i?dZddlZddlZddlZddlZddlZddlZddlZddlZddl m Z ddl m Z m Z mZmZmZmZddlmZmZddlZddlmZddlmZdd lmZmZmZmZmZdd lm Z!m"Z#erdd l$m%Z%Gd d eZ&GddeZ'GddeZ(GddeZ)GddeZ*gdZ+e,ej,dsedej,jZd<e,ej,dsedej,jZd<edej,jZd<edej,jZd<edej,jZd<edej,jZd<ddl.m/Z/m0Z0m1Z1m2Z2m3Z3m4Z4dZ5ejldZ7didjd"Z8d#Z9dkd$e:d%dfd&Z;dldmd'Zdnd)Z?dld d!d%e@eAe ffd*ZBdld d!d%e@eAe ffd+ZCdldmd,ZDdldmd-ZEd%e@eAe ffd.ZFd%e@eAe ffd/ZGdnd0ZHdnd1ZIdldmd2ZJdldmd3ZKdld d!d%eLfd4ZMdld d!d%eLfd5ZNdld d!d%eLfd6ZOdld d!d%eLfd7ZPed8eQ9dld d!d%eLfd:ZRed;eQ9dld d!d%eLfd<ZSdld=ZTdod d!d?e:d%eAfd@ZUdld d!d%eAfdAZVdld d!d%eWeLeLffdBZX dpdCe:d d!fdDZY dqdCeedFd%dfdGZZdEdEdEejdd>d>d>fdCeeAdHeeAdIeAdJeLd d!dKe:dLe:dMe:fdNZ\e e\eZ_]dOe^e&d%eLfdPZ_dQeAe*zd%e*fdRZ`dodjdSZadrdTZbdUeAfdVZcd%eAfdWZddsdXZedsdYZfedZeQ9d[eAfd\Zgd%eAfd]ZhGd^d_ZiGd`daeiZjdbeid%dfdcZkd%eifddZlGdedfe4Zmejldldgemd d!fdhZny)tzKThis package adds support for device memory management implemented in CUDA.N) signature)AnycastLiteralOptional TYPE_CHECKING TypedDict) deprecated NotRequired)_C) _dummy_type)_get_amdsmi_device_index_get_device_index_get_nvml_device_index _lazy_initis_initialized)memorysegments)DevicecpeZdZUdZeed<eed<eed<eeed<eeed<eeed<eeed<y ) _Framez1Frame information from memory profiler snapshots.filenamelinename fx_node_op fx_node_namefx_node_targetfx_original_traceN)__name__ __module__ __qualname____doc__str__annotations__intr K/var/www/html/engine/venv/lib/python3.12/site-packages/torch/cuda/memory.pyrr#s@;M I IC  c""$$"3''r(rcJeZdZUdZeed<eed<eed<eed<eeed<y)_BlockzMemory block information.sizerequested_sizeaddressstateframesN) r r!r"r#r&r%r$listrr'r(r)r+r+0s$# I L J Lr(r+c^eZdZUdZeed<eed<eed<eed<eed<eed<eeed<y ) _SegmentzMemory segment information.r. total_sizestream segment_typeallocated_size active_sizeblocksN) r r!r"r#r&r%r$r1r+r'r(r)r3r3:s0% LO K Lr(r3c`eZdZUdZeed<eeed<ee ed<eed<eed<eeed<y) _TraceEntryzMemory trace entry information.actionaddrr0r,r5 device_freeN) r r!r"r#r$r%r r&r1rr'r(r)r;r;Fs4) K c  L I KS!!r(r;c>eZdZUdZeeed<eeeeed<y) _SnapshotzMemory snapshot structure.r device_tracesN) r r!r"r#r1r3r%r r;r'r(r)r@r@Qs$$8ntD$5677r(r@)caching_allocator_alloccaching_allocator_deletecaching_allocator_enableget_per_process_memory_fractionset_per_process_memory_fraction empty_cache memory_statsmemory_stats_as_nested_dictreset_accumulated_memory_statsreset_peak_memory_statsreset_max_memory_allocatedreset_max_memory_cachedhost_memory_stats host_memory_stats_as_nested_dict#reset_accumulated_host_memory_statsreset_peak_host_memory_statsmemory_allocatedmax_memory_allocatedmemory_reservedmax_memory_reserved memory_cachedmax_memory_cachedmemory_snapshotmemory_summarylist_gpu_processes mem_get_infoget_allocator_backendCUDAPluggableAllocatorchange_current_allocatorMemPool use_mem_pool_cuda_CUDAAllocator_MemPool_cuda_beginAllocateToPool&_cuda_beginAllocateCurrentThreadToPool_cuda_endAllocateToPool_cuda_releasePool)rdrcrarerfrbcRttjjSN)rtorchr _cuda_cudaHostAllocatorr'r(r)_host_allocatorrksL 88 + + --r(c#Ktjj dtjjy#tjjwxYwwrh)rir _cuda_lock_mutex_cuda_unlock_mutexr'r(r) _free_mutexros> HH&  ##%##%sA(AA( A%%A(devicerc|tjj}t|}|tjj |}t |tjj jr |j}t |ts tdtjj|5tjj||cdddS#1swYyxYw)aPerform a memory allocation using the CUDA memory allocator. Memory is allocated for a given device and a stream, this function is intended to be used for interoperability with other frameworks. Allocated memory is released through :func:`~torch.cuda.caching_allocator_delete`. Args: size (int): number of bytes to be allocated. device (torch.device or int, optional): selected device. If it is ``None`` the default CUDA device is used. stream (torch.cuda.Stream or int, optional): selected stream. If is ``None`` then the default stream for the selected device is used. .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. NzrInvalid type for stream argument, must be `torch.cuda.Stream` or `int` representing a pointer to a existing stream)ricudacurrent_devicercurrent_stream isinstancestreamsStream cuda_streamr& TypeErrorrpr $_cuda_cudaCachingAllocator_raw_alloc)r,rpr5s r)rBrBs&~**, v &F ~**62&%**,,334## fc " #    6 "Kxx<= 1MB allocations). - ``small_pool``: statistics for the small allocation pool (as of June 2025, for size < 1MB allocations). Metric type: - ``current``: current value of this metric. - ``peak``: maximum value of this metric. - ``allocated``: historical total increase in this metric. - ``freed``: historical total decrease in this metric. In addition to the core statistics, we also provide some simple event counters: - ``"num_alloc_retries"``: number of failed ``cudaMalloc`` calls that result in a cache flush and retry. - ``"num_ooms"``: number of out-of-memory errors thrown. - ``"num_sync_all_streams"``: number of ``synchronize_and_free_events`` calls. - ``"num_device_alloc"``: number of CUDA allocation calls. This includes both cuMemMap and cudaMalloc. - ``"num_device_free"``: number of CUDA free calls. This includes both cuMemUnmap and cudaFree. The caching allocator can be configured via ENV to not split blocks larger than a defined size (see Memory Management section of the Cuda Semantics documentation). This helps avoid memory fragmentation but may have a performance penalty. Additional outputs to assist with tuning and evaluating impact: - ``"max_split_size"``: blocks above this size will not be split. - ``"oversize_allocations.{current,peak,allocated,freed}"``: number of over-size allocation requests received by the memory allocator. - ``"oversize_segments.{current,peak,allocated,freed}"``: number of over-size reserved segments from ``cudaMalloc()``. The caching allocator can be configured via ENV to round memory allocations in order to reduce fragmentation. Sometimes the overhead from rounding can be higher than the fragmentation it helps reduce. The following stat can be used to check if rounding adds too much overhead: - ``"requested_bytes.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``: memory requested by client code, compare this with allocated_bytes to check if allocation rounding adds too much overhead. Args: device (torch.device or int, optional): selected device. Returns statistics for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. .. note:: With :ref:`backend:cudaMallocAsync`, some stats are not meaningful, and are always reported as zero. ct|tr8t|dkDr|dz }|jD]\}}||z|yj ||fyNr.rudictlenitemsappendprefixobjkv_recurse_add_to_resultresults r)rz,memory_stats.._recurse_add_to_resulttZ c4 6{Q#   61&vz15 6 MM63- (r(r)rIsort collections OrderedDict)rpstatsrrs @@r)rHrHs@nF) (v 6E2u% KKM  " "6 **r(crtsiSt|d}tjj |S)zMReturn the result of :func:`~torch.cuda.memory_stats` as a nested dictionary.Toptional)rrrir _cuda_memoryStatsrs r)rIrIs.   v 5F 88 % %f --r(cZt|d}tjj|S)a}Reset the "accumulated" (historical) stats tracked by the CUDA memory allocator. See :func:`~torch.cuda.memory_stats` for details. Accumulated stats correspond to the `"allocated"` and `"freed"` keys in each individual stat dict, as well as `"num_alloc_retries"` and `"num_ooms"`. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. Tr)rrir !_cuda_resetAccumulatedMemoryStatsrs r)rJrJs% v 5F 88 5 5f ==r(cZt|d}tjj|S)aReset the "peak" stats tracked by the CUDA memory allocator. See :func:`~torch.cuda.memory_stats` for details. Peak stats correspond to the `"peak"` key in each individual stat dict. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. Tr)rrir _cuda_resetPeakMemoryStatsrs r)rKrKs%v 5F 88 . .v 66r(cgfdt}d|jtjS)aTReturn a dictionary of CUDA memory allocator statistics for a given device. The return value of this function is a dictionary of statistics, each of which is a non-negative integer. Core statistics: - ``"allocated.{current,peak,allocated,freed}"``: number of allocation requests received by the memory allocator. - ``"allocated_bytes.{current,peak,allocated,freed}"``: amount of allocated memory. - ``"segment.{current,peak,allocated,freed}"``: number of reserved segments from ``cudaMalloc()``. - ``"reserved_bytes.