fix: honor weight-check skips for quantized entries

python/sglang/srt/utils/weight_checker.py

@@ -319,7 +319,11 @@ def _build_check_entries(
             continue  # compared via its weight's comparable
         if name in quantized_set:
             qw = quantized_set[name]
-            yield CheckEntry(name, True, qw.comparable_cls(tensor, raw[qw.scale_name]))
+            yield CheckEntry(
+                name,
+                name not in skip_compare_names,
+                qw.comparable_cls(tensor, raw[qw.scale_name]),
+            )
         else:
             should_compare = name not in skip_compare_names and (
                 not _is_non_persistent_buffer_name(name)

test/registered/rl/test_weight_checker_e2e.py

@@ -160,7 +160,12 @@ def test_e_checksum_returns_ranks_with_hashes(self):
         self.assertIn("checksums", first)
         self.assertIn("parallelism_info", first)
 
-        info = first["parallelism_info"]
+        parallelism_info = first["parallelism_info"]
+        self.assertIsInstance(parallelism_info, list)
+        self.assertGreaterEqual(len(parallelism_info), 1)
+
+        info = parallelism_info[0]
+        self.assertEqual(info["role"], "target")
         for key in (
             "tp_rank",
             "tp_size",

test/registered/unit/utils/test_weight_checker.py

@@ -267,6 +267,18 @@ def test_substring_match_not_endswith(self):
             [("weird.cos_sin_cache.foo.bar", False, RawComparable(t))],
         )
 
+    def test_skip_set_marks_quantized_entry_not_compared(self):
+        qweight, sf_fp32, _ = _build_fp8_quant_pair()
+        raw = {"x.weight": qweight, "x.weight_scale_inv": sf_fp32}
+        quantized_set = {
+            "x.weight": QuantizedWeight(Fp8BlockComparable, "x.weight_scale_inv")
+        }
+        ref = Fp8BlockComparable(qweight, sf_fp32)
+        _assert_entries_close(
+            _build_check_entries(raw, {"x.weight"}, quantized_set),
+            [("x.weight", False, ref)],
+        )
+
     # --- fp8 quant pair (real dequant on real fp8 tensors) ---
 
     def test_fp8_quant_pair_yields_lazy_pair(self):