Intel Gaudi GPU Migration APIs

This document provides additional information on the process of GPU migration, which includes a full list of Python API calls and their compatibility with Intel® Gaudi® AI accelerator.

For further details on using the GPU Migration Toolkit, refer to GPU Migration Toolkit.

GPU Migration APIs Support

The table below describes which APIs are supported by the GPU Migration toolkit. They are classified as hpu_match, hpu_modified and hpu_mismatch in the HPU implementation column. These terms are described in detail below.

hpu_match

Intel Gaudi implementations replicate the functionality of API calls, and GPU Migration creates a mapping, as explained below. When the calls match functionally, no further configuration is required.

• GPU Migration maps GPU arguments to HPU arguments. For example:

  x = torch.ones(1, device="cuda")
  # GPU Migration changes the argument `device` from "cuda" to "hpu".
  

• GPU Migration maps GPU output value to HPU output value. For example:

  backend = torch.distributed.get_backend()
  # GPU Migration changes the output backend from "hccl" to "nccl".
  

• GPU Migration maps GPU calls to HPU calls. For example:

  torch.cuda.memory_summary()
  # GPU Migration maps 'torch.cuda.memory_summary' to 'torch.hpu.memory_summary'.
  

hpu_modified

Certain Gaudi-specific calls have been used to enable the replication of GPU-specific functionalities and guarantee compatibility with HPU. For example, GPU Migration toolkit maps torch.cuda.FloatTensor to torch.FloatTensor and torch.Tensor.to(“hpu”). As a result, torch.cuda.FloatTensor is no longer a tensor type, and can only be employed to create a tensor on HPU.

Please review the limitations stated in each call’s description in the table below. The number of calls in this category will decrease in future releases.

hpu_mismatch

These calls are currently incompatible with Gaudi implementation, so the GPU Migration either creates an inactive call or raises an exception.

• Inactive Call - Since the functions cannot be executed on HPU, GPU Migration converts them into inactive calls. For example:

  torch.cuda.empty_cache()
  

• NotImplementedError Exception - Certain API calls cannot be processed by either the GPU Migration toolkit or the Intel Gaudi software, causing the training process to be interrupted with a ‘NotImplementedError’ exception. For example:

  torch.cuda.comm.broadcast
  

If you encounter one of hpu_mismatch scenarios described above, it is necessary to generate a logging file, as described in Enabling GPU Migration Logging section and then proceed to manually modify the mismatching functionality and restart the training.

The table below lists all the converted calls and explain their compatibility with HPU.

