[Fix] Repair verify mask test fixture (#33087)

python/sglang/srt/layers/attention/verify_mask.py

@@ -46,10 +46,8 @@ class VerifyMask(msgspec.Struct):
     def fits(self, bs: int, num_draft_tokens: int) -> bool:
         """Whether this batch's writes stay inside the buffer.
 
-        Only the compact layout is checked. FULL_MASK keeps its pre-existing
+        Only the compact layout is checked. FULL_MASK keeps unconditional reuse
-        unconditional reuse -- its bound needs a max_context_len that composite
+        because its runtime bound depends on sequence lengths not passed to fits().
-        backends do not carry -- so a batch past max_bs can still overflow it
-        when draft * sum(seq_len) exceeds the buffer, as it could before.
         """
         if self.mode != TreeMaskMode.QLEN_ONLY:
             return True

test/registered/unit/layers/attention/test_verify_mask.py

@@ -119,6 +119,7 @@ def _make_hybrid_backend(speculative_attention_mode, prefill_mask, decode_mask):
         server_args=SimpleNamespace(
             speculative_attention_mode=speculative_attention_mode
         ),
+        model_config=SimpleNamespace(context_len=_MAX_CONTEXT_LEN),
     )
     return HybridAttnBackend(
         model_runner,
@@ -145,8 +146,6 @@ class TestHybridAttnBackendHandsOutSelectedChildMask(CustomTestCase):
         self.assertIs(backend.verify_mask, prefill_mask)
 
     def test_capacity_check_needs_nothing_from_the_backend(self):
-        """A composite backend carries no max_context_len of its own: fits()
-        reaching back through the backend would raise AttributeError here."""
         backend = _make_hybrid_backend("prefill", _mask(64, is_read=False), None)
 
         self.assertTrue(backend.verify_mask.fits(_MAX_BS, _DRAFT))