[CI] Fix baseline QSA typing and pre-commit failures

.pre-commit-config.yaml

@@ -207,6 +207,9 @@ repos:
     entry: python tools/pre_commit/check_spdx_header.py
     language: python
     types: [python]
+    # This existing test declares AGPL-3.0-only. The Apache-specific fixer
+    # must not prepend a conflicting license to it.
+    exclude: ^tests/compile/test_qwen4exp_qsa_e4m3_gate\.py$
   - id: check-root-lazy-imports
     name: Check root lazy imports
     entry: python tools/pre_commit/check_init_lazy_imports.py

README.md

@@ -1162,7 +1162,6 @@ Join the **1Cat-vLLM Open-Source Community Group 8** by scanning the latest QR c
 
 <img width="966" height="1518" alt="472f168308d5e697058d30e51ffb2f4c" src="https://github.com/user-attachments/assets/fa8c71be-6a81-491e-895a-04bca665cceb" />
 
-
 > This QR code is valid through **September 20, 2026**. WeChat group QR codes expire periodically; if it has expired, add WeChat ID **`YM_isi`** to request the latest invitation.
 
 ---

vllm/models/qwen4_exp/nvidia/model.py

@@ -165,7 +165,7 @@ def _remap_qsa_cache_scale_name(
 
 def _validate_qsa_e4m3_scale_load(
     required_scales: set[str], loaded: set[str], cache_dtype: str
-) -> None:
+) -> set[str]:
     if cache_dtype not in ("fp8", "fp8_e4m3"):
         return set()
     missing_scales = sorted(required_scales - loaded)
@@ -212,19 +212,15 @@ def _finalize_qsa_e4m3_scale_load(
     required_scales = {
         f"{name}.{kind}_scale" for name in qsa_modules for kind in ("k", "v")
     }
-    missing_scales = _validate_qsa_e4m3_scale_load(
-        required_scales, loaded, cache_dtype
-    )
+    missing_scales = _validate_qsa_e4m3_scale_load(required_scales, loaded, cache_dtype)
     if not missing_scales:
         logger.info_once(
             "QSA E4M3 calibrated scale gate passed: loaded %d/%d K/V scales.",
             len(required_scales),
             len(required_scales),
         )
     for name, module in qsa_modules.items():
-        uncovered = any(
-            f"{name}.{kind}_scale" in missing_scales for kind in ("k", "v")
-        )
+        uncovered = any(f"{name}.{kind}_scale" in missing_scales for kind in ("k", "v"))
         if uncovered:
             module.adopt_default_kv_scales()
         else: