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 callersMethodallocate_tensor
(self, swap_path, numel, aligned_numel)
deepspeed/runtime/swap_tensor/utils.py:55
↓ 3 callersMethodallreduce_and_copy_with_multiple_ranks
(self, small_bucket,
deepspeed/runtime/zero/stage_1_and_2.py:1231
↓ 3 callersFunctionapply_autotp_with_partition_config
(model, tp_size, partition_config)
tests/unit/model_parallelism/test_autotp_custom_patterns.py:156
↓ 3 callersMethodapply_rotary_pos_emb
(mixed_query, key_layer, rotary_dim, offset, num_heads, rotate_half, rope_theta)
op_builder/supa/inference.py:299
↓ 3 callersFunctionassert_ints_same_as_other_ranks
NOTE: creates both communication and synchronization overhead so should be used sparingly takes a list of ints from each rank and ensure
deepspeed/runtime/zero/utils.py:106
↓ 3 callersFunctionassert_load_balance_coeff_rejection_message
(exc: BaseException, value: object)
tests/unit/v1/moe/autoep_test_utils.py:173
↓ 3 callersMethodattention_o_mp
(self, mp_replace, reversed_dim=False)
deepspeed/module_inject/containers/base.py:249
↓ 3 callersMethodautotuning_enabled
(self)
deepspeed/runtime/engine.py:1096
↓ 3 callersFunctionbaseline_implementation
(hidden_states: torch.Tensor, seq_lens: List[int])
tests/unit/inference/v2/kernels/ragged_ops/test_logits_gather.py:17
↓ 3 callersMethodbf16_optimizer_states
(self)
deepspeed/runtime/engine.py:1309
↓ 3 callersMethodbuild_atoms
Build the atoms for this module. This is not a strict requirement for the class, so this method is a no-op by default rather than abs
deepspeed/inference/v2/modules/interfaces/attention_base.py:51
↓ 3 callersFunctioncan_send_recv
()
deepspeed/runtime/pipe/p2p.py:21
↓ 3 callersFunctioncheck_deepspeed_config
(config)
deepspeed/compression/compress.py:20
↓ 3 callersMethodcheck_using_norm
(self, norm_group, reduce_overflow=True)
deepspeed/runtime/utils.py:241
↓ 3 callersMethodcheckpoint_event_epilogue
(self)
deepspeed/runtime/zero/stage3.py:3521
↓ 3 callersMethodcheckpoint_event_prologue
(self)
deepspeed/runtime/zero/stage3.py:3518
↓ 3 callersFunctioncleanup_compiled_backward_state
Release engine-owned process-global compiled-backward state.
deepspeed/compile/backend.py:81
↓ 3 callersMethodclear_grad_attribute
(self, param)
deepspeed/runtime/zero/stage_1_and_2.py:1092
↓ 3 callersMethodclear_lp_grads
(self, set_to_none=False)
deepspeed/runtime/bf16_optimizer.py:473
↓ 3 callersMethodclear_selected_mv
(self)
deepspeed/ops/adam/zenflow_torch_adam.py:83
↓ 3 callersFunctionclone_tensors_for_torch_save
Returns a copy of ``item`` with all enclosed tensors replaced by clones on a specified device. Works on individual tensors, and tensors conta
deepspeed/checkpoint/utils.py:41
↓ 3 callersFunctioncombine_from_routed
Scatter-add expert outputs back to original token positions.
deepspeed/module_inject/auto_ep_layer.py:310
↓ 3 callersMethodcommit
(self, info: CheckpointCommitInfo)
deepspeed/runtime/checkpoint_engine/checkpoint_engine.py:43
↓ 3 callersMethodcompile_autosp
Determines if AutoSP is set in deepcompile's passes attributes.
deepspeed/runtime/engine.py:1214
↓ 3 callersMethodcomplete_backward
Release each fallback gather after its last reached forward graph completes.
