+ ܳjC^RIt^RIt^RIt^RIt^RIHt^RIt^RIH t ^RI t ^RI H t H t ^RIHtHtHtHt]P&!]4tRRlt!RR 4t!R R ]4t!R R ]4tRRltRRltRRltRRlt!RR]4tRRltRRlt ]!4t!Rt"Rt#] PHPJ'dRt"Rt#MA] PLPO4'dR t#M] PPPR'dR!t#] !4t*R# ]+dGt,]#R%9d]!R"4t*Rt,A,R#]-!],4t.]P_R#]. 2RR$7]!].4t*Rt,A,R#Rt,A,ii;i)&NPath)Optional)DYNAMIC_LIBRARY_SUFFIX PACKAGE_DIR) CUDASpecsget_cuda_specsget_cuda_version_tupleget_rocm_gpu_archc0V^8dQhR\R\/#) cuda_specsreturn)rr)formats"X/home/luca/screen-checking/.venv/lib/python3.14/site-packages/bitsandbytes/cextension.py __annotate__rs::)::ch\\PP4pV'dRMRpV'dRMRp\P P V4pVerVP4'g\V RV: R24hRV V \ 2p\PRT RT R T R V'dR MR R V R2 4\V, #\W!4pVPpV'g#\RV VP \ 2, #Wv9d Wg,#VUu.uF#q^,V^,8XgKW8gK!VNK% p pV 'd \!V 4p MeVUu.uF#q^,V^,8XgKW8gK!VNK% p pV 'd \#V 4p M"\RV VP \ 2, #\PRV'dR MR RV^, RV^, RV'dR MR RV ^, RV ^, RV R24Wj,#uupiuupi)a Get the path to the best matching CUDA/ROCm BNB native library for the given specs. When no override is set, selects from packaged libraries using the following priority: 1. Exact version match. 2. Highest packaged version <= runtime version, same major (e.g. runtime 12.9, ship 12.8). 3. Lowest packaged version > runtime version, same major (e.g. runtime 12.0, ship 12.1). No cross-major fallback: if no same-major library exists, returns the exact non-existent path so the caller raises a clear "not found" error. A warning is logged when falling back. Override env vars bypass selection entirely and load the named version with no fallback. The returned path is not guaranteed to exist when no packaged libs are found, or when an override names an absent version. rocmcudaBNB_ROCM_VERSIONBNB_CUDA_VERSION=z2: value must be digits only (e.g. '124' for 12.4).libbitsandbytes_z WARNING: z( environment variable detected; loading z. This overrides automatic ROCmCUDAzO version selection. If this was unintended clear the variable and retry: unset  zNo prebuilt binary for  .z , loading z instead. Set z to override.)booltorchversionhiposenvirongetisdigit RuntimeErrorrloggerwarningr_find_cuda_libscuda_version_tuplecuda_version_stringmaxmin) ris_hipprefix override_varoverride_value library_name availableruntime_versionvlowerselected higher_sames & rget_cuda_bnb_library_pathr;s3%--## $FV6F)/%5GLZZ^^L1N!%%'',q0BBtuv v)&.1ABXAYZ  ~Q~&66^_k^lm(28f'EFJJVWY [ \))/I 33O /x 8V8V7WXnWoppp#))! X 1qT_Q-?%?QADWQQ E X u:"+b)Qtq7I/IqaNaqq) b ;'H#3F8JH*H*&H/H/ H/cla]tRt^StoRtV3RlRlt]PR4tRt V3Rlt Rt Vt R#) BNBNativeLibraryFc4<V^8dQhRS[P/#r libctCDLL)r __classdict__s"rrBNBNativeLibrary.__annotate__WsBGGrcWnR#N_lib)selfr@s&&r__init__BNBNativeLibrary.__init__Ws rcNa\VPSR4pVeV#V3RlpV#)Nc"<\RS R24h)zMethod 'zm' not available in CPU-only version of bitsandbytes. Reinstall with GPU support or use CUDA-enabled hardware.)r()argskwargsnames*,r throw_on_call3BNBNativeLibrary.__getattr__..throw_on_callas#4&!KK r)getattrrI)rJrQfnrRs&f r __getattr__BNBNativeLibrary.__getattr__Zs, TYYd + >I  rc$VPV4#rGrV)rJitems&&r __getitem__BNBNativeLibrary.__getitem__i%%rc:<V^8dQh/S[P;R&#)r rIrA)rrDs"rrrESs ''MrrHN) __name__ __module__ __qualname____firstlineno__compiled_with_cudarK functoolscacherVr[__annotate_func____static_attributes____classdictcell__rDs@rr=r=Ss9__  &-rr=cBaa]tRt^mtoRtV3RlV3RlltRtVtV;t#)CudaBNBNativeLibraryTc4<V^8dQhRS[P/#r?rA)rrDs"rr!CudaBNBNativeLibrary.__annotate__ps33BGG3rc<\SV`V4\PVPn\PVP nR#rG)superrKrBc_void_p get_contextrestypecget_managed_ptrrJr@ __class__s&&rrKCudaBNBNativeLibrary.__init__ps2 "$++'){{$r) r_r`rarbrcrKrgrh __classcell__rurDs@@rrkrkms333rrkcBaa]tRt^vtoRtV3RlV3RlltRtVtV;t#)XpuBNBNativeLibraryz6XPU native library with SYCL USM paged memory support.c4<V^8dQhRS[P/#r?rA)rrDs"rr XpuBNBNativeLibrary.