huggingface/transformers · error · ValueError

`has_previous_state` can only be called on LinearAttention l

Error message

`has_previous_state` can only be called on LinearAttention layers, and the current Cache seem to only contain Attention layers.

What it means

Cache.has_previous_state() raises ValueError (StopIteration fallback) when layer_idx is None and no layer is a LinearAttentionCacheLayerMixin. The method scans layers from last to first for a linear attention layer to inspect its previous-state flags; on an all-attention cache there is none.

Source

Thrown at src/transformers/cache_utils.py:1539

            return max(layer.get_max_length() for layer in self.layers)
        else:
            return self.layers[layer_idx].get_max_length()

    def has_previous_state(self, layer_idx: int | None = None, state_idx: int | None = None) -> bool:
        """Returns whether the LinearAttention layer at index `layer_idx` has previous state or not."""
        if layer_idx is not None and layer_idx >= len(self.layers):
            return False

        # In this case, use last LinearAttention layer
        if layer_idx is None:
            try:
                layer_idx = next(
                    idx
                    for idx in range(len(self) - 1, -1, -1)
                    if isinstance(self.layers[idx], LinearAttentionCacheLayerMixin)
                )
            except StopIteration:
                raise ValueError(
                    "`has_previous_state` can only be called on LinearAttention layers, and the current Cache seem to "
                    "only contain Attention layers."
                )
        elif not isinstance(self.layers[layer_idx], LinearAttentionCacheLayerMixin):
            raise ValueError(
                f"You called `has_previous_state` on layer index {layer_idx}, but this layer is an Attention layer, which "
                "does not support calling it."
            )

        # We may have several conv/recurrent states in the same layers. In this case, if `state_idx` is not provided, check if all
        # of them have previous state
        if state_idx is None:
            return all(self.layers[layer_idx].has_previous_state.values())
        return self.layers[layer_idx].has_previous_state[state_idx]

    def get_mask_sizes(self, query_length: int, layer_idx: int) -> tuple[int, int]:
        """
        Return a tuple (kv_length, kv_offset) corresponding to the length and offset that will be returned for

View on GitHub (pinned to a597f97485)

Solutions

1. Only call has_previous_state on caches that contain linear attention layers
2. Guard with any(isinstance(l, LinearAttentionCacheLayerMixin) for l in cache.layers)
3. For attention-only caches use get_seq_length()/is_initialized to decide decode readiness

Example fix

# before
if cache.has_previous_state():  # attention-only cache
    ...

# after
if any(isinstance(l, LinearAttentionCacheLayerMixin) for l in cache.layers) and cache.has_previous_state():
    ...

from transformers.cache_utils import LinearAttentionCacheLayerMixin

if any(isinstance(l, LinearAttentionCacheLayerMixin) for l in cache.layers):
    ready = cache.has_previous_state()

from transformers.cache_utils import LinearAttentionCacheLayerMixin

def cache_has_linear_attention_layers(cache) -> bool:
    return any(isinstance(l, LinearAttentionCacheLayerMixin) for l in cache.layers)

When it happens

Trigger: Calling cache.has_previous_state() with no layer_idx on a standard attention-only cache (e.g. DynamicCache for a Llama-style model).

Common situations: Generic code that probes has_previous_state to decide between prefill and decode paths, run on a model without linear attention layers; sharing utilities between Mamba-style and attention models.

Related errors

• Cannot call `update_conv_state` on a non-LinearAttention lay
• You called `get_seq_length` on layer index {layer_idx}, but
• `get_seq_length` can only be called on Attention layers, and
• You called `has_previous_state` on layer index {layer_idx},
• You called `get_mask_sizes` on layer index {layer_idx}, but

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/2e5d768e64d8dd59. Report an issue: GitHub.