MCPcopy Create free account

hub / github.com/deepseek-ai/DeepSpec / functions

Functions444 in github.com/deepseek-ai/DeepSpec

⨍ Functions 444◇ Types & classes 58
↓ 2 callersFunctionprepare_4d_causal_attention_mask
( *, attention_mask: torch.Tensor, dtype: torch.dtype, q_len: int, kv_len: int, past_s
deepspec/modeling/eagle3/common.py:142
↓ 2 callersMethodrecord_dataset_metrics
( self, *, dataset_name: str, metric_summary: dict[str, int | list[int]],
deepspec/eval/base_evaluator.py:691
↓ 2 callersMethodregister
(self, name, template)
deepspec/data/parser.py:22
↓ 2 callersFunctionrender_chat_messages
( tokenizer, messages, *, add_generation_prompt: bool, enable_thinking: bool | None = None
deepspec/data/parser.py:167
↓ 2 callersMethodreport_results
(self)
deepspec/eval/base_evaluator.py:707
↓ 2 callersFunctionreset
()
deepspec/utils/metrics.py:169
↓ 2 callersMethodrun_dataset
( self, *, dataset_name: str, max_samples: int | None, )
deepspec/eval/base_evaluator.py:513
↓ 2 callersFunctionsample_anchor_positions
( *, seq_len: int, loss_mask: torch.Tensor, num_anchors: int, device: torch.device, )
deepspec/modeling/dspark/common.py:123
↓ 2 callersMethodsample_block_tokens
Autoregressive sampling with RNN state. Args: base_logits: [batch, proposal_len, vocab] first_prev_token_ids: [batch]
deepspec/modeling/dspark/markov_head.py:227
↓ 2 callersMethodsave_and_eval_checkpoint
(self)
deepspec/trainer/base_trainer.py:326
↓ 2 callersFunctionsave_checkpoint
( *, model, draft_model, optimizer, checkpoint_dir_root: str, train_config, next_m
deepspec/trainer/ckpt_manager.py:136
↓ 2 callersMethodset_embedding_head_trainable
(self, trainable: bool)
deepspec/modeling/eagle3/qwen3/modeling.py:260
↓ 2 callersFunctiontest_output_path
(args: argparse.Namespace)
scripts/data/download_and_split.py:84
↓ 2 callersFunctiontrim_output_ids
( output_ids: torch.Tensor, num_input_tokens: int, stop_token_ids: list[int] | None, )
deepspec/eval/base_evaluator.py:58
↓ 2 callersFunctionunpack_index_record
(buffer, offset: int = 0)
deepspec/data/target_cache_dataset.py:100
↓ 2 callersFunctionvalidate_conversations
(row: dict, row_number: int)
scripts/data/download_and_split.py:131
↓ 2 callersFunctionvalidate_eagle3_target_layer_ids
(layer_ids, num_target_layers: int)
deepspec/modeling/eagle3/common.py:20
↓ 2 callersFunctionvalidate_target_layer_ids
(layer_ids, num_target_layers: int)
deepspec/modeling/dspark/common.py:59
↓ 2 callersFunctionwrite_finished_result
( future, output_handle, error_handle, stats, )
scripts/data/generate_train_data.py:233
↓ 2 callersMethodwrite_sample_bytes
(self, sample: TargetCacheSampleBytes)
deepspec/data/target_cache_dataset.py:351
↓ 1 callersFunction_all_reduce_loss_denominators
( loss_terms: dict[str, torch.Tensor], *, world_size: int, )
deepspec/modeling/dspark/loss.py:11
↓ 1 callersFunction_assert_schema_consistent
()
deepspec/utils/metrics.py:72
↓ 1 callersMethod_atomic_pickle_dump
(self, obj, dst_path)
deepspec/data/jsonl_dataset.py:82
↓ 1 callersMethod_auroc_from_hist
(pos_hist: torch.Tensor, neg_hist: torch.Tensor)
deepspec/eval/dspark/confidence_head.py:106
