Skip to content

kandinsky6

Native FastVideo implementation of the Kandinsky6 T2VA/IT2VA transformer.

Kandinsky6 extends Kandinsky5's DiT (fastvideo/models/dits/kandinsky5.py) with a second, parallel "audio" tower and a fused video<->audio decoder block, so this file mirrors kandinsky5.py's structure closely for every shared piece (time/text/visual embeddings, RoPE, modulation, feed-forward, NABLA sparse attention) and only adds what's new: the audio tower and Kandinsky6FusedTransformerDecoderBlock's cross-modal attention, ported from the reference diffusers.models.transformers.transformer_kandinsky6 module.

Video and audio latents are ordinary batched tensors -- [B, T, H, W, C] for video and [B, A, D] for audio -- consistent with Kandinsky5 and the rest of FastVideo's pipeline/stage system. Kandinsky6RoPE3D.forward takes an explicit shape tuple rather than deriving T/H/W from the position tensors' lengths.

Classes

fastvideo.models.dits.kandinsky6.Kandinsky6Attention 

Kandinsky6Attention(num_channels: int, head_dim: int, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None, prefix: str = '', kv_dim: int | None = None, use_nabla: bool = False, quant_config: QuantizationConfig | None = None)

Bases: Module

Self- or cross-attention. kv_dim lets K/V come from a differently-sized stream (the video<->audio cross-modal attentions).

Source code in fastvideo/models/dits/kandinsky6.py 
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
def __init__(
    self,
    num_channels: int,
    head_dim: int,
    supported_attention_backends: tuple[AttentionBackendEnum, ...] | None,
    prefix: str = "",
    kv_dim: int | None = None,
    use_nabla: bool = False,
    quant_config: QuantizationConfig | None = None,
):
    super().__init__()
    assert num_channels % head_dim == 0
    self.num_heads = num_channels // head_dim
    kv_dim = kv_dim or num_channels

    self.to_query = ReplicatedLinear(num_channels,
                                     num_channels,
                                     bias=True,
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.to_query")
    self.to_key = ReplicatedLinear(kv_dim,
                                   num_channels,
                                   bias=True,
                                   quant_config=quant_config,
                                   prefix=f"{prefix}.to_key")
    self.to_value = ReplicatedLinear(kv_dim,
                                     num_channels,
                                     bias=True,
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.to_value")
    self.query_norm = nn.RMSNorm(head_dim)
    self.key_norm = nn.RMSNorm(head_dim)
    self.out_layer = ReplicatedLinear(num_channels,
                                      num_channels,
                                      bias=True,
                                      quant_config=quant_config,
                                      prefix=f"{prefix}.out_layer")
    self.local_attention = LocalAttention(
        num_heads=self.num_heads,
        head_size=head_dim,
        causal=False,
        supported_attention_backends=supported_attention_backends,
    )
    # Only the video self-attention gets a second NABLA-backed attention
    # layer; audio self-attention and every cross-attention are dense.
    self.nabla_attention = None
    if use_nabla:
        self.nabla_attention = LocalAttention(
            num_heads=self.num_heads,
            head_size=head_dim,
            causal=False,
            supported_attention_backends=supported_attention_backends,
            default_backend=AttentionBackendEnum.NABLA_ATTN,
        )

fastvideo.models.dits.kandinsky6.Kandinsky6FusedTransformerDecoderBlock 

Kandinsky6FusedTransformerDecoderBlock(model_dim: int, time_dim: int, ff_dim: int, head_dim: int, model_dim_a: int, time_dim_a: int, ff_dim_a: int, head_dim_a: int, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, prefix: str = '', use_nabla: bool = False, ca_rope: bool = False, cross_gates: bool = False, fix_modulation: bool = False, quant_config: QuantizationConfig | None = None)

Bases: Module

Joint video+audio decoder block: per-modality self/text-cross attention plus a dedicated bidirectional video<->audio cross-attention, all independently AdaLN-modulated. Ported from diffusers.models.transformers.transformer_kandinsky6.Kandinsky6FusedTransformerDecoderBlock.

