Dynamic VRAM fixes - Ace 1.5 performance + a VRAM leak (#12368)

* revert threaded model loader change

This change was only needed to get around the pytorch 2.7 mempool bugs,
and should have been reverted along with #12260. This fixes a different
memory leak where pytorch gets confused about cache emptying.

* load non comfy weights

* MPDynamic: Pre-generate the tensors for vbars

Apparently this is an expensive operation that slows down things.

* bump to aimdo 1.8

New features:
watermark limit feature
logging enhancements
-O2 build on linux

comfy/ldm/ace/ace_step15.py

@@ -7,6 +7,67 @@ from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
 from comfy.ldm.flux.layers import timestep_embedding
 
+def get_silence_latent(length, device):
+    head = torch.tensor([[[ 0.5707,  0.0982,  0.6909, -0.5658,  0.6266,  0.6996, -0.1365, -0.1291,
+                        -0.0776, -0.1171, -0.2743, -0.8422, -0.1168,  1.5539, -4.6936,  0.7436,
+                        -1.1846, -0.2637,  0.6933, -6.7266,  0.0966, -0.1187, -0.3501, -1.1736,
+                        0.0587, -2.0517, -1.3651,  0.7508, -0.2490, -1.3548, -0.1290, -0.7261,
+                        1.1132, -0.3249,  0.2337,  0.3004,  0.6605, -0.0298, -0.1989, -0.4041,
+                        0.2843, -1.0963, -0.5519,  0.2639, -1.0436, -0.1183,  0.0640,  0.4460,
+                        -1.1001, -0.6172, -1.3241,  1.1379,  0.5623, -0.1507, -0.1963, -0.4742,
+                        -2.4697,  0.5302,  0.5381,  0.4636, -0.1782, -0.0687,  1.0333,  0.4202],
+                        [ 0.3040, -0.1367,  0.6200,  0.0665, -0.0642,  0.4655, -0.1187, -0.0440,
+                        0.2941, -0.2753,  0.0173, -0.2421, -0.0147,  1.5603, -2.7025,  0.7907,
+                        -0.9736, -0.0682,  0.1294, -5.0707, -0.2167,  0.3302, -0.1513, -0.8100,
+                        -0.3894, -0.2884, -0.3149,  0.8660, -0.3817, -1.7061,  0.5824, -0.4840,
+                        0.6938,  0.1859,  0.1753,  0.3081,  0.0195,  0.1403, -0.0754, -0.2091,
+                        0.1251, -0.1578, -0.4968, -0.1052, -0.4554, -0.0320,  0.1284,  0.4974,
+                        -1.1889, -0.0344, -0.8313,  0.2953,  0.5445, -0.6249, -0.1595, -0.0682,
+                        -3.1412,  0.0484,  0.4153,  0.8260, -0.1526, -0.0625,  0.5366,  0.8473],
+                        [ 5.3524e-02, -1.7534e-01,  5.4443e-01, -4.3501e-01, -2.1317e-03,
+                        3.7200e-01, -4.0143e-03, -1.5516e-01, -1.2968e-01, -1.5375e-01,
+                        -7.7107e-02, -2.0593e-01, -3.2780e-01,  1.5142e+00, -2.6101e+00,
+                        5.8698e-01, -1.2716e+00, -2.4773e-01, -2.7933e-02, -5.0799e+00,
+                        1.1601e-01,  4.0987e-01, -2.2030e-02, -6.6495e-01, -2.0995e-01,
+                        -6.3474e-01, -1.5893e-01,  8.2745e-01, -2.2992e-01, -1.6816e+00,
+                        5.4440e-01, -4.9579e-01,  5.5128e-01,  3.0477e-01,  8.3052e-02,
+                        -6.1782e-02,  5.9036e-03,  2.9553e-01, -8.0645e-02, -1.0060e-01,
+                        1.9144e-01, -3.8124e-01, -7.2949e-01,  2.4520e-02, -5.0814e-01,
+                        2.3977e-01,  9.2943e-02,  3.9256e-01, -1.1993e+00, -3.2752e-01,
+                        -7.2707e-01,  2.9476e-01,  4.3542e-01, -8.8597e-01, -4.1686e-01,
+                        -8.5390e-02, -2.9018e+00,  6.4988e-02,  5.3945e-01,  9.1988e-01,
+                        5.8762e-02, -7.0098e-02,  6.4772e-01,  8.9118e-01],
+                        [-3.2225e-02, -1.3195e-01,  5.6411e-01, -5.4766e-01, -5.2170e-03,
+                        3.1425e-01, -5.4367e-02, -1.9419e-01, -1.3059e-01, -1.3660e-01,
+                        -9.0984e-02, -1.9540e-01, -2.5590e-01,  1.5440e+00, -2.6349e+00,
+                        6.8273e-01, -1.2532e+00, -1.9810e-01, -2.2793e-02, -5.0506e+00,
+                        1.8818e-01,  5.0109e-01,  7.3546e-03, -6.8771e-01, -3.0676e-01,
+                        -7.3257e-01, -1.6687e-01,  9.2232e-01, -1.8987e-01, -1.7267e+00,
+                        5.3355e-01, -5.3179e-01,  4.4953e-01,  2.8820e-01,  1.3012e-01,
+                        -2.0943e-01, -1.1348e-01,  3.3929e-01, -1.5069e-01, -1.2919e-01,
+                        1.8929e-01, -3.6166e-01, -8.0756e-01,  6.6387e-02, -5.8867e-01,
+                        1.6978e-01,  1.0134e-01,  3.3877e-01, -1.2133e+00, -3.2492e-01,
+                        -8.1237e-01,  3.8101e-01,  4.3765e-01, -8.0596e-01, -4.4531e-01,
+                        -4.7513e-02, -2.9266e+00,  1.1741e-03,  4.5123e-01,  9.3075e-01,
+                        5.3688e-02, -1.9621e-01,  6.4530e-01,  9.3870e-01]]], device=device).movedim(-1, 1)
+
+    silence_latent = torch.tensor([[[-1.3672e-01, -1.5820e-01,  5.8594e-01, -5.7422e-01,  3.0273e-02,
+                                2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01,
+                                -2.7710e-02, -1.8066e-01, -2.9688e-01,  1.6016e+00, -2.6719e+00,
+                                7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00,
+                                2.4316e-01,  4.7266e-01,  4.6387e-02, -6.6406e-01, -2.1973e-01,
+                                -6.7578e-01, -1.5723e-01,  9.5312e-01, -2.0020e-01, -1.7109e+00,
+                                5.8984e-01, -5.7422e-01,  5.1562e-01,  2.8320e-01,  1.4551e-01,
+                                -1.8750e-01, -5.9814e-02,  3.6719e-01, -1.0059e-01, -1.5723e-01,
+                                2.0605e-01, -4.3359e-01, -8.2812e-01,  4.5654e-02, -6.6016e-01,
+                                1.4844e-01,  9.4727e-02,  3.8477e-01, -1.2578e+00, -3.3203e-01,
+                                -8.5547e-01,  4.3359e-01,  4.2383e-01, -8.9453e-01, -5.0391e-01,
+                                -5.6152e-02, -2.9219e+00, -2.4658e-02,  5.0391e-01,  9.8438e-01,
+                                7.2754e-02, -2.1582e-01,  6.3672e-01,  1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, length)
+    silence_latent[:, :, :head.shape[-1]] = head
+    return silence_latent
+
+
 def get_layer_class(operations, layer_name):
     if operations is not None and hasattr(operations, layer_name):
         return getattr(operations, layer_name)
@@ -183,7 +244,7 @@ class AceStepAttention(nn.Module):
             else:
                 attn_bias = window_bias
 