__annotate__ys77BGG7rc<\SV`V4\VR4'd"\PVP nR#R#)rsN)rorKhasattrrBrprsrrrts&&rrKXpuBNBNativeLibrary.__init__ys5  3* + ++-;;C ( ,rrw) r_r`rarb__doc__rKrgrhrxrys@@rr{r{vs@777rr{c^V^8dQhR\R\R\\\3,/#)r compactr0r)strr tupleint)rs"rrrs) . . .d .uS#X .rcV'd%\VR,4\VR,43#\VR,4\VR,43#)a0Split a compact CUDA/ROCm version string from a library filename into (major, minor). CUDA: major is always 2 digits (11, 12, 13...), e.g. '118' -> (11, 8), '132' -> (13, 2). ROCm: major is always 1 digit for now (6, 7...), e.g. '72' -> (7, 2), '713' -> (7, 13). Note: revisit if ROCm major reaches 10. :NN:rNN:Nr N:r NN)r)rr0s&&r_split_cuda_versionrsA72;WR[!111 wr{ S- --rc V^8dQhR\R\R\\\\3,\ 3,/#)r r1r0r)rr dictrrr)rs"rrrs3  C  $uS#X7L2M rc /p\P!RV R\ 24FOp\P!V R2VP 4pV'gK1W2\ VP^4V4&KQ V# \\3dKii;i)zQReturn a {(major, minor): Path} mapping for all packaged CUDA/ROCm library files.r*z(\d+)) rglobrresearchrQrgroup ValueError IndexError)r1r0resultr@matchs&& rr+r+s F"26(!6BC V M +  sA88B  B c:V^8dQhR\\,/#r r)listr)rs"rrrsZZDIZrc\\PP4pV'dRMRp\ R\ W444#)z=Get formatted CUDA/ROCm versions from existing library files.rrc34"TFwrV RV 2xK R#5i)rNrw).0majorminors& r 5get_available_cuda_binary_versions..sY9XUG1UG$9Xs)r r!r"r#sortedr+)r0r1s r"get_available_cuda_binary_versionsrs5 %--## $FV6F Y9XY YYrc0V^8dQhR\R\/#)r version_strrr)rs"rrrs  C C rcVP4'd9\\PP4p\ W4wr#V RV 2#V# \ \3dT#i;i)z`Convert a raw version code string (e.g. '118', '713') to a dotted version (e.g. '11.8', '7.13').r)r'r r!r"r#rrr)rr0rrs& rparse_cuda_versionrsnemm''( .{CLEWAeW% % J'    sAA&%A&cja]tRt^toRtV3RlRltR V3RlRlltV3RlRltR tR t R t Vt R#) ErrorHandlerMockBNBNativeLibrarya Mock library handler that defers errors until native methods are called. This class serves as a fallback when the native bitsandbytes library fails to load. It captures the original error and generates detailed troubleshooting guidance. Key behaviors: - Allows attribute access and method assignment without immediate errors - Throws a RuntimeError with diagnostic information only when a native method is called, as otherwise it would error out on import, breaking backward compatibility - Handles both missing CUDA dependencies and version mismatch scenarios Error scenarios covered: 1. Missing shared library dependencies (e.g., libcudart.so not in LD_LIBRARY_PATH or through PyTorch CUDA installation) 2. CUDA version mismatch between PyTorch and available pre-compiled binaries 3. Completely missing pre-compiled binaries when CUDA is detected 4. Custom BNB_CUDA_VERSION or BNB_ROCM_VERSION override but mismatch 5. CPU-only installation attempts when GPU functionality is requested c <V^8dQhRS[/#)r error_msgr)rrDs"rr-ErrorHandlerMockBNBNativeLibrary.__annotate__s#rc&Wn\4Vn\'d \P P R4M\P P R4p\4pV'dV^, RV^, 2MRpV'd \V4MTVn RV9dVP4Vn R #TPVPTVP'dRVP R2MRVPR7Vn R #) rrrunknownzcannot open shared object filezOriginal error: r)available_versionsuser_cuda_versionoriginal_errorrequested_versionN) rrrHIP_ENVIRONMENTr$r%r&r rr_format_dependency_errorformatted_error_format_lib_error_message)rJrr3 user_versionuser_version_strs&& rrK)ErrorHandlerMockBNBNativeLibrary.__init__s""D"F?N(:;TVT^T^TbTbcuTv-/ EQl1o.a Q/@AW`GU!3N!C[k ,y 8#'#@#@#BD #'#A#A#'#:#:"2HL!1$..1AD\^"&"8"8 $B$D rNc X<V^8dQhRS[S[,RS[RS[RS[S[,RS[/#)r rrrrr)rrr)rrDs"rrrsSQsQs IQsQs Qs $C= Qs Qsrc  RpRV9p\ R2V9pV'dRpMUV'dNV'dRRPV4,MRpV'd"R\ R\ R V R \ R V R V R 2 MRpRp V'd R\ R2M R\ R2p V'dR\ R\ R2MRp V'gRM\'gRMRp R\P4 R2p V V V V V V RV 2#)z:Format detailed error message for library loading failuresrz#libbitsandbytes_cpu.so: cannot openz binary not foundu9 🚨 Failed to load CPU-only bitsandbytes library 🚨 z - NONE 🚨 u! VERSION MISMATCH 🚨 Requested z version: z Detected PyTorch z version: z" Available pre-compiled versions: zU This means: The version you're trying to use is NOT distributed with this package u 🚨 Forgot to compile the bitsandbytes library? 