Functions — github.com/EverMind-AI/MSA

hub / github.com/EverMind-AI/MSA / functions

Functions265 in github.com/EverMind-AI/MSA

⨍ Functions 265◇ Types & classes 54

↓ 1 callersMethodmerge_poolig_doc_id

(self, device)

src/msa_service.py:221

↓ 1 callersFunctionmsa_benchmark

(args, data)

src/app/benchmark.py:223

↓ 1 callersMethodnum_model_layers

(self)

src/prefill.py:55

↓ 1 callersFunctionpacking_dataset

(dataset, tokenizer, cutoff_len, num_worker, cache_dataset_overwrite)

src/utils/data_utils.py:280

↓ 1 callersFunctionparse_args

()

src/app/benchmark.py:181

↓ 1 callersFunctionparse_benchmark_file

(args)

src/app/benchmark.py:39

↓ 1 callersFunctionparse_score_result

解析 LLM 返回的评分结果，返回 0-5 的分数

src/evaluation/llm_judge.py:73

↓ 1 callersFunctionpost_process_generate_ids

(tokenizer: PreTrainedTokenizer, ids: torch.Tensor)

src/utils/common.py:28

↓ 1 callersMethodprefill_stage2

(self, layer_idx: int, query_states: torch.Tensor, query_mask: torch.Tensor)

src/msa_service.py:703

↓ 1 callersFunctionpreprocess_sp_dataset

(seq_ids, world_size, sequence_parallel_mode)

src/utils/data_utils.py:205

↓ 1 callersFunctionprint_model_stats

通用模型参数统计工具。自动识别是 Dense 还是 MoE 模型，并计算 Total vs Active 参数量。

src/utils/common.py:101

↓ 1 callersFunctionprocess_results

(requests: List[Dict], index_to_doc, doc_to_index)

src/app/benchmark.py:137

↓ 1 callersFunctionread_config_to_args

(config_path)

src/app/benchmark.py:214

↓ 1 callersMethodrecord_kwargs

(self, layer_idx, kwargs)

src/utils/cache.py:143

↓ 1 callersMethodrecv_meta

(q: mp.Queue)

src/prefill.py:108

↓ 1 callersMethodrelease

释放一个执行权限，唤醒等待的请求

src/utils/tools.py:85

↓ 1 callersMethodrespond

(self)

src/msa_service.py:153

↓ 1 callersFunctionsave_checkpoint

(model, tokenizer, save_model_path)

src/utils/resave_model.py:9

↓ 1 callersMethodsave_idx_to_doc

(self, idx_to_doc: Dict[int, str])

src/msa_service.py:494

↓ 1 callersFunctionscale_memory

(context, query, scale=3)

src/utils/scale.py:4

↓ 1 callersFunctionsearch_for_fit

r""" Finds the index of largest number that fits into the knapsack with the given capacity.

src/utils/data_utils.py:106

↓ 1 callersMethodsend_documents

(q: mp.Queue, docs)

src/prefill.py:102

↓ 1 callersMethodsequence_pooling_kv

(self, key_states, value_states, doc_indices, global_chunk_ids)

src/msa/memory_sparse_attention.py:467

↓ 1 callersMethodsequence_pooling_qkv

(self, query_states, key_states, value_states, doc_indices, global_chunk_ids)

src/msa/memory_sparse_attention.py:488

↓ 1 callersMethodserialize

安全分片序列化： 1. 创建目录 2. 保存元数据 (meta.pt) 3. 逐层保存 Tensor 到独立文件，避免内存峰值

src/msa_service.py:351

↓ 1 callersMethodset_memory_client

(self, memory_client)

src/msa/memory_sparse_attention.py:65

↓ 1 callersFunctionset_seed

固定所有随机种子以确保实验的可复现性。

src/app/benchmark.py:23

↓ 1 callersMethodsetup_memory_client

(self, model)

src/msa_service.py:1057

↓ 1 callersFunctionshould_regenerate

return a text if the response should be regenerated, or return None

src/app/benchmark.py:202

↓ 1 callersFunctionsort_requests

创建动态批次：按照input_id长度排序，然后根据max_input_length分批

src/app/benchmark.py:69

↓ 1 callersMethodsplit_docs

(docs: List[Document], block_size: int)

src/prefill.py:75

↓ 1 callersMethodstart_all

启动所有GPU监控

src/utils/gpu_monitor.py:276

↓ 1 callersMethodstart_workers

(self)

src/msa_service.py:1653

↓ 1 callersMethodstop_workers

(self)

src/msa_service.py:1668

↓ 1 callersMethodupdate

( self, key_states: torch.Tensor, value_states: torch.Tensor, layer_idx: int,

src/msa_service.py:104

↓ 1 callersMethodwait_for_ready

src/utils/gpu_monitor.py:207

Method__enter__

(self)

src/msa_service.py:1645

Method__exit__

(self, exc_type, exc_val, exc_tb)

src/msa_service.py:1648

src/msa/configuration_msa.py:12

Method__init__

(self, _distributed_cache_data=None)

src/msa_service.py:92

Method__init__

(self, gpu_id: int, generate_config: GenerateConfig, model_

src/msa_service.py:1014

Method__init__

(self, generate_config: GenerateConfig, model_config: ModelConfig,

src/msa_service.py:1369

Method__init__

(self, bench_name: str)

src/benchmarks.py:104

Method__init__

该 worker 被MemoryWorker创建并仅执行prefill stage 1获取block 的kv cache

src/prefill.py:40

Method__init__

(self, msa_config=None, **kwargs)

src/msa/configuration_msa.py:29

Method__init__

(self, config: Qwen3Config, layer_idx: int)

src/msa/memory_sparse_attention.py:21

Method__init__

(self, config: Qwen3Config, layer_idx: int, attn_type: str = "sparse_attention")

(self, *args, **kwargs)

src/utils/misc.py:10

Method__init__

(self, _distributed_cache_data=None)

src/utils/cache.py:140

Method__init__

(self, cache_config)

src/utils/cache.py:179

Method__init__

初始化请求限流器 Args: max_concurrent: 最大并发请求数，默认10

src/utils/tools.py:59

Method__init__

(self, disabled=False)

src/utils/tools.py:102

Method__init__

(self)

src/utils/callbacks.py:21

Method__init__

(self, _run=None)

src/utils/callbacks.py:52

Method__init__

(self, _run=None, multitest=None, trainer=None, gen_kwargs=None)

src/utils/callbacks.py:100

Method__init__

(self, prof)

src/utils/callbacks.py:168

Method__init__

(self, gpu_id: int, envs: dict)

src/utils/gpu_worker.py:6

Method__init__

初始化GPU显存监控器 Args: gpu_index: 要监控的GPU索引，默认0 interval: 监控间隔时间（秒），默认1秒 unit: 返回的单位，支持 'MB'

src/utils/gpu_monitor.py:9

Method__init__

初始化多GPU监控器 Args: gpu_indices: 要监控的GPU索引列表，None表示监控所有GPU interval: 监控间隔 unit: 返回的单位

src/utils/gpu_monitor.py:253

Method__post_init__

(self)

src/msa_service.py:46

Method__post_init__

(self)

src/msa_service.py:53

Method__post_init__

(self)

src/msa_service.py:63

Method__post_init__

(self)

src/msa_service.py:72

Method__post_init__

(self)

src/msa_service.py:81

Method__post_init__

(self)

src/msa_service.py:88

Method__post_init__

(self)

src/msa_service.py:130

Method__post_init__

(self)

src/config/memory_config.py:18

Method__repr__

(self)

src/benchmarks.py:119

Method__setattr__

重写 __setattr__，确保设置 msa_config 时自动转换为 DotDict

src/msa/configuration_msa.py:34

Method__setstate__

(self, state)

src/msa/configuration_msa.py:15

Function_get_cosine_schedule_with_warmup_lr_lambda

( current_step: int, *, num_warmup_steps: int, num_training_steps: int, num_cycles: float )

src/utils/misc.py:72

Method_monitor_loop

监控循环，运行在单独线程中

src/utils/gpu_monitor.py:74

Method_sample

( self, input_ids: torch.LongTensor, logits_processor: LogitsProcessorList, st

src/msa/generate.py:23

Method_setup_module

(module)

src/msa_service.py:1058

Methodadd

(self, name)

src/utils/tools.py:106

Methodclear_kvcache

(self)

src/utils/cache.py:202

Methodclear_query

清理查询相关的临时数据移除查询过程中产生的临时数据，保持缓存清洁

src/utils/cache.py:66

Methodclear_query

清理查询相关的临时数据移除查询过程中产生的临时数据，保持缓存清洁

src/utils/cache.py:300

Functionclip_grad_norm_

r"""Clips gradient norm of an iterable of parameters. The norm is computed over all gradients together, as if they were concatenated into a s

src/utils/misc.py:87

Methodcompute_pooled_states_via_cumsum

(states, counts_view, lengths)

src/msa/memory_sparse_attention.py:475

Methodcopy

创建缓存的深拷贝 Returns: CustomQuantizeDynamicCache: 缓存的新副本

src/utils/cache.py:317

Functioncopy_dict_to_cpu

(d: dict)

src/utils/cache.py:427

Functioncopy_dict_to_gpu

(d: dict, device)

src/utils/cache.py:433

Functioncopy_kv_cache_to_device

(cache_obj, cuda_device, copy_v: bool=True)

src/utils/cache.py:441

Functiondata_post_process_sequence_parallel

( dataset, training_args, sequence_parallel_size, sequence_parallel_mode, cutoff_len,

src/utils/data_utils.py:296

Functiondecode_generate_ids

(tokenizer: PreTrainedTokenizer, ids: torch.Tensor)

src/utils/common.py:36

Methoddefault_callback

(self, texts: List[str], recall_topk, userdata)

src/msa_service.py:1794

Methoddoc_query

(self, query_states: torch.Tensor, attention_mask: torch.Tensor, layer_idx: int)

src/msa_service.py:263

Functiondraw_bounding_boxes

( image: Union[torch.Tensor, PIL.Image.Image], boxes: Union[torch.Tensor, List, np.ndarray],

src/utils/common.py:49

Functionenable_explicit_format

()

src/utils/common.py:93

Functioneval_all

(predict, label)

src/app/benchmark.py:116

Methodfetch

(tensor, slot_idx)

src/msa_service.py:1250

Methodforward

( self, hidden_states: torch.Tensor, doc_ids: torch.LongTensor, attention_mask

src/msa/memory_sparse_attention.py:68

Methodforward

( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None,

src/msa/model.py:67

Methodforward

( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.T

src/msa/model.py:156

Methodforward

r""" labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked lan

src/msa/model.py:332

← previous next →101–200 of 265, ranked by callers