fix(vllm): tolerate empty 200 body in _response_json (/sleep, /wake_up)

[aoshen02]

Problem

VLLMEngine.release_memory_occupation() (colocate memory offload, via RolloutManager.offload()) does POST /sleep and then _response_json(response). vLLM's /sleep (and /wake_up) return HTTP 200 with an empty body, but _response_json unconditionally calls response.json():

def _response_json(response: requests.Response) -> dict:
    response.raise_for_status()      # 200 OK
    return response.json()           # 💥 empty body -> JSONDecodeError

So colocate offload crashes on the first offload (before rollout #1):

ray::VLLMEngine.release_memory_occupation()
  File ".../slime/backends/vllm_utils/vllm_engine.py", line 749, in release_memory_occupation
    return _response_json(response)
  File ".../slime/backends/vllm_utils/vllm_engine.py", line 41, in _response_json
    return response.json()
requests.exceptions.JSONDecodeError: Expecting value: line 1 column 1 (char 0)

Fix

Treat an empty 200 body as an empty result ({}). This covers /sleep and /wake_up, which carry no JSON payload.

How it was found

A colocate Qwen3-4B GRPO run (DP4 / TP1, --colocate) on the vLLM R3 image (vllm 0.21.1rc1.dev38), as part of a slime(SGLang)-vs-vime(vLLM) convergence A/B. The vime arm crashed at the first release_memory_occupation; the slime arm was unaffected.

Notes / follow-up

• resume_memory_occupation() (/wake_up) hits the same code path and is fixed by the same change.
• Behavior for endpoints that do return JSON is unchanged.
• Follow-up: a unit test for _response_json (empty body → {}, non-empty → parsed, non-200 → raises with response.text note) would lock this in.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the _response_json function in slime/backends/vllm_utils/vllm_engine.py to handle empty HTTP 200 response bodies from vLLM control-plane endpoints, preventing a JSONDecodeError. The reviewer suggested optimizing the check by using response.content.strip() directly on the raw bytes instead of response.text.strip(), which avoids the CPU-intensive character encoding detection overhead.

[gemini-code-assist]

Using response.text can be inefficient because it triggers automatic character encoding detection (which can be slow and CPU-intensive if charset_normalizer or chardet is invoked) and decodes the entire response body into a string.\n\nSince response.content is a bytes object, and bytes supports the .strip() method in Python, we can check if the body is empty or contains only whitespace directly on the raw bytes. This is much more efficient and avoids any encoding detection overhead.

    if not response.content.strip():\n        return {}