sgl-project/sglang · error · ValueError

DFLASH requires explicit layer ids for aux hidden capture.

Error message

DFLASH requires explicit layer ids for aux hidden capture.

What it means

DFLASH aux-hidden capture requires the caller to pass an explicit list of transformer layer indices. The model refuses to enable capture_aux_hidden_states when layers_to_capture is None because DFLASH cannot infer which layers' hidden states to capture.

Source

Thrown at python/sglang/srt/models/qwen3_5_text.py:125

    def get_input_embeddings(self) -> nn.Embedding:
        return self.model.embed_tokens

    def get_embed_and_head(self):
        return self.model.embed_tokens.weight, self.lm_head.weight

    def set_embed_and_head(self, embed, head):
        del self.model.embed_tokens.weight
        del self.lm_head.weight
        self.model.embed_tokens.weight = embed
        self.lm_head.weight = head
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

    def set_dflash_layers_to_capture(self, layers_to_capture: list[int]):
        if not self.pp_group.is_last_rank:
            return
        if layers_to_capture is None:
            raise ValueError(
                "DFLASH requires explicit layer ids for aux hidden capture."
            )
        self.capture_aux_hidden_states = True
        self.model.set_dflash_layers_to_capture(layers_to_capture)

    @torch.no_grad()
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        forward_batch: ForwardBatch,
        input_embeds: Optional[torch.Tensor] = None,
        pp_proxy_tensors: Optional[PPProxyTensors] = None,
        **kwargs,
    ) -> Union[torch.Tensor, PPProxyTensors]:
        if self.is_mrope_enabled:
            positions = forward_batch.mrope_positions

View on GitHub (pinned to 0132848349)

Solutions

1. Set the explicit layer ids flag, e.g. --dflash-capture-layers 0,2,4 or the corresponding speculative-decoding config field
2. If invoking programmatically, pass a concrete list[int] instead of None
3. Check that the caller upstream (e.g. EAGLE/DFLASH worker) is configured to compute layer indices before calling

Example fix

# before
model.set_dflash_layers_to_capture(None)
# after
model.set_dflash_layers_to_capture([0, 2, 4, 6])

layers = spec_config.dflash_capture_layers
assert layers is not None and len(layers) > 0, "set dflash capture layers"

def has_capture_layers(cfg) -> bool:
    return isinstance(getattr(cfg, 'dflash_capture_layers', None), list) and len(cfg.dflash_capture_layers) > 0

When it happens

Trigger: Calling model.set_dflash_layers_to_capture(None) (or an API like init_dflash_support that forwards a None layers argument) on a Qwen3.5 text model on the last PP rank.

Common situations: Server args enable DFLASH/speculative aux-hidden capture but --dflash-capture-layers (or equivalent spec config) was not set, so None is forwarded to the model.

Understand the failure class

Background: "missing required argument" and "the following required arguments were not provided": what required-argument errors mean and how to fix them — this error's family across 20 libraries.

Related errors

• DFLASH speculative decoding only supports CUDA and NPU devic
• Currently DFLASH speculative decoding does not support dp at
• Currently DFLASH speculative decoding only supports pp_size
• DFLASH speculative decoding requires setting --speculative-d
• DFLASH requires --speculative-dflash-block-size to be positi

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/0b779e643cedf428. Report an issue: GitHub.