MCPcopy Create free account

hub / github.com/deepspeedai/DeepSpeed / functions

Functions11,258 in github.com/deepspeedai/DeepSpeed

⨍ Functions 11,258◇ Types & classes 1,887↳ Endpoints 34
↓ 3 callersMethodget_prefix
(self)
op_builder/ragged_ops.py:60
↓ 3 callersFunctionget_rank
Returns the rank of the current process in the provided ``group`` or the default group if none was provided. Rank is a unique identifier
deepspeed/comm/comm.py:709
↓ 3 callersMethodget_result
Get result from worker process with optional event type validation. Args: expected_event_type (str, optional): Expected
deepspeed/runtime/superoffload/superoffload_utils.py:244
↓ 3 callersFunctionget_runnable_nodes
(scheduled: List[Node], unscheduled: List[Node])
deepspeed/compile/list_schedule.py:81
↓ 3 callersFunctionget_sdpa_nodes
(gm: GraphModule)
deepspeed/compile/util.py:553
↓ 3 callersMethodget_sequence_parallel_group
(self)
deepspeed/runtime/engine.py:1061
↓ 3 callersFunctionget_sized_buffers
(buffer_list, num_elems_list)
deepspeed/runtime/swap_tensor/utils.py:253
↓ 3 callersMethodget_slice_parallel_rank
(self)
deepspeed/runtime/pipe/topology.py:456
↓ 3 callersMethodget_swap_tensors
(self)
deepspeed/runtime/swap_tensor/utils.py:74
↓ 3 callersMethodget_theta
(self)
deepspeed/runtime/progressive_layer_drop.py:32
↓ 3 callersFunctionget_topology
(mp, pp, world_size)
tests/unit/model_parallelism/test_configurable_parallel_pp.py:39
↓ 3 callersMethodget_total_flops
Returns the total flops of the model. Args: as_string (bool, optional): whether to output the flops as string. Defaults to False.
deepspeed/profiling/flops_profiler/profiler.py:232
↓ 3 callersFunctionget_weight_norm
Get norm of an iterable of parameters. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model
deepspeed/runtime/utils.py:537
↓ 3 callersFunctiongraph_process
(replay_first_step, func, *args, **kwargs)
deepspeed/runtime/utils.py:92
↓ 3 callersMethodhasGatheredParam
csrc/includes/deepcompile.h:439
↓ 3 callersFunctionhas_inf_or_nan
(x)
tests/unit/runtime/half_precision/test_zero_optim_overflow.py:16
↓ 3 callersFunctionhas_moe_layers
(m: nn.Module)
deepspeed/moe/utils.py:15
↓ 3 callersMethodhas_reduce_scatter_tensor
(self)
deepspeed/comm/torch.py:162
↓ 3 callersFunctionhook
(module, input)
deepspeed/compression/helper.py:290
↓ 3 callersFunctioninit_distributed
Initialize dist backend, potentially performing MPI discovery if needed Arguments: dist_backend: Optional (str). torch distributed backe
deepspeed/comm/comm.py:792
↓ 3 callersMethodinit_process_group
(self)
deepspeed/comm/backend.py:44
↓ 3 callersMethodinit_timers
(self)
deepspeed/runtime/swap_tensor/optimizer_utils.py:217
↓ 3 callersMethodisPersistent
csrc/includes/deepcompile.h:303
↓ 3 callersFunctionis_autocast_initialized
()
deepspeed/runtime/torch_autocast.py:65
↓ 3 callersMethodis_available
(self)
accelerator/cuda_accelerator.py:232
↓ 3 callersFunctionis_compile_supported
()
deepspeed/runtime/compiler.py:32
↓ 3 callersMethodis_data_parallel_writer
(self, dp_rank)
deepspeed/runtime/checkpoint_engine/checkpoint_engine.py:47
↓ 3 callersMethodis_deepcompile_enabled
(self)
deepspeed/runtime/engine.py:5698
↓ 3 callersMethodis_first_stage
True if the configured ``stage_id`` is the first stage in the pipeline.
deepspeed/runtime/pipe/schedule.py:102
↓ 3 callersFunctionis_functorch_transforming
True when called under torch.func.grad / vmap / jacrev / etc.
deepspeed/utils/torch.py:30
↓ 3 callersMethodis_last_stage
True if the configured ``stage_id`` is the last stage in the pipeline.