Call	HPU Implementation	Description
apex.amp.frontend.initialize	hpu_modified	Uses native torch.autocast with device_type=”hpu” instead.
apex.normalization.fused_layer_norm.FusedLayerNorm	hpu_modified	Uses torch.nn.functional.layer_norm.
apex.normalization.fused_layer_norm.FusedRMSNorm	hpu_modified	Uses python manual RMSNorm implementation; ignores elementwise_affine option.
apex.normalization.fused_layer_norm.MixedFusedLayerNorm	hpu_modified	Uses torch.nn.functional.layer_norm.
apex.normalization.fused_layer_norm.MixedFusedRMSNorm	hpu_modified	Maps to FusedRMSNorm.
apex.optimizers.FusedLAMB	hpu_match	Maps apex.optimizers.FusedLamb to hpex.optimizers.FusedLamb.
apex.optimizers.FusedAdagrad	hpu_modified	Does not support adagrad_w_mode.
apex.optimizers.FusedAdam	hpu_modified	Uses hpex_FusedAdamW if adam_w_mode is True. Otherwise, it uses torch.otpim.Adam.
apex.optimizers.FusedSGD	hpu_modified	Ignores wd_after_momentum and materialize_master_grads.
apex.optimizers.FusedAdagrad.zero_grad	hpu_match	Maps apex.optimizers.FusedAdagrad.zero_grad to hpex.optimizers.FusedAdagrad.zero_grad.
apex.optimizers.FusedAdam.zero_grad	hpu_match	Maps apex.optimizers.FusedAdam.zero_grad to hpex.optimizers.FusedAdamW.zero_grad if adam_w_mode is set to True, otherwise maps to torch.optim.Adam.zero_grad.
apex.optimizers.FusedLAMB.zero_grad	hpu_match	Maps apex.optimizers.FusedLamb.zero_grad to hpex.optimizers.FusedLamb.zero_grad.
apex.optimizers.FusedSGD.zero_grad	hpu_match	Maps apex.optimizers.FusedSGD.zero_grad to hpex.optimizers.FusedSGD.zero_grad.
apex.parallel.DistributedDataParallel	hpu_mismatch	Raises NotImplementedError. Please use native torch.nn.DistributedDataParallel.
pynvml.nvml.nvmlDeviceClearCpuAffinity	hpu_match	Maps pynvml.nvmlDeviceClearCpuAffinity to pyhlml.hlmlDeviceClearCpuAffinity.
pynvml.nvml.nvmlDeviceGetClockInfo	hpu_match	Maps pynvml.nvmlDeviceGetClockInfo to pyhlml.hlmlDeviceGetClockInfo.
pynvml.nvml.nvmlDeviceGetCount	hpu_match	Maps pynvml.nvmlDeviceGetCount to pyhlml.hlmlDeviceGetCount.
pynvml.nvml.nvmlDeviceGetCpuAffinity	hpu_match	Maps pynvml.nvmlDeviceGetCpuAffinity to pyhlml.hlmlDeviceGetCpuAffinity.
pynvml.nvml.nvmlDeviceGetCurrentClocksThrottleReasons	hpu_match	Maps pynvml.nvmlDeviceGetCurrentClocksThrottleReasons to pyhlml.hlmlDeviceGetCurrentClocksThrottleReasons.
pynvml.nvml.nvmlDeviceGetHandleByIndex	hpu_match	Maps pynvml.nvmlDeviceGetHandleByIndex to pyhlml.hlmlDeviceGetHandleByIndex.
pynvml.nvml.nvmlDeviceGetHandleByUUID	hpu_match	Maps pynvml.nvmlDeviceGetHandleByUUID to pyhlml.hlmlDeviceGetHandleByUUID.
pynvml.nvml.nvmlDeviceGetMaxClockInfo	hpu_match	Maps pynvml.nvmlDeviceGetMaxClockInfo to pyhlml.hlmlDeviceGetMaxClockInfo.
pynvml.nvml.nvmlDeviceGetMemoryErrorCounter	hpu_match	Maps pynvml.nvmlDeviceGetMemoryErrorCounter to pyhlml.hlmlDeviceGetMemoryErrorCounter.
pynvml.nvml.nvmlDeviceGetMemoryInfo	hpu_match	Maps pynvml.nvmlDeviceGetMemoryInfo to pyhlml.hlmlDeviceGetMemoryInfo.
pynvml.nvml.nvmlDeviceGetMinorNumber	hpu_match	Maps pynvml.nvmlDeviceGetMinorNumber to pyhlml.hlmlDeviceGetMinorNumber.
pynvml.nvml.nvmlDeviceGetName	hpu_match	Maps pynvml.nvmlDeviceGetName to pyhlml.hlmlDeviceGetName.
pynvml.nvml.nvmlDeviceGetPerformanceState	hpu_match	Maps pynvml.nvmlDeviceGetPerformanceState to pyhlml.hlmlDeviceGetPerformanceState.
pynvml.nvml.nvmlDeviceGetPersistenceMode	hpu_match	Maps pynvml.nvmlDeviceGetPersistenceMode to pyhlml.hlmlDeviceGetPersistenceMode.
