diff --git a/python/sglang/srt/layers/radix_attention.py b/python/sglang/srt/layers/radix_attention.py
index 9d6536520..79a324302 100644
--- a/python/sglang/srt/layers/radix_attention.py
+++ b/python/sglang/srt/layers/radix_attention.py
@@ -142,8 +142,11 @@ def unified_attention_with_output(
     ret = forward_batch.attn_backend.forward(
         query, key, value, attention_layer, forward_batch, save_kv_cache
     )
-    assert output.shape == ret.shape
-    output.copy_(ret)
+    assert (
+        output.numel() == ret.numel()
+    ), f"Output tensor element mismatch: {output.numel()} != {ret.numel()}"
+
+    output.view(ret.shape).copy_(ret)
     return
 
 
diff --git a/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py b/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
index e4af95d17..7b36fd9cf 100644
--- a/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
+++ b/python/sglang/srt/model_executor/piecewise_cuda_graph_runner.py
@@ -262,9 +262,14 @@ class PiecewiseCudaGraphRunner:
 
     def can_run(self, forward_batch: ForwardBatch):
         num_tokens = len(forward_batch.input_ids)
-        # TODO(yuwei): support return logprob
+        # TODO(yuwei): support return input_ids' logprob
         if forward_batch.return_logprob:
-            return False
+            for start_len, seq_len in zip(
+                forward_batch.extend_logprob_start_lens_cpu,
+                forward_batch.extend_seq_lens_cpu,
+            ):
+                if start_len is not None and start_len < seq_len:
+                    return False
         if num_tokens <= self.max_num_tokens:
             return True
         return False
@@ -438,7 +443,7 @@ class PiecewiseCudaGraphRunner:
             out_cache_loc=out_cache_loc,
             seq_lens_sum=forward_batch.seq_lens_sum,
             encoder_lens=forward_batch.encoder_lens,
-            return_logprob=forward_batch.return_logprob,
+            return_logprob=False,
             extend_seq_lens=forward_batch.extend_seq_lens,
             extend_prefix_lens=forward_batch.extend_prefix_lens,
             extend_start_loc=forward_batch.extend_start_loc,
diff --git a/test/srt/test_piecewise_cuda_graph.py b/test/srt/test_piecewise_cuda_graph.py
index 4325d9acb..4b0042dbc 100644
--- a/test/srt/test_piecewise_cuda_graph.py
+++ b/test/srt/test_piecewise_cuda_graph.py
@@ -44,6 +44,18 @@ class TestPiecewiseCudaGraphCorrectness(CustomTestCase):
         metrics = run_eval(args)
         self.assertGreaterEqual(metrics["score"], 0.235)
 
+    def test_mmlu(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mmlu",
+            num_examples=64,
+            num_threads=32,
+        )
+
+        metrics = run_eval(args)
+        self.assertGreaterEqual(metrics["score"], 0.65)
+
 
 class TestPiecewiseCudaGraphBenchmark(CustomTestCase):