sgl-project/sglang · error · ValueError

MiniMax-H3 quality="high" is validated only for the strict 4

Error message

MiniMax-H3 quality="high" is validated only for the strict 4xH200 fl2va deployment; mismatches: {mismatches}

What it means

MiniMax-H3 quality="high" mode is only validated on a strict 4x NVIDIA H200 deployment using the fl2va stack. validate_quality_deployment collects hardware/backend mismatches (device model, compute capability) and raises with a dict of them when the actual environment differs. This guards against silently degraded or numerically wrong high-quality output on unvalidated hardware.

Source

Thrown at python/sglang/multimodal_gen/configs/pipeline_configs/minimax_h3.py:221

            name: {"expected": wanted, "actual": actual[name]}
            for name, wanted in expected.items()
            if (
                actual[name] not in wanted
                if isinstance(wanted, set)
                else actual[name] != wanted
            )
        }
        if (
            not current_platform.is_cuda()
            or "H200" not in device_name.upper()
            or capability_int != 90
        ):
            mismatches["device"] = {
                "expected": "NVIDIA H200 (compute capability 9.0)",
                "actual": f"{device_name} (compute capability {capability_int})",
            }
        if mismatches:
            raise ValueError(
                'MiniMax-H3 quality="high" is validated only for '
                f"the strict 4xH200 fl2va deployment; mismatches: {mismatches}"
            )

    def validate_server_args(self, server_args) -> None:
        # Reject known-inexact VAE modes before any large component download.
        self.vae_config.resolved_parallel_decode_mode()
        if current_platform.is_mps():
            required_components = (
                "transformer",
                "text_encoder",
                "video_vae",
                "audio_vae",
            )
            missing_components = [
                component
                for component in required_components
                if server_args.residency_mode(component) != LAYERWISE_OFFLOAD

View on GitHub (pinned to 0132848349)

Solutions

1. Use the default (non-high) quality setting on non-H200 hardware
2. Run on a strict 4x H200 deployment with the fl2va configuration if quality="high" is required
3. Check the mismatches dict in the message to see exactly which expectation (device, capability) failed and address that specific dimension

Example fix

# before
config = MiniMaxH3Config(quality="high")  # on 8xA100

# after
config = MiniMaxH3Config(quality="default")  # validated on this hardware

import torch
if quality == "high":
    ok = torch.cuda.device_count() == 4 and all(torch.cuda.get_device_capability(i)[0] == 9 for i in range(4))
    assert ok, "quality='high' requires strict 4xH200"

When it happens

Trigger: Setting quality="high" on MiniMax-H3 and running on anything other than 4x H200 GPUs (e.g. A100s, H100s, different GPU count, or non-fl2va backend) — mismatches are reported in the error dict.

Common situations: Trying quality="high" on available cluster hardware (A100/H100), CI/dev machines, or after changing attention/VAE backends away from fl2va; also triggered deliberately by a regression test rejecting transformer weight overrides.

Related errors

• MiniMax H3 SGLang backend only supports output_mode='decoded
• MiniMax H3 does not support enable_frame_interpolation: the
• MiniMax H3 does not support enable_upscaling: the accepted d
• MiniMax H3 video generation produced {len(output_paths)} out
• Video generation failed: {error_msg}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/793778300a09e632. Report an issue: GitHub.