sgl-project/sglang · error · ValueError

SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC is no longer supporte

Error message

SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC is no longer supported. Use --expert-balancedness-report-mode with one of: off, server_log, prometheus, both.

What it means

Raised during expert-metrics resolution when the legacy environment variable SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC is set. The env-var toggle was replaced by the --expert-balancedness-report-mode flag (off | server_log | prometheus | both); the resolver hard-fails so users notice the migration instead of silently losing metrics.

Source

Thrown at python/sglang/srt/server_args.py:8105

            )

    def _validate_experimental_sgl_marlin(self):
        view = self._resolved()
        if view.moe_runner_backend != "experimental_sgl_marlin":
            return

        # ===== TO BE REFACTORED ====
        from sglang.srt.lora.marlin_lora_temp.policy import (
            validate_experimental_sgl_marlin_server_args,
        )

        validate_experimental_sgl_marlin_server_args(self, view)
        # ===== END TO BE REFACTORED ====

    def _handle_expert_distribution_metrics(self):
        cfg = resolving_view(self)
        if "SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC" in os.environ:
            raise ValueError(
                "SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC is no longer supported. Use "
                "--expert-balancedness-report-mode with one of: off, server_log, "
                "prometheus, both."
            )

        if self.should_report_expert_balancedness() and (
            cfg.expert_distribution_recorder_mode is None
        ):
            self._declare(
                "_handle_expert_distribution_metrics",
                expert_distribution_recorder_mode="stat",
            )

        if cfg.expert_distribution_recorder_buffer_size is None:
            if (x := cfg.eplb_rebalance_num_iterations) is not None:
                self._declare(
                    "_handle_expert_distribution_metrics",
                    expert_distribution_recorder_buffer_size=x,

View on GitHub (pinned to 0132848349)

Solutions

1. Unset the variable: unset SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC (or remove it from the container/unit spec)
2. Replace it with --expert-balancedness-report-mode server_log|prometheus|both as needed

Example fix

# before
export SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC=1
python -m sglang.launch_server ...
# after
python -m sglang.launch_server ... --expert-balancedness-report-mode both

import os
os.environ.pop("SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC", None)
if want_balancedness:
    cmd.append("--expert-balancedness-report-mode=both")

When it happens

Trigger: Having SGLANG_ENABLE_EPLB_BALANCEDNESS_METRIC exported (to any value, including 0/empty) in the environment of the server process.

Common situations: Old deployment scripts, Docker images, or Helm charts that export the legacy env var; CI environments carrying stale variables after upgrading SGLang.

Related errors

• --ep-dispatch-algorithm {cfg.ep_dispatch_algorithm} picks a
• SGLANG_LOG_SCHEDULER_STATUS_TARGET is set but --enable-metri
• unknown q-prep variant {variant!r} (SGLANG_OPT_Q8KV8_QPREP_V
• MiniMax-H3 on MPS requires synchronous layerwise offload for
• MiniMax-H3 MPS execution does not support torch.compile; pas

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/49aa03a6fd747fd6. Report an issue: GitHub.