File tree Expand file tree Collapse file tree
knowledge/repos/vllm-omni/components Expand file tree Collapse file tree Original file line number Diff line number Diff line change @@ -83,3 +83,4 @@ shared RMSNorm 的empty-input与fused精度边界沿 tensor dtype rules 的DIFF-
8383
8484- [ Diffusion request batch 的 pipeline 合同] ( rules-request-batch.md ) 。
8585- [ Diffusion JIT CUDA 构建 include 顺序规则] ( rules-jit-cuda-includes.md )
86+ - [ vLLM-Omni Diffusion AMD CI 规则] ( rules-amd-ci.md )
Original file line number Diff line number Diff line change 1+ ---
2+ title : " vLLM-Omni Diffusion AMD CI 规则"
3+ created : 2026-10-13
4+ updated : 2026-10-13
5+ type : rule
6+ tags : [vllm-omni, components, diffusion]
7+ sources : ["PR # 8849"]
8+ ---
9+
10+ # vLLM-Omni Diffusion AMD CI 规则
11+
12+ ## VLLM-OMNI-PR8849-AMD-SLOW-CPU — 慢速 diffusion CPU 用例只属于专用 slow job,不回填 hash 分片
13+
14+ - 触发:修改 ` .buildkite/amd/test-amd-merge.yml ` 或 ` .buildkite/amd/test-amd-ready.yml ` 中五路 diffusion CPU hash 分片的命令或 ` --deselect ` 列表、` Simple · Diffusion Slow CPU Test ` job 定义,或 ` tests/buildkite/test_amd_pipeline.py ` 的 ` DIFFUSION_SLOW_CPU_* ` 常量;向该慢速集合增删用例。
15+ - 强制:六个慢速 node ID 由 blocking job ` Simple · Diffusion Slow CPU Test ` (mi300_1、timeout 40 分钟、JUnit ` artifacts/amd-diffusion-slow-cpu/pytest.xml ` 、exit 134 自动重试一次、` VLLM_ROCM_USE_AITER=0 ` 、沿用 ` core_model and cpu ` marker 过滤)唯一运行:` tests/diffusion/distributed/test_wan_vae_fastpath_install.py::test_omni_wan_vae_decode_override_is_bitwise_exact[dtype0] ` /` [dtype1] ` 、` tests/diffusion/distributed/test_wan_decoder_fast_path.py::test_sharded_fast_path_on_every_rank[True-exact] ` 、` tests/diffusion/models/lingbot_world/test_pipeline_lingbot_world.py::test_condition_continues_real_causal_vae_beyond_initial_horizon[realtime] ` /` [stepwise] ` ,以及同文件的 ` test_condition_encoder_stops_at_its_fixed_point_and_restarts_with_the_session ` 。merge 与 ready 两份 yml 的五路分片必须 ` --deselect ` 同一组 node ID;该集合在两份 yml 与 pipeline 测试三处同步维护。
16+ - 禁止:把这组 node ID 留在或加回五路 hash 分片(重复执行并拖慢关键路径,实测曾占 shard 2/5 约 15–23 分钟);用删用例替代迁移;改集合时只同步 merge/ready 之一或漏改 ` test_amd_pipeline.py ` ;用改 marker、单卡池或 Blocking 等级规避上述同步。
17+ - 验收:` pytest tests/buildkite/test_amd_pipeline.py ` 通过;` pytest --collect-only ` 能收集全部六个 node ID;AMD READY/MERGE 中 slow job 执行全部六例且 hash 分片最慢用例不再包含它们;slow job 超时保持在实测耗时(约 23 分钟)之上。 ^[ PR #8849 ]
18+
19+ <!-- kb:rule status=active since=v0.30.0 -->
Original file line number Diff line number Diff line change @@ -61,3 +61,4 @@ sources: []
6161- [ AutoRound NVFP4 method dispatch 合同] ( rules-autoround-dispatch.md ) 。
6262- [ 数值精确性与平台可移植性规则] ( rules-numeric-exactness.md )
6363- [ Model Executor — NPU KV cache 布局规则] ( rules-npu-kv-layout.md )
64+ - [ Model Executor — duplex per-unit turn metadata 规则] ( rules-duplex-turn-meta.md )
Original file line number Diff line number Diff line change 1+ ---
2+ title : " Model Executor — duplex per-unit turn metadata 规则"
3+ created : 2026-10-13
4+ updated : 2026-10-13
5+ type : rule
6+ tags : [vllm-omni, components, model-executor]
7+ sources : ["PR # 8822"]
8+ ---
9+
10+ # Model Executor — duplex per-unit turn metadata 规则
11+
12+ ## VLLM-OMNI-PR8822-MINICPMO-TURN-END — MiniCPM-o 4.5 llm2tts 必须在每次 native duplex handoff 写入 turn_end
13+
14+ - 触发:修改 ` vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py::llm2tts ` 中 ` turn_end ` /` turn_start ` 标志,或 ` is_native_duplex_handoff ` /` native_turn_end_handoff ` 分支写入 ` model_intermediate_buffer["meta"] ` 的逻辑。
15+ - 强制:每次 ` is_native_duplex_handoff ` 都要执行 ` model_intermediate_buffer.setdefault("meta", {})["turn_end"] = native_turn_end_handoff ` ,True 与 False 均显式写入,与 ` turn_start ` 的每-unit 写法一致;` turn_end ` 只能描述当前 unit,不得沿用上一 unit 留下的值。
16+ - 禁止:只在 ` native_turn_end_handoff ` 为真的 unit 上写 ` turn_end = True ` 、其余 unit 不写。connector 会跨 unit 合并 per-request meta,上一 turn 最后一个 unit 留下的 True 会粘到之后每个 unit,使其都拿到 multi-unit turn-end Talker budget:语音越过 unit 边界多生成帧并复述上一 unit 尾部文字,从第二个 turn 起表现为重复词和语速变慢。
17+ - 验收:MiniCPM-o 4.5 多 turn、每 turn 多 unit 的 duplex 用例须断言非结尾 unit 的 meta 携带 ` turn_end=False ` ,并确认第二个 turn 起不再出现越过文本多生成帧或复述前文的现象。^[ PR #8822 ]
18+
19+ <!-- kb:rule status=active since=v0.30.0 -->
You can’t perform that action at this time.
0 commit comments