hiyouga/LlamaFactory · error · RuntimeError

FLA operator `{op_name}` did not match any model module attr

Error message

FLA operator `{op_name}` did not match any model module attributes.

What it means

For each selected FLA kernel, the plugin scans the model's modules for a callable attribute matching FLA_MODULE_ATTRIBUTES[op_name] and patches them via patch_model_members. If zero modules match, the patch would be a no-op, so it raises RuntimeError naming the operator that found no targets.

Source

Thrown at src/llamafactory/v1/plugins/model_plugins/kernels/ops/linear_attention/fla.py:99

            raise ValueError(f"chunk_size must be one of {SUPPORTED_CHUNK_SIZES}, got {chunk_size!r}.")

        from fsdp_turbo.ops.registry import get_op
        from fsdp_turbo.utils.patch import patch_model_members

        patched = 0
        named_modules = tuple(model.named_modules())
        for op_name in selected:
            module_attribute = FLA_MODULE_ATTRIBUTES[op_name]
            op = get_op(op_name)
            configured_op = partial(op, chunk_size=chunk_size) if op_name == CHUNK_GATED_DELTA_RULE else op
            targets = {
                f"{type(module).__module__}.{type(module).__name__}.{module_attribute}"
                for _, module in named_modules
                if callable(getattr(module, module_attribute, None))
            }
            matched = patch_model_members(model, sorted(targets), configured_op) if targets else 0
            if matched == 0:
                raise RuntimeError(f"FLA operator `{op_name}` did not match any model module attributes.")
            patched += matched

        logger.info_rank0(f"Flash Linear Attention kernels updated {patched} module callables: {selected}.")
        return model

View on GitHub (pinned to f28afaf635)

Solutions

Confirm the model actually uses Flash Linear Attention layers before enabling the fla plugin
Use include_kernels: "auto" only on FLA-based models; remove the plugin for standard attention models
Check FLA_MODULE_ATTRIBUTES against type(module).__module__/__name__ of your model's layers; align transformers/fsdp_turbo versions with what the mapping expects
If a specific op never matches your architecture, exclude it from include_kernels

Example fix

# before (Llama model, no FLA layers)
kernels: [fla]
kernel_config:
  include_kernels: "auto"

# after
kernels: []  # fla only for FLA-architecture models

Defensive patterns

Strategy: validation

Validate before calling

has_fla = any(
    callable(getattr(m, attr, None))
    for _, m in model.named_modules()
    for attr in FLA_MODULE_ATTRIBUTES.values()
)
if not has_fla:
    raise ValueError("model has no FLA layers; remove the fla kernel plugin")

Try / catch

try:
    apply_fla_kernel(model, config=kernel_config)
except RuntimeError as e:
    if "did not match any model module attributes" in str(e):
        logger.warning("skipping fla plugin for %s", type(model).__name__)
    else:
        raise

Prevention

Only enable fla kernels for linear-attention architectures
Smoke-test kernel application on one batch before long runs

When it happens

Trigger: Running the FLA kernel plugin on a model that does not use Flash Linear Attention layers (e.g. a standard transformer with softmax attention), or a linear-attention model whose class/attribute layout differs from what FLA_MODULE_ATTRIBUTES expects (transformers version mismatch).

Common situations: Enabling fla kernels in kernel_config for a non-FLA model (Llama, Qwen, etc.); upgrading transformers so FLA layer class names or module paths changed; selecting an op whose attribute only exists on some FLA architectures.

Related errors

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/0eb15cd7f20ca68d. Report an issue: GitHub.