-        attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True)
+        attn_output = optimized_attention(query_states, key_states, value_states, self.num_heads, attn_bias, skip_reshape=True, low_precision_attention=False)
         attn_output = self.o_proj(attn_output)
 
         return attn_output
@@ -677,7 +738,7 @@ class AttentionPooler(nn.Module):
     def forward(self, x):
         B, T, P, D = x.shape
         x = self.embed_tokens(x)
-        special = self.special_token.expand(B, T, 1, -1)
+        special = comfy.model_management.cast_to(self.special_token, device=x.device, dtype=x.dtype).expand(B, T, 1, -1)
         x = torch.cat([special, x], dim=2)
         x = x.view(B * T, P + 1, D)
 
@@ -728,7 +789,7 @@ class FSQ(nn.Module):
         self.register_buffer('implicit_codebook', implicit_codebook, persistent=False)
 
     def bound(self, z):
-        levels_minus_1 = (self._levels - 1).to(z.dtype)
+        levels_minus_1 = (comfy.model_management.cast_to(self._levels, device=z.device, dtype=z.dtype) - 1)
         scale = 2. / levels_minus_1
         bracket = (levels_minus_1 * (torch.tanh(z) + 1) / 2.) + 0.5
 
@@ -743,8 +804,8 @@ class FSQ(nn.Module):
         return codes_non_centered.float() * (2. / (self._levels.float() - 1)) - 1.
 
     def codes_to_indices(self, zhat):
-        zhat_normalized = (zhat + 1.) / (2. / (self._levels.to(zhat.dtype) - 1))
-        return (zhat_normalized * self._basis.to(zhat.dtype)).sum(dim=-1).round().to(torch.int32)
+        zhat_normalized = (zhat + 1.) / (2. / (comfy.model_management.cast_to(self._levels, device=zhat.device, dtype=zhat.dtype) - 1))
+        return (zhat_normalized * comfy.model_management.cast_to(self._basis, device=zhat.device, dtype=zhat.dtype)).sum(dim=-1).round().to(torch.int32)
 
     def forward(self, z):
         orig_dtype = z.dtype
@@ -826,7 +887,7 @@ class ResidualFSQ(nn.Module):
         x = self.project_in(x)
 
         if hasattr(self, 'soft_clamp_input_value'):
-            sc_val = self.soft_clamp_input_value.to(x.dtype)
+            sc_val = comfy.model_management.cast_to(self.soft_clamp_input_value, device=x.device, dtype=x.dtype)
             x = (x / sc_val).tanh() * sc_val
 
         quantized_out = torch.tensor(0., device=x.device, dtype=x.dtype)
@@ -834,7 +895,7 @@ class ResidualFSQ(nn.Module):
         all_indices = []
 
         for layer, scale in zip(self.layers, self.scales):
-            scale = scale.to(residual.dtype)
+            scale = comfy.model_management.cast_to(scale, device=x.device, dtype=x.dtype)
 
             quantized, indices = layer(residual / scale)
             quantized = quantized * scale
@@ -1035,28 +1096,26 @@ class AceStepConditionGenerationModel(nn.Module):
                     audio_codes = torch.nn.functional.pad(audio_codes, (0, math.ceil(src_latents.shape[1] / 5) - audio_codes.shape[1]), "constant", 35847)
                 lm_hints_5Hz = self.tokenizer.quantizer.get_output_from_indices(audio_codes, dtype=text_hidden_states.dtype)
             else:
-                assert False
-                # TODO ?
+                lm_hints_5Hz, indices = self.tokenizer.tokenize(refer_audio_acoustic_hidden_states_packed)
 
             lm_hints = self.detokenizer(lm_hints_5Hz)
 
