From 77b9e4f80912ebbf1b9b4f521b56b1962a23bef1 Mon Sep 17 00:00:00 2001 From: zuiho <2324465096@qq.com> Date: Sun, 23 Aug 2026 15:41:10 +0800 Subject: [PATCH 1/2] Learn owner rules from human reviews --- .../vllm-omni/components/diffusion/_index.md | 4 +-- .../vllm-omni/components/diffusion/rules.md | 33 ++++++++++++++++--- .../vllm-omni/components/serving/_index.md | 4 +-- .../vllm-omni/components/serving/rules.md | 19 +++++++++-- 4 files changed, 49 insertions(+), 11 deletions(-) diff --git a/knowledge/repos/vllm-omni/components/diffusion/_index.md b/knowledge/repos/vllm-omni/components/diffusion/_index.md index 1d073c227..d7e546b75 100644 --- a/knowledge/repos/vllm-omni/components/diffusion/_index.md +++ b/knowledge/repos/vllm-omni/components/diffusion/_index.md @@ -1,7 +1,7 @@ --- title: "Diffusion" created: 2026-07-10 -updated: 2026-08-13 +updated: 2026-08-23 type: index tags: [vllm-omni, components, diffusion] sources: [] @@ -28,7 +28,7 @@ sources: [] | 遇到什么 | 查看哪里 | |---|---| | 理解共享职责和数据流 | [architecture](architecture.md) | -| 根据 PR 描述直达 execution parity、checkpoint/distributed、quality evidence 或 paged-cache system-runtime 规则组与第一批源码 | [rules 与代码地图](rules.md) | +| 根据 PR 描述直达 execution parity、checkpoint/artifact identity、quality evidence 或 system-runtime 异常清理规则组与第一批源码 | [rules 与代码地图](rules.md) | | diffusion step 与 request/continuous batching | [step and batching](step-and-batching.md) | | Cache-DiT、TeaCache 和 prefix cache | [cache acceleration](cache-acceleration.md) | | TP/PP/SP/CFG/VAE/HSDP 等并行策略 | [parallelism](parallelism.md) | diff --git a/knowledge/repos/vllm-omni/components/diffusion/rules.md b/knowledge/repos/vllm-omni/components/diffusion/rules.md index e26607084..9f9559598 100644 --- a/knowledge/repos/vllm-omni/components/diffusion/rules.md +++ b/knowledge/repos/vllm-omni/components/diffusion/rules.md @@ -1,10 +1,10 @@ --- title: "Diffusion 共享规则" created: 2026-07-20 -updated: 2026-08-13 +updated: 2026-08-23 type: rule tags: [vllm-omni, components, diffusion] -sources: ["PR #4341", "PR #5001", "PR #5087", "PR #5088", "PR #5136", "PR #6094", vllm_omni/diffusion/worker/diffusion_model_runner.py, vllm_omni/diffusion/model_loader/diffusers_loader.py, vllm_omni/diffusion/distributed/hsdp.py] +sources: ["PR #4341", "PR #5001", "PR #5087", "PR #5088", "PR #5136", "PR #6094", "PR #6385", "PR #6445", vllm_omni/diffusion/worker/diffusion_model_runner.py, vllm_omni/diffusion/model_loader/diffusers_loader.py, vllm_omni/diffusion/distributed/hsdp.py] confidence: high --- @@ -23,18 +23,20 @@ confidence: high | CUDA Graph、compile、fused scheduler/solver、cache path、eager parity | `execution-parity`：`DIFF-1a` | `vllm_omni/diffusion/compile.py::regionally_compile` → `vllm_omni/diffusion/worker/diffusion_model_runner.py::{DiffusionModelRunner.execute_model,execute_model_batch}` → 命中模型的 denoise/solver consumer | | seed、request-local generator、guidance=0、并发 