do not use cached chunks for prompt_logprobs

PygmalionAI · Dec 16, 2024 · bc1a2bd · bc1a2bd
1 parent d69273b
commit bc1a2bd
Showing 1 changed file with 5 additions and 4 deletions.
diff --git a/aphrodite/processing/scheduler.py b/aphrodite/processing/scheduler.py
@@ -1046,9 +1046,6 @@ def schedule(self) -> Tuple[List[SequenceGroupMetadata], SchedulerOutputs]:
         scheduler_outputs = self._schedule()
         now = time.time()
 
-        if not self.cache_config.enable_prefix_caching:
-            common_computed_block_nums = []
-
         # Create input data structures.
         seq_group_metadata_list: List[SequenceGroupMetadata] = []
         for i, scheduled_seq_group in enumerate(
@@ -1079,10 +1076,14 @@ def schedule(self) -> Tuple[List[SequenceGroupMetadata], SchedulerOutputs]:
                 block_tables[seq_id] = self.block_manager.get_block_table(seq)
                 self.block_manager.access_all_blocks_in_seq(seq, now)
 
-            if self.cache_config.enable_prefix_caching:
+
+            if (self.cache_config.enable_prefix_caching and
+                not seq_group.sampling_params.prompt_logprobs):
                 common_computed_block_nums = (
                     self.block_manager.get_common_computed_block_ids(
                         seq_group.get_seqs(status=SequenceStatus.RUNNING)))
+            else:
+                common_computed_block_nums = []
 
             do_sample = True
             is_prompt = seq_group.is_prefill()