huggingface/candle · error

only TorchAttn is supported

Error message

only TorchAttn is supported

What it means

SelfAttention::new in the metavoice model only implements the TorchAttn attention kernel; any other attn_kernel_type in the model Config causes this bail. The comment notes the other variants (like FusedAttn) are probably identical but are intentionally rejected until verified. This is an explicit unsupported-feature guard.

Source

Thrown at candle-transformers/src/models/metavoice.rs:446

                Self::LayerNorm(m) => m.forward(xs),
            }
        }
    }

    // https://github.com/metavoiceio/metavoice-src/blob/11550bb4e8a1ad032cc1556cc924f7a4e767cbfa/fam/llm/layers/attn.py#L18
    struct SelfAttention {
        c_attn: Linear,
        c_proj: Linear,
        n_head: usize,
        span: tracing::Span,
    }

    impl SelfAttention {
        fn new(cfg: &Config, vb: VarBuilder) -> Result<Self> {
            // The different attention variants are likely to be identical but still we only accept
            // TorchAttn for now.
            if cfg.attn_kernel_type != AttnKernelType::TorchAttn {
                candle::bail!("only TorchAttn is supported")
            }
            if cfg.kv_cache_enabled {
                candle::bail!("kv_cache_enabled=true is not supported")
            }
            let c_attn = linear_b(cfg.n_embd, cfg.n_embd * 3, cfg.bias, vb.pp("c_attn"))?;
            let c_proj = linear_b(cfg.n_embd, cfg.n_embd, cfg.bias, vb.pp("c_proj"))?;
            Ok(Self {
                c_attn,
                c_proj,
                n_head: cfg.n_head,
                span: tracing::span!(tracing::Level::TRACE, "self-attn"),
            })
        }
    }

    impl Module for SelfAttention {
        fn forward(&self, xs: &Tensor) -> Result<Tensor> {
            let _enter = self.span.enter();

View on GitHub (pinned to d5fee525bf)

Solutions

1. Set attn_kernel_type to AttnKernelType::TorchAttn in the Config (or in the source config file) before building the model
2. Use a checkpoint/config that was exported with the TorchAttn kernel
3. Patch SelfAttention::new to implement (or verify) the other kernel if you truly need it

Example fix

// before
let cfg = Config { attn_kernel_type: AttnKernelType::FusedAttn, .. };
// after
let cfg = Config { attn_kernel_type: AttnKernelType::TorchAttn, .. };

if cfg.attn_kernel_type != AttnKernelType::TorchAttn {
    return Err("metavoice in candle requires attn_kernel_type = TorchAttn");
}

let model = Metavoice::new(&cfg, vb)
    .map_err(|e| format!("unsupported attention kernel: {e}"))?;

When it happens

Trigger: Constructing the MetaVoice self-attention layer with a Config parsed from a checkpoint whose config.json sets attn_kernel_type to anything other than "TORCH_ATTN" (e.g. FAVA, or another kernel variant).

Common situations: Loading a MetaVoice variant exported with different attention kernels; a converted config that preserves the original kernel type; hand-writing a Config with a non-default attn_kernel_type.

Related errors

• kv_cache_enabled=true is not supported
• swiglu-multiple-of has to be set
• sliding window is not supported
• Unsupported activation function: {}
• Unsupported projector activation: {}

AI-assisted analysis of huggingface/candle@d5fee525bf (2026-09-02). Data as JSON: /api/errors/2185ad43fc3a4688. Report an issue: GitHub.