deepspeed/compile/z3_eager_fallback.py:215
↓ 3 callersFunctionconcat_to_compat_param
(quantized_weight: Tensor, quant_scale: Tensor, quant_mi
deepspeed/inference/quantization/utils.py:186
↓ 3 callersMethodconfig_requires_grad
(self, weight)
deepspeed/module_inject/layers.py:356
↓ 3 callersFunctionconvert_size
(size_bytes)
deepspeed/utils/comms_logging.py:22
↓ 3 callersMethodconvert_to_zero_parameters
(param_list)
deepspeed/runtime/zero/partition_parameters.py:1615
↓ 3 callersFunctioncount_tokens_per_expert
Count routed tokens per expert. Because the output shape is known up front, it avoids the device-to-host synchronization that ``torch.bincoun
deepspeed/moe/ep_count.py:10
↓ 3 callersFunctioncount_used_parameters_in_backward
Count the number of parameters that participate in the currently running backward graph. This helper is designed to be invoked from within a
deepspeed/runtime/utils.py:1444
↓ 3 callersMethodcurrent_device
(self)
tests/unit/v1/compile/test_graph_profile.py:30
↓ 3 callersMethodcxx_args
Returns optional list of compiler flags to forward to the build
op_builder/builder.py:301
↓ 3 callersFunctiondebug_module2name
(module)
deepspeed/utils/debug.py:32
↓ 3 callersFunctiondebug_param2name_id_shape
(param)
deepspeed/utils/debug.py:72
↓ 3 callersMethoddequantize
(self, input_q, fp_out=None, q_bits=8, q_mantisa_bits=3, scale=None)
deepspeed/ops/fp_quantizer/quantize.py:39
↓ 3 callersMethoddequantize
(self, input_q, fp_out=None, q_bits=8, q_mantisa_bits=3, scale=None)
deepspeed/ops/fp_quantizer/quantize.py:110
↓ 3 callersMethoddequantized
Return a tensor containing the dequantized weights of this parameter.
deepspeed/linear/quantization.py:67
↓ 3 callersFunctiondimension_problem
csrc/spatial/csrc/pt_binding.cpp:12
↓ 3 callersMethoddisable
(self)
deepspeed/linear/optimized_linear.py:120
↓ 3 callersFunctiondispatch_counters
(ctx: RestoreContext)
deepspeed/moe/ep_tp_dispatch.py:451
↓ 3 callersFunctionds_shape
(param)
deepspeed/utils/debug.py:61
↓ 3 callersFunctiondummy_init_engine
(config)
tests/unit/model_parallelism/test_autotp_training.py:640
↓ 3 callersMethoddynamic_loss_scale_args
(self)
deepspeed/runtime/engine.py:1418
↓ 3 callersMethodelapsed
(self)
tests/unit/modeling.py:110
↓ 3 callersMethodelasticity_enabled
(self)
deepspeed/runtime/engine.py:954
↓ 3 callersFunctionenable
Decorator factory to enable compiling of a function if the minimum PyTorch version requirement is met. Args: min_version (str, optio
deepspeed/runtime/compiler.py:42
↓ 3 callersMethodend_profile
Ends profiling. The added attributes and handles are removed recursively on all the modules.
deepspeed/profiling/flops_profiler/profiler.py:203
↓ 3 callersMethodeval
r
deepspeed/runtime/engine.py:2625
↓ 3 callersMethodextra_ldflags
(self)
op_builder/builder.py:313
↓ 3 callersFunctionfill_autoep_config_from_hf
Back-fill AutoEPConfig fields from HF model config when user hasn't set them. HF field names (e.g. n_group, topk_group, routed_scaling_factor) di
deepspeed/module_inject/auto_ep_config.py:331
↓ 3 callersFunctionfill_with_head_ids
Fills a tensor with the associated head ids. All columns should have the same value.
tests/unit/inference/v2/model_implementations/sharding/test_attn_out_sharding.py:16
↓ 3 callersFunctionfilter_args
(node: Node)
deepspeed/compile/list_schedule.py:48
↓ 3 callersFunctionfind_fit_int_dtype
(min_value, max_value)
deepspeed/runtime/data_pipeline/data_sampling/utils.py:12
↓ 3 callersFunctionfind_node_by_tag
(gm: GraphModule, tag: str)
deepspeed/compile/fx.py:178
↓ 3 callersMethodfork_rng
(self, devices)
tests/unit/v1/compile/test_graph_profile.py:15
↓ 3 callersMethodforward
(self, input)
tests/unit/runtime/zero/test_zero_context_return.py:88
↓ 3 callersMethodfp16_enabled
(self)
deepspeed/autotuning/autotuner.py:265
↓ 3 callersMethodfree
csrc/aio/py_lib/deepspeed_pin_tensor.cpp:51
↓ 3 callersMethodgK_strideM
csrc/deepspeed4science/evoformer_attn/kernel_backward.h:270
↓ 3 callersMethodgV_strideM
csrc/deepspeed4science/evoformer_attn/kernel_backward.h:274
↓ 3 callersMethodgen_workspace
(self, num_layers, num_heads, batch_size, prompt_len, hidden_dim, mp_size, external_cache,
deepspeed/ops/transformer/inference/op_binding/workspace.py:37
↓ 3 callersMethodgetGradBuffer
csrc/includes/deepcompile.h:294
↓ 3 callersFunctionget_annotations
Retrieves annotations from a Python object. In python >=3.14 this is a thin wrapper around the `annotationlib.get_annotations` function
deepspeed/compat.py:31
↓ 3 callersFunctionget_attn_mask_stride
infer stride of attention mask memory layout based on the model type.