pynvml.nvml.nvmlDeviceGetPowerManagementDefaultLimit	hpu_match	Maps pynvml.nvmlDeviceGetPowerManagementDefaultLimit to pyhlml.hlmlDeviceGetPowerManagementDefaultLimit.
pynvml.nvml.nvmlDeviceGetPowerUsage	hpu_match	Maps pynvml.nvmlDeviceGetPowerUsage to pyhlml.hlmlDeviceGetPowerUsage.
pynvml.nvml.nvmlDeviceGetSerial	hpu_match	Maps pynvml.nvmlDeviceGetSerial to pyhlml.hlmlDeviceGetSerial.
pynvml.nvml.nvmlDeviceGetTemperature	hpu_match	Maps pynvml.nvmlDeviceGetTemperature to pyhlml.hlmlDeviceGetTemperature.
pynvml.nvml.nvmlDeviceGetTemperatureThreshold	hpu_match	Maps pynvml.nvmlDeviceGetTemperatureThreshold to pyhlml.hlmlDeviceGetTemperatureThreshold.
pynvml.nvml.nvmlDeviceGetTotalEnergyConsumption	hpu_match	Maps pynvml.nvmlDeviceGetTotalEnergyConsumption to pyhlml.hlmlDeviceGetTotalEnergyConsumption.
pynvml.nvml.nvmlDeviceGetUUID	hpu_match	Maps pynvml.nvmlDeviceGetUUID to pyhlml.hlmlDeviceGetUUID.
pynvml.nvml.nvmlDeviceGetUtilizationRates	hpu_match	Maps pynvml.nvmlDeviceGetUtilizationRates to pyhlml.hlmlDeviceGetUtilizationRates.
pynvml.nvml.nvmlDeviceGetViolationStatus	hpu_match	Maps pynvml.nvmlDeviceGetViolationStatus to pyhlml.hlmlDeviceGetViolationStatus.
pynvml.nvml.nvmlDeviceRegisterEvents	hpu_match	Maps pynvml.nvmlDeviceRegisterEvents to pyhlml.hlmlDeviceRegisterEvents.
pynvml.nvml.nvmlDeviceSetCpuAffinity	hpu_match	Maps pynvml.nvmlDeviceSetCpuAffinity to pyhlml.hlmlDeviceSetCpuAffinity.
pynvml.nvml.nvmlEventSetCreate	hpu_match	Maps pynvml.nvmlEventSetCreate to pyhlml.hlmlEventSetCreate.
pynvml.nvml.nvmlEventSetFree	hpu_match	Maps pynvml.nvmlEventSetFree to pyhlml.hlmlEventSetFree.
pynvml.nvml.nvmlEventSetWait	hpu_match	Maps pynvml.nvmlEventSetWait to pyhlml.hlmlEventSetWait.
pynvml.nvml.nvmlInit	hpu_match	Maps pynvml.nvmlInit to pyhlml.hlmlInit.
pynvml.nvml.nvmlShutdown	hpu_match	Maps pynvml.nvmlShutdown to pyhlml.hlmlShutdown.
torch.autocast	hpu_match	Changes device_type to “hpu” and dtype to None, leaving the data type casting decision to hpu autocast engine.
torch.compile	hpu_match	Maps inductor backend to hpu_backend. For further details, refer to PyTorch.
torch.empty	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.empty_like	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.randn	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.rand_like	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.rand	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.randint	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.randn_like	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.zeros	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.zeros_like	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.arange	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.range	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.full	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.full_like	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.eye	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.empty_strided	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.ones	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.ones_like	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.Tensor.cuda	hpu_match	Changes the function to torch.Tensor.to(“hpu”).
torch.Tensor.half	hpu_match	Changes the function to torch.Tensor.to(torch.bfloat16)
torch.Tensor.pin_memory	hpu_match	Pins memory to HPU device instead of GPU.
torch.Tensor.to	hpu_match	Changes device arguments from “cuda” to “hpu” and dtype from torch.float16 to torch.bfloat16.
torch.Tensor.type	hpu_match	Maps to torch.Tensor.type(“torch.<Scalar>Tensor”) and torch.Tensor.to(“hpu”) if dtype is torch.hpu.<Scalar>Tensor. If dtype is None this changes the tensor type to the same type as CUDA tensor.
torch.Tensor.record_stream	hpu_mismatch	Inactive Call.
torch.Tensor.numpy	hpu_modified	If tensor datatype is torch.bfloat16, change the function to torch.Tensor.to(torch.float16) and torch.Tensor.numpy.
torch.cuda.StreamContext	hpu_match	Maps torch.cuda.StreamContext to torch.hpu.StreamContext.
torch.cuda.can_device_access_peer	hpu_match	Maps torch.cuda.can_device_access_peer to torch.hpu.can_device_access_peer.
torch.cuda.current_stream	hpu_match	Maps torch.cuda.current_stream to torch.hpu.current_stream.
torch.cuda.default_stream	hpu_match	Maps torch.cuda.default_stream to torch.hpu.default_stream.
torch.cuda.get_device_name	hpu_match	Maps torch.cuda.get_device_name to torch.hpu.get_device_name.
torch.cuda.get_gencode_flags	hpu_match	Maps torch.cuda.get_gencode_flags to torch.hpu.get_gencode_flags.
torch.cuda.init	hpu_match	Maps torch.cuda.init to torch.hpu.init.
torch.cuda.is_available	hpu_match	Maps torch.cuda.is_available to torch.hpu.is_available.
torch.cuda.is_bf16_supported	hpu_match	Maps torch.cuda.is_bf16_supported to torch.hpu.is_bf16_supported.
torch.cuda.is_initialized	hpu_match	Maps torch.cuda.is_initialized to torch.hpu.is_initialized.
torch.cuda.memory_usage	hpu_match	Maps to torch.hpu.memory_usage().
torch.cuda.set_device	hpu_match	Maps to torch.hpu.set_device
torch.cuda.set_stream	hpu_match	Maps torch.cuda.set_stream to torch.hpu.set_stream.
torch.cuda.set_sync_debug_mode	hpu_match	Maps torch.cuda.set_sync_debug_mode to torch.hpu.set_sync_debug_mode.
torch.cuda.synchronize	hpu_match	Maps torch.cuda.synchronize to torch.hpu.synchronize.
torch.cuda.utilization	hpu_match	Maps to torch.hpu.utilization().
torch.cuda.BFloat16Tensor	hpu_modified	Maps torch.cuda.BFloat16Tensor to torch.BFloat16Tensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.BoolTensor	hpu_modified	Maps torch.cuda.BoolTensor to torch.BoolTensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.ByteTensor	hpu_modified	Maps torch.cuda.ByteTensor to torch.ByteTensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.CharTensor	hpu_modified	Maps torch.cuda.CharTensor to torch.CharTensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.CudaError	hpu_modified	Uses RuntimeError.
torch.cuda.DoubleTensor	hpu_modified	Maps torch.cuda.DoubleTensor to torch.DoubleTensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.FloatTensor	hpu_modified	Maps torch.cuda.FloatTensor to torch.FloatTensor + torch.Tensor.to(“hpu”). torch.cuda.FloatTensor is no longer the tensor type, and can only be used to create a tensor on HPU.
torch.cuda.HalfTensor	hpu_modified	Maps torch.cuda.HalfTensor to torch.HalfTensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.IntTensor	hpu_modified	Maps torch.cuda.IntTensor to torch.IntTensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.LongTensor	hpu_modified	Maps torch.cuda.LongTensor to torch.LongTensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.ShortTensor	hpu_modified	Maps torch.cuda.ShortTensor to torch.ShortTensor + torch.Tensor.to(“hpu”). See torch.cuda.FloatTensor for details.
torch.cuda.check_error	hpu_modified	Uses RuntimeError.
torch.cuda.cudart	hpu_modified	Returns None.
