Fix TokenizerManager crash on top_logprobs with tensor values (#26825)

python/sglang/srt/managers/tokenizer_manager.py

@@ -2188,7 +2188,7 @@ class TokenizerManager(TokenizerControlMixin, TokenizerManagerScoreMixin):
         # We should batch all top-k tokens in all positions.
         ret = []
         for i in range(len(token_logprobs_val)):
-            if token_logprobs_val[i]:
+            if token_logprobs_val[i] is not None:
                 ret.append(
                     self.detokenize_logprob_tokens(
                         token_logprobs_val[i], token_logprobs_idx[i], decode_to_text

test/registered/unit/managers/test_tokenizer_manager_top_logprobs_tensor.py

@@ -0,0 +1,82 @@
+"""
+Unit tests for TokenizerManager.detokenize_top_logprobs_tokens.
+
+Regression coverage for a crash where a per-position logprob value is a
+multi-element torch.Tensor instead of the annotated List[float]. The old
+truthiness check ``if token_logprobs_val[i]:`` raises::
+
+    RuntimeError: Boolean value of Tensor with more than one value is ambiguous.
+
+which propagates out of the detokenization handler, gets caught by
+print_exception_wrapper, and SIGKILLs the whole (prefill) process. The fix
+uses an explicit ``is not None`` test that matches the actual sentinel for
+skipped positions and works for both lists and tensors.
+"""
+
+import unittest
+
+import torch
+
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import CustomTestCase, maybe_stub_sgl_kernel
+
+maybe_stub_sgl_kernel()
+
+from sglang.srt.managers.tokenizer_manager import TokenizerManager
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+def _make_tokenizer_manager() -> TokenizerManager:
+    """Create a bare TokenizerManager, bypassing __init__.
+
+    detokenize_top_logprobs_tokens / detokenize_logprob_tokens only touch
+    self.tokenizer when decode_to_text=True, so an uninitialised instance is
+    sufficient for the decode_to_text=False paths exercised here.
+    """
+    return TokenizerManager.__new__(TokenizerManager)
+
+
+class TestDetokenizeTopLogprobsTensor(CustomTestCase):
+    def test_multi_element_tensor_value_does_not_crash(self):
+        """A multi-element tensor position must be detokenized, not raise."""
+        tm = _make_tokenizer_manager()
+
+        val = [torch.tensor([-0.1, -0.2, -0.3])]
+        idx = [[10, 20, 30]]
+
+        ret = tm.detokenize_top_logprobs_tokens(val, idx, decode_to_text=False)
+
+        self.assertEqual(len(ret), 1)
+        self.assertEqual(
+            ret[0],
+            [(-0.1, 10, None), (-0.2, 20, None), (-0.3, 30, None)],
+        )
+
+    def test_none_position_yields_none(self):
+        """None is the sentinel for skipped positions and must stay None."""
+        tm = _make_tokenizer_manager()
+
+        val = [None, torch.tensor([-0.5, -0.6])]
+        idx = [None, [1, 2]]
+
+        ret = tm.detokenize_top_logprobs_tokens(val, idx, decode_to_text=False)
+
+        self.assertEqual(len(ret), 2)
+        self.assertIsNone(ret[0])
+        self.assertEqual(ret[1], [(-0.5, 1, None), (-0.6, 2, None)])
+
+    def test_plain_list_values_still_work(self):
+        """The ordinary List[float] path is unaffected by the fix."""
+        tm = _make_tokenizer_manager()
+
+        val = [[-0.1, -0.2]]
+        idx = [[7, 8]]
+
+        ret = tm.detokenize_top_logprobs_tokens(val, idx, decode_to_text=False)
+
+        self.assertEqual(ret, [[(-0.1, 7, None), (-0.2, 8, None)]])
+
+
+if __name__ == "__main__":
+    unittest.main(verbosity=2)