csrc/transformer/inference/csrc/pt_binding.cpp:41
↓ 3 callersMethodget_axis_comm_lists
Construct lists suitable for a communicator group along axis ``axis``. Example: >>> topo = Topo(axes=['pipe', 'data', 'model'],
deepspeed/runtime/pipe/topology.py:127
↓ 3 callersMethodget_axis_list
Returns the list of global ranks whose coordinate in an axis is idx. For example: >>> X = ProcessTopology(axes=['x', 'y'], dims=[
deepspeed/runtime/pipe/topology.py:193
↓ 3 callersMethodget_buffer
(self)
deepspeed/io/base_io_buffer.py:28
↓ 3 callersFunctionget_ccl_reduce_op
csrc/cpu/comm/ccl.cpp:168
↓ 3 callersMethodget_checkpoint_info
(self, info_key=CHECKPOINT_INFO_KEY)
deepspeed/checkpoint/deepspeed_checkpoint.py:169
↓ 3 callersMethodget_cmd
(self, environment, active_resources)
deepspeed/launcher/multinode_runner.py:150
↓ 3 callersMethodget_compile_backend
(self)
accelerator/hpu_accelerator.py:319
↓ 3 callersMethodget_data_parallel_rank
Which pipeline this rank resides in.
deepspeed/runtime/pipe/topology.py:428
↓ 3 callersFunctionget_deepspeed_model
(model)
tests/unit/model_parallelism/test_configurable_parallel_pp.py:24
↓ 3 callersMethodget_difficulty
(self, global_steps)
deepspeed/runtime/data_pipeline/curriculum_scheduler.py:143
↓ 3 callersFunctionget_file_size
csrc/aio/common/deepspeed_aio_utils.cpp:104
↓ 3 callersMethodget_gas_from_user_config
(self)
deepspeed/autotuning/autotuner.py:983
↓ 3 callersFunctionget_global_norm_of_tensors
Get norm of an iterable of tensors. This is adapted from torch.nn.utils.clip_grad.clip_grad_norm_ and added functionality to handle model par
deepspeed/runtime/utils.py:878
↓ 3 callersFunctionget_gradient_clipping
(param_dict)
deepspeed/runtime/config.py:246
↓ 3 callersMethodget_hp_fragment_address
(self)
deepspeed/utils/tensor_fragment.py:49
↓ 3 callersFunctionget_inactive_params
(param_list)
deepspeed/runtime/utils.py:1104
↓ 3 callersFunctionget_index_by_graph_id
(graph_order, target_graph_id)
deepspeed/compile/util.py:481
↓ 3 callersMethodget_kv_requirements
(self, sequence: DSSequenceDescriptor, max_new_tokens: int, max_new_blocks: int)
deepspeed/inference/v2/model_implementations/exaone4_5/model.py:158
↓ 3 callersFunctionget_label_id_node
(gm: GraphModule)
deepspeed/compile/util.py:568
↓ 3 callersFunctionget_list
(val)
deepspeed/autotuning/utils.py:74
↓ 3 callersMethodget_locks
(size, dev)
deepspeed/ops/sparse_attention/matmul.py:259
↓ 3 callersFunctionget_log_level
(param_dict)
deepspeed/runtime/config.py:656
↓ 3 callersMethodget_lr
(self)
deepspeed/runtime/lr_schedules.py:719
↓ 3 callersMethodget_lr
(self)
deepspeed/runtime/lr_schedules.py:906
↓ 3 callersFunctionget_mapping_to_flat_buffer
(tensors: List[torch.Tensor])
deepspeed/runtime/zero/utils.py:194
↓ 3 callersMethodget_mask
(self, pruning_type='sparse')
deepspeed/compression/basic_layer.py:528
↓ 3 callersFunctionget_module_duration
(module)
deepspeed/profiling/flops_profiler/profiler.py:1196
↓ 3 callersFunctionget_module_macs
(module)
deepspeed/profiling/flops_profiler/profiler.py:1188
↓ 3 callersFunctionget_no_copy_ops
()
deepspeed/compile/util.py:94
↓ 3 callersMethodget_offset
(self)
deepspeed/io/base_io_buffer.py:31
↓ 3 callersMethodget_param_comm_dtype
(self, param)
deepspeed/runtime/base_optimizer.py:376
↓ 3 callersFunctionget_partition_weights
Return the amount of weight in each partition.
tests/unit/runtime/utils/test_partition.py:91
↓ 3 callersMethodget_pipe_parallel_world_size
The number of stages in the pipeline.
deepspeed/runtime/pipe/topology.py:417
↓ 3 callersFunctionget_position_id_node
(gm: GraphModule)
deepspeed/compile/util.py:576
↓ 3 callersMethodget_prefix
(self)
op_builder/inference_cutlass_builder.py:59
← previousnext →1,001–1,100 of 11,258, ranked by callers