↓ 1 callersMethod_build_all_line_starts
(self)
deepspec/data/jsonl_dataset.py:90
↓ 1 callersMethod_build_confidence_head_recorder
(self)
deepspec/eval/dspark/evaluator.py:44
↓ 1 callersMethod_build_draft_model
(self, *, target_config, model_args)
deepspec/trainer/base_trainer.py:277
↓ 1 callersMethod_build_draft_model
(self, *, target_config, model_args)
deepspec/trainer/eagle3_trainer.py:54
↓ 1 callersFunction_build_fsdp_kwargs
( *, sharding_strategy_name: str, precision_dtype, world_size: int )
deepspec/trainer/base_trainer.py:55
↓ 1 callersFunction_build_loss
( *, loss_terms: dict[str, torch.Tensor], global_denominators: dict[str, torch.Tensor], ce_los
deepspec/modeling/dspark/loss.py:227
↓ 1 callersFunction_build_loss_weight_mask
( *, eval_mask: torch.Tensor, block_size: int, device: torch.device, loss_decay_gamma: Opt
deepspec/modeling/dspark/loss.py:25
↓ 1 callersFunction_build_next_token_position_mask
( *, loss_mask: torch.Tensor, attention_mask: torch.Tensor, )
deepspec/modeling/eagle3/loss.py:77
↓ 1 callersFunction_build_padded_next_token_target_probs
( target_logits: torch.Tensor, ttt_length: int, )
deepspec/modeling/eagle3/loss.py:92
↓ 1 callersMethod_build_train_dataloader
(self, start_offset_samples=0, num_samples=None)
deepspec/trainer/base_trainer.py:288
↓ 1 callersMethod_cache_path_from_key
(self, file_key)
deepspec/data/jsonl_dataset.py:78
↓ 1 callersFunction_collect_local_terms
( *, outputs: DSparkForwardOutput, loss_decay_gamma: Optional[float], l1_loss_alpha: float, )
deepspec/modeling/dspark/loss.py:90
↓ 1 callersFunction_compute_accept_rate_3d
( *, outputs: DSparkForwardOutput, aligned_target_logits: Optional[torch.Tensor], )
deepspec/modeling/dspark/loss.py:60
↓ 1 callersFunction_compute_gradient_accumulation_steps
( *, world_size: int, local_batch_size: int, global_batch_size: int )
deepspec/trainer/base_trainer.py:77
↓ 1 callersFunction_compute_local_l1_term
( *, outputs: DSparkForwardOutput, aligned_target_logits: Optional[torch.Tensor], loss_weight_
deepspec/modeling/dspark/loss.py:73
↓ 1 callersFunction_compute_local_probabilistic_stats
( *, outputs: DSparkForwardOutput, accept_rate_3d: Optional[torch.Tensor], valid_block_weights
deepspec/modeling/dspark/loss.py:40
↓ 1 callersFunction_compute_loss_normalizers
( *, position_masks: list[torch.Tensor], loss_normalization: str = "local_mean", )
deepspec/modeling/eagle3/loss.py:57
↓ 1 callersFunction_compute_samples_per_epoch
(*, dataset_size: int, global_batch_size: int)
deepspec/trainer/base_trainer.py:89
↓ 1 callersFunction_compute_training_schedule
( *, world_size: int, dataset_size: int, local_batch_size: int, global_batch_size: int,
deepspec/trainer/base_trainer.py:98
↓ 1 callersFunction_confident_prefix_length
( confidence_logits: torch.Tensor, *, block_size: int, threshold: float, )
deepspec/eval/dspark/draft_ops.py:82
↓ 1 callersMethod_ensure_index_mmap
(self)
deepspec/data/target_cache_dataset.py:668
↓ 1 callersMethod_ensure_shard
(self, sample_nbytes: int)
deepspec/data/target_cache_dataset.py:341
↓ 1 callersMethod_fetch_and_transfer
Pop the next CPU batch from the DataLoader and queue H2D on the side stream.
