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
↓ 4 callersFunctionadd_postprocess
(graph: Graph, node: Node, fn: Callable[..., Any],
deepspeed/compile/fx.py:84
↓ 4 callersFunctionalign_up
csrc/deepspeed4science/evoformer_attn/gemm_kernel_utils.h:127
↓ 4 callersMethodall_gather_coalesced
(self, output_tensors, input_tensors, group=None, async_op=False)
deepspeed/comm/torch.py:287
↓ 4 callersFunctionall_gather_dp_groups
(groups_flat, partitioned_param_groups, dp_process_group, start_alignment_factor, all
deepspeed/runtime/utils.py:1016
↓ 4 callersMethodallocate
(self, num_elems, count, dtype)
deepspeed/runtime/swap_tensor/utils.py:201
↓ 4 callersMethodallocate_all
(self, num_elems, dtype)
deepspeed/runtime/swap_tensor/utils.py:217
↓ 4 callersMethodallreduce_gradients
(self, bucket_size=MEMORY_OPT_ALLREDUCE_SIZE)
deepspeed/runtime/engine.py:2838
↓ 4 callersFunctionargs_from_dict
(tmpdir, config_dict)
tests/unit/simple_model.py:309
↓ 4 callersFunctionassignment_ordinals_by_expert
Return stable ordinals within each contiguous expert segment.
deepspeed/moe/ep_tp_dispatch.py:43
↓ 4 callersMethodautotuning_profile_model_info
(self)
deepspeed/runtime/engine.py:1120
↓ 4 callersMethodaverage_tensor
(self, tensor: torch.Tensor, communication_data_type: torch.dtype)
deepspeed/runtime/zero/stage_1_and_2.py:1291
↓ 4 callersMethodbuild
Build the stored specification.
deepspeed/runtime/pipe/module.py:69
↓ 4 callersFunctioncheck_and_handle_empty_buffer
( buffer_m: torch.Tensor, original_shape: torch.Size, original_size: int, worker_error: torch.
deepspeed/runtime/comm/utils.py:11
↓ 4 callersFunctioncheck_injection
(model)
tests/unit/inference/test_inference.py:249
↓ 4 callersMethodclear
Clear experiment queues, does not reset self.experiment_count
deepspeed/autotuning/scheduler.py:246
↓ 4 callersMethodclear
csrc/deepspeed4science/evoformer_attn/kernel_backward.h:1009
↓ 4 callersFunctionclear_backward_inputs
Drop captured inputs and pending capture registrations for selected frames.
deepspeed/compile/patch_compiled_func.py:124
↓ 4 callersMethodclose
(self)
deepspeed/io/fast_file_writer.py:94
↓ 4 callersFunctioncollect_deepspeed_parameters
Collect parameters from DeepSpeed engine (handles ZeRO-3 gathering)
tests/unit/v1/zero/test_zero_user_backward.py:190
↓ 4 callersFunctioncompare_loss
(model_cls, enable, zero_stage, model_dtype,
tests/unit/v1/zero/test_zero_autocast.py:67
↓ 4 callersFunctioncompare_parameters
Compare parameters between DDP and DeepSpeed
tests/unit/v1/zero/test_zero_user_backward.py:203
↓ 4 callersMethodconfig_tp_params
Configures the weight tensor for training with tensor parallelism. This includes enabling gradients and associating necessary methods
deepspeed/module_inject/layers.py:364
↓ 4 callersMethodconfigure
(self, comms_config)
deepspeed/utils/comms_logging.py:78
↓ 4 callersFunctioncreate_deepspeed_engine
(model_class, zero_stage, seed=42, gradient_accumulation_steps=1, **model_kwargs)
tests/unit/v1/zero/test_zero_user_backward.py:117
↓ 4 callersFunctioncreate_file
(filename, num_bytes)
deepspeed/nvme/test_ds_aio_utils.py:78
↓ 4 callersFunctioncreate_file
(filename, num_bytes)
csrc/aio/py_test/test_ds_aio_utils.py:78
↓ 4 callersFunctioncreate_filename
(folder, read_op, size, tid)
deepspeed/nvme/test_ds_aio_utils.py:73
↓ 4 callersFunctioncreate_filename
(folder, read_op, size, tid)
csrc/aio/py_test/test_ds_aio_utils.py:73
↓ 4 callersMethodcreate_op_builder
(self, class_name)
accelerator/hpu_accelerator.py:286
↓ 4 callersFunctioncreate_page_locked_tensor