{current,peak,allocated,freed}"``: amount of reserved memory. For these core statistics, values are broken down as follows. Metric type: - ``current``: current value of this metric. - ``peak``: maximum value of this metric. - ``allocated``: historical total increase in this metric. - ``freed``: historical total decrease in this metric. In addition to the core statistics, we also provide some simple event counters: - ``"num_host_alloc"``: number of CUDA allocation calls. This includes both cudaHostAlloc and cudaHostRegister. - ``"num_host_free"``: number of CUDA free calls. This includes both cudaHostFree and cudaHostUnregister. Finally, we also provide some simple timing counters: - ``"host_alloc_time.{total,max,min,count,avg}"``: timing of allocation requests going through CUDA calls. - ``"host_free_time.{total,max,min,count,avg}"``: timing of free requests going through CUDA calls. For these timing statistics, values are broken down as follows. Metric type: - ``total``: total time spent. - ``max``: maximum value per call. - ``min``: minimum value per call. - ``count``: number of times it was called. - ``avg``: average time per call. ct|tr8t|dkDr|dz }|jD]\}}||z|yj ||fyrrrs r)rz1host_memory_stats.._recurse_add_to_resultrr(r)rOrrr)rrrs @@r)rNrNs>fF) - .E2u% KKM  " "6 **r(cVtsiStjjS)zRReturn the result of :func:`~torch.cuda.host_memory_stats` as a nested dictionary.)rrir _cuda_hostMemoryStatsr'r(r)rOrOs   88 ) ) ++r(c>tjjS)zReset the "accumulated" (historical) stats tracked by the host memory allocator. See :func:`~torch.cuda.host_memory_stats` for details. Accumulated stats correspond to the `"allocated"` and `"freed"` keys in each individual stat dict. )rir %_cuda_resetAccumulatedHostMemoryStatsr'r(r)rPrPs 88 9 9 ;;r(c>tjjS)zReset the "peak" stats tracked by the host memory allocator. See :func:`~torch.cuda.host_memory_stats` for details. Peak stats correspond to the `"peak"` key in each individual stat dict. )rir _cuda_resetPeakHostMemoryStatsr'r(r)rQrQs 88 2 2 44r(cRtjdtdt|S)aReset the starting point in tracking maximum GPU memory occupied by tensors for a given device. See :func:`~torch.cuda.max_memory_allocated` for details. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. warning:: This function now calls :func:`~torch.cuda.reset_peak_memory_stats`, which resets /all/ peak memory stats. .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. zytorch.cuda.reset_max_memory_allocated now calls torch.cuda.reset_peak_memory_stats, which resets /all/ peak memory stats. stacklevelrwarningswarn FutureWarningrKrs r)rLrL($ MM 0  #& 11r(cRtjdtdt|S)aReset the starting point in tracking maximum GPU memory managed by the caching allocator for a given device. See :func:`~torch.cuda.max_memory_cached` for details. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. warning:: This function now calls :func:`~torch.cuda.reset_peak_memory_stats`, which resets /all/ peak memory stats. .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. zvtorch.cuda.reset_max_memory_cached now calls torch.cuda.reset_peak_memory_stats, which resets /all/ peak memory stats.rrrrrs r)rMrM,rr(c:t|jddS)a[Return the current GPU memory occupied by tensors in bytes for a given device. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. note:: This is likely less than the amount shown in `nvidia-smi` since some unused memory can be held by the caching allocator and some context needs to be created on GPU. See :ref:`cuda-memory-management` for more details about GPU memory management. rzallocated_bytes.all.currentrrHgetrs r)rRrRGs v & * *+H! LLr(c:t|jddS)aReturn the maximum GPU memory occupied by tensors in bytes for a given device. By default, this returns the peak allocated memory since the beginning of this program. :func:`~torch.cuda.reset_peak_memory_stats` can be used to reset the starting point in tracking this metric. For example, these two functions can measure the peak allocated memory usage of each iteration in a training loop. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. rzallocated_bytes.all.peakrrrs r)rSrSXs$ v & * *+Eq IIr(c:t|jddS)aReturn the current GPU memory managed by the caching allocator in bytes for a given device. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. rzreserved_bytes.all.currentrrrs r)rTrTms v & * *+G KKr(c:t|jddS)aReturn the maximum GPU memory managed by the caching allocator in bytes for a given device. By default, this returns the peak cached memory since the beginning of this program. :func:`~torch.cuda.reset_peak_memory_stats` can be used to reset the starting point in tracking this metric. For example, these two functions can measure the peak cached memory amount of each iteration in a training loop. Args: device (torch.device or int, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. rzreserved_bytes.all.peakrrrs r)rUrU|s$ v & * *+Da HHr(zK`torch.cuda.memory_cached` has been renamed to `torch.cuda.memory_reserved`)categoryct|S)z4Deprecated; see :func:`~torch.cuda.memory_reserved`.r)rTrs r)rVrVs & ))r(zS`torch.cuda.max_memory_cached` has been renamed to `torch.cuda.max_memory_reserved`ct|S)z8Deprecated; see :func:`~torch.cuda.max_memory_reserved`.r)rUrs r)rWrWs f --r(cFtjj|dS)a(Return a snapshot of the CUDA memory allocator state across all devices. Interpreting the output of this function requires familiarity with the memory allocator internals. .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. r)rir _cuda_memorySnapshot) mempool_ids r)rXrXs 88 ( ( 4Z @@r(F abbreviatedct|d}t|}d}d}dd|fdd |fd d |fd d |fdd|fdd|fdd|fdd|fdd|fg }g}|jd|jd|jd|jd|jd|jd|D]\}}} |jdd|fg} |s"| jd| jdd \} } } }| D]x\}}|d!z|zd!z}||d"z}||d#z}||d$z}||d%z}| |} |} |} |}|jd&|d'd(| || d(| || d(| || d(| ||d& zd)d*|fd+d,|fg}|D]z\}}} |jd|d!z}||d"z}||d#z}||d$z}||d%z}|jd&|d'd(| ||d(| ||d(| ||d(| ||d& ||jdd-|d.}|jD]\}}|||j d!d/<d0d1j |j d3i|zd2zS)4aReturn a human-readable printout of the current memory allocator statistics for a given device. This can be useful to display periodically during training, or when handling out-of-memory exceptions. Args: device (torch.device or int, optional): selected device. Returns printout for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). abbreviated (bool, optional): whether to return an abbreviated summary (default: False). .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. Trrc^gd}|d}|ddD]}|dkrn|}|dz}|dz}|dd|S)N)zB KiBMiBGiBTiBPiBrri i6d r')szpref_szprefixesr new_prefixs r) _format_sizez$memory_summary.._format_sizes\=!"12, J#F 4KB tOG   R&""r(c`gd}|d}|ddD]}|dkrn|}|dz}|dz}|dd|dS)N)rKMrriq i7drr')cntpref_cntrrrs r) _format_countz%memory_summary.._format_counts_"!"12, J*$F DLC  H   b6(!$$r(allocated_byteszAllocated memory active_bytesz Active memoryrequested_byteszRequested memoryreserved_byteszGPU reserved memoryinactive_split_byteszNon-releasable memory allocation Allocationsactivez Active allocssegmentzGPU reserved segmentsinactive_splitzNon-releasable allocszK===========================================================================z= {_:16} PyTorch CUDA memory summary, device ID {device:<17d} zK---------------------------------------------------------------------------zX {_:9} CUDA OOMs: {num_ooms:<12d} | {_:6} cudaMalloc retries: {num_alloc_retries:<8d} zK Metric | Cur Usage | Peak Usage | Tot Alloc | Tot Freed all) large_poolz from large pool) small_poolz from small pool)NNNNrcurrentpeak allocatedfreedrz<21z | oversize_allocationszOversize allocationsoversize_segmentszOversize GPU segmentsr)_rp-|z| |z| r')rrHrrreplacejoinformat)rprrrrmetrics_to_displaylines metric_key metric_name formatter submetricscurrent_prefval peak_prefvalallocated_prefval freed_prefval submetric_keysubmetric_namerrrrrfmt_dictrrs r)rYrYs}"v 5F  'E # % . = ,7 . = 0,? !8,G }m4 ?M2 +]; 2MB  E LL LLPQ LL LLb LL LLU/A* K Xk*+    E F   E FK G'8-.8  )M>#% 5;FFY./G&)Df{23I&7*+E&")# $-! % LLN3's9Wo+N*OsS\]acoSpRqqtY(9:;3yP]?^>__`b  D !7G 5}E /A  * K Xc!*+Vf_%&;./ fw&'  C Igw$?