torch.cuda.current_blas_handle	hpu_modified	Returns 0.
torch.cuda.current_device	hpu_modified	Uses HABANA_VISIBLE_MODULES_VAR and HLS_MODULE_ID_VAR to get the device ID.
torch.cuda.device_count	hpu_modified	Uses pyhlml to get the device count.
torch.cuda.get_arch_list	hpu_modified	Returns a constructed list of all nvcc archs.
torch.cuda.get_device_capability	hpu_modified	Returns latest cuda capability.
torch.cuda.get_device_properties	hpu_modified	Returns a constructed _CudaDeviceProperties.
torch.cuda.get_sync_debug_mode	hpu_modified	Uses PT_ENABLE_HABANA_STREAMASYNC environment variable to get get_sync_debug_mode.
torch.cuda.ipc_collect	hpu_modified	Returns None.
torch.cuda.stream	hpu_modified	Returns a torch.cuda.StreamContext.
torch.cuda.amp.GradScaler	hpu_modified	Sets enabled argument to False. GradScaler prevents gradient values “underflow”, and used for Ops with FP16 inputs. However, HPU uses BF16 in the training..
torch.cuda.amp.GradScaler.state_dict	hpu_match	Returns the state of the scaler with values in disable mode.
torch.cuda.comm.broadcast	hpu_mismatch	Raises NotImplementedError. Please use torch.distributed.broadcast instead.
torch.cuda.comm.broadcast_coalesced	hpu_mismatch	Raises NotImplementedError. Please use torch.distributed.broadcast instead.
torch.cuda.comm.gather	hpu_mismatch	Raises NotImplementedError. Please use torch.distributed.gather instead.
torch.cuda.comm.reduce_add	hpu_mismatch	Raises NotImplementedError. Please use torch.distributed.reduce instead.
torch.cuda.comm.reduce_add_coalesced	hpu_mismatch	Raises NotImplementedError. Please use torch.distributed.reduce instead.
torch.cuda.comm.scatter	hpu_mismatch	Raises NotImplementedError. Please use torch.distributed.scatter instead.
torch.cuda.graphs.CUDAGraph	hpu_match	Maps torch.cuda.CUDAGraph to habana_frameworks.torch.hpu.graphs.HPUGraph.
torch.cuda.graphs.graph	hpu_match	Maps torch.cuda.graph to habana_frameworks.torch.hpu.graphs.graph.
torch.cuda.graphs.is_current_stream_capturing	hpu_match	Maps torch.cuda.graphs.is_current_stream_capturing to torch.hpu.graphs.is_current_stream_capturing.
torch.cuda.graphs.make_graphed_callables	hpu_match	Maps torch.cuda.make_graphed_callables to habana_frameworks.torch.hpu.graphs.make_graphed_callables.
torch.cuda.graphs.graph_pool_handle	hpu_mismatch	Returns None.
torch.cuda.memory.max_memory_allocated	hpu_match	Maps torch.cuda.max_memory_allocated to torch.hpu.max_memory_allocated.
torch.cuda.memory.max_memory_cached	hpu_match	Maps torch.cuda.max_memory_cached to torch.hpu.max_memory_reserved.
torch.cuda.memory.max_memory_reserved	hpu_match	Maps torch.cuda.max_memory_reserved to torch.hpu.max_memory_reserved.
torch.cuda.memory.mem_get_info	hpu_match	Maps torch.cuda.mem_get_info to torch.hpu.mem_get_info.
torch.cuda.memory.memory_allocated	hpu_match	Maps torch.cuda.memory_allocated to torch.hpu.memory_allocated.
torch.cuda.memory.memory_cached	hpu_match	Maps torch.cuda.memory_cached to torch.hpu.memory_reserved.
torch.cuda.memory.memory_reserved	hpu_match	Maps torch.cuda.memory_reserved to torch.hpu.memory_reserved.
torch.cuda.memory.memory_stats	hpu_match	Maps torch.cuda.memory_stats to torch.hpu.memory_stats.
torch.cuda.memory.memory_summary	hpu_match	Maps torch.cuda.memory_summary to torch.hpu.memory_summary.
torch.cuda.memory.reset_max_memory_allocated	hpu_match	Maps torch.cuda.reset_max_memory_allocated to torch.hpu.reset_max_memory_allocated.
