+ "ݳj<aRt0t^RIt^RIt^RIHt^RIHtHt^RIt^RI H t ]P!4t ]^k!RR]P4t!RR]P4t]!R R 44t]!4t]P$t]P&t]!R R 7!R R44t]!R R 7!RR44t!RR]P4t]!R R 7!RR44tR#)N) dataclass)OptionalUnion)is_MI300c2a]tRt^toRtRtRtRtRtVt R#) ScalingTypez Defines the type of scaling to use for casting to float8. Values: * ``DYNAMIC``: Compute scaling factor dynamically based on the tensor's values. * ``DISABLED``: Skip scaling for this tensor, leave it in its original precision. dynamicdisabledc^V\PJdR#V\PJgQhR#)dyndis)rDYNAMICDISABLEDselfs&V/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/float8/config.py short_strScalingType.short_str"s* ;&& &;/// //N) __name__ __module__ __qualname____firstlineno____doc__rrr__static_attributes____classdictcell__ __classdict__s@rrrs"GHrrc2a]tRt^*toRtRtRtRtRtVt R#)ScalingGranularityz Defines the granularity of scaling strategies for casting to float8. Values: * ``TENSORWISE``: A single scaling factor for the entire tensor. * ``AXISWISE``: Scaling factors computed along one axis of the tensor (rowwise scaling). tensorwiseaxiswisec^V\PJdR#V\PJgQhR#)tenaxs)r! TENSORWISEAXISWISErs&rrScalingGranularity.short_str:s, %00 0-666 66rrN) rrrrrr'r(rrrrs@rr!r!*s"JHrr!cZa]tRt^BtoRt]P t]Pt Rt Rt Vt R#)Float8TypeConfigz Configuration for selecting the preferred float8 type pair, either e4m3fn/e5m2 or e4m3fnuz/e5m2fnuz. Currently, ROCm supports 1. fnuz variants in MI300. 2. OCP F8 variants in MI350/Navi4. c\PP'de\PP 4'd?\ 4'd-\P Vn\PVn R#R#R#R#)N) torchversionhipcuda is_availablerfloat8_e4m3fnuz e4m3_dtypefloat8_e5m2fnuz e5m2_dtypers&r __post_init__Float8TypeConfig.__post_init__PsQ ==   !8!8!:!:xzz#33DO#33DO@J!: r)r3r5N) rrrrrr- float8_e4m3fnr3 float8_e5m2r5r6rrrs@rr+r+Bs.$$J""J44rr+T)frozencpa]tRt^\toRt]P t]Pt Rt Rt Rt V3RltRtVtR#) CastConfiga Configuration for casting a single tensor to float8. Args: scaling_type: The type of scaling to use. See :class:`ScalingType`. Default: ``ScalingType.DYNAMIC`` scaling_granularity: The granularity of scaling. See :class:`ScalingGranularity`. Default: ``ScalingGranularity.TENSORWISE`` target_dtype: The target float8 dtype (e.g., ``torch.float8_e4m3fn``). Default: ``None`` (will be set based on the recipe) Nc\R\R/VP,pVPP 4 RVP P 4 RV 2#)e4m3e5m2_)r3r5 target_dtype scaling_typerscaling_granularity)rdtypes& rrCastConfig.short_strnsUVZ89J9JK##--/0$2J2J2T2T2V1WWXY^X_``rc(VP\PJd&VP\P JgQR4hVP eAVP P'dVP P^8XgQR4hR#R#)zGonly dynamic scaling type is supported for axiswise scaling granularityNz)must specify a 8-bit floating-point dtype) rCr!r(rBrrrAis_floating_pointitemsizers&rr6CastConfig.__post_init__rs  # #'9'B'B B$$ (;(;; Y ;  (    / / /D4E4E4N4NRS4S 7 6 7 4S)rcb<V^8dQh/S[;R&S[;R&S[S[P,;R&#)rBrCrA)rr!rr-rD)formatrs"r __annotate__CastConfig.__annotate__\s53,K 5;;'.!rr)rrrrrrrrBr!r'rCrArr6__annotate_func__rrrs@rr<r<\s: !, 3 3L.@.K.K*.La7-rr<c4a]tRt^|toRtRtV3RltRtVtR#)Float8GemmConfigz Configuration for a float8 gemm. Args: use_fast_accum: If True, use fast accumulation in lower precision. This can improve performance but may reduce numerical accuracy. Default: ``False`` Fc&<V^8dQh/S[;R&#)rKuse_fast_accum)bool)rLrs"rrMFloat8GemmConfig.__annotate__|s rrN) rrrrrrSrOrrrs@rrQrQ|s!NrrQc&]tRt^tRtRtRtRtRtR#)Float8LinearRecipeNamea. Pre-made recipes for common float8 training configurations. Values: * ``TENSORWISE``: Default, dynamic per-tensor scaling with the cuBLAS tensorwise kernel. Fastest option. * ``ROWWISE``: Dynamic rowwise scaling with the CUTLASS rowwise kernel. Uses e4m3 for activations, weights, gradients. Scales are rounded (floor) to the nearest power of two for increased accuracy. * ``ROWWISE_WITH_GW_HP``: A modification on rowwise scaling with increased accuracy for grad_weight by keeping grad_weight computation in high precision. Most accurate option. r"rowwiserowwise_with_gw_hprN) rrrrrr'ROWWISEROWWISE_WITH_GW_HPrrrrrWrWs J G.rrWca]tRt^toRt]!4tRt]!4tRt ]!4t Rt ] !RR7t ] !4t] !4tRtRtRtRtRtRt]V3RlRl4tV3R ltR tVtR#) Float8LinearConfigzQ Configuration for converting a `torch.nn.Linear` module to float8 for training. NT)rSFcVPf"\PVRVP4VPf"\PVRVP 4VP f"\PVRVP4VP P\P8wd1VP'dQRVP P 24hVPpVP pVPpVPpVPpVP pWR3W5R3WFR33FNwrxp VP\PJp VP\PJp W8XdKFQRV 24h WR \3W%R \3W6R \ 33FuwrxrVP"f\PVR V 4VP"f\PVR V 4VP"VP"8XdKmQV R 24h VP$'d\&P)R4R#R#)N!cast_config_input_for_grad_weight!cast_config_weight_for_grad_input'cast_config_grad_output_for_grad_weightzPenable_fsdp_float8_all_gather only supports tensorwise scaling granularity, got output grad_input grad_weightz#incompatible operand precision for inputweight grad_outputrAz< must be cast to the same dtype in both matmuls it's used inz`config.force_recompute_fp8_weight_in_bwd` is deprecated and will be removed in a future release. Please see https://github.com/pytorch/ao/issues/2251 for more details.)r_object __setattr__cast_config_inputr`cast_config_weightracast_config_grad_outputrCr!r'enable_fsdp_float8_all_gatherrBrrr3r5rA!force_recompute_fp8_weight_in_bwdloggerwarning)rcc_icc_wcc_gocc_i_gwcc_w_gicc_go_gwcc1cc2 gemm_name is_disabled_1 is_disabled_2 operand_name default_dtypes& rr6 Float8LinearConfig.__post_init__sX  1 1 9   94;Q;Q   1 1 9   94;R;R   7 7 ?   9,,   " " 6 6:L:W:W W999 bcgczczdOdOcPQ 9 %%&&,,8888??  " \ *  .$ Ci  ,, 0D0DDM,, 0D0DDM 1 5i[A 1$ GZ 0 Hj 1 mZ 86 1Cl '""3 F'""3 F##s'7'77 . \] 76   1 1 1 NN{  2rc:<V^8dQhRS[S[S[3,RR/#)rK recipe_namereturnr])rrWstr)rLrs"rrMFloat8LinearConfig.__annotate__=s0@G@G1367@G @Grc  \V4\8Xd=\Uu.uFqPNK ppW9gQRV RV 24h\V4pV\PJd \ 4#V\P Jdm\\P\R7p\\P\R7p\\P\R7p\ VVVRR7#V\PJd\\PR7p\\PR7p\\P\R7p\\PR7p\\PR7p\\P\R7p\ VVVVVVRR 7#\R V 24huupi) z Input: `Float8LinearRecipeName` value, or a string representing a `Float8LinearRecipeName` value Output: a `Float8LinearConfig` configured to implement the specified recipe z recipe_name z not in valid names )rCrAT)rjrkrlround_scales_to_power_of_2)rC)rB)rBrA)rjrkrlr_r`rarzunknown recipe_name )typerrWvaluer'r]rZr<r!r(r3r[rrAssertionError) rn valid_namesrqrrrsrurtrvs & rfrom_recipe_name#Float8LinearConfig.from_recipe_name<s   #,BC,Bq77,BKC- {m+? }M -1=K 0;; ;%' ' 2:: :$6$?$?jD$6$?$?jD$6$?$?jE&"&#'(-+/  2EE E2D2M2MND2D2M2MND$6$?$?jE!5G5R5RSG!k.B.BCG!(11 H&"&#'(-29298@+/ !#7 }!EF FqDsG c<V^8dQh/S[;R&S[S[,;R&S[;R&S[S[,;R&S[;R&S[S[,;R&S[;R&S[;R&S[;R &S[;R &S[;R &S[;R &S[;R &S[;R&#)rKrjr_rkr`rlragemm_config_outputgemm_config_grad_inputgemm_config_grad_weightrm pad_inner_dimemulaternr)r<rrQrT)rLrs"rrMrs."0/0(0 ';B14#156(0 ';B7:(6;<.6j-AH=H)PIJ-AKL.BM\$(/]hinox(,3yF!%,Grr)rrrrrr<rjr_rkr`rlrarQrrrrmrrrnrr6 staticmethodrrOrrrs@rr]r]s*%/L>B%%/\>B%*4,DH+ ,<4+P/?/A0@0B+0!  MG /4%(-BH@G@GUrr]cTV^8dQh/^\9d\P;R&#)rKro)__conditional_annotations__loggingLogger)rLs"rrMrMs -,, r)renumr dataclassesrtypingrrr- torchao.utilsr getLoggerroEnumrr!r+ type_configr3r5r<rQrWr]rM)rs@rrs !" " **,,$)).0 44 4(  # #  # #  $777> $ ! ! ! #.TYY#.L $IGIGIGr