#@IdTXDYCZZ]I./s9UE3J2K1 N    LL6*H *1()3$%* *U#**6X6 6 >>r(ctjjsJ ddl}ddlm} |j t|}|j|}|j|}nF ddl } |jt|} |j|}|j|}g}|j!d |t#|dk(r|j!d |D]u}tjjs|j$d z }|j&} n# j)||} | d d d z }| d} |j!d| dd|ddwdj-|S#t$rYywxYw#|$rYywxYw#t$rYywxYw#|j$rYywxYw#|j$rYywxYw#t*$r|} YwxYw)aReturn a human-readable printout of the running processes and their GPU memory use for a given device. This can be useful to display periodically during training, or when handling out-of-memory exceptions. Args: device (torch.device or int, optional): selected device. Returns printout for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). rNz4pynvml module not found, please install nvidia-ml-py)NVMLError_DriverNotLoadedz-cuda driver can't be loaded, is cuda enabled?z.amdsmi module not found, please install amdsmiz1amdsmi driver can't be loaded, is ROCm installed?z-amdsmi cannot list processes from other userszGPU:zno processes are runningi memory_usagevram_mempidzprocess z>10dz uses z>12.3fz MB GPU memory )riversionhippynvmlModuleNotFoundErrorrnvmlInitrnvmlDeviceGetHandleByIndex$nvmlDeviceGetComputeRunningProcessesamdsmi amdsmi_initAmdSmiExceptionramdsmi_get_processor_handlesamdsmi_get_gpu_process_listrr usedGpuMemoryramdsmi_get_gpu_process_infoAttributeErrorr) rpr rhandleprocsrrpmemr proc_infos r)rZrZ/s ==   J  5 C OO (/226:;;FC D  G    *&1 C88:6BF66v>E E LL4x! 5zQ /0  L}}  //[1C%%C ">>vqI N+J7;GCE"C xDzF|>JK L 99U c# JI J) CB C# DC D%% GF G%% CB C"   sjE+E:&F+F$F)F>+ E76E7:FF FFF&%F&)F;:F;> G  G c|tjj}t|d}tjj j |S)aReturn the global free and total GPU memory for a given device using cudaMemGetInfo. Args: device (torch.device or int or str, optional): selected device. Returns statistic for the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default) or if the device index is not specified. .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. Tr)rirrrsrcudartcudaMemGetInfors r)r[r[qsE~**, v 5F ::    - -f 55r(enabledc <tj||||||||yrh)r "_cuda_record_memory_history_legacy) r!record_contexttrace_alloc_max_entriestrace_alloc_record_contextrprecord_context_cpp clear_historycompile_contextglobal_record_annotationss r)_record_memory_history_legacyr+s*))"! r(r)r/rc^t|trt|g|i|St|g|i|S)a Enable recording of stack traces associated with memory allocations, so you can tell what allocated any piece of memory in :func:`torch.cuda.memory._snapshot()`. In addition to keeping stack traces with each current allocation and free, this will also enable recording of a history of all alloc/free events. Use :func:`torch.cuda.memory._snapshot()` to retrieve this information, and the tools in `_memory_viz.py` to visualize snapshots. Buffer behavior --------------- This will store up to `max_entries` instances of `TraceEntry` when enabled. Python trace collection defaults to `sys.maxsize`, meaning long-running or indefinitely running jobs should set a reasonable limit to avoid excessive memory use. Expect each entry to be several KB. Longer running workflows or those with smaller `max_entries` values will only store the last accumulated `max_entries` entries, meaning new entries overwrite older entries. C++ implementation for reference to ring buffer implementation: .. code-block:: cpp if (record_history) { if (alloc_trace->size() < alloc_trace_max_entries_) { alloc_trace->emplace_back(te); } else { (*alloc_trace)[alloc_trace_next++] = te; if (alloc_trace_next == alloc_trace_max_entries_) { alloc_trace_next = 0; } } } Latency impact -------------- The Python trace collection is fast (2us per trace), so you may consider enabling this on production jobs if you anticipate ever having to debug memory issues. C++ trace collection is also fast (~50ns/frame), which for many typical programs works out to ~2us per trace, but can vary depending on stack depth. Args: enabled (Literal[None, "state", "all"], optional): `None`, disable recording memory history. `"state"`, keep information for currently allocated memory. `"all"`, additionally keep a history of all alloc/free calls. Defaults to "all". context (Literal[None, "state", "alloc", "all"], optional): `None`, Do not record any tracebacks. `"state"`, Record tracebacks for currently allocated memory. `"alloc"`, additionally keep tracebacks for alloc calls. `"all"`, additionally keep tracebacks for free calls. Defaults to "all". stacks (Literal["python", "all"], optional): `"python"`, include Python, TorchScript, and inductor frames in tracebacks `"all"`, additionally include C++ frames Defaults to "all". max_entries (int, optional): Keep a maximum of `max_entries` alloc/free events in the recorded history recorded. )ruboolr+_record_memory_history_impl)r!argskwargss r)_record_memory_historyr1s9J'4 ,WFtFvFF*7DTDVDDr(contextstacks max_entriesr(r)r*c :tj|||||||yrh)r _cuda_record_memory_history)r!r2r3r4rpr(r)r*s r)r.r.s'""!r(r0cddlm}tjtj|d}d}|s|S|D]}d|vs d|vs|d}|d}|j t jj|sGddl m }|j|}|a|jdi} |jdi} |jd d} | j|| z } | | | vs| | } | jd }| jd }| jd }| jd }||d<||d<t||d<|r||d<|dz }"|S)z Augment a list of frames with FX debug information. Args: frames: List of frame dictionaries to augment Returns: The count of frames that were augmented. r)FX_GRAPH_MODULE_FILE_PREFIXz.*\.py$rr)_FX_METADATA_REGISTRY lineno_map node_metadataprologue_start stack_traceoprtargetrrrrr) torch.fx.graph_moduler8recompileescapesearchospathbasenametorch.fx.tracebackr9rr$)r0r8_FX_GENERATED_PATTERNcountframerlinenor9metadatar:r;r<node_idx node_infooriginal_tracenode_op node_name node_targets r)_augment_framesrTsBJJ II1 2 37; E  '  6U?Z(H6]F)//0@0@0JK A,00:H!lB7J$LL"=M%\\*:A>N"~~f~&=>H#M(A)(3 !*}!=#---%MM&1 'mmH5 '.l#(1n%*-k*:&'"1?E-. O'R Lr(snapshotct|tr9t|d5}ttt j |}dddn|}d}dvr/|dD]'}d|vs|dD]}d|vs|t|dz })d|vr8|dD]0}|D])}t|tsd|vs|t|dz }+2|S#1swY|xYw)a Augment a memory snapshot with original source stack traces from FX metadata. IMPORTANT: This function reads from a global in-memory registry (_FX_METADATA_REGISTRY) that is populated during graph module compilation. It must be called in the same Python process where the FX graphs were compiled. It cannot be used to augment snapshots loaded from disk in a different process. Args: snapshot: Either a memory snapshot dict or path to a snapshot pickle file Returns: The augmented snapshot dictionary with fx_node_op, fx_node_name, fx_original_trace, and fx_node_info fields added to frames rbNrrr9r0rA) rur$openrr@pickleloadrTr)rUf snapshot_dictaugmented_countrblock trace_list trace_entrys r)%_augment_memory_snapshot_stack_tracesra@s ((C (D ! 1MB) allocated_size: int # size of memory in use active_size: int # size of memory in use or in active_awaiting_free state blocks: List[Block] class Block(TypedDict): # A piece of memory returned from the allocator, or # current cached but inactive. size: int requested_size: int # size requested during malloc, may be smaller than # size due to rounding address: int state: Literal[ "active_allocated", # used by a tensor "active_awaiting_free", # waiting for another stream to finish using # this, then it will become free "inactive", ] # free for reuse frames: List[Frame] # stack trace from where the allocation occurred class Frame(TypedDict): filename: str line: int name: str # Optional FX debug fields (present when augment_with_fx_traces=True # and the frame corresponds to FX-generated code) fx_node_op: str # FX node operation type (e.g., 'call_function', 'output') fx_node_name: str # FX node name (e.g., 'linear', 'relu_1') fx_original_trace: str # Original model source code stack trace class TraceEntry(TypedDict): # When `torch.cuda.memory._record_memory_history()` is enabled, # the snapshot will contain TraceEntry objects that record each # action the allocator took. action: Literal[ "alloc" # memory allocated "free_requested", # the allocated received a call to free memory "free_completed", # the memory that was requested to be freed is now # able to be used in future allocation calls "segment_alloc", # the caching allocator ask cudaMalloc for more memory # and added it as a segment in its cache "segment_free", # the caching allocator called cudaFree to return memory # to cuda possibly trying free up memory to # allocate more segments or because empty_caches was called "oom", # the allocator threw an OOM exception. 