RNG、batched generators | `execution-parity`：`DIFF-1b` | `vllm_omni/inputs/data.py::OmniDiffusionSamplingParams` → `diffusion_model_runner.py::DiffusionModelRunner._initialize_generator` → `request_batch.py::DiffusionRequestBatch.collate_sampling_param_generators` | | ModelOpt/checkpoint adapter、weight/scale remap、unknown tensor、resolution path | `checkpoint-distributed`：`DIFF-2a` | `vllm_omni/diffusion/model_loader/diffusers_loader.py::{DiffusersPipelineLoader._get_checkpoint_adapter,load_weights}` → `checkpoint_adapters/modelopt.py::{ModelOptFp8CheckpointAdapter._resolve_target_and_output_names,adapt}` | +| host-weight artifact、source identity、layout/dtype policy、warm restore | `checkpoint-distributed`：`DIFF-2d` | `vllm_omni/diffusion/model_loader/host_weights/source_identity.py` → `contracts.py` / `identity_adapter.py` → 具体 representation policy 与 restorer | | HSDP/FSDP、`fully_shard`、DeviceMesh、packed/scalar parameter、FP8 | `checkpoint-distributed`：`DIFF-2b` | `vllm_omni/diffusion/distributed/hsdp.py::{apply_hsdp_to_model,shard_model}` → `model_loader/diffusers_loader.py::DiffusersPipelineLoader._load_model_with_hsdp` → `quantization/hsdp_fp8.py::prepare_fp8_layers_for_fsdp` | | component quantization、text encoder/transformer/VAE 独立配置、owner prefix、meta/offload | `checkpoint-distributed`：`DIFF-2c` | `vllm_omni/diffusion/data.py::OmniDiffusionConfig._propagate_quantization_from_tf_config` → `model_loader/diffusers_loader.py::{DiffusersPipelineLoader._get_weight_sources,_process_weights_after_loading}` → 命中 component 的真实 linear consumer | | LPIPS/PSNR/相似度阈值、CPU offload、量化质量证据 | `quality-evidence`：`DIFF-3a` | changed quality test 的 exact case → `vllm_omni/diffusion/worker/diffusion_model_runner.py::DiffusionModelRunner.execute_model` → 命中 pipeline；baseline/candidate 必须复用同一路径 | | paged KV/cache、block manager、显存预算、warmup/profile、scheduler admission | `system-runtime`：`DIFF-4a`, `DIFF-4b`, `DIFF-4c` | `diffusion_engine.py` 初始化顺序 → `diffusion_kv/initialization.py` → `worker/diffusion_worker.py::determine_available_kv_memory` → native config/allocator → `sched/base_scheduler.py` 的 enabled 请求、异常和清理路径 | +| worker/RPC 执行异常、rank-status envelope、traceback 或 device cache 清理 | `system-runtime`：`DIFF-4d` | `worker/diffusion_worker.py::{WorkerProc._execute_rpc,WorkerProc._worker_busy_loop}` → raise/reply/status-envelope 各终止路径 | | 审查组 | 什么时候触发 | 规则 ID | |---|---|---| | `core` | 每次共享 diffusion 审查 | `DIFF-1a`, `DIFF-1b` | | `execution-parity` | graph/eager、solver、RNG、generator、zero/default | `DIFF-1a`, `DIFF-1b` | -| `checkpoint-distributed` | checkpoint、quantization、HSDP/FSDP | `DIFF-2a`, `DIFF-2b`, `DIFF-2c` | +| `checkpoint-distributed` | checkpoint、quantization、HSDP/FSDP、host-weight artifact identity | `DIFF-2a`, `DIFF-2b`, `DIFF-2c`, `DIFF-2d` | | `quality-evidence` | 质量阈值、offload、A/B case | `DIFF-3a` | -| `system-runtime` | cache/资源预算、native 边界、feature gate、异常或并发调度 | `DIFF-4a`, `DIFF-4b`, `DIFF-4c` | +| `system-runtime` | cache/资源预算、native 边界、feature gate、异常或并发调度 | `DIFF-4a`, `DIFF-4b`, `DIFF-4c`, `DIFF-4d` | | `author-routing` | 只供 Direct reviewer 导航，不作为 finding 规则 | `DIFF-0a`, `DIFF-0b` | ## 优化路径与 eager 的等价合同 @@ -91,6 +93,18 @@ confidence: high 命中/排除集合，并验证 meta-device parameter 不会被提前 move。FLUX.2 的具体边界见 [FLUX.2 规则](../../models/flux2/rules.md)。 ^[PR #5136] +### DIFF-2d — Artifact identity 只信任已验证的不可变来源 + +- 触发：新增或修改 host-weight artifact、warm restore、source identity、layout/dtype policy。 +- 强制：共享 identity、restore 和 model ABI 保持 representation-neutral，dtype/layout 只属于 + 具体 policy/producer；只有同时验证 source kind、snapshot revision、repo topology 和 blob + 位置的内容寻址仓库，才可用 blob 名代替内容哈希，其他本地文件和 symlink 必须按内容定身份。 +- 禁止：因 symlink 目标名看似 40/64 位十六进制就假设不可变；把 `bf16`、具体 layout 或首个 + producer 名写进共享 identity/restorer API；用进程内 inode/mtime guard 代替跨启动身份校验。 +- 验收：第二种 synthetic representation 复用共享 identity/restorer；任意本地十六进制命名 + symlink 在同大小内容被替换后产生不同 identity；合法 Hub snapshot→blob 拓扑仍走受控快捷路径。 + ^[PR #6445] + ## 质量阈值与资源辅助 ### DIFF-3a — 质量阈值必须由完全相同的测试 case 产生 @@ -140,5 +154,16 @@ confidence: high - 验收：path/lifecycle matrix 每行有 production-path 测试或明确 `MISSING_EVIDENCE`；enabled 关键行缺失时审查只能是 partial，不能输出 clean/no findings。 ^[PR #6094] +### DIFF-4d — 异常清理覆盖每种错误传递形态 + +- 触发：worker/RPC 在异常后清 traceback、释放 device tensor/cache，或把异常转换成 + reply、rank-status envelope、`DiffusionOutput`。 +- 强制：先保存对外所需的字符串化错误，再对 raise、普通错误回复和 collect-status 正常返回 + 三条路径逐一执行同一清理；调用方把异常折叠成成功返回形态时仍必须释放原异常引用。 +- 禁止：只在最外层 `except` 清理，却让内部状态封装路径正常返回；清理后继续从异常对象读取 + message/traceback；用单一路径 mock 证明所有 RPC mode 都不会保留 device tensor。 +- 验收：测试分别覆盖 `collect_rank_status=true` 与抛出/错误回复路径，断言错误文本仍可见、 + `exc.__traceback__ is None`、cache cleanup 被调用，后续健康 RPC 仍成功。 ^[PR #6385] + 相关执行流见 [Diffusion architecture](architecture.md)；benchmark 证据合同见 [performance evidence](../../benchmark/guides/performance-evidence.md)。 diff --git a/knowledge/repos/vllm-omni/components/serving/_index.md b/knowledge/repos/vllm-omni/components/serving/_index.md index 072bf422c..1cdeb8f6e 100644 --- a/knowledge/repos/vllm-omni/components/serving/_index.md +++ b/knowledge/repos/vllm-omni/components/serving/_index.md @@ -1,7 +1,7 @@ --- title: "Serving" created: 2026-07-10 -updated: 2026-07-31 +updated: 2026-08-23 type: index tags: [vllm-omni, components, serving] sources: [] @@ -28,4 +28,4 @@ sources: [] | 遇到什么 | 查看哪里 | |---|---| | 理解入口到 engine 的边界 | [architecture](architecture.md) | -| 根据 PR 描述直达 request contract、streaming format、artifact readiness 或 metrics 的规则组与第一批源码 | [rules 与代码地图](rules.md) | +| 根据 PR 描述直达 request contract、streaming format、artifact readiness、factory lifecycle 或 metrics 的规则组与第一批源码 | [rules 与代码地图](rules.md) | diff --git a/knowledge/repos/vllm-omni/components/serving/rules.md b/knowledge/repos/vllm-omni/components/serving/rules.md index 8aa99a5fa..3affc400c 100644 --- a/knowledge/repos/vllm-omni/components/serving/rules.md +++ b/knowledge/repos/vllm-omni/components/serving/rules.md @@ -1,10 +1,10 @@ --- title: "Serving 规则" created: 2026-07-20 -updated: 2026-07-31 +updated: 2026-08-23 type: rule tags: [vllm-omni, components, serving] -sources: ["PR #3576", "PR #4718", "PR #4834", "PR #4905", "PR #4912", "PR #5157", "claude-workflow-starter-private@09dca46", "zuiho-kai/claude-workflow-starter@c217fc6", vllm_omni/entrypoints/async_omni.py, vllm_omni/entrypoints/openai/diffusion_request_utils.py, vllm_omni/entrypoints/openai/serving_speech.py, vllm_omni/metrics/prometheus.py] +sources: ["PR #3576", "PR #4718", "PR #4834", "PR #4905", "PR #4912", "PR #5157", "PR #6138", "claude-workflow-starter-private@09dca46", "zuiho-kai/claude-workflow-starter@c217fc6", vllm_omni/entrypoints/async_omni.py, vllm_omni/entrypoints/openai/diffusion_request_utils.py, vllm_omni/entrypoints/openai/serving_speech.py, vllm_omni/metrics/prometheus.py] confidence: high --- @@ -23,6 +23,7 @@ confidence: high | `chat_template_kwargs`、raw HTTP/SDK `extra_body`、text/audio modalities、choices、空音频 | `chat-multimodal-contract`：`SERV-4c` + 命中模型规则 | upstream `ChatCompletionRequest` → `serving_chat.py::{OmniOpenAIServingChat._preprocess_chat,OmniOpenAIServingChat.chat_completion_full_generator,OmniOpenAIServingChat._create_text_choice,OmniOpenAIServingChat._create_audio_choice}` | | endpoint restriction、unsupported route、capability、completions/chat/speech 400 | `endpoint-capability`：`SERV-4c`, `SERV-4d` | `config/endpoint_policy.py::{OmniServingCapability,shutdown_unsupported_routes}` → `config/config_factory.py::StageConfigFactory.get_pipeline_endpoint_restrictions` → `engine/async_omni_engine.py::AsyncOmniEngine.__init__` → `entrypoints/openai/api_server.py::build_app` | | sleep/wake、partial stage/tag、idempotency、ACK、generation admission | `engine-lifecycle`：`SERV-5a`, `SERV-5b` | `entrypoints/async_omni.py::{AsyncOmni.sleep,AsyncOmni.wake_up,AsyncOmni.generate}` → `worker/base.py::{handle_sleep_task,handle_wake_task}` / `diffusion/worker/diffusion_worker.py` | +| serving class/factory 重构、optional adapter、diffusion/no-TTS 实例、warmup | `engine-lifecycle`：`SERV-5c` | `entrypoints/openai/serving_speech.py` 的所有 factory/`__new__` 路径 → `warmup`、voice upload/list、speech request caller | | SSE/streaming speech、audio format、PCM/WAV、speed、首 chunk 前校验 | `streaming-format`：`SERV-1a`, `SERV-1b` | `vllm_omni/entrypoints/openai/protocol/audio.py::{OpenAICreateSpeechRequest.validate_streaming_constraints,StreamingSpeechSessionConfig.validate_streaming_constraints}` → `serving_speech.py::{OmniOpenAIServingSpeech._validate_speech_streaming_request,OmniOpenAIServingSpeech.create_speech}` | | `ref_audio`、x-vector/ICL、artifact cache、readiness、失败后 engine 存活 | `artifact-readiness`：`SERV-3a`, `SERV-3b` | `vllm_omni/entrypoints/openai/serving_speech.py::{_qwen3_tts_can_use_ref_audio_artifact_only,_track_ref_audio_artifact_warmup,_mark_ref_audio_artifact_ready_for_request,_discard_ref_audio_artifact_ready_if_unreferenced}` | | Prometheus、waiting/running gauge、replica stats、throttle、collector lifecycle | `metrics-lifecycle`：`SERV-2a`, `SERV-2b` | `vllm_omni/entrypoints/omni_base.py::{OmniBase._log_summary_and_cleanup,OmniBase._process_stage_metrics_message}` → `vllm_omni/metrics/prometheus.py::{OmniPrometheusMetrics.