[bugfix] fix gdn conv1d

[Jintao-Huang]

No description provided.

[gemini-code-assist]

Code Review

This pull request updates the 1D convolution in the gated delta network module to include the groups parameter, calculated using the tensor parallel and context parallel sizes. The reviewer suggests using conv1d_weight.shape[0] instead of manually computing this value, as it is cleaner and more robust.

[gemini-code-assist]

Using conv1d_weight.shape[0] is more robust and cleaner than manually computing self.conv_dim // self.tp_size // cp_size. Since conv1d_weight is already sliced according to the TP and CP configurations, its first dimension dynamically represents the correct local channel size.

Suggested change

	groups=self.conv_dim // self.tp_size // cp_size,
	groups=conv1d_weight.shape[0],