(num_elem, use_accelerator, aio_handle=None)
deepspeed/nvme/test_ds_aio_utils.py:86
↓ 4 callersFunctioncreate_page_locked_tensor
(num_elem, use_accelerator, aio_handle=None)
csrc/aio/py_test/test_ds_aio_utils.py:86
↓ 4 callersFunctioncuFileGetErrorString
csrc/gds/py_lib/deepspeed_gds_utils.h:76
↓ 4 callersMethodcurriculum_enabled_legacy
(self)
deepspeed/runtime/engine.py:1001
↓ 4 callersMethodcurriculum_learning_enabled
(self)
deepspeed/runtime/engine.py:1019
↓ 4 callersFunctiondense_to_sparse
Converts dense matrix with explicit zeros to sparse matrix
tests/unit/ops/sparse_attention/test_sparse_attention.py:22
↓ 4 callersMethoddequantize
(self, tensor: Tensor, quant_scale: Tensor, quant_min: Tensor)
deepspeed/inference/quantization/utils.py:105
↓ 4 callersFunctiondp_index_to_str
(dp_index)
deepspeed/checkpoint/ds_to_universal.py:207
↓ 4 callersFunctionduration_to_string
(duration, units=None, precision=DEFAULT_PRECISION)
deepspeed/profiling/flops_profiler/profiler.py:1174
↓ 4 callersMethoddynamic_loss_scale
(self)
deepspeed/runtime/engine.py:1412
↓ 4 callersMethodeval_batch
Evaluate the pipeline on a batch of data from ``data_iter``. The engine will evaluate ``self.train_batch_size()`` total samples collec
deepspeed/runtime/pipe/engine.py:431
↓ 4 callersFunctionextract_tensors
Separate objects in list/tuple into tensors and non-tensors and create a mapping to enable re-aggregation. The order of tensors and non-tenso
deepspeed/runtime/activation_checkpointing/checkpointing.py:305
↓ 4 callersFunctionf
(size)
deepspeed/compile/profilers/comm_profile.py:147
↓ 4 callersMethodfind_matching_spec
Find the first matching spec for a parameter.
deepspeed/module_inject/autotp_config.py:266
↓ 4 callersFunctionfind_node_by_name
(gm: GraphModule, name: str)
deepspeed/compile/fx.py:167
↓ 4 callersMethodflatten_dense_tensors_aligned
(self, tensor_list, alignment, use_cpu_data=False)
deepspeed/runtime/zero/stage_1_and_2.py:1134
↓ 4 callersMethodflops_profiler_profile_step
(self)
deepspeed/runtime/engine.py:1073
↓ 4 callersMethodforward
(self, x, y)
tests/unit/v1/zero/test_zero.py:259
↓ 4 callersMethodfree
Return a list of blocks to the free pool. If a single invalid block is provided (i.e., one that is out of range of the allocator or i
deepspeed/inference/v2/ragged/blocked_allocator.py:74
↓ 4 callersMethodfrom_dict
Constructs a `BertConfig` from a Python dictionary of parameters.
tests/unit/modeling.py:238
↓ 4 callersMethodfrom_meta
(cls, meta, local_part, group, device=get_accelerator().device_name())
deepspeed/runtime/utils.py:682
↓ 4 callersMethodgather_params
(self, params_list)
deepspeed/module_inject/layers.py:755
↓ 4 callersMethodgen_output_name
(self, test_config, prefix, baseline_config=False)
tests/model/Megatron_GPT2/test_common.py:21
↓ 4 callersMethodget
deepspeed/inference/v2/kernels/includes/memory_access_utils.h:74
↓ 4 callersMethodgetSendBuf
csrc/includes/deepcompile.h:127
↓ 4 callersFunctionget_all_subclasses
(cls, include_root=True)
deepspeed/runtime/zero/partition_parameters.py:285
↓ 4 callersFunctionget_allgather_dtype
(param, param_ds_tensor)
deepspeed/runtime/zero/partition_parameters.py:55
↓ 4 callersMethodget_axis_names
Return a list of the axis names in the ordering of the topology.
deepspeed/runtime/pipe/topology.py:65
↓ 4 callersFunctionget_ccl_datatype
csrc/cpu/comm/ccl.cpp:153
↓ 4 callersFunctionget_compression_config
(param_dict)
deepspeed/compression/config.py:11
↓ 4 callersFunctionget_cuda_rng_tracker
Get cuda rng tracker.
