sgl-project/sglang · error · NotImplementedError

{beta_schedule} is not implemented for {self.__class__}

Error message

{beta_schedule} is not implemented for {self.__class__}

What it means

Raised in UniPCMultistepScheduler.__init__ when the beta_schedule config value is not one of the supported schedules ('linear', 'scaled_linear', 'squaredcos_cap_v2'). The scheduler maps the schedule name to a betas array; unknown names cannot be mapped, so construction fails.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/schedulers/scheduling_unipc_multistep.py:280

            self.betas = torch.linspace(
                beta_start, beta_end, num_train_timesteps, dtype=torch.float32
            )
        elif beta_schedule == "scaled_linear":
            # this schedule is very specific to the latent diffusion model.
            self.betas = (
                torch.linspace(
                    beta_start**0.5,
                    beta_end**0.5,
                    num_train_timesteps,
                    dtype=torch.float32,
                )
                ** 2
            )
        elif beta_schedule == "squaredcos_cap_v2":
            # Glide cosine schedule
            self.betas = betas_for_alpha_bar(num_train_timesteps)
        else:
            raise NotImplementedError(
                f"{beta_schedule} is not implemented for {self.__class__}"
            )

        if rescale_betas_zero_snr:
            self.betas = rescale_zero_terminal_snr(self.betas)

        self.alphas = 1.0 - self.betas
        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)

        if rescale_betas_zero_snr:
            # Close to 0 without being 0 so first sigma is not inf
            # FP16 smallest positive subnormal works well here
            self.alphas_cumprod[-1] = 2**-24

        # Currently we only support VP-type noise schedule
        self.alpha_t = torch.sqrt(self.alphas_cumprod)
        self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
        self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)

View on GitHub (pinned to 0132848349)

Solutions

1. Set beta_schedule to 'scaled_linear' (the default) or 'linear' or 'squaredcos_cap_v2'
2. Check the saved scheduler_config.json of the model you are loading for the exact original value
3. If you need a cosine schedule, use 'squaredcos_cap_v2' which is the Glide cosine schedule

Example fix

// before
sched = UniPCMultistepScheduler.from_config(cfg, beta_schedule="cosine")
// after
sched = UniPCMultistepScheduler.from_config(cfg, beta_schedule="squaredcos_cap_v2")

from sglang.multimodal_gen.runtime.models.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
allowed = {"linear", "scaled_linear", "squaredcos_cap_v2"}
assert cfg["beta_schedule"] in allowed, f"beta_schedule must be one of {allowed}"

When it happens

Trigger: Instantiating UniPCMultistepScheduler(from_config(...)) or from_pretrained with config beta_schedule set to a typo or unsupported value like 'cosine' (instead of 'squaredcos_cap_v2') or 'linear_beta'.

Common situations: Hand-written scheduler configs copied from other libraries (e.g. 'cosine' used in k-diffusion/other schedulers), typos, or config JSONs migrated from a diffusers version with different schedule names.

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

• {solver_type} is not implemented for {self.__class__}
• {self.config.timestep_spacing} is not supported. Please make
• `final_sigmas_type` must be one of 'zero', or 'sigma_min', b
• missing `sample` as a required keyword argument
• prediction_type given as {self.config.prediction_type} must

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/1a77fa15c90aeedf. Report an issue: GitHub.