         lm_hints = lm_hints[:, :src_latents.shape[1], :]
-        if is_covers is None:
+        if is_covers is None or is_covers is True:
             src_latents = lm_hints
-        else:
-            src_latents = torch.where(is_covers.unsqueeze(-1).unsqueeze(-1) > 0, lm_hints, src_latents)
+        elif is_covers is False:
+            src_latents = refer_audio_acoustic_hidden_states_packed
 
         context_latents = torch.cat([src_latents, chunk_masks.to(src_latents.dtype)], dim=-1)
 
         return encoder_hidden, encoder_mask, context_latents
 
-    def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, **kwargs):
+    def forward(self, x, timestep, context, lyric_embed=None, refer_audio=None, audio_codes=None, is_covers=None, replace_with_null_embeds=False, **kwargs):
         text_attention_mask = None
         lyric_attention_mask = None
         refer_audio_order_mask = None
         attention_mask = None
         chunk_masks = None
-        is_covers = None
         src_latents = None
         precomputed_lm_hints_25Hz = None
         lyric_hidden_states = lyric_embed
@@ -1068,7 +1127,7 @@ class AceStepConditionGenerationModel(nn.Module):
         if refer_audio_order_mask is None:
             refer_audio_order_mask = torch.zeros((x.shape[0],), device=x.device, dtype=torch.long)
 
-        if src_latents is None and is_covers is None:
+        if src_latents is None:
             src_latents = x
 
         if chunk_masks is None:
@@ -1081,6 +1140,9 @@ class AceStepConditionGenerationModel(nn.Module):
             src_latents, chunk_masks, is_covers, precomputed_lm_hints_25Hz=precomputed_lm_hints_25Hz, audio_codes=audio_codes
         )
 
+        if replace_with_null_embeds:
+            enc_hidden[:] = self.null_condition_emb.to(enc_hidden)
+
         out = self.decoder(hidden_states=x,
                            timestep=timestep,
                            timestep_r=timestep,

comfy/ldm/cosmos/predict2.py

@@ -335,7 +335,7 @@ class FinalLayer(nn.Module):
         device=None, dtype=None, operations=None
     ):
         super().__init__()
-        self.layer_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.layer_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
         self.linear = operations.Linear(
             hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, device=device, dtype=dtype
         )
@@ -463,6 +463,8 @@ class Block(nn.Module):
         extra_per_block_pos_emb: Optional[torch.Tensor] = None,
         transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
+        residual_dtype = x_B_T_H_W_D.dtype
+        compute_dtype = emb_B_T_D.dtype
         if extra_per_block_pos_emb is not None:
             x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
 
@@ -512,7 +514,7 @@ class Block(nn.Module):
         result_B_T_H_W_D = rearrange(
             self.self_attn(
                 # normalized_x_B_T_HW_D,
-                rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                rearrange(normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
                 None,
                 rope_emb=rope_emb_L_1_1_D,
                 transformer_options=transformer_options,
@@ -522,7 +524,7 @@ class Block(nn.Module):
             h=H,
             w=W,
         )
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D * result_B_T_H_W_D
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
 
         def _x_fn(
             _x_B_T_H_W_D: torch.Tensor,
@@ -536,7 +538,7 @@ class Block(nn.Module):
             )
             _result_B_T_H_W_D = rearrange(
                 self.cross_attn(
-                    rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                    rearrange(_normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
                     crossattn_emb,
                     rope_emb=rope_emb_L_1_1_D,
                     transformer_options=transformer_options,
@@ -555,7 +557,7 @@ class Block(nn.Module):
             shift_cross_attn_B_T_1_1_D,
             transformer_options=transformer_options,
         )
-        x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
+        x_B_T_H_W_D = result_B_T_H_W_D.to(residual_dtype) * gate_cross_attn_B_T_1_1_D.to(residual_dtype) + x_B_T_H_W_D
 
         normalized_x_B_T_H_W_D = _fn(
             x_B_T_H_W_D,
@@ -563,8 +565,8 @@ class Block(nn.Module):
             scale_mlp_B_T_1_1_D,
             shift_mlp_B_T_1_1_D,
         )
-        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D)
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D * result_B_T_H_W_D
+        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D.to(compute_dtype))
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
         return x_B_T_H_W_D
 
 
@@ -876,6 +878,14 @@ class MiniTrainDIT(nn.Module):
             "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
             "transformer_options": kwargs.get("transformer_options", {}),
         }
+
+        # The residual stream for this model has large values. To make fp16 compute_dtype work, we keep the residual stream
+        # in fp32, but run attention and MLP modules in fp16.
+        # An alternate method that clamps fp16 values "works" in the sense that it makes coherent images, but there is noticeable
+        # quality degradation and visual artifacts.
+        if x_B_T_H_W_D.dtype == torch.float16:
+            x_B_T_H_W_D = x_B_T_H_W_D.float()
+
         for block in self.blocks:
             x_B_T_H_W_D = block(
                 x_B_T_H_W_D,
@@ -884,6 +894,6 @@ class MiniTrainDIT(nn.Module):
                 **block_kwargs,
             )
 
-        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
+        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D.to(crossattn_emb.dtype), t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
         x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)[:, :, :orig_shape[-3], :orig_shape[-2], :orig_shape[-1]]
         return x_B_C_Tt_Hp_Wp

comfy/ldm/modules/attention.py

@@ -524,6 +524,9 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
 
 @wrap_attn
 def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    if kwargs.get("low_precision_attention", True) is False:
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=skip_reshape, skip_output_reshape=skip_output_reshape, **kwargs)
+
     exception_fallback = False
     if skip_reshape:
         b, _, _, dim_head = q.shape

comfy/model_base.py

@@ -147,11 +147,11 @@ class BaseModel(torch.nn.Module):
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
             logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
+            comfy.model_management.archive_model_dtypes(self.diffusion_model)
+
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
-        comfy.model_management.archive_model_dtypes(self.diffusion_model)
-
         self.adm_channels = unet_config.get("adm_in_channels", None)
         if self.adm_channels is None:
             self.adm_channels = 0
@@ -1548,9 +1548,12 @@ class ACEStep15(BaseModel):
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
         device = kwargs["device"]
+        noise = kwargs["noise"]
 