Source code in fastvideo/models/dits/kandinsky6.py 
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
def __init__(
    self,
    model_dim: int,
    time_dim: int,
    ff_dim: int,
    head_dim: int,
    model_dim_a: int,
    time_dim_a: int,
    ff_dim_a: int,
    head_dim_a: int,
    supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None,
    prefix: str = "",
    use_nabla: bool = False,
    ca_rope: bool = False,
    cross_gates: bool = False,
    fix_modulation: bool = False,
    quant_config: QuantizationConfig | None = None,
):
    super().__init__()
    self.videoT = Kandinsky6TransformerDecoderBlock(model_dim,
                                                    time_dim,
                                                    ff_dim,
                                                    head_dim,
                                                    supported_attention_backends,
                                                    prefix=f"{prefix}.videoT",
                                                    use_nabla=use_nabla,
                                                    quant_config=quant_config)
    self.audioT = Kandinsky6TransformerDecoderBlock(model_dim_a,
                                                    time_dim_a,
                                                    ff_dim_a,
                                                    head_dim_a,
                                                    supported_attention_backends,
                                                    prefix=f"{prefix}.audioT",
                                                    use_nabla=False,
                                                    quant_config=quant_config)
    self.va_cross_attention = Kandinsky6Attention(model_dim,
                                                  head_dim,
                                                  supported_attention_backends=supported_attention_backends,
                                                  prefix=f"{prefix}.va_cross_attention",
                                                  kv_dim=model_dim_a,
                                                  quant_config=quant_config)
    self.av_cross_attention = Kandinsky6Attention(model_dim_a,
                                                  head_dim_a,
                                                  supported_attention_backends=supported_attention_backends,
                                                  prefix=f"{prefix}.av_cross_attention",
                                                  kv_dim=model_dim,
                                                  quant_config=quant_config)
    self.va_modulation = Kandinsky6Modulation(
        time_dim,
        model_dim if not cross_gates else model_dim * 2 + model_dim_a,
        1 if cross_gates else 3,
    )
    self.av_modulation = Kandinsky6Modulation(
        time_dim_a,
        model_dim_a if not cross_gates else model_dim_a * 2 + model_dim,
        1 if cross_gates else 3,
    )
    self.va_normalization = nn.LayerNorm(model_dim, elementwise_affine=False)
    self.av_normalization = nn.LayerNorm(model_dim_a, elementwise_affine=False)
    self.ca_rope = ca_rope
    self.cross_gates = cross_gates
    self.fix_modulation = fix_modulation
    self.model_dim = model_dim
    self.model_dim_a = model_dim_a

Methods:

fastvideo.models.dits.kandinsky6.Kandinsky6FusedTransformerDecoderBlock.forward 
forward(vis: Tensor | None, aud: Tensor | None, text_v: Tensor, text_a: Tensor | None, time_embed: tuple[Tensor, Tensor | None], vis_rope: Tensor | None, aud_rope: Tensor | None, sparse_params: dict[str, Any] | None, va_gate_scale: float = 1.0, av_gate_scale: float = 1.0) -> tuple[Tensor | None, Tensor | None]

vis/aud may each be None (matches the diffusers reference's guarded Kandinsky6FusedTransformerDecoderBlock): every stage below only runs for a present modality, and the video<->audio cross-modal mixing only runs when both are present.

Source code in fastvideo/models/dits/kandinsky6.py 
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
def forward(
    self,
    vis: torch.Tensor | None,
    aud: torch.Tensor | None,
    text_v: torch.Tensor,
    text_a: torch.Tensor | None,
    time_embed: tuple[torch.Tensor, torch.Tensor | None],
    vis_rope: torch.Tensor | None,
    aud_rope: torch.Tensor | None,
    sparse_params: dict[str, Any] | None,
    va_gate_scale: float = 1.0,
    av_gate_scale: float = 1.0,
) -> tuple[torch.Tensor | None, torch.Tensor | None]:
    """``vis``/``aud`` may each be ``None`` (matches the diffusers reference's guarded
    ``Kandinsky6FusedTransformerDecoderBlock``): every stage below only runs for a present
    modality, and the video<->audio cross-modal mixing only runs when both are present."""
    t_v, t_a = time_embed
    gate_v = vis_out_t = None

    if vis is not None:
        sa_p, ca_p, ff_p = torch.chunk(self.videoT.visual_modulation(t_v).unsqueeze(dim=1), 3, dim=-1)
        shift, scale, gate = torch.chunk(sa_p, 3, dim=-1)
        vis = _apply_gate_sum(
            vis,
            self.videoT.self_attention(
                self.videoT.self_attention_norm(vis.float(), shift=shift, scale=scale,
                                                convert_modulation_dtype=True).type_as(vis),
                rotary_emb=vis_rope,
                sparse_params=sparse_params,
            ),
            gate,
        )
        shift, scale, gate_v = torch.chunk(ca_p, 3, dim=-1)
        vis_pre_ca = self.videoT.cross_attention_norm(vis.float(), shift=shift, scale=scale,
                                                      convert_modulation_dtype=True).type_as(vis)
        vis_out_t = self.videoT.cross_attention(vis_pre_ca, encoder_hidden_states=text_v)