__init__,set_running,set_waiting}` | @@ -35,7 +36,7 @@ confidence: high | `artifact-readiness` | artifact cache、capability、ready/mark/discard | `SERV-3a`, `SERV-3b` | | `chat-multimodal-contract` | chat template kwargs、SDK flatten、text/audio response shape | `SERV-4c` + 命中模型规则 | | `endpoint-capability` | endpoint restriction、unsupported route、公开 400 | `SERV-4c`, `SERV-4d` | -| `engine-lifecycle` | sleep/wake、partial stage/tag、ACK、generation admission | `SERV-5a`, `SERV-5b` | +| `engine-lifecycle` | sleep/wake、partial stage/tag、ACK、generation admission、factory 状态矩阵 | `SERV-5a`, `SERV-5b`, `SERV-5c` | | `request-contract` | 请求字段、来源、冲突、dispatcher、consumer view | `SERV-4a`, `SERV-4b`, `SERV-4c`, `SERV-4d`, `SERV-4e`, `SERV-4f`, `SERV-4g`, `SERV-4h` | | `author-routing` | 只供 Direct reviewer 导航，不作为 finding 规则 | `SERV-0a`, `SERV-0b` | @@ -192,5 +193,17 @@ confidence: high sleep → wake → generate，不支持的 stage 在调用 worker 前明确拒绝。 ^[PR #4834] ^[PR #4905] ^[PR #4912] +### SERV-5c — 共享服务类重构必须闭合全部构造状态 + +- 触发：移动 model-specific capability 到 adapter、增加 factory/`__new__` 快路径，或让同一 + serving class 同时承载普通、无对应 stage、diffusion-only 等实例。 +- 强制：列出每个生产构造入口及其必有/可选属性；所有公共 caller 通过统一 accessor 或显式 + `None` 状态访问 optional capability，startup warmup 与 request route 使用同一状态合同。 +- 禁止：只测试目标 adapter 的正常构造；在 factory 实例上直接读取未初始化属性；用旧的 + model-type early return 掩盖无 adapter 状态。 +- 验收：至少覆盖普通 adapter、无 adapter 和 bypass-`__init__` factory 三类实例，并实际调用 + startup warmup、voice list/upload 与对应 request route；每条路径保持重构前的成功或结构化 + 错误合同，不得出现 `AttributeError`。 ^[PR #6138] + 请求到 engine 的边界见 [Serving architecture](architecture.md)；公开协议通用检查见 [review contracts](../../../../general/review/guides/reviewer-lens-contracts.md)。 From bdb6ac06d9bd99c198377b05231738b9608bf3af Mon Sep 17 00:00:00 2001 From: zuiho <2324465096@qq.com> Date: Sun, 23 Aug 2026 15:43:59 +0800 Subject: [PATCH 2/2] Keep diffusion quick map within Direct budget --- .../vllm-omni/components/diffusion/rules.md | 24 +++++++++---------- 1 file changed, 12 insertions(+), 12 deletions(-) diff --git a/knowledge/repos/vllm-omni/components/diffusion/rules.md b/knowledge/repos/vllm-omni/components/diffusion/rules.md index 9f9559598..70e51020b 100644 --- a/knowledge/repos/vllm-omni/components/diffusion/rules.md +++ b/knowledge/repos/vllm-omni/components/diffusion/rules.md @@ -15,26 +15,26 @@ confidence: high ## Direct 代码快速入口 -- **DIFF-0a — PR 描述先选共享执行地图。** Direct review 先按 title/body 声明的 graph、RNG、checkpoint、HSDP、quantization 或质量语义命中下表，再用 pinned changed files 验证是否确属共享 diffusion owner。 -- **DIFF-0b — 共享 runner 与模型 pipeline 分层。** 先打开命中行的共享 producer；只有 live 调用链证明偏差属于某个模型 pipeline、checkpoint 或常量时才进入模型 owner，不能横向遍历所有 diffusion 模型。 +- **DIFF-0a — PR 描述先选共享执行地图。** 按 title/body 的 graph、RNG、checkpoint、HSDP、quantization 或质量信号命中下表，再用 pinned files 验证 owner。 +- **DIFF-0b — 共享 runner 与模型 pipeline 分层。** 先查命中行的共享 producer；只有 live 调用链落到单模型 pipeline/checkpoint/常量时才进入模型 owner，不横扫其他模型。 | PR 描述在做什么 | 精确规则组 | 第一批 live 源码 | |---|---|---| -| CUDA Graph、compile、fused scheduler/solver、cache path、eager parity | `execution-parity`：`DIFF-1a` | `vllm_omni/diffusion/compile.py::regionally_compile` → `vllm_omni/diffusion/worker/diffusion_model_runner.py::{DiffusionModelRunner.execute_model,execute_model_batch}` → 命中模型的 denoise/solver consumer | -| seed、request-local generator、guidance=0、并发 RNG、batched generators | `execution-parity`：`DIFF-1b` | `vllm_omni/inputs/data.py::OmniDiffusionSamplingParams` → `diffusion_model_runner.py::DiffusionModelRunner._initialize_generator` → `request_batch.py::DiffusionRequestBatch.collate_sampling_param_generators` | -| ModelOpt/checkpoint adapter、weight/scale remap、unknown tensor、resolution path | `checkpoint-distributed`：`DIFF-2a` | `vllm_omni/diffusion/model_loader/diffusers_loader.py::{DiffusersPipelineLoader._get_checkpoint_adapter,load_weights}` → `checkpoint_adapters/modelopt.py::{ModelOptFp8CheckpointAdapter._resolve_target_and_output_names,adapt}` | -| host-weight artifact、source identity、layout/dtype policy、warm restore | `checkpoint-distributed`：`DIFF-2d` | `vllm_omni/diffusion/model_loader/host_weights/source_identity.py` → `contracts.py` / `identity_adapter.py` → 具体 representation policy 与 restorer | -| HSDP/FSDP、`fully_shard`、DeviceMesh、packed/scalar parameter、FP8 | `checkpoint-distributed`：`DIFF-2b` | `vllm_omni/diffusion/distributed/hsdp.py::{apply_hsdp_to_model,shard_model}` → `model_loader/diffusers_loader.py::DiffusersPipelineLoader._load_model_with_hsdp` → `quantization/hsdp_fp8.py::prepare_fp8_layers_for_fsdp` | -| component quantization、text encoder/transformer/VAE 独立配置、owner prefix、meta/offload | `checkpoint-distributed`：`DIFF-2c` | `vllm_omni/diffusion/data.py::OmniDiffusionConfig._propagate_quantization_from_tf_config` → `model_loader/diffusers_loader.py::{DiffusersPipelineLoader._get_weight_sources,_process_weights_after_loading}` → 命中 component 的真实 linear consumer | -| LPIPS/PSNR/相似度阈值、CPU offload、量化质量证据 | `quality-evidence`：`DIFF-3a` | changed quality test 的 exact case → `vllm_omni/diffusion/worker/diffusion_model_runner.py::DiffusionModelRunner.execute_model` → 命中 pipeline；baseline/candidate 必须复用同一路径 | -| paged KV/cache、block manager、显存预算、warmup/profile、scheduler admission | `system-runtime`：`DIFF-4a`, `DIFF-4b`, `DIFF-4c` | `diffusion_engine.py` 初始化顺序 → `diffusion_kv/initialization.py` → `worker/diffusion_worker.py::determine_available_kv_memory` → native