Ë ÌÓ^j~#ãóÀ—UdZddlZddlmZeejdd„«Zdaejjed<dejjfd„Z de fd „Zd d „Zd!de e zejzdefd„Zd!d ede e zejzddfd„Zde ddfd„Zd d„Zd d„Zd d„Zde fd„Zde fd„Zde fd„Zde fd„Zd„Zdefd„Zd"d„ZddlmZddlm Z gd¢Z!y)#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. éN)ÚTensorÚ_mps_is_in_bad_forkcó—y)NF©róú]/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/torch/mps/__init__.pyúr srÚ_default_mps_generatorÚreturncóV—t€tjj«atS)N)r ÚtorchÚ_CÚ_mps_get_default_generatorrrrÚ_get_default_mps_generatorrs!€äÐ%Ü!&§¡×!DÑ!DÓ!FÐÜ!Ð!rcóˆ—ttjjxrtjj ««S)z,Returns the number of available MPS devices.)Úintr rÚ_has_mpsÚ_mps_is_availablerrrÚdevice_countrs)€äŒux‰x× Ñ ÒA¤U§X¡X×%?Ñ%?Ó%AÓBÐBrcó>—tjj«S)zAWaits for all kernels in all streams on a MPS device to complete.)r rÚ_mps_deviceSynchronizerrrÚsynchronizers€ä8‰8×*Ñ*Ó,Ð,rÚdevicecó2—t«j«S)zùReturns 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). )rÚ get_state)rs rÚ get_rng_stater#s€ô&Ó'×1Ñ1Ó3Ð3rÚ new_statecóv—|jtj¬«}t«j |«y)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). )Ú memory_formatN)Úcloner Úcontiguous_formatrÚ set_state)rrÚnew_state_copys rÚ set_rng_stater$-s+€ð—_‘_´5×3JÑ3J_ÓK€NÜÓ ×*Ñ*¨>Õ:rÚseedcó‚—tjjsyt|«}t «j|«y)zbSets the seed for generating random numbers. Args: seed (int): The desired seed. N)r rrrrÚmanual_seed)r%s rr'r'9s.€ô8‰8×ÒØÜˆt‹9€DÜÓ ×,Ñ,¨TÕ2rcó4—t«j«y)z?Sets the seed for generating random numbers to a random number.N)rr%rrrr%r%Is€äÓ ×%Ñ%Õ'rcó@—tjj«y)z‹Releases all unoccupied cached memory currently held by the caching allocator so that those can be used in other GPU applications. N)r rÚ_mps_emptyCacherrrÚempty_cacher+Ns€ô ‡HH×ÑÕrcóª—t|t«std«‚|dks|dkDrtd|›d«‚tj j |«y)aÅSet 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`rézInvalid fraction value: z. Allowed range: 0~2N)Ú isinstanceÚfloatÚ TypeErrorÚ ValueErrorr rÚ_mps_setMemoryFraction)Úfractions rÚset_per_process_memory_fractionr4UsO€ô"h¤Ô&ÜÐMÓNÐNØ!‚|x !’|ÜÐ3°H°:Ð=QÐRÓSÐSä ‡HH×#Ñ# HÕ-rcó>—tjj«S)zµReturns the current GPU memory occupied by tensors in bytes. .. note:: The returned size does not include cached allocations in memory pools of MPSAllocator. )r rÚ_mps_currentAllocatedMemoryrrrÚcurrent_allocated_memoryr7ns€ô8‰8×/Ñ/Ó1Ð1rcó>—tjj«S)zçReturns 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 rÚ_mps_driverAllocatedMemoryrrrÚdriver_allocated_memoryr:xs€ô8‰8×.Ñ.Ó0Ð0rcó>—tjj«S)zÂReturns recommended max Working set size for GPU memory in bytes. .. note:: Recommended max working set size for Metal. returned from device.recommendedMaxWorkingSetSize. )r rÚ_mps_recommendedMaxMemoryrrrÚrecommended_max_memoryr=‚s€ô8‰8×-Ñ-Ó/Ð/rÚsourcecóT—ddlm}ddlm}t t jd«std«‚||jd«Dcgc]}|dz‘Œ c}|t«jjdzgt««}t jj|«Scc}w)aÎCompiles 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) r)ÚPath)Ú_embed_headersÚ_mps_compileShaderúMPS is not availableú Úinclude) Úpathlibr@Útorch.utils._cpp_embed_headersrAÚhasattrr rÚRuntimeErrorÚsplitÚ__file__ÚparentÚsetrB)r>r@rAÚls rÚcompile_shaderrOŒs†€õå=ä”5—8‘8Ð1Ô2ÜÐ1Ó2Ð2Ù Ø!Ÿ<™<¨Ó-Ö.aˆˆT‹Ò.Ù Œh‹× Ñ × %Ñ %¨ Ñ 1Ð2Ü‹ó€Fô 8‰8×&Ñ& vÓ.Ð.ùò /sÁB%cóä—ddl}t|ttf«rMt t jd«std«‚t jjt|««St|t|jf«rMt t jd«std«‚t jjt|««Stdt|«j›«‚)aƒLoads 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) rNÚ_mps_loadMetalllibrCÚ_mps_loadMetallibFromPathzexpected bytes or path, got )Úosr.ÚbytesÚ bytearrayrHr rrIrQÚstrÚPathLikerRr0ÚtypeÚ__name__)r>rSs rÚ load_metallibrZ¦s¯€ó$ä&œ5¤)Ð,Ô-Ü”u—x‘xÐ!5Ô6ÜÐ5Ó6Ð6Üx‰x×*Ñ*¬5°«=Ó9Ð9Ü FœS "§+¡+Ð.Ô /Ü”u—x‘xÐ!<Ô=ÜÐ5Ó6Ð6Üx‰x×1Ñ1´#°f³+Ó>Ð>äÐ6´t¸F³|×7LÑ7LÐ6MÐNÓOÐOrcó—t«dkDS)Nr)rrrrÚis_availabler\Æs€Ü‹>˜AÑÐrcó@—tjj|«S)a Returns a CPU :class:`torch.UntypedStorage` that aliases the host-visible contents of the MTLBuffer backing ``storage``. The returned storage shares memory with ``storage``: writes through the CPU alias land directly in the MPS-allocated MTLBuffer, avoiding a CPU->MPS 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 rÚ_mps_host_alias_storage)Ústorages rÚ_host_alias_storager`Ês€ô:8‰8×+Ñ+¨GÓ4Ð4ré)Úprofiler)ÚEvent)rOrZrrr'r%r$rr+r4r7r:rcrbr=r\)rN)Úmps)r_útorch.UntypedStoragerre)"Ú__doc__r rÚgetattrrÚ_is_in_bad_forkr Ú GeneratorÚ__annotations__rrrrrVrrr$r'r%r+r4r7r:r=rOrZÚboolr\r`ÚrbÚeventrcÚ__all__rrrúros'ðòó Ýñ˜%Ÿ(™(Ð$9¹=ÓI€Ø-1Ð˜Ÿ™×*Ñ*Ó1ð" E§H¡H×$6Ñ$6ó"ðCcóCó -ñ 4˜# ™) e§l¡lÑ2ð4¸vó4ñ ;˜Vð ;¨S°3©Y¸¿¹Ñ-Eð ;ÐRVó ;ð 3cð 3˜dó 3ó (ó ó.ð22 #ó2ð1 ó1ð0 ó0ð/˜3ó/ò4Pð@dóó5õ@Ýòr