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
+            if torch.count_nonzero(cross_attn) == 0:
+                out['replace_with_null_embeds'] = comfy.conds.CONDConstant(True)
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
 
         conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
@@ -1559,27 +1562,26 @@ class ACEStep15(BaseModel):
 
         refer_audio = kwargs.get("reference_audio_timbre_latents", None)
         if refer_audio is None or len(refer_audio) == 0:
-            refer_audio = torch.tensor([[[-1.3672e-01, -1.5820e-01,  5.8594e-01, -5.7422e-01,  3.0273e-02,
-                                        2.7930e-01, -2.5940e-03, -2.0703e-01, -1.6113e-01, -1.4746e-01,
-                                        -2.7710e-02, -1.8066e-01, -2.9688e-01,  1.6016e+00, -2.6719e+00,
-                                        7.7734e-01, -1.3516e+00, -1.9434e-01, -7.1289e-02, -5.0938e+00,
-                                        2.4316e-01,  4.7266e-01,  4.6387e-02, -6.6406e-01, -2.1973e-01,
-                                        -6.7578e-01, -1.5723e-01,  9.5312e-01, -2.0020e-01, -1.7109e+00,
-                                        5.8984e-01, -5.7422e-01,  5.1562e-01,  2.8320e-01,  1.4551e-01,
-                                        -1.8750e-01, -5.9814e-02,  3.6719e-01, -1.0059e-01, -1.5723e-01,
-                                        2.0605e-01, -4.3359e-01, -8.2812e-01,  4.5654e-02, -6.6016e-01,
-                                        1.4844e-01,  9.4727e-02,  3.8477e-01, -1.2578e+00, -3.3203e-01,
-                                        -8.5547e-01,  4.3359e-01,  4.2383e-01, -8.9453e-01, -5.0391e-01,
-                                        -5.6152e-02, -2.9219e+00, -2.4658e-02,  5.0391e-01,  9.8438e-01,
-                                        7.2754e-02, -2.1582e-01,  6.3672e-01,  1.0000e+00]]], device=device).movedim(-1, 1).repeat(1, 1, 750)
+            refer_audio = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            pass_audio_codes = True
         else:
-            refer_audio = refer_audio[-1]
+            refer_audio = refer_audio[-1][:, :, :noise.shape[2]]
+            out['is_covers'] = comfy.conds.CONDConstant(True)
+            pass_audio_codes = False
+
+        if pass_audio_codes:
+            audio_codes = kwargs.get("audio_codes", None)
+            if audio_codes is not None:
+                out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
+                refer_audio = refer_audio[:, :, :750]
+            else:
+                out['is_covers'] = comfy.conds.CONDConstant(False)
+
+        if refer_audio.shape[2] < noise.shape[2]:
+            pad = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            refer_audio = torch.cat([refer_audio.to(pad), pad[:, :, refer_audio.shape[2]:]], dim=2)
+
         out['refer_audio'] = comfy.conds.CONDRegular(refer_audio)
-
-        audio_codes = kwargs.get("audio_codes", None)
-        if audio_codes is not None:
-            out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
-
         return out
 
 class Omnigen2(BaseModel):

comfy/model_management.py

@@ -19,7 +19,7 @@
 import psutil
 import logging
 from enum import Enum
-from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
+from comfy.cli_args import args, PerformanceFeature
 import threading
 import torch
 import sys
@@ -651,7 +651,7 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
                 soft_empty_cache()
     return unloaded_models
 
-def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
     cleanup_models_gc()
     global vram_state
 
@@ -747,26 +747,6 @@ def load_models_gpu_orig(models, memory_required=0, force_patch_weights=False, m
         current_loaded_models.insert(0, loaded_model)
     return
 
-def load_models_gpu_thread(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load):
-    with torch.inference_mode():
-        load_models_gpu_orig(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
-        soft_empty_cache()
-
-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):
-    #Deliberately load models outside of the Aimdo mempool so they can be retained accross
-    #nodes. Use a dummy thread to do it as pytorch documents that mempool contexts are
-    #thread local. So exploit that to escape context
-    if enables_dynamic_vram():
-        t = threading.Thread(
-            target=load_models_gpu_thread,
-            args=(models, memory_required, force_patch_weights, minimum_memory_required, force_full_load)
-        )
-        t.start()
-        t.join()
-    else:
-        load_models_gpu_orig(models, memory_required=memory_required, force_patch_weights=force_patch_weights,
-                             minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
-
 def load_model_gpu(model):
     return load_models_gpu([model])
 
@@ -1226,21 +1206,16 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
         if dtype is None:
             dtype = weight._model_dtype
 
-        r = torch.empty_like(weight, dtype=dtype, device=device)
-
         signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
         if signature is not None:
-            raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
-            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0]
+            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, weight._v_tensor)[0]
             if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
                 weight._v_signature = signature
                 #Send it over
                 v_tensor.copy_(weight, non_blocking=non_blocking)
-            #always take a deep copy even if _v is good, as we have no reasonable point to unpin
-            #a non comfy weight
-            r.copy_(v_tensor)
-            comfy_aimdo.model_vbar.vbar_unpin(weight._v)
-            return r
+            return v_tensor.to(dtype=dtype)
+
+        r = torch.empty_like(weight, dtype=dtype, device=device)
 
         if weight.dtype != r.dtype and weight.dtype != weight._model_dtype:
             #Offloaded casting could skip this, however it would make the quantizations

comfy/model_patcher.py

@@ -1400,7 +1400,7 @@ class ModelPatcher:
                 continue
             key = "diffusion_model." + k
             unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key))
-        return self.model.state_dict_for_saving(unet_state_dict)
+        return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
 
     def __del__(self):
         self.unpin_all_weights()
@@ -1492,7 +1492,9 @@ class ModelPatcherDynamic(ModelPatcher):
             if vbar is not None:
                 vbar.prioritize()
 
