[Bugfix] Handle empty FlatLogprobs slices and delta output (#53704)

Signed-off-by: tony <864832769@qq.com>

tests/test_logprobs.py

@@ -208,3 +208,10 @@ def test_flat_logprobs_access() -> None:
     assert logprobs_last2.logprobs == [0.4, 0.5, 0.6, 0.1]
     assert logprobs_last2.ranks == [40, 50, 60, 10]
     assert logprobs_last2.decoded_tokens == ["40", "50", "60", "10"]
+
+    for empty_slice in (slice(0, 0), slice(3, 3), slice(10, 10)):
+        empty = logprobs[empty_slice]
+        assert isinstance(empty, FlatLogprobs)
+        assert len(empty) == 0
+        assert empty.start_indices == []
+        assert empty.end_indices == []

tests/v1/engine/test_output_processor.py

@@ -3,6 +3,7 @@
 
 import math
 import time
+from unittest.mock import MagicMock
 
 import pytest
 
@@ -14,7 +15,7 @@ from tests.v1.engine.utils import (
     MockEngineCore,
 )
 from vllm import PoolingParams
-from vllm.logprobs import PromptLogprobs, SampleLogprobs
+from vllm.logprobs import FlatLogprobs, Logprob, PromptLogprobs, SampleLogprobs
 from vllm.lora.request import LoRARequest
 from vllm.outputs import CompletionOutput, RequestOutput
 from vllm.sampling_params import RequestOutputKind, SamplingParams
@@ -26,10 +27,39 @@ from vllm.v1.engine import (
     EngineCoreRequest,
     FinishReason,
 )
-from vllm.v1.engine.output_processor import OutputProcessor, RequestOutputCollector
+from vllm.v1.engine.output_processor import (
+    OutputProcessor,
+    RequestOutputCollector,
+    RequestState,
+)
 from vllm.v1.metrics.stats import IterationStats, SchedulerStats
 
 
+@pytest.mark.parametrize("flat_logprobs", [False, True])
+def test_delta_output_without_new_tokens_returns_empty_logprobs(
+    flat_logprobs: bool,
+) -> None:
+    accumulated = FlatLogprobs() if flat_logprobs else []
+    accumulated.append({1: Logprob(logprob=-0.5, rank=1)})
+
+    state = RequestState.__new__(RequestState)
+    state.detokenizer = MagicMock()
+    state.detokenizer.get_next_output_text.return_value = ""
+    state.logprobs_processor = MagicMock()
+    state.logprobs_processor.logprobs = accumulated
+    state.logprobs_processor.cumulative_logprob = -0.5
+    state.output_kind = RequestOutputKind.DELTA
+    state.request_index = 0
+    state.sampling_mask_chunks = []
+    state.routed_experts_chunks = []
+    state.spec_decode_metrics = None
+
+    output = state._new_completion_output([], None, None)
+
+    assert isinstance(output.logprobs, FlatLogprobs if flat_logprobs else list)
+    assert len(output.logprobs) == 0
+
+
 def _ref_convert_id_to_token(
     tokenizer: TokenizerLike,
     token_id: int,

vllm/logprobs.py

@@ -119,13 +119,18 @@ class FlatLogprobs(MutableSequence[LogprobsOnePosition | None]):
                 for i in range(self.start_indices[index], self.end_indices[index])
             }
         elif isinstance(index, slice):
-            min_index = self.start_indices[index][0]
-            max_index = self.end_indices[index][-1]
+            selected_starts = self.start_indices[index]
+            selected_ends = self.end_indices[index]
+            # Empty slices have no source offset to normalize.
+            if not selected_starts:
+                return FlatLogprobs()
+            min_index = selected_starts[0]
+            max_index = selected_ends[-1]
             return FlatLogprobs(
                 # Shift updated start_indices and end_indices to
                 # be 0-indexed
-                start_indices=[i - min_index for i in self.start_indices[index]],
-                end_indices=[i - min_index for i in self.end_indices[index]],
+                start_indices=[i - min_index for i in selected_starts],
+                end_indices=[i - min_index for i in selected_ends],
                 token_ids=self.token_ids[min_index:max_index],
                 logprobs=self.logprobs[min_index:max_index],
                 ranks=self.ranks[min_index:max_index],

vllm/v1/engine/output_processor.py

@@ -411,7 +411,11 @@ class RequestState:
         # Prepare logprobs, based on delta mode
         logprobs = self.logprobs_processor.logprobs
         if delta and logprobs:
-            logprobs = logprobs[-len(token_ids) :]
+            num_new_tokens = len(token_ids)
+            # Avoid [-0:], which returns the full accumulated history when a
+            # delta contains no new token IDs. [:0] preserves the concrete
+            # list or FlatLogprobs representation while returning no entries.
+            logprobs = logprobs[-num_new_tokens:] if num_new_tokens else logprobs[:0]
 
         sampling_mask = None
         if finished and self.sampling_mask_chunks: