"""Shared test setup.

HF's Qwen3.5 / LFM2 models dispatch their causal depthwise conv to the ``causal_conv1d``
hub kernel whenever that package is installed; it is CUDA-only, so the HF reference
forwards the CPU tests compare against would fail on CPU tensors. CPU calls take HF's
own pure-torch fallback (the same arithmetic); CUDA calls keep the kernel.
"""

from __future__ import annotations

import pytest


def _cpu_fallbacks(monkeypatch) -> None:
    import torch
    import torch.nn.functional as F

    def conv_fn(x, weight, bias=None, activation=None, **_):
        out = F.conv1d(x.to(weight.dtype), weight.unsqueeze(1), bias,
                       padding=weight.shape[-1] - 1, groups=x.shape[1])[:, :, :x.shape[-1]]
        return (F.silu(out) if activation else out).to(x.dtype)

    def conv_update(x, conv_state, weight, bias=None, activation=None):
        state_len = conv_state.shape[-1]
        cat = torch.cat([conv_state, x], -1).to(weight.dtype)
        conv_state.copy_(cat[:, :, -state_len:])
        out = F.conv1d(cat, weight.unsqueeze(1), bias, groups=x.shape[1])[:, :, -x.shape[-1]:]
        return (F.silu(out) if activation else out).to(x.dtype)

    for mod_name in ("transformers.models.qwen3_5.modeling_qwen3_5",
                     "transformers.models.lfm2.modeling_lfm2"):
        try:
            mod = __import__(mod_name, fromlist=["_"])
        except ImportError:
            continue
        for name, ref in (("causal_conv1d_fn", conv_fn), ("causal_conv1d_update", conv_update)):
            orig = getattr(mod, name, None)
            if orig is None:
                continue

            def wrapped(x, *a, _orig=orig, _ref=ref, **kw):
                return _orig(x, *a, **kw) if x.is_cuda else _ref(x, *a, **kw)

            monkeypatch.setattr(mod, name, wrapped)


@pytest.fixture(autouse=True)
def _hf_conv_cpu_fallback(monkeypatch):
    try:
        import torch  # noqa: F401
    except ImportError:
        yield
        return
    _cpu_fallbacks(monkeypatch)
    yield