    if aud is not None:
        sa_p, ca_p, ff_p_a = torch.chunk(self.audioT.visual_modulation(t_a).unsqueeze(dim=1), 3, dim=-1)
        shift, scale, gate = torch.chunk(sa_p, 3, dim=-1)
        aud = _apply_gate_sum(
            aud,
            self.audioT.self_attention(
                self.audioT.self_attention_norm(aud.float(), shift=shift, scale=scale,
                                                convert_modulation_dtype=True).type_as(aud),
                rotary_emb=aud_rope,
            ),
            gate,
        )
        shift, scale, gate_a = torch.chunk(ca_p, 3, dim=-1)
        aud_pre_ca = self.audioT.cross_attention_norm(aud.float(), shift=shift, scale=scale,
                                                       convert_modulation_dtype=True).type_as(aud)
        aud_out_t = self.audioT.cross_attention(aud_pre_ca, encoder_hidden_states=text_a)
        aud = _apply_gate_sum(aud, aud_out_t, gate_a)

        if vis is not None:
            t_va_mod = t_a if not self.fix_modulation else t_v
            t_av_mod = t_v if not self.fix_modulation else t_a
            va_params = self.va_modulation(t_va_mod).unsqueeze(dim=1)
            av_params = self.av_modulation(t_av_mod).unsqueeze(dim=1)
            if self.cross_gates:
                va_shift, va_scale, va_gate = torch.split(
                    va_params, [self.model_dim, self.model_dim, self.model_dim_a], dim=-1)
                av_shift, av_scale, av_gate = torch.split(
                    av_params, [self.model_dim_a, self.model_dim_a, self.model_dim], dim=-1)
            else:
                va_shift, va_scale, va_gate = torch.chunk(va_params, 3, dim=-1)
                av_shift, av_scale, av_gate = torch.chunk(av_params, 3, dim=-1)

            vis = _apply_gate_sum(vis, vis_out_t, gate_v)
            vis_for_va = _apply_scale_shift(self.va_normalization, vis, va_scale, va_shift)
            aud_for_av = _apply_scale_shift(self.av_normalization, aud, av_scale, av_shift)
            rq_v = vis_rope if self.ca_rope else None
            rk_a = aud_rope if self.ca_rope else None
            vis_from_aud = self.va_cross_attention(vis_for_va, encoder_hidden_states=aud_pre_ca, rotary_emb=rq_v,
                                                   rotary_emb_kv=rk_a)
            aud_from_vis = self.av_cross_attention(aud_for_av, encoder_hidden_states=vis_pre_ca, rotary_emb=rk_a,
                                                   rotary_emb_kv=rq_v)
            vis = _apply_gate_sum(vis, vis_from_aud, (va_gate if not self.cross_gates else av_gate) * va_gate_scale)
            aud = _apply_gate_sum(aud, aud_from_vis, (av_gate if not self.cross_gates else va_gate) * av_gate_scale)
    elif vis is not None:
        vis = _apply_gate_sum(vis, vis_out_t, gate_v)

    if vis is not None:
        shift, scale, gate = torch.chunk(ff_p, 3, dim=-1)
        vis = _apply_gate_sum(
            vis,
            self.videoT.feed_forward(
                self.videoT.feed_forward_norm(vis.float(), shift=shift, scale=scale,
                                              convert_modulation_dtype=True).type_as(vis)),
            gate,
        )
    if aud is not None:
        shift, scale, gate = torch.chunk(ff_p_a, 3, dim=-1)
        aud = _apply_gate_sum(
            aud,
            self.audioT.feed_forward(
                self.audioT.feed_forward_norm(aud.float(), shift=shift, scale=scale,
                                              convert_modulation_dtype=True).type_as(aud)),
            gate,
        )
    return vis, aud

fastvideo.models.dits.kandinsky6.Kandinsky6OutLayer 

Kandinsky6OutLayer(model_dim: int, time_dim: int, visual_dim: int, patch_size: tuple[int, int, int])

Bases: Module

Projects visual hidden states back to packed latent patches.

