From 6289cbc43475c467227d981327f542b70aaa0caa Mon Sep 17 00:00:00 2001
From: chen <103103266+ckl117@users.noreply.github.com>
Date: Thu, 11 Dec 2025 10:46:27 +0800
Subject: [PATCH] [BugFix] fix hung when n>1 and --enable-logprob (#5492)

---
 fastdeploy/worker/gpu_model_runner.py | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
index a7ee4be8f..ee64fb247 100644
--- a/fastdeploy/worker/gpu_model_runner.py
+++ b/fastdeploy/worker/gpu_model_runner.py
@@ -2396,6 +2396,19 @@ class GPUModelRunner(ModelRunnerBase):
                     self.sampling_metadata,
                     p_done_idxs,
                 )
+
+                if (
+                    self.enable_logprob
+                    and not envs.FD_USE_GET_SAVE_OUTPUT_V1
+                    and sampler_output.logprobs_tensors is None
+                ):
+                    sampler_output.logprobs_tensors = LogprobsTensors(
+                        logprob_token_ids=sampler_output.sampled_token_ids,
+                        logprobs=paddle.empty_like(sampler_output.sampled_token_ids, device="cpu", dtype="float32"),
+                        selected_token_ranks=paddle.empty(
+                            [sampler_output.sampled_token_ids.shape[0]], device="cpu", dtype="int64"
+                        ),
+                    )
                 if self.parallel_config.tensor_parallel_size > 1:
                     paddle.distributed.broadcast(
                         sampler_output.sampled_token_ids,