🚨 1. You're not using the package but checked-out the source code 2. You MUST compile from source zTAttempted to use bitsandbytes native library functionality but it's not available. z^This typically happens when: 1. bitsandbytes doesn't ship with a pre-compiled binary for your zR version 2. The library wasn't compiled properly during installation from source zlThis typically happens when you checked the code out from source and your torch installation doesn't detect z on your machine. z(bitsandbytes tried to find a compatible z* binary but none could be loaded. If your z_ version isn't among the available pre-compiled versions above, you must compile from source. zPCOMPILE FROM SOURCE for CPU-only: `cmake -DCOMPUTE_BACKEND=cpu -S . && make` aNYou have two options: 1. COMPILE FROM SOURCE (required if no binary exists): https://huggingface.co/docs/bitsandbytes/main/en/installation#cuda-compile 2. Use BNB_CUDA_VERSION to specify a DIFFERENT CUDA version from the detected one, which is installed on your machine and matching an available pre-compiled version listed above aYou have two options: 1. COMPILE FROM SOURCE as mentioned here: https://huggingface.co/docs/bitsandbytes/main/en/installation?backend=AMD+ROCm#amd-gpu 2. Use BNB_ROCM_VERSION to specify a DIFFERENT ROCm version from the detected one, matching the version the library was built with. u🔍 Run this command for detailed diagnostics: python -m bitsandbytes If you've tried everything and still have issues: 1. Include ALL version info (operating system, bitsandbytes, pytorch, , python) 2. Describe what you've tried in detail 3. Open an issue with this information: https://github.com/bitsandbytes-foundation/bitsandbytes/issues r) BNB_BACKENDjoinrr8)rJrrrranalysisno_cpu_lib_foundno_cuda_lib_foundversion_list_strbase_msgtroubleshootingnotecompile_instructions diagnosticss&&&&& rr:ErrorHandlerMockBNBNativeLibrary._format_lib_error_messages@NR*m+<=O UH Oax(--8J*KKgm &k]+!!, -@AR@ST((3}LAR@ST88H7IJ``6 l! TT_S`a^_ @AL@MMbc !;;-H&-(IJ %c#? HZ &UVaUfUfUhTijT U H:o%6tf=Q.throw_on_callEs&$"6"6!77abfaggijk krrw)rJrQrRsff rrV,ErrorHandlerMockBNBNativeLibrary.__getattr__Bs lrc$VPV4#rGrY)rJrQs&&rr[,ErrorHandlerMockBNBNativeLibrary.__getitem__Jr]r)rrrr)rN) r_r`rarbrrKrrrVr[rgrhris@rrrs<(&QsQsf  8&&rrc$V^8dQhR\/#rr)rs"rrrNs H H$ Hrc\\PRR4pV'd6\RVR, \ 2, pVP 4'dV#\R\ 2, #)zGet the path to the XPU native library matching the oneAPI toolchain. Prefers the versioned library (e.g. libbitsandbytes_xpu2026) matching the first 4 digits of torch.version.xpu, falling back to an unversioned libbitsandbytes_xpu. xpuNlibbitsandbytes_xpu:NN)rTr!r"rrexists) xpu_version versioneds rget_xpu_bnb_library_pathrNs` %--5K$7 B7HI_H`"aa       ./E.FG GGrc$V^8dQhR\/#r)r=)rs"rrr\s!!,!rc8\4p\R\ 2, pV'd9\V4pVP 4'g\ R\ RV 24hTp\PP'd \4p\PRV 24\PP\!V44p\#VR4'd \%V4#\PP'd \'V4#\)V4#)zV Load CUDA library XOR CPU, as the latter contains a subset of symbols of the former. libbitsandbytes_cpuzNo compatible z binary found at z*Loading bitsandbytes native library from: rq)r rrr;rr(rr!_C_has_xpurr)debugrBcdll LoadLibraryrrrkr{r=)r binary_pathcuda_binary_pathdlls rget_native_libraryr\s !J"56L5M NNK4Z@&&(( }rsU   C   8 $:z&&43+37*7 . Z `&'7`&F H!@"#  ==OK ZZK XXK :  C :n$./XYF  / { ;  /y9 :sDE$E3&EE$