[Test] Correctly skip tests on non-Hopper GPUs

The previous logic for skipping hardware-specific tests was incorrect,
causing tests to run on unsupported platforms.

This patch ensures that tests with `D=128`, `CombaConfig`, or
`GatedDeltaNetConfig` are now properly skipped on all non-Hopper
architectures to prevent spurious failures.

Author: zhiyuan1i

fla/modules/convolution.py

@@ -584,6 +584,14 @@ def __init__(
                     "Consider installing `causal_conv1d` to enable the CUDA implementation."
                 )
                 self.use_fast_conv1d = False
+        if bias is False:
+            # There is a bug in https://github.com/Dao-AILab/causal-conv1d/blob/main/causal_conv1d/cpp_functions.py#L135
+            self.use_fast_conv1d = False
+            warnings.warn(
+                "The `use_fast_conv1d` parameter is set to `True`, but bias is set to `False`. "
+                "Switching to the Triton implementation instead. "
+                "Since there is a bug in causal_conv1d that does not support bias during backward pass."
+            )
 
     def extra_repr(self):
         s = ('{in_channels}, {out_channels}, kernel_size={kernel_size}'

tests/test_model.py

@@ -80,8 +80,10 @@ def test_model(
     dtype: torch.dtype,
     use_l2warp: bool
 ):
-    if not is_nvidia_hopper and D == 128 or config_class in [CombaConfig, GatedDeltaNetConfig]:
-        # Due to lack of shared memory
+    if not is_nvidia_hopper and (D == 128 or config_class in [CombaConfig, GatedDeltaNetConfig]):
+        # Skip D=128 for non-Hopper GPUs
+        # CombaConfig and GatedDeltaNetConfig are not supported on non-Hopper GPUs
+        # as they require specific shared memory
         pytest.skip("D=128 is only Tested on Hopper GPUs")
     if config_class in [
         ABCConfig, ForgettingTransformerConfig, LinearAttentionConfig, LightNetConfig,