Skip to content

Commit b9bd40a

Browse files
authored
Merge pull request #363 from JiusiServe/kb/vllm-omni/vllm-omni-intake-f6717f6dec9b
knowledge(vllm-omni): 1 change(s) from intake
2 parents b2654e2 + 702d2ec commit b9bd40a

2 files changed

Lines changed: 20 additions & 0 deletions

File tree

‎knowledge/repos/vllm-omni/components/model-executor/_index.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -62,3 +62,4 @@ sources: []
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)
6464
- [Model Executor — Dynamo/cudagraph compile guard rules](rules-compile-guard.md)
65+
- [Model Executor — INC/AutoRound 量化配置规则](rules-inc-autoround.md)
Lines changed: 19 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,19 @@
1+
---
2+
title: "Model Executor — INC/AutoRound 量化配置规则"
3+
created: 2026-10-13
4+
updated: 2026-10-13
5+
type: rule
6+
tags: [vllm-omni, components, model-executor]
7+
sources: ["PR #8870"]
8+
---
9+
10+
# Model Executor — INC/AutoRound 量化配置规则
11+
12+
## VLLM-OMNI-PR8870-INC-PARSER — OmniINCConfig 升级后必须重建 config_parser,使阶段前缀重映射对层解析可见
13+
14+
- 触发:改动 `vllm_omni/quantization/inc_config.py` 中 `OmniINCConfig.from_inc_config` / `maybe_upgrade` 的属性复制或 `config_parser` 接线,或改动 `apply_vllm_mapper` 之后的层名解析路径;也覆盖 Qwen2.5-Omni AutoRound W4A16 的加载与 `tests/e2e/offline_inference/test_qwen2_5_omni_autoround_w4a16_expansion.py` 中的 e2e 测试。
15+
- 强制:`from_inc_config` 以 `omni.__dict__.update(inc.__dict__)` 复制原 INCConfig 属性后,必须用升级后的实例重建 parser(`omni.config_parser = type(inc.config_parser)(omni)`)。否则被复制的 parser 仍指向原 vanilla `INCConfig`,`apply_vllm_mapper` 对升级配置上 `block_name_to_quantize` 的前缀重映射(`thinker.model.layers` → `thinker.language_model.model.layers`)在层解析时不可见,受影响 thinker 层退化为未量化方法,加载 checkpoint 打包 `qweight` 张量时报 `There is no module or parameter named 'layers.0.mlp.down_proj.qweight'`(issue #5652)。升级与重映射行为由 `tests/diffusion/quantization/test_component_routing.py::test_upgraded_config_resolves_remapped_blocks` 回归守护:重映射后的 thinker 层解析为 (4, 128, True)、量化块之外的层保持未量化 (16, -1, True)、原始 vanilla 配置的 `block_name_to_quantize` 不被改写。issue #5652 专属的 e2e skip 已随本修复移除,Qwen2.5-Omni AutoRound W4A16 五条 e2e 测试须保持启用;Qwen3-Omni 同路径测试因独立的部署后端限制(默认 `moe_backend: triton` 被 vLLM 0.31.0 对该 AutoRound MoE checkpoint 拒绝)仍保持跳过,不得以本规则要求启用。
16+
- 禁止:令升级后的 `OmniINCConfig` 沿用指向原 `INCConfig` 的 `config_parser`,或以其他方式使 `get_layer_config` 读不到重映射后的块名;删除或绕过 `test_upgraded_config_resolves_remapped_blocks`;以本次修复已解决为由给 Qwen2.5-Omni AutoRound e2e 重新加回 issue-5652 skip,或把 Qwen3-Omni 的跳过误判为本规则违反。
17+
- 验收:`from_inc_config` 在复制 `__dict__` 后重建 `config_parser` 指向新实例;`pytest tests/diffusion/quantization/test_component_routing.py` 通过;`test_qwen2_5_omni_autoround_w4a16_expansion.py` 不再含 `_SKIP_ISSUE_5652` 且五个测试均无该 skip 装饰。^[PR #8870]
18+
19+
<!-- kb:rule status=active since=v0.30.0 -->

0 commit comments

Comments
 (0)