-            #We have way more tools for acceleration on comfy weight offloading, so always
+            #We force reserve VRAM for the non comfy-weight so we dont have to deal
+            #with pin and unpin syncrhonization which can be expensive for small weights
+            #with a high layer rate (e.g. autoregressive LLMs).
             #prioritize the non-comfy weights (note the order reverse).
             loading = self._load_list(prio_comfy_cast_weights=True)
             loading.sort(reverse=True)
@@ -1541,6 +1543,7 @@ class ModelPatcherDynamic(ModelPatcher):
 
                     if vbar is not None and not hasattr(m, "_v"):
                         m._v = vbar.alloc(v_weight_size)
+                        m._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(m._v, device_to)
                     allocated_size += v_weight_size
 
                 else:
@@ -1555,8 +1558,10 @@ class ModelPatcherDynamic(ModelPatcher):
                         weight_size = geometry.numel() * geometry.element_size()
                         if vbar is not None and not hasattr(weight, "_v"):
                             weight._v = vbar.alloc(weight_size)
+                            weight._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device_to)
                             weight._model_dtype = model_dtype
                         allocated_size += weight_size
+                    vbar.set_watermark_limit(allocated_size)
 
             logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.")
 

comfy/ops.py

@@ -54,6 +54,8 @@ try:
             SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
 
             def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+                if q.nelement() < 1024 * 128:  # arbitrary number, for small inputs cudnn attention seems slower
+                    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
                 with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
                     return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
         else:
@@ -85,7 +87,7 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
 
     signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
     if signature is not None:
-        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+        xfer_dest = s._v_tensor
     resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
 
     if not resident:

comfy/sd.py

@@ -976,7 +976,7 @@ class VAE:
         if overlap is not None:
             args["overlap"] = overlap
 
-        if dims == 1:
+        if dims == 1 or self.extra_1d_channel is not None:
             args.pop("tile_y")
             output = self.decode_tiled_1d(samples, **args)
         elif dims == 2:

comfy/supported_models.py

@@ -993,7 +993,7 @@ class CosmosT2IPredict2(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
@@ -1023,11 +1023,7 @@ class Anima(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
-
-    def __init__(self, unet_config):
-        super().__init__(unet_config)
-        self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.Anima(self, device=device)
@@ -1038,6 +1034,12 @@ class Anima(supported_models_base.BASE):
         detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_06b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.anima.AnimaTokenizer, comfy.text_encoders.anima.te(**detect))
 
+    def set_inference_dtype(self, dtype, manual_cast_dtype, **kwargs):
+        self.memory_usage_factor = (self.unet_config.get("model_channels", 2048) / 2048) * 0.95
+        if dtype is torch.float16:
+            self.memory_usage_factor *= 1.4
+        return super().set_inference_dtype(dtype, manual_cast_dtype, **kwargs)
+
 class CosmosI2VPredict2(CosmosT2IPredict2):
     unet_config = {
         "image_model": "cosmos_predict2",

comfy/text_encoders/ace15.py

@@ -3,6 +3,7 @@ import comfy.text_encoders.llama
 from comfy import sd1_clip
 import torch
 import math
+import yaml
 import comfy.utils
 
 
@@ -101,9 +102,7 @@ def sample_manual_loop_no_classes(
     return output_audio_codes
 
 
-def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0):
-    cfg_scale = 2.0
-
+def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0, cfg_scale=2.0, temperature=0.85, top_p=0.9, top_k=0):
     positive = [[token for token, _ in inner_list] for inner_list in positive]
     negative = [[token for token, _ in inner_list] for inner_list in negative]
     positive = positive[0]
@@ -120,34 +119,80 @@ def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=102
         positive = [model.special_tokens["pad"]] * pos_pad + positive
 
     paddings = [pos_pad, neg_pad]
-    return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
+    return sample_manual_loop_no_classes(model, [positive, negative], paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
 
 
 class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer)
 
+    def _metas_to_cot(self, *, return_yaml: bool = False, **kwargs) -> str:
+        user_metas = {
+            k: kwargs.pop(k)
+            for k in ("bpm", "duration", "keyscale", "timesignature", "language", "caption")
+            if k in kwargs
+        }
+        timesignature = user_metas.get("timesignature")
+        if isinstance(timesignature, str) and timesignature.endswith("/4"):
+            user_metas["timesignature"] = timesignature.rsplit("/", 1)[0]
+        user_metas = {
+            k: v if not isinstance(v, str) or not v.isdigit() else int(v)
+            for k, v in user_metas.items()
+            if v not in {"unspecified", None}
+        }
+        if len(user_metas):
+            meta_yaml = yaml.dump(user_metas, allow_unicode=True, sort_keys=True).strip()
+        else:
+            meta_yaml = ""
+        return f"<think>\n{meta_yaml}\n</think>" if not return_yaml else meta_yaml
+
+    def _metas_to_cap(self, **kwargs) -> str:
+        use_keys = ("bpm", "duration", "keyscale", "timesignature")
+        user_metas = { k: kwargs.pop(k, "N/A") for k in use_keys }
+        duration = user_metas["duration"]
+        if duration == "N/A":
+            user_metas["duration"] = "30 seconds"
+        elif isinstance(duration, (str, int, float)):
+            user_metas["duration"] = f"{math.ceil(float(duration))} seconds"
+        else:
+            raise TypeError("Unexpected type for duration key, must be str, int or float")
+        return "\n".join(f"- {k}: {user_metas[k]}" for k in use_keys)
+
     def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
         out = {}
         lyrics = kwargs.get("lyrics", "")
-        bpm = kwargs.get("bpm", 120)
         duration = kwargs.get("duration", 120)
-        keyscale = kwargs.get("keyscale", "C major")
-        timesignature = kwargs.get("timesignature", 2)
-        language = kwargs.get("language", "en")
+        language = kwargs.get("language")
         seed = kwargs.get("seed", 0)
 
