This commit is contained in:
Tong Li
2025-02-28 10:16:42 +08:00
parent f736d747e3
commit 070907dd7f
6 changed files with 74 additions and 26 deletions

View File

@@ -51,13 +51,17 @@ if __name__ == "__main__":
elif args.backend == "vllm":
inference_model_config.update(
dict(
gpu_memory_utilization=0.6,
gpu_memory_utilization=0.7,
)
)
generate_config.update(
dict(
max_tokens=256,
max_tokens=2048,
ignore_eos=True,
include_stop_str_in_output=True,
stop=["</answer>"],
temperature=0.2,
top_p=0.95,
)
)
else: