sgl-project/sglang · error · ValueError

Unknown forward method: {forward_method}

Error message

Unknown forward method: {forward_method}

What it means

Raised by SarvamMoE attention forward when the attention layer is dispatched with a forward method name that matches none of the supported branches (e.g. full/MLA-style paths selected by forward_method). The model implementation only implements a fixed set of forward paths, so any other string is a programming or configuration error rather than a runtime data problem.

Source

Thrown at python/sglang/srt/models/sarvam_moe.py:817

            attn_output = self.attn_mqa(
                q_nope_out,
                k_nope,
                k_nope,
                forward_batch,
                q_rope=q_pe,
                k_rope=k_pe,
            )
        elif forward_method == AttnForwardMethod.MLA_CONCAT_ROPE:
            q = torch.cat([q_nope_out, q_pe], dim=-1)
            k = torch.cat([k_nope, k_pe], dim=-1)
            attn_output = self.attn_mqa(
                q,
                k,
                k_nope,
                forward_batch,
            )
        else:
            raise ValueError(f"Unknown forward method: {forward_method}")
        attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)

        attn_bmm_output = self._maybe_fp8_bmm(
            attn_output.transpose(0, 1), self.w_vc, zero_allocator
        )
        attn_bmm_output = attn_bmm_output.transpose(0, 1).flatten(1, 2)

        output, _ = self.o_proj(attn_bmm_output)
        return output

    def forward_prepare(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        forward_batch: ForwardBatch,
        zero_allocator: Optional[BumpAllocator] = None,
        llama_4_scaling: Optional[torch.Tensor] = None,
    ) -> Tuple[Optional[torch.Tensor], ForwardBatch, Optional[Tuple]]:

View on GitHub (pinned to 0132848349)

Solutions

1. Check the server/attention backend flags (--attention-backend) and switch to one supported by sarvam-moe
2. Update sglang to a version where sarvam_moe.py supports the forward method being requested
3. If running a fork, inspect the if/elif chain above line 817 and add a branch for the missing forward_method

Example fix

# before
python -m sglang.launch_server --model sarvam-moe --attention-backend some_new_backend
# after
python -m sglang.launch_server --model sarvam-moe --attention-backend fa3

from sglang.srt.models.sarvam_moe import SUPPORTED  # if exposed; otherwise read source
# before launch: assert chosen attention backend is in the model's supported set
assert attention_backend in {"fa3", "flashinfer", "triton", "torch_native"}, attention_backend

When it happens

Trigger: Calling the SarvamMoE attention module's forward with a forward_method value not recognized by the if/elif chain in python/sglang/srt/models/sarvam_moe.py:817 (e.g. a new backend name passed from model config or a custom patch that renames methods).

Common situations: Using an unsupported attention backend with sarvam-moe, mixing model code versions where a newer scheduler passes a method name the installed model file doesn't know, or typos in custom forks that add a forward method selection string.

Related errors

• Sparse Video Gen 2 attention does not support causal attenti
• {selection_error}{component_suffix}
• Attention backend name must be a string
• Invalid attention backend '{backend}'. Available options are
• component_attention_backends must be a dict or a comma-separ

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/7978a9f71dca6772. Report an issue: GitHub.