fix(vllm): unify weight-transfer HTTP handling

[aoshen02]

Follow-up to #22 and rebased over #49 / current main. PR #49 merged the R3 routing-replay work and partially centralized the vLLM weight-transfer timeout path, but current main still had two issues in the vLLM control plane:

• Response parsing failures were converted to {"ok": True, "raw": ...} in several endpoints, which can hide bad/plain-text responses as success.
• init_weight_transfer_engine / init_weights_update_group still used a separate timeout path from the rest of the weight-transfer calls.

Why this is not duplicating #49

#49 was a routing replay feature PR. It did touch vllm_engine.py, and it left start_weight_update / finish_weight_update using a timeout helper, but it did not make response parsing fail loud and did not move all weight-transfer control-plane calls to a config-backed timeout. This PR is now scoped to those remaining control-plane fixes.

Changes

• Replace the old fallback parser with a slime-style _response_json(response) helper:
  • raise_for_status() first.
  • Attach response.text to HTTP errors with Exception.add_note, matching the SGLang engine pattern.
  • Parse JSON directly and let invalid JSON fail instead of returning ok: True.
• Add --vllm-weight-transfer-timeout-sec (default 900.0) as the vime config knob for vLLM weight-transfer HTTP control-plane calls.
• Exclude that vime-only config knob from forwarding to vllm serve.
• Route update_weights, start_weight_update, finish_weight_update, init_weight_transfer_engine, and init_weights_update_group through the same timeout helper.
• Remove the old SLIME_VLLM_WEIGHT_TRANSFER_*_TIMEOUT_SEC env-var path from this code.

Tests

uv run --active python -m pytest tests/unit/backends/vllm_utils/test_vllm_engine.py tests/unit/backends/vllm_utils/test_arguments.py -q
# 66 passed, 2 warnings in 85.43s

uv run --active pre-commit run --files slime/backends/vllm_utils/arguments.py slime/backends/vllm_utils/vllm_engine.py tests/unit/backends/vllm_utils/test_arguments.py tests/unit/backends/vllm_utils/test_vllm_engine.py
# all selected hooks passed

AI assistance

AI assistance was used to inspect related PRs/reference implementations and update this PR. The human submitter should review every changed line and validate the behavior before merge.

[gemini-code-assist]

Code Review

This pull request refactors the vLLM engine backend by introducing a helper function _response_json_or_fallback to safely parse JSON responses and handle decoding failures or non-dictionary shapes consistently. It also consolidates the timeout retrieval logic into a helper method _weight_transfer_http_timeout and adds corresponding unit tests. The reviewer recommended making the environment variable parsing within this timeout helper more robust by gracefully handling potential ValueError exceptions and falling back to a default value.

[gemini-code-assist]

Parsing environment variables directly with float() can raise a ValueError if the variable is set to an empty string or a malformed value (e.g., by a user configuration error). It is safer to use defensive programming to handle empty or invalid values gracefully and fall back to the default timeout of 900.0 seconds while logging a warning.

        val = os.environ.get("SLIME_VLLM_WEIGHT_TRANSFER_UPDATE_TIMEOUT_SEC")
        if not val:
            val = os.environ.get("SLIME_VLLM_WEIGHT_TRANSFER_HTTP_TIMEOUT_SEC", "900")
        try:
            return float(val)
        except ValueError:
            logger.warning(
                "Invalid timeout value %r for weight transfer. Falling back to default 900.0s.",
                val,
            )
            return 900.0

[CalvinXKY]

LGTM. Rebase to solve the conflicts

[CalvinXKY]

remove “as SGLangEngine.”

[CalvinXKY]

LGTM