hiyouga/LlamaFactory · error · ImportError

Qwen3.5 packing-seq forwarding requires `flash-linear-attent

Error message

Qwen3.5 packing-seq forwarding requires `flash-linear-attention>=0.4.1` (provides `fla.modules.convolution.causal_conv1d` and `fla.ops.gated_delta_rule.{chunk,fused_recurrent}_gated_delta_rule`). Please install/upgrade it.

What it means

Error "Qwen3.5 packing-seq forwarding requires `flash-linear-attention>=0.4.1` (provides `fla.modules.convolution.causal_conv1d` and `fla.ops.gated_delta_rule.{chunk,fused_recurrent}_gated_delta_rule`). Please install/upgrade it." thrown in hiyouga/LlamaFactory.

Source

Thrown at src/llamafactory/model/patcher.py:80

def _check_fla_dependencies() -> None:
    """Check that the FLA dependencies required for varlen GDN forwarding are available.

    Requires ``flash-linear-attention >= 0.4.1`` (which exposes the varlen
    ``causal_conv1d`` under ``fla.modules.convolution`` and the
    ``chunk_gated_delta_rule`` / ``fused_recurrent_gated_delta_rule`` kernels
    under ``fla.ops.gated_delta_rule``). Raises ``ImportError`` with an
    actionable message otherwise.
    """
    try:
        from fla.modules.convolution import causal_conv1d  # noqa: F401
        from fla.ops.gated_delta_rule import (  # noqa: F401
            chunk_gated_delta_rule,
            fused_recurrent_gated_delta_rule,
        )
    except ImportError as exc:
        raise ImportError(
            "Qwen3.5 packing-seq forwarding requires `flash-linear-attention>=0.4.1` "
            "(provides `fla.modules.convolution.causal_conv1d` and "
            "`fla.ops.gated_delta_rule.{chunk,fused_recurrent}_gated_delta_rule`). "
            "Please install/upgrade it."
        ) from exc


def patch_qwen3_5_forward_npu(model: "PreTrainedModel") -> None:
    """Patch for Qwen3.5 models on NPU by importing torch_npu to enable torch.cuda compatibility.

    On NPU, torch.cuda operations will fail unless torch_npu is imported.
    torch_npu provides compatibility layer that maps torch.cuda calls to NPU operations.

    Also replaces chunk_gated_delta_rule with NPU-compatible implementation.
    """
    import importlib.metadata

    if "Ascend910" not in torch.npu.get_device_name(0):

View on GitHub (pinned to f28afaf635)

Solutions

1. Install or upgrade the flash-linear-attention package: pip install 'flash-linear-attention>=0.4.1'.
2. If installation fails, disable Qwen3.5 packing-seq forwarding (do not enable packed/packing training for Qwen3.5) so the fla imports are not required.

Example fix

pip install 'flash-linear-attention>=0.4.1'

When it happens

Trigger: Thrown at src/llamafactory/model/patcher.py:80 when the library encounters an invalid state.

Common situations: See trigger scenarios.

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/e60468cabe68f6b8. Report an issue: GitHub.