Source code in fastvideo/models/dits/kandinsky6.py 
433
434
435
436
437
438
def __init__(self, model_dim: int, time_dim: int, visual_dim: int, patch_size: tuple[int, int, int]):
    super().__init__()
    self.patch_size = patch_size
    self.modulation = Kandinsky6Modulation(time_dim, model_dim, 2)
    self.norm = nn.LayerNorm(model_dim, eps=1e-5, elementwise_affine=False)
    self.out_layer = ReplicatedLinear(model_dim, math.prod(patch_size) * visual_dim, bias=True)

fastvideo.models.dits.kandinsky6.Kandinsky6OutLayerAudio 

Kandinsky6OutLayerAudio(model_dim: int, time_dim: int, audio_dim: int)

Bases: Module

Projects audio hidden states back to audio latent channels.

Source code in fastvideo/models/dits/kandinsky6.py 
464
465
466
467
468
def __init__(self, model_dim: int, time_dim: int, audio_dim: int):
    super().__init__()
    self.modulation = Kandinsky6Modulation(time_dim, model_dim, 2)
    self.norm = nn.LayerNorm(model_dim, eps=1e-5, elementwise_affine=False)
    self.out_layer = ReplicatedLinear(model_dim, audio_dim, bias=True)

fastvideo.models.dits.kandinsky6.Kandinsky6RoPE1D 

Kandinsky6RoPE1D(dim: int, max_pos: int = 2048, max_period: float = 10000.0, freqs_scaling: float = 1.0)

Bases: Module

1D rotary embedding for text and audio sequences.

Source code in fastvideo/models/dits/kandinsky6.py 
204
205
206
207
208
209
210
211
212
def __init__(self, dim: int, max_pos: int = 2048, max_period: float = 10000.0, freqs_scaling: float = 1.0):
    super().__init__()
    self.max_period = max_period
    self.dim = dim
    self.max_pos = max_pos
    self.freqs_scaling = freqs_scaling
    freq = _build_rotary_freqs(dim // 2, max_period) * freqs_scaling
    pos = torch.arange(max_pos, dtype=freq.dtype)
    self.register_buffer("args", torch.outer(pos, freq), persistent=False)

fastvideo.models.dits.kandinsky6.Kandinsky6RoPE3D 

Kandinsky6RoPE3D(axes_dims: tuple[int, int, int], max_pos: tuple[int, int, int] = (128, 128, 128), max_period: float = 10000.0)

Bases: Module

3D rotary embedding for video spatial-temporal (T, H, W) tokens.

Source code in fastvideo/models/dits/kandinsky6.py 
226
227
228
229
230
231
232
233
234
235
236
237
238
def __init__(self,
             axes_dims: tuple[int, int, int],
             max_pos: tuple[int, int, int] = (128, 128, 128),
             max_period: float = 10000.0):
    super().__init__()
    self.axes_dims = axes_dims
    self.max_pos = max_pos
    self.max_period = max_period

    for i, (axes_dim, ax_max_pos) in enumerate(zip(axes_dims, max_pos, strict=True)):
        freq = _build_rotary_freqs(axes_dim // 2, max_period)
        pos = torch.arange(ax_max_pos, dtype=freq.dtype)
        self.register_buffer(f"args_{i}", torch.outer(pos, freq), persistent=False)

fastvideo.models.dits.kandinsky6.Kandinsky6TextEmbeddings 

Kandinsky6TextEmbeddings(in_dim: int, model_dim: int)

Bases: Module

Linear + LayerNorm projection, reused for text tokens and audio latents.

Source code in fastvideo/models/dits/kandinsky6.py 
168
169
170
171
def __init__(self, in_dim: int, model_dim: int):
    super().__init__()
    self.in_layer = ReplicatedLinear(in_dim, model_dim, bias=True)
    self.norm = nn.LayerNorm(model_dim, elementwise_affine=True)

fastvideo.models.dits.kandinsky6.Kandinsky6Transformer3DModel 

Kandinsky6Transformer3DModel(config: Kandinsky6VideoAudioConfig, hf_config: dict[str, Any])

Bases: BaseDiT

Native FastVideo implementation of the Kandinsky6 T2VA/IT2VA transformer.

