附录 B 核心符号速查
符号按黄金路径顺序排列。路径缩写约定见第 2 章（
srt/…=python/sglang/srt/…），行号锚定 commit65824258549671404188450748b90ab7cacab2ae。
入口与进程
| 符号 | 职责 | 位置 | 相关章节 |
|---|---|---|---|
run_server | CLI 模式分派 | python/sglang/launch_server.py:19 | 3 |
launch_server | SRT 服务器总装 | srt/entrypoints/http_server.py:2871 | 3 |
Engine._launch_subprocesses | 拉起三进程 | srt/entrypoints/engine.py:1065 | 3 |
lifespan | FastAPI 运行时装配 | srt/entrypoints/http_server.py:281 | 3、4 |
generate_request（HTTP） | /generate 路由 | srt/entrypoints/http_server.py:915 | 4 |
OpenAIServingChat | OpenAI 协议翻译 | srt/entrypoints/openai/serving_chat.py:265 | 4 |
请求与状态
| 符号 | 职责 | 位置 | 相关章节 |
|---|---|---|---|
GenerateReqInput | 原生请求对象 | srt/managers/io_struct.py:177 | 4、5 |
TokenizerManager.generate_request | 请求总入口（async generator） | srt/managers/tokenizer_manager.py:854 | 5 |
TokenizerManager.handle_loop | 结果事件循环 | srt/managers/tokenizer_manager.py:2332 | 5 |
_handle_batch_output | meta 组装与流式唤醒 | srt/managers/tokenizer_manager.py:2347 | 5、12 |
TokenizedGenerateReqInput | 下行 ZMQ 消息 | srt/managers/io_struct.py:1041 | 5 |
Req | 单请求调度状态 | srt/managers/schedule_batch.py:999 | 8 |
调度
| 符号 | 职责 | 位置 | 相关章节 |
|---|---|---|---|
Scheduler.event_loop_normal | 同步主循环 | srt/managers/scheduler.py:1906 | 6 |
Scheduler.event_loop_overlap | 重叠主循环 | srt/managers/scheduler.py:1941 | 6 |
get_next_batch_to_run | 选批决策汇总 | srt/managers/scheduler.py:3641 | 6 |
update_running_batch | decode 批维护与 retract 触发 | srt/managers/scheduler.py:4191 | 6 |
get_new_batch_prefill | prefill 装批入口 | srt/managers/scheduler.py:3804 | 7 |
PrefillAdder / add_one_req | 装批与预算 | srt/managers/schedule_policy.py:621 / :1347 | 7 |
ScheduleBatch.retract_decode | 收回请求 | srt/managers/schedule_batch.py:3291 | 6 |
prepare_for_extend / prepare_for_decode | 批张量化 | srt/managers/schedule_batch.py:2722 / :3551 | 8 |
ForwardMode | 前向语义枚举 | srt/model_executor/forward_batch_info.py:193 | 8 |
ForwardBatch | 执行期张量视图 | srt/model_executor/forward_batch_info.py:488 | 8 |
BatchTokenIDOutput / BatchStrOutput | 上行 ZMQ 消息 | srt/managers/io_struct.py:1526 / :1627 | 12 |
内存与缓存
| 符号 | 职责 | 位置 | 相关章节 |
|---|---|---|---|
RadixCache.match_prefix | 前缀命中 | srt/mem_cache/radix_cache.py:354 | 9 |
RadixCache.insert | 前缀写回 | srt/mem_cache/radix_cache.py:414 | 9 |
RadixCache.cache_unfinished_req | chunk/在途写回 + 锁切换 | srt/mem_cache/radix_cache.py:502 | 9 |
RadixCache.evict | 叶子优先淘汰 | srt/mem_cache/radix_cache.py:553 | 9 |
inc_lock_ref / dec_lock_ref | 树节点锁 | srt/mem_cache/radix_cache.py:583 | 9 |
MHATokenToKVPool | KV 物理池 | srt/mem_cache/memory_pool.py:1971 | 9 |
ReqToToken.alloc/free | 请求行槽位 | srt/mem_cache/memory_pool.py:317 / :360 | 8、9 |
执行与采样
| 符号 | 职责 | 位置 | 相关章节 |
|---|---|---|---|
TpModelWorker.forward_batch_generation | 执行门面 | srt/managers/tp_worker.py:598 | 8、10 |
ModelRunner.forward / _forward_raw | 前向与分派 | srt/model_executor/model_runner.py:1706 / :1856 | 10 |
ATTENTION_BACKENDS | 注意力后端注册表 | srt/layers/attention/attention_registry.py:40 | 10 |
Sampler | 采样 | srt/layers/sampler.py:109 | 11 |
GrammarManager | 受限解码 FSM 管理 | srt/constrained/grammar_manager.py:30 | 11 |
输出与集群
| 符号 | 职责 | 位置 | 相关章节 |
|---|---|---|---|
process_batch_result | 结果分派 | srt/managers/scheduler.py:4765 | 12 |
process_batch_result_decode | decode 账本结算 | srt/managers/scheduler_components/batch_result_processor.py:921 | 12 |
stream_output | 攒批发往 Detokenizer | srt/managers/scheduler_components/output_streamer.py:119 | 12 |
DetokenizerManager.event_loop | 反分词循环 | srt/managers/detokenizer_manager.py:179 | 12 |
DataParallelController | DP 请求分发 | srt/managers/data_parallel_controller.py:140 | 13 |
PrefillBootstrapQueue | PD bootstrap | srt/disaggregation/prefill.py:156 | 14 |
CacheAwarePolicy | 缓存感知路由 | sgl-model-gateway/src/policies/cache_aware.rs | 15 |