From 59261df527c2ad92dcfaeeb68f2e23fa9780f4bd Mon Sep 17 00:00:00 2001 From: SimengLiu-nv Date: Tue, 6 Jan 2026 13:34:28 -0800 Subject: [PATCH] Resolve pipeline failures. Signed-off-by: SimengLiu-nv --- tensorrt_llm/executor/result.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tensorrt_llm/executor/result.py b/tensorrt_llm/executor/result.py index aabe0318a3..c02fd09fde 100644 --- a/tensorrt_llm/executor/result.py +++ b/tensorrt_llm/executor/result.py @@ -625,7 +625,7 @@ class DetokenizedGenerationResultBase(GenerationResultBase): def _handle_response(self, response: "GenerationExecutor.Response"): # Save token lengths before processing to detect which outputs received new tokens - prev_token_lens = {o.index: len(o.token_ids) for o in self._outputs} + prev_token_lens = {id(o): len(o.token_ids) for o in self._outputs} GenerationResultBase._handle_response(self, response) @@ -647,7 +647,7 @@ class DetokenizedGenerationResultBase(GenerationResultBase): # to prevent re-decoding the same tokens multiple times output_received_new_tokens = len( beam_output.token_ids) != prev_token_lens.get( - beam_output.index, 0) + id(beam_output), 0) if not output_received_new_tokens: continue if hasattr(