Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.
• 4. If the issue you raised is not a bug but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 5. Please use English, otherwise it will be closed.

Describe the bug

As described in the title. I am wondering if users are required to do an explicit weight update before generating. Is the following expected behavior? Thank you!

Reproduction

import torch
import sglang as sgl

def main_sglang():
    server_args = {
        "model_path": "Qwen/Qwen3-0.6B",
        "disable_cuda_graph": True,  # does not affect behavior
        "tp_size": 1,
        "enable_memory_saver": True,
    }
    llm = sgl.Engine(**server_args)

    print(f"Free GPU memory before sleep: {torch.cuda.mem_get_info()[0] / 1024**2:.1f} MB")
    llm.release_memory_occupation()
    print(f"Free GPU memory after sleep: {torch.cuda.mem_get_info()[0] / 1024**2:.1f} MB")
    llm.resume_memory_occupation()
    print(f"Free GPU memory after wake up: {torch.cuda.mem_get_info()[0] / 1024**2:.1f} MB")

    outputs = llm.generate("Hello, my name is", {"max_new_tokens": 10})
    print(outputs['text'])

if __name__ == "__main__":
    main_sglang()

Output:

Loading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  4.48it/s]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  4.48it/s]

Free GPU memory before sleep: 8338.2 MB
Free GPU memory after sleep: 38830.2 MB
Free GPU memory after wake up: 8318.2 MB
هذهn屋쨌养老保险 admins Aerospace.lex العسكري والتي

Environment

Environment:

• A100 40GB
• Tried SGLang build-from-source at head (766392c), pip-installed 0.4.9.post1, 0.4.8.post1