@@ -15,16 +15,18 @@ sources: ["PR #7648", "claude-workflow-starter-private@296ea45", "PR #4281", "PR
1515
1616## Direct 代码快速入口
1717
18+ 源码路径相对 ` vllm_omni/ ` ;测试路径相对仓库根。
19+
1820- ** VOMNI-CFG-0a — PR 描述先选代码地图。** Direct review 先用 PR title/body 声明的配置语义命中下表,再一次性用 pinned changed files 验证真实范围。描述只负责导航;冲突时以 live diff 和 consumer 为准。
1921- ** VOMNI-CFG-0b — 命中函数后停止文档导航。** 打开命中行的第一批源码后,沿 live producer→consumer 审查;只有调用链跨 owner 或具体未知量阻塞时才增加一个 owner 或 guide。
2022
2123| PR 描述在做什么 | 精确规则组 | 第一批 live 源码 |
2224| ---| ---| ---|
23- | 环境变量 inventory | ` VOMNI-CFG-3a ` | 失败路径对应源码与测试 |
24- | strict schema、unknown field、alias、flat→nested、structured/legacy/direct parity、typed projection | ` strict-normalization ` :` VOMNI-CFG-1a ` –` 1h ` ;[ 1s/1t] ( rules-legacy-engine-args.md ) | ` vllm_omni/ config/stage_config.py::{build_stage_runtime_overrides,strip_parent_engine_args}` → ` vllm_omni/ config/omni_config.py::{_build_diffusion_config_projection,VllmOmniConfig.from_pipeline_config}` → ` vllm_omni/ engine/stage_init_utils.py::{build_engine_args_dict,build_engine_args_dict_from_omni_stage_config}` |
25- | deploy YAML、` base_config ` 、pipeline/stage overlay、pipeline-owned alias/hook、headless/offline parity、最终逐 stage config | ` deploy-topology ` :` CONF-3a ` , ` CONF-4b ` , ` CONF-5a ` , [ CONF-5i] ( rules-topology-profiles.md#conf-5i-pipeline-专属-alias-与-hook-必须在通用配置层声明归一并消费 ) , [ CONF-5j] ( rules-topology-profiles.md#conf-5j-pipelineconfig-在构造时必须验证终端-topology ) | ` vllm_omni/ config/stage_config.py::{resolve_deploy_yaml,load_deploy_config,normalize_pipeline_cli_overrides,merge_pipeline_deploy,build_stage_runtime_overrides,_build_engine_args}` → ` vllm_omni/ config/config_factory.py::StageConfigFactory.create_from_model` → ` vllm_omni/ engine/stage_init_utils.py::_resolve_model_path` |
26- | composable strategy、axis、routing、load balancing、` strategy-config ` | ` composable-strategy ` :` CONF-4a ` | ` vllm_omni/ config/composable_parallel/strategy_loader.py::{parse_strategy_specs,load_strategy_specs}` → ` translator.py::translate_strategy_stack ` → ` apply.py::apply_strategy_specs ` → ` config_factory.py::{StageConfigFactory._apply_strategy_specs,StageConfigFactory._reconcile_strategy_with_cli} ` |
27- | ` gpu_memory_utilization ` 、` kv_cache_memory_bytes ` 、多 stage 共卡、小显存 OOM | ` deploy-memory ` :` CONF-1a ` , ` CONF-2a ` | ` vllm_omni/ config/stage_config.py::{build_stage_runtime_overrides,_build_engine_args}` → ` vllm_omni/ config/omni_config.py::{_build_runtime_config,_build_parallel_config,VllmOmniConfig.from_pipeline_config}` |
25+ | 环境变量 inventory | ` VOMNI-CFG-3a ` | config/environment_variable_inventory.py → tests/config/test_environment_variables.py |
26+ | strict schema、unknown field、alias、flat→nested、structured/legacy/direct parity、typed projection | ` strict-normalization ` :` VOMNI-CFG-1a ` –` 1h ` ;[ 1s/1t] ( rules-legacy-engine-args.md ) | ` config/stage_config.py::{build_stage_runtime_overrides,strip_parent_engine_args} ` → ` config/omni_config.py::{_build_diffusion_config_projection,VllmOmniConfig.from_pipeline_config} ` → ` engine/stage_init_utils.py::{build_engine_args_dict,build_engine_args_dict_from_omni_stage_config} ` |
27+ | deploy YAML、` base_config ` 、pipeline/stage overlay、pipeline-owned alias/hook、headless/offline parity、最终逐 stage config | ` deploy-topology ` :` CONF-3a ` , ` CONF-4b ` , ` CONF-5a ` , [ CONF-5i] ( rules-topology-profiles.md#conf-5i-pipeline-专属-alias-与-hook-必须在通用配置层声明归一并消费 ) , [ CONF-5j] ( rules-topology-profiles.md#conf-5j-pipelineconfig-在构造时必须验证终端-topology ) | ` config/stage_config.py::{resolve_deploy_yaml,load_deploy_config,normalize_pipeline_cli_overrides,merge_pipeline_deploy,build_stage_runtime_overrides,_build_engine_args} ` → ` config/config_factory.py::StageConfigFactory.create_from_model ` → ` engine/stage_init_utils.py::_resolve_model_path ` |
28+ | composable strategy、axis、routing、load balancing、` strategy-config ` | ` composable-strategy ` :` CONF-4a ` | ` config/composable_parallel/strategy_loader.py::{parse_strategy_specs,load_strategy_specs} ` → ` translator.py::translate_strategy_stack ` → ` apply.py::apply_strategy_specs ` → ` config_factory.py::{StageConfigFactory._apply_strategy_specs,StageConfigFactory._reconcile_strategy_with_cli} ` |
29+ | ` gpu_memory_utilization ` 、` kv_cache_memory_bytes ` 、多 stage 共卡、小显存 OOM | ` deploy-memory ` :` CONF-1a ` , ` CONF-2a ` | ` config/stage_config.py::{build_stage_runtime_overrides,_build_engine_args} ` → ` config/omni_config.py::{_build_runtime_config,_build_parallel_config,VllmOmniConfig.from_pipeline_config} ` |
2830| multi-stage HF sub-config、stage quantization view、` hf_config_name ` | ` stage-model-config ` :` VOMNI-CFG-1i ` | pipeline stage declaration → ` OmniModelConfig::{draw_hf_text_config,get_model_arch_config} ` → vLLM quantization selection |
2931| pipeline ` sampling_constraints ` 、` StageConfig.to_omegaconf() ` 、runtime stage config | ` stage-config-propagation ` :` VOMNI-CFG-1o ` | ` merge_pipeline_deploy ` → ` StageConfig.to_omegaconf() ` → ` engine.stage_configs ` |
3032| stage transport、` requires_full_payload_input ` 、topology projection 或 override rejection | ` stage-transport ` :` VOMNI-CFG-1p ` | ` stage_config.py::{StagePipelineConfig,_build_engine_args} ` → ` omni_config.py::_build_model_config ` → ` engine/arg_utils.py::OmniEngineArgs ` |
0 commit comments