config/allocator → `sched/base_scheduler.py` 的 enabled 请求、异常和清理路径 | -| worker/RPC 执行异常、rank-status envelope、traceback 或 device cache 清理 | `system-runtime`：`DIFF-4d` | `worker/diffusion_worker.py::{WorkerProc._execute_rpc,WorkerProc._worker_busy_loop}` → raise/reply/status-envelope 各终止路径 | +| CUDA Graph、compile、fused scheduler/solver、cache path、eager parity | `execution-parity`：`DIFF-1a` | `compile.py::regionally_compile` → `diffusion_model_runner.py::{execute_model,execute_model_batch}` → model denoise/solver | +| seed、request-local generator、guidance=0、并发 RNG、batched generators | `execution-parity`：`DIFF-1b` | `inputs/data.py::OmniDiffusionSamplingParams` → runner `_initialize_generator` → request-batch generator collate | +| ModelOpt/checkpoint adapter、weight/scale remap、unknown tensor、resolution path | `checkpoint-distributed`：`DIFF-2a` | `diffusers_loader.py::{_get_checkpoint_adapter,load_weights}` → `modelopt.py::{_resolve_target_and_output_names,adapt}` | +| host-weight artifact、source identity、layout/dtype、warm restore | `checkpoint-distributed`：`DIFF-2d` | `model_loader/host_weights/{source_identity,contracts,identity_adapter}.py` → policy/restorer | +| HSDP/FSDP、`fully_shard`、DeviceMesh、packed/scalar parameter、FP8 | `checkpoint-distributed`：`DIFF-2b` | `distributed/hsdp.py::{apply_hsdp_to_model,shard_model}` → loader `_load_model_with_hsdp` → `hsdp_fp8.py` | +| component quantization、text encoder/transformer/VAE 独立配置、owner prefix、meta/offload | `checkpoint-distributed`：`DIFF-2c` | `data.py::_propagate_quantization_from_tf_config` → loader weight sources/post-load → component linear consumer | +| LPIPS/PSNR/相似度阈值、CPU offload、量化质量证据 | `quality-evidence`：`DIFF-3a` | changed exact case → runner `execute_model` → model pipeline；A/B 同路径 | +| paged KV/cache、block manager、显存预算、warmup/profile、scheduler admission | `system-runtime`：`DIFF-4a`, `DIFF-4b`, `DIFF-4c` | engine init → `diffusion_kv/initialization.py` → worker memory probe → native allocator → scheduler lifecycle | +| worker/RPC 异常、rank-status、traceback/device cache 清理 | `system-runtime`：`DIFF-4d` | `diffusion_worker.py::{_execute_rpc,_worker_busy_loop}` 的 raise/reply/status 路径 | | 审查组 | 什么时候触发 | 规则 ID | |---|---|---| | `core` | 每次共享 diffusion 审查 | `DIFF-1a`, `DIFF-1b` | | `execution-parity` | graph/eager、solver、RNG、generator、zero/default | `DIFF-1a`, `DIFF-1b` | -| `checkpoint-distributed` | checkpoint、quantization、HSDP/FSDP、host-weight artifact identity | `DIFF-2a`, `DIFF-2b`, `DIFF-2c`, `DIFF-2d` | +| `checkpoint-distributed` | checkpoint、quantization、HSDP/FSDP、artifact identity | `DIFF-2a`, `DIFF-2b`, `DIFF-2c`, `DIFF-2d` | | `quality-evidence` | 质量阈值、offload、A/B case | `DIFF-3a` | | `system-runtime` | cache/资源预算、native 边界、feature gate、异常或并发调度 | `DIFF-4a`, `DIFF-4b`, `DIFF-4c`, `DIFF-4d` | | `author-routing` | 只供 Direct reviewer 导航，不作为 finding 规则 | `DIFF-0a`, `DIFF-0b` |