Source code in fastvideo/models/dits/kandinsky6.py 
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
def __init__(self, config: Kandinsky6VideoAudioConfig, hf_config: dict[str, Any]) -> None:
    super().__init__(config=config, hf_config=hf_config)
    arch = config.arch_config
    quant_config = config.quant_config
    self.quant_config = quant_config

    head_dim = sum(arch.axes_dims)
    head_dim_a = sum(arch.axes_dims_a)
    self.in_visual_dim = arch.in_visual_dim
    self.in_audio_dim = arch.in_audio_dim
    self.model_dim = arch.model_dim
    self.patch_size = arch.patch_size
    self.visual_cond = arch.visual_cond
    self.attention_engine = arch.attention_engine
    self.visual_token_type_num_embeddings = arch.visual_token_type_num_embeddings
    self.scale_factor = tuple(float(v) for v in arch.scale_factor)

    visual_embed_dim = (2 * arch.in_visual_dim + 1) if arch.visual_cond else arch.in_visual_dim

    self.visual_embeddings = Kandinsky6VisualEmbeddings(visual_embed_dim, arch.model_dim, arch.patch_size)
    if self.visual_token_type_num_embeddings > 0:
        self.visual_token_type_embeddings = nn.Embedding(self.visual_token_type_num_embeddings, arch.model_dim)
    self.visual_rope_embeddings = Kandinsky6RoPE3D(arch.axes_dims)
    self.out_layer = Kandinsky6OutLayer(arch.model_dim, arch.time_dim, arch.out_visual_dim, arch.patch_size)

    use_nabla = arch.attention_engine == "nabla"

    # The model is always the joint video+audio DiT; a video-only call passes
    # hidden_states_audio=None instead of using a different model shape.

    # Registered before the (much smaller) text towers below: enable_layerwise_offload
    # (fastvideo/hooks/layerwise_offload.py) hooks only the first top-level nn.ModuleList in
    # registration order, and with dit_layerwise_offload=True (the default) that has to be these
    # blocks rather than the 4-entry video_text_transformer_blocks.
    self.visual_transformer_blocks = nn.ModuleList([
        Kandinsky6FusedTransformerDecoderBlock(
            arch.model_dim,
            arch.time_dim,
            arch.ff_dim,
            head_dim,
            arch.model_dim_a,
            arch.time_dim_a,
            arch.ff_dim_a,
            head_dim_a,
            self._supported_attention_backends,
            prefix=f"{config.prefix}.visual_transformer_blocks.{i}",
            use_nabla=use_nabla,
            ca_rope=arch.ca_rope,
            cross_gates=arch.cross_gates,
            fix_modulation=arch.fix_modulation,
            quant_config=quant_config) for i in range(arch.num_visual_blocks)
    ])

    self.audio_embeddings = Kandinsky6TextEmbeddings(arch.in_audio_dim, arch.model_dim_a)
    self.audio_rope_embeddings = Kandinsky6RoPE1D(head_dim_a, freqs_scaling=arch.audio_freqs_scaling)
    self.audio_out_layer = Kandinsky6OutLayerAudio(arch.model_dim_a, arch.time_dim_a, arch.out_audio_dim or arch.in_audio_dim)

    for tower_prefix, model_dim, time_dim, hd in (
        ("video", arch.model_dim, arch.time_dim, head_dim),
        ("audio", arch.model_dim_a, arch.time_dim_a, head_dim_a),
    ):
        setattr(self, f"{tower_prefix}_time_embeddings", Kandinsky6TimeEmbeddings(model_dim, time_dim))
        setattr(self, f"{tower_prefix}_text_embeddings", Kandinsky6TextEmbeddings(arch.in_text_dim, model_dim))
        setattr(self, f"{tower_prefix}_pooled_text_embeddings",
                Kandinsky6TextEmbeddings(arch.in_text_dim2, time_dim))
        setattr(self, f"{tower_prefix}_text_rope_embeddings", Kandinsky6RoPE1D(hd))
        setattr(
            self, f"{tower_prefix}_text_transformer_blocks",
            nn.ModuleList([
                Kandinsky6TransformerEncoderBlock(
                    model_dim,
                    time_dim,
                    arch.ff_dim if tower_prefix == "video" else arch.ff_dim_a,
                    hd,
                    self._supported_attention_backends,
                    prefix=f"{config.prefix}.{tower_prefix}_text_transformer_blocks.{i}",
                    quant_config=quant_config) for i in range(arch.num_text_blocks)
            ]))

    self.gradient_checkpointing = False
    self.hidden_size = arch.hidden_size
    self.num_attention_heads = arch.num_attention_heads
    self.num_channels_latents = arch.num_channels_latents
    self.__post_init__()

fastvideo.models.dits.kandinsky6.Kandinsky6TransformerDecoderBlock 

Kandinsky6TransformerDecoderBlock(model_dim: int, time_dim: int, ff_dim: int, head_dim: int, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, prefix: str = '', use_nabla: bool = False, quant_config: QuantizationConfig | None = None)

Bases: Module

Self-attention + text cross-attention + feed-forward block.

