[ethan-scitix]

Type

• fix — bug fix or alignment correction

Summary

• A scaled serving group (<model>@<window>) on SGLang failed at startup when its context.scaling YaRN block named original_max_position_embeddings. With that key, SGLang derives the unscaled window (40960), and _derive_context_length refuses the larger planned --context-length unless SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 is set. --dry-run passed, so the failure only showed after the GPUs were held.
• This PR follows the issue's first suggestion: a scaled group's SGLang engines now launch with that variable. It is defined per backend in SCALING_ENV (sieval/cli/serving.py) and applied in run._serving_groups, which the dry-run also uses, so the dry-run's launch[].env matches what the run sets.
• Trade-off: the variable turns off SGLang's derived-window check for every scaling block, not just the YaRN shape above. So context.max is now the only bound on a scaled window. If max overstates what the scaling reaches, SGLang no longer catches it at startup. docs/guide/infer.md now says this.
• The default group is unchanged. The model's infer.env takes precedence (a user can set the variable to 0). vLLM has no entry: no failure has been observed there, and a test pins this.
• The variable is set on the launch command only. engine_params, infer_plans*.yaml and the inference identity are unchanged, so resume is unaffected. No RESULT_SCHEMA_VERSION bump.

Related Issues

Closes #264

Test Plan

Automated

• Lint/format clean (ruff check sieval tests && ruff format --check sieval tests)
• Type check clean (ty check)
• pytest tests/unit/cli tests/integration -n 8 → 1744 passed
• TestServingGroups::test_only_a_scaled_sglang_engine_may_exceed_its_derived_window has three cases:
  • the scaled SGLang engine gets =1 and the default engine gets nothing;
  • infer.env overrides it;
  • a scaled vLLM engine gets nothing.
• Reverse checks:
  • dropping scaled=True at the call site fails the first case;
  • adding a vllm entry to SCALING_ENV fails the third.

Manual

• Not re-run on GPUs here. The issue reports that the same launch succeeds with this variable set (Qwen3-0.6B, context_length 64096).

Checklist

Required (all PRs)

• PR title follows conventional format (type(scope): description)
• No internal paths, credentials, or personal info in committed files
• AI-generated code has AI-Generated Code - <model> (<provider>) in module docstring
• No new upper-layer dependencies added to core/
• Deleted code verified — no remaining call sites depend on it

🤖 Generated with Claude Code