'size' is # the requested number of bytes that did not succeed "snapshot", # the allocator generated a memory snapshot # useful to coorelate a previously taken # snapshot with this trace ] addr: int # not present for OOM frames: List[Frame] size: int stream: int device_free: int # only present for OOM, the amount of # memory cuda still reports to be free Args: device: Device to capture snapshot for. If None, captures for current device. augment_with_fx_traces: If True, augment stack trace frames with FX debug information that maps generated FX code back to original model source code. This adds fx_node_op, fx_node_name, fx_original_trace, and fx_node_info fields to Frame objects. Default: False. Returns: The Snapshot dictionary object N)r rra)rpaugment_with_fx_tracesss r) _snapshotreps'x %A 1! 4 Hr(ct|}t|d5}tj||dddy#1swYyxYw)a Save a pickled version of the `torch.memory._snapshot()` dictionary to a file. This file can be opened by the interactive snapshot viewer at pytorch.org/memory_viz Snapshot file sizes scale with `max_entries` and stack trace depth per entry, with several KB per entry. These can easily be in the GB range for longer running workflows with large `max_entries`. Args: filename (str, optional): Name of the file to create. Defaults to "dump_snapshot.pickle". augment_with_fx_traces (bool, optional): If True, augment the snapshot with FX debug information before dumping. This maps generated FX code stack traces back to original model source code. Defaults to False. verbose (bool, optional): If True and augment_with_fx_traces is True, print verbose debug output during augmentation. Defaults to False. )rcwbN)rerXrYdump)rrcrdr[s r)_dump_snapshotris>$ )?@A h  Aqs 9ArMcBtjj|y)a Set custom metadata that will be attached to all subsequent CUDA memory allocations. This metadata will be recorded in the memory snapshot for all allocations made after this call until the metadata is cleared or changed. Args: metadata (str): Custom metadata string to attach to allocations. Pass an empty string to clear the metadata. N)rir _cuda_setMemoryMetadata)rMs r)_set_memory_metadatarls HH$$X.r(c>tjjS)z Get the current custom metadata that is being attached to CUDA memory allocations. Returns: str: The current metadata string, or empty string if no metadata is set. )rir _cuda_getMemoryMetadatar'r(r)_get_memory_metadataros 88 + + --r(c| t}t|d5}|jt|dddy#1swYyxYwNw)rerXwrite _segmentsrrUr[s r)_save_segment_usagervsA; h %  (#$%%% =Ac| t}t|d5}|jt|dddy#1swYyxYwrq)rerXrs_memoryrus r)_save_memory_usagerz sA; h # !"###rwzitorch.cuda._set_allocator_settings is deprecated. Use torch._C._accelerator_setAllocatorSettings instead.envc@tjj|Srh)rir !_accelerator_setAllocatorSettings)r{s r)_set_allocator_settingsr~s 88 5 5c ::r(c>tjjS)adReturn a string describing the active allocator backend as set by ``PYTORCH_ALLOC_CONF``. Currently available backends are ``native`` (PyTorch's native caching allocator) and `cudaMallocAsync`` (CUDA's built-in asynchronous allocator). .. note:: See :ref:`cuda-memory-management` for details on choosing the allocator backend. )rir _cuda_getAllocatorBackendr'r(r)r\r\s 88 - - //r(cJeZdZdZdej j fdZdZy)_CUDAAllocatorz-Wrapper over internal CUDA memory allocators. allocatorc||_yrh _allocator)selfrs r)__init__z_CUDAAllocator.__init__*s #r(c|jSrhr)rs r)rz_CUDAAllocator.allocator-s r(N) r r!r"r#rir rarrr'r(r)rr's 8$%((">">$r(rc$eZdZdZdededefdZy)r]z,CUDA memory allocator loaded from a so file.path_to_so_file alloc_fn_name free_fn_namechtj|}tjt||tjj }tjt||tjj }|J|Jt jj|||_ y)aMemory allocators are compiled in .so files and loaded dynamically using ctypes. To change the active allocator use the :func:`torch.memory.cuda.change_current_allocator` function. Args: path_to_so_file(str): Path in the filesystem to the `.so` file containing the allocator functions alloc_fn_name(str): Name of the function to perform the memory allocation in the so file. The signature must be: void* alloc_fn_name(ssize_t size, int device, cudaStream_t stream); free_fn_name(str): Name of the function to perform the memory release in the so file. The signature must be: void free_fn_name(void* ptr, size_t size, cudaStream_t stream); .. warning:: This is currently supported only in unix OSs .. note:: See :ref:`cuda-memory-management` for details on creating and using a custom allocator N) ctypesCDLLrgetattrc_void_pr~rir _cuda_customAllocatorr)rrrrralloc_fnfree_fns r)rzCUDAPluggableAllocator.__init__4s*KK0 ;;wy-@&//RXX++gi>PVV###"""((887Kr(N)r r!r"r#r$rr'r(r)r]r]1s#7LLCLsLr(r]rc^tjj|jy)axChange the currently used memory allocator to be the one provided. If the current allocator has already been used/initialized, this function will error. Args: allocator (torch.cuda.memory._CUDAAllocator): allocator to be set as the active one. .. note:: See :ref:`cuda-memory-management` for details on creating and using a custom allocator N)rir _cuda_changeCurrentAllocatorr)rs r)r^r^Qs HH)))*=*=*?@r(cPttjjS)zReturn the allocator being currently used. .. note:: See :ref:`cuda-memory-management` for details on creating and using a custom allocator )rrir _cuda_getAllocatorr'r(r)_get_current_allocatorr_s %((557 88r(ceZdZdZ d deededeffd Zede e e fffd Z edeeffd Z de ffd Z d ZxZS) r_aOMemPool represents a pool of memory in a caching allocator. Currently, it's just the ID of the pool object maintained in the CUDACachingAllocator. Args: allocator(torch._C._cuda_CUDAAllocator, optional): a torch._C._cuda_CUDAAllocator object that can be used to define how memory gets allocated in the pool. If :attr:`allocator` is ``None`` (default), memory allocation follows the default/ current configuration of the CUDACachingAllocator. use_on_oom(bool): a bool that indicates if this pool can be used as a last resort if a memory allocation outside of the pool fails due to Out Of Memory. This is False by default. no_split(bool): a bool that indicates if this pool should not split a segment. This is False by default. r use_on_oomno_splitc*t||d||y)NT)superr)rrrr __class__s r)rzMemPool.__init__ys D*h?r(rct|S)z3Returns the ID of this pool as a tuple of two ints.)ridrrs r)rz MemPool.idswzr(ct|S)z9Returns the allocator this MemPool routes allocations to.)rrrs r)rzMemPool.allocatorsw  r(c t|S)z)Returns the reference count of this pool.)r use_countrs r)rzMemPool.use_countsw ""r(cXtjj|j}|S)aMReturn a snapshot of the CUDA memory allocator pool state across all devices. Interpreting the output of this function requires familiarity with the memory allocator internals. .. note:: See :ref:`cuda-memory-management` for more details about GPU memory management. )rirrrXr)rrUs r)rUzMemPool.snapshots!::--dgg6r()NFF)r r!r"r#rrar-rpropertytupler&rrrrU __classcell__)rs@r)r_r_hs$48  @/0@@ @E#s(O!8$78!!#3# r(r_poolc#RK|tjjn t|}t ||j  dt ||j t||j y#t ||j t||j wxYww)aA context manager that routes allocations to a given pool. Args: pool(torch.cuda.MemPool): a MemPool object to be made active so that allocations route to this pool. device (torch.device or int, optional): selected device. Uses MemPool on the current device, given by :func:`~torch.cuda.current_device`, if :attr:`device` is ``None`` (default). .. note:: This context manager makes only current thread's allocations route to the given pool. If a new thread is spawned inside the context manager (e.g. by calling backward) the allocations in that thread will not route to the given pool. N)rirrrsrrdrrerf)rrp device_indexs r)r`r`s|$(.~ !!#;LV;T+<A1  dgg6,0 dgg6,0sAB'A6 -B'6.B$$B')NN)rpr)Trh)rprrN)rN)NF)TrFNFFFF)r)zdump_snapshot.pickleF)z output.svgN)or#r contextlibrrErYrAsysrinspectrtypingrrrrrr typing_extensionsr r rir torch._utilsr rrrrrr _memory_vizrryrrt torch.typesrrr+r3r;r@__all__hasattr__dict__torch._CrdrcrarerfrbrkcontextmanagerrorBrCr-rDrFrrErGrr$rHrIrJrKrNrOrPrQrLrMr&rRrSrTrUrrVrWrXrYrZrr[r+r1maxsizer. __signature__r1rTrarerirlrorvrzr~r\rr]r^rr_r`r'r(r)rs5R II5 $B" (Y (Y y ")"8 8 Fuxx.//:;P/QEHH+,uxx$$/ $;EHHj!5@#6EHH12CN0CEHH>?4?!4EHH/0.99L-MEHH)*.  &&!KH<":D:D: 76 4H 4 4 $f+f+T#s(^f+R..DcN.>(7&B+4S>B+J,$sCx.,<52626MXMM"JJSJ* LH L LIICI* Q *(*c* *  Y .h.#. . A|?8|?|?#|?~?x?3?D66U38_6*$#   227HE gn- .HE HEX#"{{!&+ c] c]       $*(11L'M$>DL>S>B-Io--`_ D0 /3 /.c.%# o ;; ; 0s 0L^L@ A A4 A993h3l 1w111r(