fix(vllm_engine): actually export VLLM_BATCH_INVARIANT=1 for deterministic inference

[aoshen02]

Problem

--vllm-enable-deterministic-inference's help text promises it:

Forwards a per-sample seed … AND exports VLLM_BATCH_INVARIANT=1 to the vLLM subprocess so attention / comm / MM kernels pick batch-invariant variants. Both are required for true determinism — seed alone does not control kernel selection.

But no code ever set VLLM_BATCH_INVARIANT. Grepping the tree, the only occurrence was that help string:

• the flag's only runtime effect is the per-sample seed (vime/rollout/vllm_rollout.py);
• it's an orchestration-only dest (_VIME_ORCHESTRATION_DESTS), so it is not forwarded to vllm serve;
• build_vllm_subprocess_env never set it.

The only thing that ever exported the env was a line in scripts/run-qwen2.5-0.5B-reproducibility.sh's runtime-env-json (added in #39), which #180 dropped when it re-created that script — leaving nothing to set it. So --vllm-enable-deterministic-inference did not actually give batch-invariant kernels, and the reproducibility recipe lost bitwise determinism silently.

Fix

Set VLLM_BATCH_INVARIANT=1 in build_vllm_subprocess_env when the flag is on, so the flag delivers its documented behavior regardless of launch script. Seed + kernel selection now both hold.

if getattr(args, "vllm_enable_deterministic_inference", False):
    env["VLLM_BATCH_INVARIANT"] = "1"

Tests

Two unit tests in tests/unit/backends/vllm_utils/test_vllm_engine.py:

• flag on → VLLM_BATCH_INVARIANT == "1"
• flag off → not set

Both pass (pytest -k "batch_invariant or subprocess_env" → 3 passed).

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request adds logic to set the VLLM_BATCH_INVARIANT environment variable to '1' in the vLLM subprocess environment when vllm_enable_deterministic_inference is enabled. It also adds unit tests to verify this behavior. There are no review comments, so I have no feedback to provide.