deepspeed/runtime/activation_checkpointing/checkpointing.py:196
↓ 4 callersMethodget_data_parallel_partitions
(self, tensor, group_id)
deepspeed/runtime/zero/stage_1_and_2.py:1904
↓ 4 callersMethodget_data_types
(self)
deepspeed/runtime/engine.py:1430
↓ 4 callersFunctionget_deepspeed_model
(model)
tests/unit/model_parallelism/test_configurable_parallel_mp.py:23
↓ 4 callersFunctionget_files
(dir)
deepspeed/checkpoint/reshape_utils.py:34
↓ 4 callersFunctionget_fp32_state_dict_from_zero_checkpoint
Convert ZeRO 2 or 3 checkpoint into a single fp32 consolidated state_dict that can be loaded with ``load_state_dict()`` and used for training
deepspeed/utils/zero_to_fp32.py:563
↓ 4 callersFunctionget_global_norm
Compute total from a list of norms
deepspeed/runtime/utils.py:349
↓ 4 callersMethodget_hp_fragment
(self, optim_state_key=None)
deepspeed/utils/tensor_fragment.py:55
↓ 4 callersFunctionget_last_uses
(graph: Graph)
deepspeed/compile/util.py:253
↓ 4 callersMethodget_mask
(self, pruning_type='row')
deepspeed/compression/basic_layer.py:296
↓ 4 callersMethodget_module
(self, sd)
deepspeed/runtime/state_dict_factory.py:149
↓ 4 callersFunctionget_nccl_data_type
csrc/compile/deepcompile.cpp:63
↓ 4 callersFunctionget_node_shape_meta
(node: Node)
deepspeed/compile/fx.py:174
↓ 4 callersMethodget_overlap_step_state
(self)
deepspeed/runtime/zenflow/zenflow_stage_1_and_2.py:639
↓ 4 callersMethodget_pipe_parallel_group
The group of ranks within the same pipeline.
deepspeed/runtime/pipe/topology.py:424
↓ 4 callersMethodget_sd_loader
(ckpt_list, checkpoint_engine, sd_type='Megatron', version=None)
deepspeed/runtime/state_dict_factory.py:41
↓ 4 callersMethodget_states
Get rng states. Copy the dictionary so we have direct pointers to the states, not just a pointer to the dictionary.
deepspeed/runtime/activation_checkpointing/checkpointing.py:144
↓ 4 callersFunctionget_test_path
(filename)
tests/unit/common.py:506
↓ 4 callersFunctionget_torch_optimizer
(optimizer)
deepspeed/runtime/lr_schedules.py:268
↓ 4 callersMethodgradient_predivide_factor
(self)
deepspeed/runtime/engine.py:1370
↓ 4 callersFunctionhas_comm_dtype
(param: torch.nn.Parameter)
deepspeed/runtime/torch_autocast.py:77
↓ 4 callersMethodhas_gradients
(self)
deepspeed/runtime/swap_tensor/optimizer_utils.py:63
↓ 4 callersMethodhas_space
(self, numel)
deepspeed/runtime/swap_tensor/utils.py:71
↓ 4 callersMethodhas_value
(self, name: str)
deepspeed/compile/util.py:221
↓ 4 callersMethodinclude_paths
Returns list of include paths, relative to root of deepspeed package (i.e., DeepSpeed/deepspeed)
op_builder/builder.py:289
↓ 4 callersMethodinput_ids
The input ids tensor for the ragged batch. If the device Tensor is requested, the Tensor is truncated to the number of tokens in the
deepspeed/inference/v2/ragged/ragged_wrapper.py:220
↓ 4 callersMethodinsert_tensor
(self, tensor, swap_path, aligned_numel)
deepspeed/runtime/swap_tensor/utils.py:50
↓ 4 callersMethodirecv
(self, tensor, src=None, group=None, tag=0)
deepspeed/comm/torch.py:360
↓ 4 callersFunctionis_activation_to_checkpoint
Is an activation to be checkpointed
deepspeed/runtime/activation_checkpointing/checkpointing.py:367
↓ 4 callersMethodis_decoupled
(self)
deepspeed/runtime/checkpoint_engine/checkpoint_engine.py:50
↓ 4 callersFunctionis_deepcompile_supported
()
deepspeed/compile/util.py:31
↓ 4 callersFunctionis_moe_param_group
(param_group)
deepspeed/moe/utils.py:151
↓ 4 callersMethodis_swappable_tensor
(self, tensor=None, numel=None)
deepspeed/runtime/swap_tensor/optimizer_utils.py:211
↓ 4 callersMethodis_zenflow_select_boundary
(self)
deepspeed/runtime/zenflow/zenflow_stage_1_and_2.py:136
↓ 4 callersMethodisend
(self, tensor, dst, group=None, tag=0)
deepspeed/comm/torch.py:356
↓ 4 callersMethodload_hp_checkpoint_state
(self, folder, key, param=None)
deepspeed/runtime/zero/stage3.py:3445
↓ 4 callersMethodload_universal_checkpoint
(self)
deepspeed/runtime/engine.py:1343
↓ 4 callersFunctionlocal_folding_ranks
(global_rank: int, spec: ParallelFoldingSpec)
deepspeed/module_inject/auto_ep_folding.py:161
↓ 4 callersMethodlog_level
(self)
deepspeed/runtime/engine.py:1346
↓ 4 callersFunctionmake_cpp_batch_wrapper
deepspeed/inference/v2/kernels/ragged_ops/ragged_helpers/ragged_kernel_helpers.cpp:8
↓ 4 callersFunctionmark_autoep_folding_router_parameter
Tag a router/gate parameter as the *replicated* folded family (AVERAGE). This is the ONLY family marker applied on the live forward path today:
deepspeed/module_inject/auto_ep_folding.py:265
← previousnext →701–800 of 11,258, ranked by callers