deepspeed/runtime/pipe/schedule.py:107
↓ 3 callersMethodis_moe_group
(self, group)
deepspeed/runtime/zero/stage_1_and_2.py:744
↓ 3 callersFunctionis_profile_incomplete
(graph: Graph)
deepspeed/compile/profilers/graph_profile.py:74
↓ 3 callersFunctionis_rocm_pytorch
()
tests/unit/common.py:35
↓ 3 callersMethodis_synchronized_device
(self)
accelerator/supa_accelerator.py:21
↓ 3 callersMethodis_synchronized_device
(self)
accelerator/sdaa_accelerator.py:53
↓ 3 callersMethodis_synchronized_device
(self)
accelerator/cpu_accelerator.py:42
↓ 3 callersMethodis_synchronized_device
(self)
accelerator/mlu_accelerator.py:23
↓ 3 callersMethodis_synchronized_device
(self)
accelerator/npu_accelerator.py:29
↓ 3 callersMethodis_synchronized_device
(self)
accelerator/xpu_accelerator.py:32
↓ 3 callersMethodis_synchronized_device
(self)
accelerator/cuda_accelerator.py:52
↓ 3 callersMethodjit_load
(self, verbose=True)
op_builder/builder.py:545
↓ 3 callersFunctionlayer_norm
(a, weight, bias, eps)
deepspeed/ops/transformer/inference/triton/layer_norm.py:167
↓ 3 callersFunctionlayer_norm_residual
(a, input_bias, residual, weight, bias, eps)
deepspeed/ops/transformer/inference/triton/layer_norm.py:199
↓ 3 callersMethodlazy_initialization
(self, key_states: torch.Tensor, value_states: torch.Tensor)
deepspeed/utils/static_cache.py:54
↓ 3 callersFunctionload
(module, prefix)
deepspeed/module_inject/load_checkpoint.py:46
↓ 3 callersMethodload
csrc/deepspeed4science/evoformer_attn/gemm/mma_from_smem.h:276
↓ 3 callersFunctionload_module_recursive
(module, prefix='', level=0)
deepspeed/module_inject/load_checkpoint.py:222
↓ 3 callersMethodload_moe_state_dict
(checkpoint_path, tag, state_dict,
deepspeed/runtime/engine.py:3938
↓ 3 callersMethodload_with_byte_offset
Loads a fragment from memory
csrc/deepspeed4science/evoformer_attn/iterators/predicated_tile_iterator_residual_last.h:819
↓ 3 callersMethodloss_scale
(self)
deepspeed/runtime/engine.py:1331
↓ 3 callersMethodmake_attn_layer
Builds the attention layer for the model. This sets the `self.attn` attribute.
deepspeed/inference/v2/model_implementations/inference_transformer_base.py:317
↓ 3 callersMethodmake_attn_out_layer
Instantiates the linear projection layer for the attention output linear layer. This sets the `self.attn_out` attribute.
deepspeed/inference/v2/model_implementations/inference_transformer_base.py:399
↓ 3 callersFunctionmake_graph_from_schedule
(scheduled: List[Node])
deepspeed/compile/list_schedule.py:24
↓ 3 callersFunctionmake_metadata_filename
Make a filename for a metadata file. Arguments: rank: Rank of the file. n_ranks: Total number of ranks. Returns:
deepspeed/inference/v2/model_implementations/flat_model_helpers.py:73
↓ 3 callersMethodmake_qkv_layer
Instantiates the linear projection layer for the QKV linear layer. This sets the `self.qkv` attribute.
deepspeed/inference/v2/model_implementations/inference_transformer_base.py:286
↓ 3 callersMethodmake_unembedding_layer
Performs setup and creates an unembedding layer. This implementation assumes normalization prior to the LM head projection. If this d
deepspeed/inference/v2/model_implementations/inference_transformer_base.py:248
↓ 3 callersFunctionmask_nan_or_inf_with_val_inplace
(input, device=None, val=-1.)
deepspeed/runtime/utils.py:870
↓ 3 callersMethodmax_memory_cached
(self, device_index=None)
accelerator/hpu_accelerator.py:135
↓ 3 callersFunctionmaybe_copy_geglu
(module, sd, weight_quantizer, mp_replace, dst_name, src_names)
deepspeed/module_inject/policy.py:189
↓ 3 callersMethodmemory_allocated
(self)
tests/unit/v1/compile/test_graph_profile.py:33
↓ 3 callersMethodmemory_cached
(self, device_index=None)
accelerator/hpu_accelerator.py:132
↓ 3 callersMethodmemory_stats
(self, device_index=None)
accelerator/hpu_accelerator.py:141
↓ 3 callersMethodmodule_state_dict
(self, destination=None, prefix="", keep_vars=False, exclude_frozen_parameters=False)
deepspeed/runtime/engine.py:3818
↓ 3 callersFunctionmoe_test_helper
Helper function for validating the GEMM kernel for a single expert.
tests/unit/inference/v2/kernels/cutlass_ops/test_moe_gemm.py:48
↓ 3 callersMethodnum_tensors
(self)
deepspeed/runtime/swap_tensor/optimizer_utils.py:74
↓ 3 callersFunctionoffload_adam_states_sync
()
deepspeed/compile/passes/offload_adam_states.py:144
↓ 3 callersMethodoperand_B_ref
Returns a TensorRef to the B operand
csrc/deepspeed4science/evoformer_attn/gemm/mma_from_smem.h:214
↓ 3 callersMethodoverlapping_partition_gradients_reduce_epilogue
(self)
deepspeed/runtime/zero/stage3.py:1358
↓ 3 callersFunctionpatch_compiled_func
()
deepspeed/compile/patch_compiled_func.py:74
↓ 3 callersMethodpread
csrc/aio/py_lib/deepspeed_py_io_handle.cpp:279
↓ 3 callersFunctionprepare_tp_model
(hidden_dim, nlayers, linear_indices, allreduce_indices, group, return_global_copy=False)
tests/unit/model_parallelism/test_autotp_training.py:656
↓ 3 callersFunctionpreset_name_for_hf_model_type
Return the AutoEP preset name for a supported HF model_type.
deepspeed/module_inject/auto_ep_presets/registry.py:84
↓ 3 callersMethodprint_allocation
(self, resolution=200)
deepspeed/runtime/zero/contiguous_memory_allocator.py:121
↓ 3 callersMethodprint_model_profile
Prints the model graph with the measured profile attached to each module. Args: profile_step (int, optional): The global training
deepspeed/profiling/flops_profiler/profiler.py:286
↓ 3 callersFunctionpytest_hack
This is an ugly hack to get the unittest suites to play nicely with pytest. Otherwise failed tests are not reported by pytest for some reason.
tests/model/run_sanity_check.py:20
↓ 3 callersFunctionqtorch_quantize
(input, exp_bits=4, man_bits=3, rounding="nearest", group_size=1024)
tests/unit/ops/fp_quantizer/test_fp_quant.py:23
↓ 3 callersMethodquantize
(self, tensor: Tensor)
deepspeed/inference/quantization/utils.py:51
↓ 3 callersMethodrandom
(self)
tests/unit/v1/compile/test_graph_profile.py:50
↓ 3 callersFunctionrandom_dataset
(total_samples, hidden_dim, device, dtype=preferred_dtype())
tests/unit/simple_model.py:268
↓ 3 callersMethodrange_pop
(self)
tests/unit/utils/test_nvtx.py:122
↓ 3 callersMethodrange_push
(self, msg)
tests/unit/utils/test_nvtx.py:118
↓ 3 callersMethodreenter_backward_if_needed
Re-enter backward context for subsequent phases in reentrant checkpointing.
deepspeed/runtime/base_optimizer.py:455
↓ 3 callersMethodregisterGatheredParam
csrc/includes/deepcompile.h:418
↓ 3 callersFunctionregister_grad_hook
(param, hook)
deepspeed/utils/torch.py:40
↓ 3 callersMethodrelease_memory
(self)
deepspeed/runtime/swap_tensor/optimizer_utils.py:35
↓ 3 callersMethodrelease_swap_buffers
(self, parameter)
deepspeed/runtime/swap_tensor/partitioned_optimizer_swapper.py:63
↓ 3 callersMethodrelease_workspace
(self)
deepspeed/ops/transformer/inference/op_binding/workspace.py:184
↓ 3 callersFunctionremove_random_ltd_state_dict
(state_dict)
deepspeed/runtime/data_pipeline/data_routing/helper.py:38
↓ 3 callersFunctionreplace_dict
Replaces values in dict d with values in dict u. Args: d (dict): the target dict to overwrite u (dict): the dict containing the v
deepspeed/autotuning/utils.py:110
↓ 3 callersFunctionreplace_reduce_outputs_with_none
(graph: Graph)
deepspeed/compile/fx.py:38
↓ 3 callersFunctionreplace_transformer_layer
Replace bert-style transformer layers with DeepSpeed's transformer layer Arguments: orig_layer_impl (torch.nn.Module): the original trans
deepspeed/module_inject/replace_module.py:189
↓ 3 callersMethodreset
csrc/includes/deepcompile.h:170
↓ 3 callersMethodreset_rss
(self)
accelerator/cpu_accelerator.py:147
↓ 3 callersMethodrestore_slots
(self, slots: list)
deepspeed/autotuning/scheduler.py:270
↓ 3 callersFunctionrun_forward
(ds_config, seq_len, atol=1e-2, verbose=False, test_bsz=None)
tests/unit/ops/accelerators/test_accelerator_forward.py:165
↓ 3 callersFunctionsafe_get_local_fp32_param
Get the local partition of a ZeRO-3 partitioned parameter in fp32 precision. Args: param (``torch.nn.Parameter``): A model parame
deepspeed/utils/tensor_fragment.py:284
↓ 3 callersFunctionsafe_get_local_grad
Get the local gradient partition of a ZeRO-3 partitioned parameter. The return data type is that used for gradient accumulation. This
deepspeed/utils/tensor_fragment.py:254
↓ 3 callersFunctionsafe_get_local_optimizer_state
Get the local optimizer state partition of ZeRO-3 partitioned parameter in fp32 precision. Args: param (``torch.nn.Parameter``):
deepspeed/utils/tensor_fragment.py:297
↓ 3 callersMethodscale_loss
Scales the loss by the current loss scale. We need this function to scale loss without calling backward on it.
deepspeed/runtime/fp16/loss_scaler.py:151
← previousnext →1,101–1,200 of 11,258, ranked by callers