torch.cuda.memory.reset_peak_memory_stats	hpu_match	Maps torch.cuda.reset_peak_memory_stats to torch.hpu.reset_peak_memory_stats.nnDescription:Resets the peak stats tracked by the memory allocator.
torch.cuda.memory.caching_allocator_alloc	hpu_mismatch	Inactive Call.
torch.cuda.memory.caching_allocator_delete	hpu_mismatch	Inactive Call.
torch.cuda.memory.empty_cache	hpu_mismatch	Inactive Call.
torch.cuda.memory.memory_snapshot	hpu_mismatch	Inactive Call.
torch.cuda.memory.reset_max_memory_cached	hpu_mismatch	Inactive Call.
torch.cuda.memory.set_per_process_memory_fraction	hpu_mismatch	Inactive Call.
torch.cuda.memory.list_gpu_processes	hpu_modified	Prints out the running processes and their HPU memory use by calling torch.hpu.memory_reserved for the given device.
torch.cuda.nvtx.range	hpu_mismatch	Inactive Call.
torch.cuda.random.get_rng_state	hpu_match	Maps torch.cuda.random.get_rng_state to torch.hpu.random.get_rng_state.
torch.cuda.random.get_rng_state_all	hpu_match	Maps torch.cuda.random.get_rng_state_all to torch.hpu.random.get_rng_state_all.
torch.cuda.random.initial_seed	hpu_match	Maps torch.cuda.random.initial_seed to torch.hpu.random.initial_seed.
torch.cuda.random.manual_seed	hpu_match	Maps torch.cuda.random.manual_seed to torch.hpu.random.manual_seed.
torch.cuda.random.manual_seed_all	hpu_match	Maps torch.cuda.random.manual_seed_all to torch.hpu.random.manual_seed_all.
torch.cuda.random.seed	hpu_match	Maps torch.cuda.random.seed to torch.hpu.random.seed.
torch.cuda.random.seed_all	hpu_match	Maps torch.cuda.random.seed_all to torch.hpu.random.seed_all.
torch.cuda.random.set_rng_state	hpu_match	Maps torch.cuda.random.set_rng_state to torch.hpu.random.set_rng_state.
torch.cuda.random.set_rng_state_all	hpu_match	Maps torch.cuda.random.set_rng_state_all to torch.hpu.random.set_rng_state_all.
torch.cuda.streams.Stream	hpu_match	Maps torch.cuda.Stream to torch.hpu.Stream.
torch.cuda.streams.ExternalStream	hpu_modified	Maps torch.cuda.Stream.ExternalStream to torch.hpu.Stream.
torch.distributed.barrier	hpu_match	Ignores device_ids, which is valid only for NCCL backend.
torch.distributed.get_backend	hpu_match	Returns hccl as a backend.
torch.distributed.init_process_group	hpu_match	Changes backend from nccl to hccl.
torch.distributed.is_nccl_available	hpu_match	Checks import habana_frameworks.torch.distributed.hccl. If success, return True; else return False.
torch.nn.DataParallel	hpu_mismatch	Raises NotImplementedError. Please use torch.nn.DistributedDataParallel instead.
torch.nn.Embedding	hpu_modified	When using sparse gradients, the whole layer and inputs are moved to CPU.
torch.nn.EmbeddingBag	hpu_modified	When using sparse gradients, the whole layer and inputs are moved to CPU.
torch.nn.functional.scaled_dot_product_attention	hpu_modified	Uses Intel Gaudi FusedSDPA when all of math_sdp_enabled, mem_efficient_sdp_enabled and flash_sdp_enabled are True and native torch.nn.scaled_dot_product_attention otherwise. Supports only FusedSDPA no-recompute mode. For further details, refer to Using Fused Scaled Dot Product Attention (FusedSDPA).
torch.nn.parallel.DistributedDataParallel	hpu_match	Sets device_ids and output_device to None.
torch.serialization.load	hpu_match	Changes map_location to enable HPU mapping instead of CUDA mapping.
torch.utils.data.DataLoader	hpu_match	If pin_memory is True and pin_memory_device is None/CUDA, sets pin_memory_device to “hpu”.