Method forward

models/ltx2.py:613–624

(self, inputs)

    @torch.autocast('cuda', dtype=AUTOCAST_DTYPE)
    @torch.compiler.disable()
    def forward(self, inputs):
        (
            vx, ax, v_context, a_context, attention_mask, v_timestep, a_timestep, v_pe_cos, v_pe_sin, a_pe_cos, a_pe_sin, av_cross_video_freq_cos, av_cross_video_freq_sin, av_cross_audio_freq_cos, av_cross_audio_freq_sin,
            av_ca_video_scale_shift_timestep, av_ca_audio_scale_shift_timestep, av_ca_a2v_gate_noise_timestep, av_ca_v2a_gate_noise_timestep, v_prompt_timestep, a_prompt_timestep, v_embedded_timestep, a_embedded_timestep, orig_shape
        ) = inputs
        # TODO: this will return a list with 2 elements when audio is enabled (currently single tensor)
        return self._process_output(
            [vx, ax],
            [v_embedded_timestep, a_embedded_timestep],
            keyframe_idxs=None,
            orig_shape=orig_shape.tolist(),
        )