Used standalone for plain (non-multimodal) T2V/I2V-parity checkpoints, and as the videoT/audioT sub-block inside Kandinsky6FusedTransformerDecoderBlock for T2VA/IT2VA.

Source code in fastvideo/models/dits/kandinsky6.py 
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
def __init__(self,
             model_dim: int,
             time_dim: int,
             ff_dim: int,
             head_dim: int,
             supported_attention_backends: tuple[AttentionBackendEnum, ...]
             | None = None,
             prefix: str = "",
             use_nabla: bool = False,
             quant_config: QuantizationConfig | None = None):
    super().__init__()
    self.visual_modulation = Kandinsky6Modulation(time_dim, model_dim, 9)

    self.self_attention_norm = LayerNormScaleShift(model_dim,
                                                   norm_type="layer",
                                                   eps=1e-5,
                                                   elementwise_affine=False,
                                                   dtype=torch.float32,
                                                   compute_dtype=torch.float32)
    self.self_attention = Kandinsky6Attention(model_dim,
                                              head_dim,
                                              supported_attention_backends=supported_attention_backends,
                                              prefix=f"{prefix}.self_attention",
                                              use_nabla=use_nabla,
                                              quant_config=quant_config)

    self.cross_attention_norm = LayerNormScaleShift(model_dim,
                                                    norm_type="layer",
                                                    eps=1e-5,
                                                    elementwise_affine=False,
                                                    dtype=torch.float32,
                                                    compute_dtype=torch.float32)
    self.cross_attention = Kandinsky6Attention(model_dim,
                                               head_dim,
                                               supported_attention_backends=supported_attention_backends,
                                               prefix=f"{prefix}.cross_attention",
                                               quant_config=quant_config)

    self.feed_forward_norm = LayerNormScaleShift(model_dim,
                                                 norm_type="layer",
                                                 eps=1e-5,
                                                 elementwise_affine=False,
                                                 dtype=torch.float32,
                                                 compute_dtype=torch.float32)
    self.feed_forward = Kandinsky6FeedForward(model_dim,
                                              ff_dim,
                                              prefix=f"{prefix}.feed_forward",
                                              quant_config=quant_config)

fastvideo.models.dits.kandinsky6.Kandinsky6TransformerEncoderBlock 

Kandinsky6TransformerEncoderBlock(model_dim: int, time_dim: int, ff_dim: int, head_dim: int, supported_attention_backends: tuple[AttentionBackendEnum, ...] | None = None, prefix: str = '', quant_config: QuantizationConfig | None = None)

Bases: Module

Text-only self-attention + feed-forward block (video/audio text towers).

Source code in fastvideo/models/dits/kandinsky6.py 
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
def __init__(self,
             model_dim: int,
             time_dim: int,
             ff_dim: int,
             head_dim: int,
             supported_attention_backends: tuple[AttentionBackendEnum, ...]
             | None = None,
             prefix: str = "",
             quant_config: QuantizationConfig | None = None):
    super().__init__()
    self.text_modulation = Kandinsky6Modulation(time_dim, model_dim, 6)

    self.self_attention_norm = LayerNormScaleShift(model_dim,
                                                   norm_type="layer",
                                                   eps=1e-5,
                                                   elementwise_affine=False,
                                                   dtype=torch.float32,
                                                   compute_dtype=torch.float32)
    self.self_attention = Kandinsky6Attention(model_dim,
                                              head_dim,
                                              supported_attention_backends=supported_attention_backends,
                                              prefix=f"{prefix}.self_attention",
                                              quant_config=quant_config)

    self.feed_forward_norm = LayerNormScaleShift(model_dim,
                                                 norm_type="layer",
                                                 eps=1e-5,
                                                 elementwise_affine=False,
                                                 dtype=torch.float32,
                                                 compute_dtype=torch.float32)
    self.feed_forward = Kandinsky6FeedForward(model_dim,
                                              ff_dim,
                                              prefix=f"{prefix}.feed_forward",
                                              quant_config=quant_config)

Functions: