sgl-project/sglang · error · ValueError

Unknown Pi05 Gemma variant: {variant}

Error message

Unknown Pi05 Gemma variant: {variant}

What it means

The Pi05 (pi05_core.py) model factory maps a Gemma variant name string to a hardcoded transformer config (depth, mlp_dim, num_heads, num_kv_heads, head_dim). If the supplied variant string matches none of the known keys, construction fails with 'Unknown Pi05 Gemma variant'. This is a config-name validation error raised during model __init__, before any weights load.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vlas/pi05_core.py:823

    if variant == "gemma_300m":
        return GemmaVariantConfig(
            width=1024,
            depth=18,
            mlp_dim=4096,
            num_heads=8,
            num_kv_heads=1,
            head_dim=256,
        )
    if variant == "gemma_2b":
        return GemmaVariantConfig(
            width=2048,
            depth=18,
            mlp_dim=16_384,
            num_heads=8,
            num_kv_heads=1,
            head_dim=256,
        )
    raise ValueError(f"Unknown Pi05 Gemma variant: {variant}")


def create_sinusoidal_pos_embedding(
    time: torch.Tensor,
    dimension: int,
    min_period: float,
    max_period: float,
) -> Tensor:
    if dimension % 2 != 0:
        raise ValueError(f"dimension ({dimension}) must be divisible by 2")
    if time.ndim != 1:
        raise ValueError("time must have shape [batch]")
    fraction = torch.linspace(
        0.0,
        1.0,
        dimension // 2,
        dtype=torch.float64,
        device=time.device,

View on GitHub (pinned to 0132848349)

Solutions

1. Check the error message for the exact variant string and compare it against the keys accepted in get_gemma_variant_config (pi05_core.py around line 790-823)
2. Fix the typo/rename in your config to a supported variant name
3. If the variant is legitimately new, add a config entry to the mapping in get_gemma_variant_config and rebuild
4. Upgrade sglang to a version that supports the variant

Example fix

// before
model = Pi05Model(config=Pi05Config(gemma_variant="gemma_3"))

// after
model = Pi05Model(config=Pi05Config(gemma_variant="gemma3"))

import inspect
from sglang.multimodal_gen.runtime.models.vlas import pi05_core

src = inspect.getsource(pi05_core.get_gemma_variant_config)
# fast structural check: any string literal compared to `variant`
import re
supported = set(re.findall(r"variant\s*==\s*[\"']([\w-]+)[\"']", src))
assert my_config.gemma_variant in supported, f"use one of {supported}"

def is_supported_gemma_variant(v: str) -> bool:
    import re, inspect
    from sglang.multimodal_gen.runtime.models.vlas import pi05_core
    src = inspect.getsource(pi05_core.get_gemma_variant_config)
    return v in set(re.findall(r'variant\s*==\s*["\']([\w-]+)["\']', src))

try:
    model = Pi05Model(cfg)
except ValueError as e:
    if "Unknown Pi05 Gemma variant" in str(e):
        raise SystemExit(f"Bad gemma_variant in config: {cfg.gemma_variant}. Check supported variants.") from e
    raise

When it happens

Trigger: Calling the Pi05 model constructor (directly or via a config object) with a variant key not in the supported set, e.g. a typo like 'gemma_3' instead of 'gemma3', or a new/renamed variant string from a checkpoint config that this code version does not know.

Common situations: Upgrading/downgrading sglang versions where variant names changed; hand-editing a model config JSON; loading a community-finetuned Pi05 checkpoint that names its language model variant differently.

Understand the failure class

Background: "Invalid value" and "allowed values are" config errors: what your library rejected and how to fix it — this error's family across 41 libraries.

Related errors

• Invalid Pi05 precision: {precision}
• Unsupported Pi05 dtype: {dtype_name}
• VLA action expert should not share the prefix TP layout. Use
• This browser cannot encode H.264 MP4
• H.264 encoder did not return MP4 decoder config

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d407dd69f8fe6f68. Report an issue: GitHub.