+        generate_audio_codes = kwargs.get("generate_audio_codes", True)
+        cfg_scale = kwargs.get("cfg_scale", 2.0)
+        temperature = kwargs.get("temperature", 0.85)
+        top_p = kwargs.get("top_p", 0.9)
+        top_k = kwargs.get("top_k", 0.0)
+
+
         duration = math.ceil(duration)
-        meta_lm = 'bpm: {}\nduration: {}\nkeyscale: {}\ntimesignature: {}'.format(bpm, duration, keyscale, timesignature)
-        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n{}\n<|im_end|>\n<|im_start|>assistant\n<think>\n{}\n</think>\n\n<|im_end|>\n"
+        kwargs["duration"] = duration
 
-        meta_cap = '- bpm: {}\n- timesignature: {}\n- keyscale: {}\n- duration: {}\n'.format(bpm, timesignature, keyscale, duration)
-        out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, meta_lm), disable_weights=True)
-        out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, ""), disable_weights=True)
+        cot_text = self._metas_to_cot(caption = text, **kwargs)
+        meta_cap = self._metas_to_cap(**kwargs)
 
-        out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric{}<|endoftext|><|endoftext|>".format(language, lyrics), return_word_ids, disable_weights=True, **kwargs)
-        out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}# Metas\n{}<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs)
-        out["lm_metadata"] = {"min_tokens": duration * 5, "seed": seed}
+        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n# Lyric\n{}\n<|im_end|>\n<|im_start|>assistant\n{}\n<|im_end|>\n"
+
+        out["lm_prompt"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, cot_text), disable_weights=True)
+        out["lm_prompt_negative"] = self.qwen3_06b.tokenize_with_weights(lm_template.format(text, lyrics, "<think>\n</think>"), disable_weights=True)
+
+        out["lyrics"] = self.qwen3_06b.tokenize_with_weights("# Languages\n{}\n\n# Lyric\n{}<|endoftext|><|endoftext|>".format(language if language is not None else "", lyrics), return_word_ids, disable_weights=True, **kwargs)
+        out["qwen3_06b"] = self.qwen3_06b.tokenize_with_weights("# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}\n# Metas\n{}\n<|endoftext|>\n<|endoftext|>".format(text, meta_cap), return_word_ids, **kwargs)
+        out["lm_metadata"] = {"min_tokens": duration * 5,
+                              "seed": seed,
+                              "generate_audio_codes": generate_audio_codes,
+                              "cfg_scale": cfg_scale,
+                              "temperature": temperature,
+                              "top_p": top_p,
+                              "top_k": top_k,
+                              }
         return out
 
 
@@ -203,10 +248,14 @@ class ACE15TEModel(torch.nn.Module):
         self.qwen3_06b.set_clip_options({"layer": [0]})
         lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics)
 
-        lm_metadata = token_weight_pairs["lm_metadata"]
-        audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"])
+        out = {"conditioning_lyrics": lyrics_embeds[:, 0]}
 
-        return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]}
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        if lm_metadata["generate_audio_codes"]:
+            audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"], cfg_scale=lm_metadata["cfg_scale"], temperature=lm_metadata["temperature"], top_p=lm_metadata["top_p"], top_k=lm_metadata["top_k"])
+            out["audio_codes"] = [audio_codes]
+
+        return base_out, None, out
 
     def set_clip_options(self, options):
         self.qwen3_06b.set_clip_options(options)

comfy/text_encoders/anima.py

@@ -23,7 +23,7 @@ class AnimaTokenizer:
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
         qwen_ids = self.qwen3_06b.tokenize_with_weights(text, return_word_ids, **kwargs)
-        out["qwen3_06b"] = [[(token, 1.0) for token, _ in inner_list] for inner_list in qwen_ids]  # Set weights to 1.0
+        out["qwen3_06b"] = [[(k[0], 1.0, k[2]) if return_word_ids else (k[0], 1.0) for k in inner_list] for inner_list in qwen_ids]  # Set weights to 1.0
         out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 

comfy/text_encoders/llama.py

@@ -651,10 +651,10 @@ class Llama2_(nn.Module):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1])
-            mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min)
+            mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min / 4)
 
         if seq_len > 1:
-            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min).triu_(1)
+            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min / 4).triu_(1)
             if mask is not None:
                 mask += causal_mask
             else:

comfy/text_encoders/lt.py

@@ -25,7 +25,7 @@ def ltxv_te(*args, **kwargs):
 class Gemma3_12BTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}

comfy/utils.py

@@ -82,14 +82,12 @@ _TYPES = {
 def load_safetensors(ckpt):
     f = open(ckpt, "rb")
     mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
+    mv = memoryview(mapping)
 
     header_size = struct.unpack("<Q", mapping[:8])[0]
     header = json.loads(mapping[8:8+header_size].decode("utf-8"))
 
