A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 02:15 UTC
[Bugfix] Fix LLM priority normalization for single-string prompts (#40011)
Signed-off-by: daiyu1111 <2356690121@qq.com>
daiyu1111 committed
Apr 16, 2026 at 22:56 UTC
a302a8fd1b2033fd3ea3981556af525174b6ff16
2 files changed
+7
-1
tests/entrypoints/llm/test_generate.py
+6
@@ -91,6 +91,12 @@ def test_multiple_priority(llm: LLM):
91
outputs = llm.generate(PROMPTS, sampling_params=None, priority=[])
92
93
94
+def test_single_prompt_priority(llm: LLM):
95
+ # Single string prompts should be normalized to one request.
96
+ outputs = llm.generate(PROMPTS[0], sampling_params=None, priority=[0])
97
+ assert len(outputs) == 1
98
+
99
+
100
def test_max_model_len():
101
max_model_len = 20
102
llm = LLM(
vllm/entrypoints/llm.py
+1
-1
@@ -1575,7 +1575,7 @@ class LLM:
1575
seq_prompts = prompt_to_seq(prompts)
1576
seq_params = self._params_to_seq(params, len(seq_prompts))
1577
seq_lora_requests = self._lora_request_to_seq(lora_request, len(seq_prompts))
1578
- seq_priority = self._priority_to_seq(priority, len(prompts))
1578
+ seq_priority = self._priority_to_seq(priority, len(seq_prompts))
1579
1580
return self._render_and_add_requests(
1581
prompts=(