+ j~#BaR&t0tRt^RIt^RIHt]!]P RR4tRs]^kRRltRRlt R R lt R'R R llt R'R Rllt RRlt RRltRRltRRltRRltRRltRRltRRltRtR R!ltR"R#lt^R$IHt^R%IHt.R(OtR#))ag This package enables an interface for accessing MPS (Metal Performance Shaders) backend in Python. Metal is Apple's API for programming metal GPU (graphics processor unit). Using MPS means that increased performance can be achieved, by running work on the metal GPU(s). See https://developer.apple.com/documentation/metalperformanceshaders for more details. N)Tensor_mps_is_in_bad_forkcR#)FrS/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/mps/__init__.pyr s5rcLV^8dQhR\PP/#return)torch_C Generator)formats"r __annotate__rs""EHH$6$6"rcZ\f\PP4s\#)N)_default_mps_generatorr r_mps_get_default_generatorrrr_get_default_mps_generatorrs!%!&!D!D!F !!rc$V^8dQhR\/#r int)rs"rrrsCCcCrc\\PP;'d\PP 44#)z,Returns the number of available MPS devices.)rr r_has_mps_mps_is_availablerrr device_countrs, uxx  AAUXX%?%?%A BBrcV^8dQhRR/#r r Nr)rs"rrrs--T-rc>\PP4#)zAWaits for all kernels in all streams on a MPS device to complete.)r r_mps_deviceSynchronizerrr synchronizer!s 88 * * ,,rcpV^8dQhR\\,\P,R\/#)r devicer )rstrr r#r)rs"rrr#s'44#)ell24v4rc2\4P4#)zReturns the random number generator state as a ByteTensor. Args: device (torch.device or int, optional): The device to return the RNG state of. Default: ``'mps'`` (i.e., ``torch.device('mps')``, the current MPS device). )r get_state)r#s&r get_rng_stater'#s & ' 1 1 33rctV^8dQhR\R\\,\P,RR/#)r new_stater#r N)rrr$r r#)rs"rrr-s/ ; ;V ;S3Y-E ;RV ;rcxVP\PR7p\4P V4R#)aSets the random number generator state. Args: new_state (torch.ByteTensor): The desired state device (torch.device or int, optional): The device to set the RNG state. Default: ``'mps'`` (i.e., ``torch.device('mps')``, the current MPS device). ) memory_formatN)cloner contiguous_formatr set_state)r)r#new_state_copys&& r set_rng_stater0-s+__53J3J_KN **>:rc(V^8dQhR\RR/#)r seedr Nr)rs"rrr9s 3 3c 3d 3rc\PP'gR#\V4p\ 4P V4R#)zVSets the seed for generating random numbers. Args: seed (int): The desired seed. N)r rrrr manual_seed)r2s&rr4r49s1 88    t9D ,,T2rcV^8dQhRR/#rr)rs"rrrIs((d(rc6\4P4R#)z?Sets the seed for generating random numbers to a random number.N)rr2rrrr2r2Is %%'rcV^8dQhRR/#rr)rs"rrrNsTrcB\PP4R#)zReleases all unoccupied cached memory currently held by the caching allocator so that those can be used in other GPU applications. N)r r_mps_emptyCacherrr empty_cacher:Ns HHrcV^8dQhRR/#rr)rs"rrrUs...rc\V\4'g \R4hV^8gV^8d\RV R24h\P P V4R#)aSet memory fraction for limiting process's memory allocation on MPS device. The allowed value equals the fraction multiplied by recommended maximum device memory (obtained from Metal API device.recommendedMaxWorkingSetSize). If trying to allocate more than the allowed value in a process, it will raise an out of memory error in allocator. Args: fraction(float): Range: 0~2. Allowed memory equals total_memory * fraction. .. note:: Passing 0 to fraction means unlimited allocations (may cause system failure if out of memory). Passing fraction greater than 1.0 allows limits beyond the value returned from device.recommendedMaxWorkingSetSize. z3Invalid type for fraction argument, must be `float`zInvalid fraction value: z. Allowed range: 0~2N) isinstancefloat TypeError ValueErrorr r_mps_setMemoryFraction)fractions&rset_per_process_memory_fractionrCUsR" h & &MNN!|x!|3H:=QRSS HH##H-rc$V^8dQhR\/#r r)rs"rrrns22#2rc>\PP4#)zReturns the current GPU memory occupied by tensors in bytes. .. note:: The returned size does not include cached allocations in memory pools of MPSAllocator. )r r_mps_currentAllocatedMemoryrrrcurrent_allocated_memoryrGns 88 / / 11rc$V^8dQhR\/#r r)rs"rrrxs111rc>\PP4#)zReturns total GPU memory allocated by Metal driver for the process in bytes. .. note:: The returned size includes cached allocations in MPSAllocator pools as well as allocations from MPS/MPSGraph frameworks. )r r_mps_driverAllocatedMemoryrrrdriver_allocated_memoryrKxs 88 . . 00rc$V^8dQhR\/#r r)rs"rrrs000rc>\PP4#)zReturns recommended max Working set size for GPU memory in bytes. .. note:: Recommended max working set size for Metal. returned from device.recommendedMaxWorkingSetSize. )r r_mps_recommendedMaxMemoryrrrrecommended_max_memoryrOs 88 - - //rc$V^8dQhR\/#)r source)r$)rs"rrrs//3/rcr^RIHp^RIHp\ \ P R4'g \R4hT!VPR4Uu.uF q3R,NK upV!\4PPR, .\44p\ P PV4#uupi)aCompiles compute shader from source and allows one to invoke kernels defined there from the comfort of Python runtime Example:: >>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_MPS) >>> lib = torch.mps.compile_shader( ... "kernel void full(device float* out, constant float& val, uint idx [[thread_position_in_grid]]) { out[idx] = val; }" ... ) >>> x = torch.zeros(16, device="mps") >>> lib.full(x, 3.14) )Path)_embed_headers_mps_compileShaderMPS is not available include) pathlibrStorch.utils._cpp_embed_headersrThasattrr r RuntimeErrorsplit__file__parentsetrU)rQrSrTls& rcompile_shaderrbs= 5881 2 2122 !<<-.-aT-. h   % % 12 F 88 & &v .. /s B4c^RIp\V\\34'dT\ \ P R4'g \R4h\ P P\V44#\V\VP34'dT\ \ P R4'g \R4h\ P P\V44#\R\V4P 24h)aSLoads a precompiled Metal library (.metallib) and returns a shader library object that allows invoking kernels defined in it. Args: source: Either raw metallib bytes (``bytes``/``bytearray``) or a filesystem path (``str``/``os.PathLike``) to a ``.metallib`` file. This is useful for loading Metal libraries compiled ahead of time or generated by external tools (e.g. Triton, MetalASM). Example:: >>> # xdoctest: +SKIP("requires external .metallib file") >>> lib = torch.mps.load_metallib("kernels.metallib") >>> x = torch.ones(16, device="mps") >>> lib.square(x) N_mps_loadMetalllibrV_mps_loadMetallibFromPathzexpected bytes or path, got )osr=bytes bytearrayr[r rr\rdr$PathLikerer?type__name__)rQrfs& r load_metallibrls$&5),--uxx!56656 6xx**5=99 FS"++. / /uxx!<==56 6xx11#f+>>6tF|7L7L6MNOOrc$V^8dQhR\/#r )bool)rs"rrrsdrc\4^8#))rrrr is_availablerqs >A rc V^8dQhRRRR/#)r storageztorch.UntypedStorager r)rs"rrrs55!75MPS staging copy. This is intended for advanced interop with bulk loaders (e.g. safetensors) that already know how to write into CPU memory. The alias storage retains a reference to the source MPS storage, so the host pointer remains valid for the alias's lifetime even if the original tensor is freed. Raises an exception if ``storage`` is not backed by a shared-storage ``id`` allocated by the MPS allocator. .. warning:: Use with caution. This bypasses the cache-coherence guarantees that the higher-level PyTorch APIs (:meth:`torch.Tensor.cpu`, :meth:`torch.Tensor.to`, ``copy_``) provide for you, and makes the caller responsible for ordering CPU and GPU accesses to the same memory. You should **always** call :func:`torch.mps.synchronize` both **before** issuing host reads/writes through the alias (to drain any in-flight GPU work that may still be touching the buffer) and **after** (before launching GPU work that depends on the host writes). Failure to do so can produce stale reads, torn writes, or data corruption. )r r_mps_host_alias_storage)rss&r_host_alias_storagervs: 88 + +G 44r)profiler)EventchV^8dQh/^\9d\PP;R&#)r r)__conditional_annotations__r rr)rs"rrrs(21**1r)mps)rbrlrr'r4r2r0r!r:rCrGrKrxrwrOrq)rz__doc__r rgetattrr_is_in_bad_forkrrrr!r'r0r4r2r:rCrGrKrOrbrlrqrvrweventrx__all__r)rzs@rrs %(($9=I-11"C - 4 ; 3 ( .2210/4P@5@ r