sgl-project/sglang · error · ValueError

Unsupported latent_log_var: {latent_log_var}

Error message

Unsupported latent_log_var: {latent_log_var}

What it means

The LTX-2.3 condition encoder sizes its conv_out head according to latent_log_var: 'per_channel' doubles the channels, 'uniform'/'constant' add one channel, and 'none' leaves them unchanged. Any other string raises this ValueError in __init__.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vaes/ltx_2_3_condition_encoder.py:168

            )
            block, feature_channels = _make_ltx23_encoder_block(
                block_name=block_name,
                block_config=block_config,
                in_channels=feature_channels,
                spatial_padding_mode=spatial_padding_mode,
            )
            self.down_blocks.append(block)

        self.conv_norm_out = LTX23VideoPixelNorm(dim=1, eps=1e-8)
        self.conv_act = nn.SiLU()

        conv_out_channels = latent_channels
        if latent_log_var == "per_channel":
            conv_out_channels *= 2
        elif latent_log_var in {"uniform", "constant"}:
            conv_out_channels += 1
        elif latent_log_var != "none":
            raise ValueError(f"Unsupported latent_log_var: {latent_log_var}")

        self.conv_out = LTX2VideoCausalConv3d(
            in_channels=feature_channels,
            out_channels=conv_out_channels,
            kernel_size=3,
            stride=1,
            spatial_padding_mode=spatial_padding_mode,
        )

    def forward(self, sample: torch.Tensor) -> torch.Tensor:
        frames_count = int(sample.shape[2])
        if (frames_count - 1) % 8 != 0:
            frames_to_crop = (frames_count - 1) % 8
            sample = sample[:, :, :-frames_to_crop, ...]

        hidden_states = _patchify_video(sample, self.patch_size)
        hidden_states = self.conv_in(hidden_states, causal=True)

View on GitHub (pinned to 0132848349)

Solutions

1. Set latent_log_var to one of 'per_channel', 'uniform', 'constant', or 'none' in the config
2. Upgrade sglang if the checkpoint uses a newly introduced mode
3. If writing a custom config, copy the value verbatim from the official checkpoint's config

Example fix

// before
"latent_log_var": "per-channel"

// after
"latent_log_var": "per_channel"

VALID_LOG_VAR = {'per_channel', 'uniform', 'constant', 'none'}
assert cfg['latent_log_var'] in VALID_LOG_VAR, f"latent_log_var must be one of {VALID_LOG_VAR}"

def is_valid_latent_log_var(v: str) -> bool:
    return v in {'per_channel', 'uniform', 'constant', 'none'}

When it happens

Trigger: Constructing LTX2ConditionEncoder with latent_log_var set to something outside {per_channel, uniform, constant, none} — e.g. 'per-channel', 'disabled', None handled as a string, or a value from an incompatible checkpoint config.

Common situations: Loading a checkpoint whose config records a newer latent_log_var mode not supported by this sglang version; renaming/typo issues in converted configs; mixing diffusers LTX config keys with this implementation.

Related errors

• Unsupported LTX-2.3 encoder block: {block_name}
• Unsupported down_block_type: {down_block_type}
• Unsupported time_compression_ratio: {temporal_compression_ra
• Unsupported up_block_type: {up_block_type}
• Unsupported time_compression_ratio: {time_compression_ratio}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/32c38f47a2608002. Report an issue: GitHub.