[BugFix] preserve cached token details in multi-tokenizer output (#26590)

python/sglang/srt/managers/multi_tokenizer_mixin.py

@@ -246,6 +246,9 @@ def _handle_output_by_index(output, i):
             completion_tokens=_extract_field_by_index(output, "completion_tokens", i),
             reasoning_tokens=_extract_field_by_index(output, "reasoning_tokens", i),
             cached_tokens=_extract_field_by_index(output, "cached_tokens", i),
+            cached_tokens_details=_extract_field_by_index(
+                output, "cached_tokens_details", i
+            ),
             input_token_logprobs_val=_extract_field_by_index(
                 output, "input_token_logprobs_val", i, check_length=False
             ),

test/registered/unit/managers/test_multi_tokenizer_mixin.py

@@ -0,0 +1,68 @@
+import unittest
+
+from sglang.test.ci.ci_register import register_cpu_ci
+from sglang.test.test_utils import maybe_stub_sgl_kernel
+
+maybe_stub_sgl_kernel()
+
+from sglang.srt.managers.io_struct import BatchStrOutput
+from sglang.srt.managers.multi_tokenizer_mixin import _handle_output_by_index
+
+register_cpu_ci(est_time=5, suite="base-a-test-cpu")
+
+
+def _make_batch_str_output() -> BatchStrOutput:
+    return BatchStrOutput(
+        rids=["rid-0", "rid-1"],
+        spec_verify_ct=[0, 0],
+        spec_num_correct_drafts=[0, 0],
+        spec_correct_drafts_histogram=[[], []],
+        finished_reasons=[None, {"type": "length"}],
+        output_strs=["first", "second"],
+        output_ids=[[1], [2]],
+        prompt_tokens=[10, 20],
+        completion_tokens=[1, 2],
+        reasoning_tokens=[0, 0],
+        cached_tokens=[3, 4],
+        cached_tokens_details=[
+            {"device": 3, "host": 0},
+            {"device": 1, "host": 3},
+        ],
+        input_token_logprobs_val=[[], []],
+        input_token_logprobs_idx=[[], []],
+        output_token_logprobs_val=[[], []],
+        output_token_logprobs_idx=[[], []],
+        input_top_logprobs_val=[[], []],
+        input_top_logprobs_idx=[[], []],
+        output_top_logprobs_val=[[], []],
+        output_top_logprobs_idx=[[], []],
+        input_token_ids_logprobs_val=[[], []],
+        input_token_ids_logprobs_idx=[[], []],
+        output_token_ids_logprobs_val=[[], []],
+        output_token_ids_logprobs_idx=[[], []],
+        output_token_entropy_val=[0.0, 0.0],
+        output_hidden_states=[None, None],
+        routed_experts=[None, None],
+        indexer_topk=[None, None],
+        placeholder_tokens_idx=[None, None],
+        placeholder_tokens_val=[None, None],
+        retraction_counts=[0, 0],
+    )
+
+
+class TestMultiTokenizerMixin(unittest.TestCase):
+    def test_batch_str_output_preserves_cached_tokens_details(self):
+        output = _make_batch_str_output()
+
+        single_output = _handle_output_by_index(output, 1)
+
+        self.assertEqual(single_output.rids, ["rid-1"])
+        self.assertEqual(single_output.cached_tokens, [4])
+        self.assertEqual(
+            single_output.cached_tokens_details,
+            [{"device": 1, "host": 3}],
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()