+ j4^RIHt^RIHt]'d ^RIHt^RIHt^RIH t H t H t H t H t ^RIHtHt^RIHt] !4'd ^RIt^R IHt] P,!]4tRt!R R ]4tR#) ) TYPE_CHECKING) HfQuantizer)PreTrainedModel) Mxfp4Config)is_accelerate_availableis_kernels_availableis_torch_availableis_triton_availablelogging)KERNELS_MAX_VERSIONKERNELS_MIN_VERSION)get_module_from_nameN)WeightConvertercaa]tRt^+toRtRtV3RltRtRtV3RlRlt V3RlR lt RV3R lR llt R t R t RtRt]V3RlRl4tRtRtV3RltRtVtV;t#)Mxfp4HfQuantizerz' FP4 quantization using fbgemm kernels Fc 8<\SV`!V3/VBRVnR#N)super__init__triton_kernels_hub)selfquantization_configkwargs __class__s&&,h/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/quantizers/quantizer_mxfp4.pyrMxfp4HfQuantizer.__init__3s ,77"&c VPf#^RIHpV!R^R7VnVP#VP# \d \R4hi;i)z3Lazy import and initialize kernels only when needed) get_kernelz(kernels-community/gpt-oss-triton-kernels)versionz2kernels package is required for MXFP4 quantization)rintegrations.hub_kernelsr ImportError)rrs& r_lazy_import_kernels%Mxfp4HfQuantizer._lazy_import_kernels7s`  " " * XA*45_ij*k'&&&t&&& X!"VWW Xs =Ac `\4'g \R4hVPP'dR#\ 4'g \R4h\ P P4;'g\ P!R4pVPR9dNVP'd-\PRV R24RVPnR#\RV R24h\ PP4'dRp\!R 4p\#4pM\ P$P4'd:\ P$P'4pVR8p\!R 4p\#4pM/VPR8XdRp\!R 4p\#4pMR pR pR pVP'dV'g)\PR 4RVPnR#V'g)\PR 4RVPnR#V'g?\PR\( R\* R\( R24RVPnR#M9V'g \-R4hV'g \-R4hV'g \-R4hVP'gVP/4VP1R4pVeO\3V\44'd7VP'g#RVP749d \-R4hR#R#R#R#)zqUsing mxfp4 quantization requires torchPlease install the latest version of torch ( pip install --upgrade torch )Nz9Using mxfp4 requires Accelerate: `pip install accelerate`cpuzGUsing MXFP4 quantized models requires model on cuda/xpu/cpu, but found zj, we will default to dequantizing the model to bf16. To use mxfp4, please disable the current accelerator.TzIQuantizing a model using MXFP4 requires model on cuda/xpu/cpu, but found z7. To use mxfp4, please disable the current accelerator.z3.5.0z3.4.0FuMXFP4 quantization is only supported on GPUs with compute capability >= 7.5 (e.g T4, A100, L4, H100, or B200) or XPUs (e.g Intel® Data Center GPU Max Series). We will default to dequantizing the model to bf16.zMXFP4 quantization requires Triton: CUDA requires Triton >= 3.4.0, XPU/CPU requires Triton >= 3.5.0. Please install triton: `pip install triton`. We will default to dequantizing the model to bf16.zXMXFP4 quantization requires the `kernels` package: Please install a compatible version (z <= version < z), e.g. `pip install kernels==z3`We will default to dequantizing the model to bf16.uMXFP4 quantization is only supported on GPUs with compute capability >= 7.5 (e.g T4, A100, L4, H100, or B200) or XPUs (e.g Intel® Data Center GPU Max Series) or CPUzMXFP4 quantization requires Triton: CUDA requires Triton >= 3.4.0, XPU/CPU requires Triton >= 3.5.0. Please install triton: `pip install triton`zPMXFP4 quantization requires the `kernels` package: `pip install kernels>=0.12.0` device_mapdiskzYou are attempting to load an FP4 model with a device_map that contains a disk device.This is not supported when the model is quantized on the fly. Please use a quantized checkpoint or remove the disk device from the device_map.)cudaxpur&))r r"r dequantizertorch acceleratorcurrent_acceleratordevicetype pre_quantizedlogger warning_once RuntimeErrorr* is_availabler rr)get_device_capabilityr r ValueErrorr#get isinstancedictvalues) rargsrr1is_device_supported_mxfp4triton_availablekernels_installedcompute_capabilityr's &*, rvalidate_environment%Mxfp4HfQuantizer.validate_environmentBs!##]   # # . . . &((YZ Z""668OOELL)#^^,=r2rJF#^^,=BPRSF)66>>EETT$1199>>)B#7=dV1TF'236rs`!07J24   H %R {R r