deepspec/data/cuda_prefetcher.py:36
↓ 1 callersMethod_file_key
(self, path)
deepspec/data/jsonl_dataset.py:72
↓ 1 callersMethod_forward_backbone
( self, *, position_ids: torch.LongTensor, attention_mask: Optional[torch.Tens
deepspec/modeling/dspark/gemma4/modeling.py:419
↓ 1 callersFunction_full_model_state_dict
(model)
deepspec/trainer/ckpt_manager.py:222
↓ 1 callersFunction_get_hook_tensor
(output)
scripts/data/prepare_target_cache.py:73
↓ 1 callersMethod_get_shard_mmap
(self, shard_id: int)
deepspec/data/target_cache_dataset.py:677
↓ 1 callersFunction_get_target_backbone
(target_model)
scripts/data/prepare_target_cache.py:55
↓ 1 callersFunction_get_target_hidden_size
(target_model)
scripts/data/prepare_target_cache.py:66
↓ 1 callersMethod_iter_stream
(self)
deepspec/utils/distributed.py:120
↓ 1 callersFunction_launch_eval
( *, target_model_name_or_path: str, checkpoint_dir: str, step: int, tensorboard_dir: str,
deepspec/trainer/base_trainer.py:138
↓ 1 callersFunction_local_reduce
(values, reduction: str)
deepspec/utils/metrics.py:46
↓ 1 callersFunction_log_eagle3_prefix_metrics
( *, correct_masks: list[torch.Tensor], accept_rate_masks: list[torch.Tensor], valid_masks: li
deepspec/modeling/eagle3/loss.py:138
↓ 1 callersFunction_log_eagle3_step_metrics
( *, step_idx: int, draft_logits: torch.Tensor, target_probs: torch.Tensor, position_mask:
deepspec/modeling/eagle3/loss.py:107
↓ 1 callersMethod_map_global_to_local
(self, global_idx)
deepspec/data/jsonl_dataset.py:59
↓ 1 callersMethod_open_file_mmap
(self, index)
deepspec/data/jsonl_dataset.py:64
↓ 1 callersFunction_pad_hidden_batch
(features: List[Dict], key: str)
deepspec/data/target_cache_dataset.py:812
↓ 1 callersFunction_parse_scalar
(value)
deepspec/utils/config.py:80
↓ 1 callersFunction_predict_confidence_logits
( model: DSparkModel, *, proposal_hidden_states: torch.Tensor, draft_input_ids: torch.Tensor,
deepspec/eval/dspark/draft_ops.py:57
↓ 1 callersMethod_prepare_attention_mask
( self, *, attention_mask: Optional[torch.Tensor], hidden_states: torch.Tensor
deepspec/modeling/eagle3/qwen3/modeling.py:274
↓ 1 callersMethod_prepare_attention_mask
( self, *, attention_mask: Optional[torch.Tensor], hidden_states: torch.Tensor
deepspec/modeling/eagle3/gemma4/modeling.py:348
↓ 1 callersMethod_prepare_render_messages
(self, messages)
deepspec/data/parser.py:146
↓ 1 callersFunction_print_summary
( *, summary, global_step: int, next_micro_step: int, micro_batches_per_epoch: int, ma
deepspec/utils/training_logger.py:73
↓ 1 callersMethod_process_feature
(self, item)
deepspec/data/target_cache_dataset.py:837
↓ 1 callersMethod_read_record
(self, index: int)
deepspec/data/target_cache_dataset.py:697
↓ 1 callersFunction_render_opt_assignment
(opt: str)
deepspec/trainer/ckpt_manager.py:48
↓ 1 callersFunction_safe_div
(numerator: torch.Tensor, denominator: torch.Tensor)
deepspec/utils/metrics.py:84
↓ 1 callersMethod_save_and_suspend
(self)
deepspec/trainer/base_trainer.py:339
↓ 1 callersFunction_save_model_checkpoint
