[Fix] missing max_context_len on HybridAttnBackend (#32690)

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

@@ -42,6 +42,7 @@ class HybridAttnBackend(AttentionBackend):
         self.needs_cpu_seq_lens = (
             prefill_backend.needs_cpu_seq_lens or decode_backend.needs_cpu_seq_lens
         )
+        self.max_context_len = model_runner.model_config.context_len
 
     def _select_backend(self, forward_mode: ForwardMode) -> AttentionBackend:
         """

test/registered/attention/test_trtllm_mha_graph_metadata.py

@@ -146,6 +146,7 @@ def test_hybrid_wrappers_forward_in_graph_hook():
             token_to_kv_pool=None,
             req_to_token_pool=None,
             server_args=SimpleNamespace(speculative_attention_mode="decode"),
+            model_config=SimpleNamespace(context_len=2048),
         ),
         prefill_backend=make_fake("prefill", calls),
         decode_backend=make_fake("decode", calls),

test/registered/unit/model_executor/model_runner_components/test_attention_backend_setup.py

@@ -26,6 +26,7 @@ class _FakeBackend:
 def test_split_full_attention_applies_model_wrapper_once():
     runner = SimpleNamespace(
         server_args=SimpleNamespace(speculative_attention_mode="prefill"),
+        model_config=SimpleNamespace(context_len=2048),
         kv_cache_dtype=None,
         token_to_kv_pool=object(),
         req_to_token_pool=object(),

test/registered/unit/spec/test_dflash_overlap_hostsync.py

@@ -231,6 +231,7 @@ class TestHybridNeedsCpuSeqLens(CustomTestCase):
             kv_cache_dtype=torch.bfloat16,
             token_to_kv_pool=None,
             req_to_token_pool=None,
+            model_config=SimpleNamespace(context_len=2048),
         )
         return HybridAttnBackend(runner, backend(prefill_flag), backend(decode_flag))