Hp\P !V4\P !V4,p\P!.RO\PVPR7p\P!^^.\PVPR7p.p.p .p VFp V ^8XdZ^p VFOp V R,pV^V ,,pVPV4V PV 4V PV 4KQ KcV P4^, p VFVp RV R,,pV^V ,,pVPV4V PV 4V PV 4KX K \P!V\PVPR7p\P!V \PVPR7p \P!V \PVPR7p VPR 4pVP^,pVP^4pVP4P4pVVVPR 4&\P!VVP^4, 4p\P !V^RR7wppVV8HpVP#4^8dTV P^4P%VR 4pV^8HP\P&4pVVR,, p\P(!V^R7pV V,pV V,pVPVP4p VPVP4p ^W&^W&V^,V ^,,V ,P\P4#) a Convert float32 numbers to mxf4 e2m1 format. * No encodings are reserved for Inf or NaN in mxf4. * Conversion from float supports roundTiesToEven rounding mode. * If a value exceeds the mxf4 representable range after rounding, clamps to the maximum mxf4 magnitude, preserving the sign. * If a value has magnitude less than the minimum subnormal magnitude in mxf4 after rounding, converts to zero. Parameters: - values: A torch tensor of float32 numbers to convert to fp4 format. rrr$r%T)dimkeepdimgư>r:)rr&)r signbitrrabsisnanisinftensorrappenditemr'viewshape unsqueezemaxminsumexpandint32argmin)rvaluesr abs_valuesr- is_invalidE_bitsM_bitscandidate_values candidate_E candidate_Mrr3r significandr, candidatesabs_values_flatNabs_values_expandedmax_candidate_valueerrors min_errors_is_tieM_bits_expanded tie_breaker best_indices E_selected M_selecteds&& rr MXFP4Tensor._from_floatNs$ MM& ! & &u{{ 3YYv& ?[[(5;;v+>> l%++dkkRq!fEKK L  AAvA"#c'K'1h;7E$++E2&&q)&&q)  668a<A"%C-K'1h;7E$++E2&&q)&&q)  (\\"2%--PTP[P[\ ll;ekk$++V ll;ekk$++V $//"-  ! !! $-77:)nn.335/B +,.1E1Ea1HHI  &a> AJ& ::"22U78_D  *   "a 'T{mdii/G N$78GLg'Dyy%%rr rrNNN) __name__ __module__ __qualname____firstlineno__rr!r6r r{r__static_attributes____classdictcell__ __classdict__s@rrr s-O*!)FV;p!F&&rrc@a]tRt^toRRltRRltRtRtRtVt R#) MXScaleTensorNc W0nVeQ\V\P4'gQR4hVPVnVP V4VnR#Ve#\V\ 4'dTMV3VnR#\R4h)a Tensor class for working with microscaling E8M0 block scale factors. Parameters: - data: A torch tensor of float32 numbers to convert to fp8e8m0 microscaling format. - size: The size of the tensor to create. - device: The device on which to create the tensor. Nrrrrs&&&&rrMXScaleTensor.__init__sr  dELL11 Z3Z Z1++DK((.DI   *4 7 7dXDIMN Nrc  ^pVf^MC\^\\P!\P!V444V,4pVf^MM\ ^\^\\P!\P!V444V,44pWE8:gQR4h\P !WE^,VP\PVPR7pW`n V#)zX Generate random E8M0 data within a specified range. * Excludes the NaN encoding (255). z&Low must be less than or equal to highr) rJintr log2rDrKrrrrr )rrxrybias min_exponent max_exponentrs&&& rr!MXScaleTensor.randoms KqSC 5<>&A+N  k* KK <0 1s7}}U[[1  ;;|Q<.33EKK@{ rrr)NN) rrrrrr!r6r rrrs@rrrsO&  !rr)__doc__r rrrrrs( Z&Z&zDDr