(*, model, draft_model, checkpoint_dir: str)
deepspec/trainer/ckpt_manager.py:233
↓ 1 callersFunction_schema
()
deepspec/utils/metrics.py:61
↓ 1 callersFunction_serialize_training_state
( *, optimizer, next_micro_step: int, gradient_accumulation_steps: int, global_rank: int,
deepspec/trainer/ckpt_manager.py:195
↓ 1 callersFunction_validate_required_text_fields
(text_config)
deepspec/modeling/eagle3/gemma4/config.py:22
↓ 1 callersFunction_validate_required_text_fields
(text_config)
deepspec/modeling/dspark/gemma4/config.py:22
↓ 1 callersMethod_wrap_with_fsdp
(self, model)
deepspec/trainer/base_trainer.py:280
↓ 1 callersFunction_write_manifest
( *, output_dir: str, config, train_data_paths, target_layer_ids, hidden_size: int,
scripts/data/prepare_target_cache.py:157
↓ 1 callersFunction_write_scalars
(summary, *, global_step: int)
deepspec/utils/training_logger.py:66
↓ 1 callersFunctionapply_rotary_pos_emb
(q, k, cos, sin, unsqueeze_dim=1)
deepspec/modeling/eagle3/qwen3/modeling.py:33
↓ 1 callersFunctionapply_rotary_pos_emb
(q, k, cos, sin, unsqueeze_dim=1)
deepspec/modeling/dspark/qwen3/modeling.py:34
↓ 1 callersMethodbackward
(ctx, grad_output)
deepspec/modeling/eagle3/loss.py:328
↓ 1 callersMethodbuild_dataset_payload
( self, *, metrics_row: dict[str, object], confidence_row: dict, args_
deepspec/eval/dspark/confidence_head.py:442
↓ 1 callersMethodbuild_dataset_row
( self, *, dataset_name: str, metric_summary: dict[str, int | list[int]],
deepspec/eval/dspark/confidence_head.py:399
↓ 1 callersFunctionbuild_dspark_proposal
( model: DSparkModel, *, draft_input_ids: torch.Tensor, block_hidden: torch.Tensor, block_
deepspec/eval/dspark/draft_ops.py:96
↓ 1 callersFunctionbuild_global_target_cache_shard_map
(summaries)
deepspec/data/target_cache_dataset.py:510
↓ 1 callersMethodbuild_metrics_row
( self, *, dataset_name: str, metric_summary: dict[str, int | list[int]],
deepspec/eval/base_evaluator.py:469
↓ 1 callersMethodbuild_models
(self)
deepspec/eval/base_evaluator.py:458
↓ 1 callersMethodbuild_models
(self)
deepspec/trainer/base_trainer.py:244
↓ 1 callersFunctionbuild_query_kwargs
(args, messages, max_tokens=None)
scripts/data/generate_train_data.py:70
↓ 1 callersMethodbuild_table
(self)
deepspec/eval/dspark/confidence_head.py:541
↓ 1 callersFunctionbuild_target_cache_manifest
( *, num_samples: int, shards, target_layer_ids, hidden_size: int, extra_fields=None,
deepspec/data/target_cache_dataset.py:580
↓ 1 callersFunctioncall_sglang
(args, server_address, sample, max_tokens=None)
scripts/data/generate_train_data.py:106
↓ 1 callersFunctioncheck_output_paths
(args: argparse.Namespace)
scripts/data/download_and_split.py:88
↓ 1 callersFunctioncleanup_target_cache_tmp_dir
(output_dir: str)
deepspec/data/target_cache_dataset.py:609
↓ 1 callersMethodclose
(self)
deepspec/data/target_cache_dataset.py:490
↓ 1 callersFunctioncompact_row
(row: dict)
deepspec/eval/dspark/confidence_head.py:222
↓ 1 callersMethodcompute
(self)
deepspec/eval/dspark/confidence_head.py:116
← previousnext →101–200 of 444, ranked by callers