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 3, 21:30 UTC
[CI] De-flake Language Models Test (Extended Generation) test_models(False-False-5-32-bigcode/starcoder2-3b) (#42392)
Signed-off-by: haosdent <haosdent@gmail.com>
haosdent committed
May 12, 2026 at 18:46 UTC
fc8bf6eedb357f79145cb472d93bd9fc65e7996d
1 file changed
+9
tests/models/language/generation/test_common.py
+9
@@ -143,6 +143,15 @@ def test_models(
143
# in parts of the operators
144
pytest.skip(f"Skipping '{model}' model test with AITER kernel.")
145
146
+ if model == "bigcode/starcoder2-3b":
147
+ # Replace example.txt's Test1 (an NL prompt) with a code prompt:
148
+ # starcoder2-3b is a code model, so NL prompts give near-uniform
149
+ # digit logits where HF<->vLLM bf16 drift can reorder top-K.
150
+ example_prompts = list(example_prompts)
151
+ example_prompts[1] = (
152
+ "def add(a, b):\n return a + b\n\ndef sub(a, b):\n return a - "
153
+ )
154
+
155
with hf_runner(model) as hf_model:
156
hf_outputs = hf_model.generate_greedy_logprobs_limit(
157
example_prompts, max_tokens, num_logprobs