-    with warnings.catch_warnings():
-        #We are working with read-only RAM by design
-        warnings.filterwarnings("ignore", message="The given buffer is not writable")
-        data_area = torch.frombuffer(mapping, dtype=torch.uint8)[8 + header_size:]
+    mv = mv[8 + header_size:]
 
     sd = {}
     for name, info in header.items():
@@ -97,7 +95,13 @@ def load_safetensors(ckpt):
             continue
 
         start, end = info["data_offsets"]
-        sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"])
+        if start == end:
+            sd[name] = torch.empty(info["shape"], dtype =_TYPES[info["dtype"]])
+        else:
+            with warnings.catch_warnings():
+                #We are working with read-only RAM by design
+                warnings.filterwarnings("ignore", message="The given buffer is not writable")
+                sd[name] = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
 
     return sd, header.get("__metadata__", {}),
 

comfy_extras/nodes_ace.py

@@ -44,13 +44,18 @@ class TextEncodeAceStepAudio15(io.ComfyNode):
                 io.Combo.Input("timesignature", options=['2', '3', '4', '6']),
                 io.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]),
                 io.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]),
+                io.Boolean.Input("generate_audio_codes", default=True, tooltip="Enable the LLM that generates audio codes. This can be slow but will increase the quality of the generated audio. Turn this off if you are giving the model an audio reference.", advanced=True),
+                io.Float.Input("cfg_scale", default=2.0, min=0.0, max=100.0, step=0.1, advanced=True),
+                io.Float.Input("temperature", default=0.85, min=0.0, max=2.0, step=0.01, advanced=True),
+                io.Float.Input("top_p", default=0.9, min=0.0, max=2000.0, step=0.01, advanced=True),
+                io.Int.Input("top_k", default=0, min=0, max=100, advanced=True),
             ],
             outputs=[io.Conditioning.Output()],
         )
 
     @classmethod
-    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale) -> io.NodeOutput:
-        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed)
+    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale, generate_audio_codes, cfg_scale, temperature, top_p, top_k) -> io.NodeOutput:
+        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed, generate_audio_codes=generate_audio_codes, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k)
         conditioning = clip.encode_from_tokens_scheduled(tokens)
         return io.NodeOutput(conditioning)
 
@@ -100,14 +105,15 @@ class EmptyAceStep15LatentAudio(io.ComfyNode):
         latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device())
         return io.NodeOutput({"samples": latent, "type": "audio"})
 
-class ReferenceTimbreAudio(io.ComfyNode):
+class ReferenceAudio(io.ComfyNode):
     @classmethod
     def define_schema(cls):
         return io.Schema(
             node_id="ReferenceTimbreAudio",
+            display_name="Reference Audio",
             category="advanced/conditioning/audio",
             is_experimental=True,
-            description="This node sets the reference audio for timbre (for ace step 1.5)",
+            description="This node sets the reference audio for ace step 1.5",
             inputs=[
                 io.Conditioning.Input("conditioning"),
                 io.Latent.Input("latent", optional=True),
@@ -131,7 +137,7 @@ class AceExtension(ComfyExtension):
             EmptyAceStepLatentAudio,
             TextEncodeAceStepAudio15,
             EmptyAceStep15LatentAudio,
-            ReferenceTimbreAudio,
+            ReferenceAudio,
         ]
 
 async def comfy_entrypoint() -> AceExtension:

comfy_extras/nodes_audio.py

@@ -94,6 +94,19 @@ class VAEEncodeAudio(IO.ComfyNode):
     encode = execute  # TODO: remove
 
 
+def vae_decode_audio(vae, samples, tile=None, overlap=None):
+    if tile is not None:
+        audio = vae.decode_tiled(samples["samples"], tile_y=tile, overlap=overlap).movedim(-1, 1)
+    else:
+        audio = vae.decode(samples["samples"]).movedim(-1, 1)
+
+    std = torch.std(audio, dim=[1, 2], keepdim=True) * 5.0
+    std[std < 1.0] = 1.0
+    audio /= std
+    vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+    return {"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]}
+
+
 class VAEDecodeAudio(IO.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -111,16 +124,33 @@ class VAEDecodeAudio(IO.ComfyNode):
 
     @classmethod
     def execute(cls, vae, samples) -> IO.NodeOutput:
-        audio = vae.decode(samples["samples"]).movedim(-1, 1)
-        std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0
-        std[std < 1.0] = 1.0
-        audio /= std
-        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
-        return IO.NodeOutput({"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]})
+        return IO.NodeOutput(vae_decode_audio(vae, samples))
 
     decode = execute  # TODO: remove
 
 
+class VAEDecodeAudioTiled(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="VAEDecodeAudioTiled",
+            search_aliases=["latent to audio"],
+            display_name="VAE Decode Audio (Tiled)",
+            category="latent/audio",
+            inputs=[
+                IO.Latent.Input("samples"),
+                IO.Vae.Input("vae"),
+                IO.Int.Input("tile_size", default=512, min=32, max=8192, step=8),
+                IO.Int.Input("overlap", default=64, min=0, max=1024, step=8),
+            ],
+            outputs=[IO.Audio.Output()],
+        )
+
+    @classmethod
+    def execute(cls, vae, samples, tile_size, overlap) -> IO.NodeOutput:
+        return IO.NodeOutput(vae_decode_audio(vae, samples, tile_size, overlap))
+
+
 class SaveAudio(IO.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -675,6 +705,7 @@ class AudioExtension(ComfyExtension):
             EmptyLatentAudio,
             VAEEncodeAudio,
             VAEDecodeAudio,
+            VAEDecodeAudioTiled,
             SaveAudio,
             SaveAudioMP3,
             SaveAudioOpus,

comfy_extras/nodes_custom_sampler.py

@@ -622,6 +622,7 @@ class SamplerSASolver(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerSASolver",
+            search_aliases=["sde"],
             category="sampling/custom_sampling/samplers",
             inputs=[
                 io.Model.Input("model"),
@@ -666,6 +667,7 @@ class SamplerSEEDS2(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerSEEDS2",
+            search_aliases=["sde", "exp heun"],
             category="sampling/custom_sampling/samplers",
             inputs=[
                 io.Combo.Input("solver_type", options=["phi_1", "phi_2"]),

comfy_extras/nodes_easycache.py

@@ -9,6 +9,14 @@ if TYPE_CHECKING:
     from uuid import UUID
 
 
+def _extract_tensor(data, output_channels):
+    """Extract tensor from data, handling both single tensors and lists."""
+    if isinstance(data, list):
+        # LTX2 AV tensors: [video, audio]
+        return data[0][:, :output_channels], data[1][:, :output_channels]
+    return data[:, :output_channels], None
+
+
 def easycache_forward_wrapper(executor, *args, **kwargs):
     # get values from args
     transformer_options: dict[str] = args[-1]
@@ -17,7 +25,7 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
         if not transformer_options:
             transformer_options = args[-2]
     easycache: EasyCacheHolder = transformer_options["easycache"]
-    x: torch.Tensor = args[0][:, :easycache.output_channels]
+    x, ax = _extract_tensor(args[0], easycache.output_channels)
     sigmas = transformer_options["sigmas"]
     uuids = transformer_options["uuids"]
     if sigmas is not None and easycache.is_past_end_timestep(sigmas):
@@ -35,7 +43,11 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
         if easycache.skip_current_step and can_apply_cache_diff:
             if easycache.verbose:
                 logging.info(f"EasyCache [verbose] - was marked to skip this step by {easycache.first_cond_uuid}. Present uuids: {uuids}")
-            return easycache.apply_cache_diff(x, uuids)
+            result = easycache.apply_cache_diff(x, uuids)
+            if ax is not None:
+                result_audio = easycache.apply_cache_diff(ax, uuids, is_audio=True)
+                return [result, result_audio]
+            return result
         if easycache.initial_step:
             easycache.first_cond_uuid = uuids[0]
             has_first_cond_uuid = easycache.has_first_cond_uuid(uuids)
@@ -51,13 +63,18 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
                         logging.info(f"EasyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
                     # other conds should also skip this step, and instead use their cached values
                     easycache.skip_current_step = True
-                    return easycache.apply_cache_diff(x, uuids)
+                    result = easycache.apply_cache_diff(x, uuids)
+                    if ax is not None:
+                        result_audio = easycache.apply_cache_diff(ax, uuids, is_audio=True)
+                        return [result, result_audio]
+                    return result
                 else:
                     if easycache.verbose:
                         logging.info(f"EasyCache [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
                     easycache.cumulative_change_rate = 0.0
 
-    output: torch.Tensor = executor(*args, **kwargs)
+    full_output: torch.Tensor = executor(*args, **kwargs)
+    output, audio_output = _extract_tensor(full_output, easycache.output_channels)
     if has_first_cond_uuid and easycache.has_output_prev_norm():
         output_change = (easycache.subsample(output, uuids, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean()
         if easycache.verbose:
@@ -74,13 +91,15 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
             logging.info(f"EasyCache [verbose] - output_change_rate: {output_change_rate}")
     # TODO: allow cache_diff to be offloaded
     easycache.update_cache_diff(output, next_x_prev, uuids)
+    if audio_output is not None:
+        easycache.update_cache_diff(audio_output, ax, uuids, is_audio=True)
     if has_first_cond_uuid:
         easycache.x_prev_subsampled = easycache.subsample(next_x_prev, uuids)
         easycache.output_prev_subsampled = easycache.subsample(output, uuids)
         easycache.output_prev_norm = output.flatten().abs().mean()
         if easycache.verbose:
             logging.info(f"EasyCache [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}")
-    return output
+    return full_output
 
 def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
     # get values from args
@@ -89,8 +108,8 @@ def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
     easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"]
     if easycache.is_past_end_timestep(timestep):
         return executor(*args, **kwargs)
-    # prepare next x_prev
     x: torch.Tensor = args[0][:, :easycache.output_channels]
+    # prepare next x_prev
     next_x_prev = x
     input_change = None
     do_easycache = easycache.should_do_easycache(timestep)
@@ -197,6 +216,7 @@ class EasyCacheHolder:
         self.output_prev_subsampled: torch.Tensor = None
         self.output_prev_norm: torch.Tensor = None
         self.uuid_cache_diffs: dict[UUID, torch.Tensor] = {}
+        self.uuid_cache_diffs_audio: dict[UUID, torch.Tensor] = {}
         self.output_change_rates = []
         self.approx_output_change_rates = []
         self.total_steps_skipped = 0
@@ -245,20 +265,21 @@ class EasyCacheHolder:
     def can_apply_cache_diff(self, uuids: list[UUID]) -> bool:
         return all(uuid in self.uuid_cache_diffs for uuid in uuids)
 
-    def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID]):
-        if self.first_cond_uuid in uuids:
+    def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID], is_audio: bool = False):
+        if self.first_cond_uuid in uuids and not is_audio:
             self.total_steps_skipped += 1
+        cache_diffs = self.uuid_cache_diffs_audio if is_audio else self.uuid_cache_diffs
         batch_offset = x.shape[0] // len(uuids)
         for i, uuid in enumerate(uuids):
             # slice out only what is relevant to this cond
             batch_slice = [slice(i*batch_offset,(i+1)*batch_offset)]
             # if cached dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
-            if x.shape[1:] != self.uuid_cache_diffs[uuid].shape[1:]:
+            if x.shape[1:] != cache_diffs[uuid].shape[1:]:
                 if not self.allow_mismatch:
                     raise ValueError(f"Cached dims {self.uuid_cache_diffs[uuid].shape} don't match x dims {x.shape} - this is no good")
                 slicing = []
                 skip_this_dim = True
-                for dim_u, dim_x in zip(self.uuid_cache_diffs[uuid].shape, x.shape):
+                for dim_u, dim_x in zip(cache_diffs[uuid].shape, x.shape):
                     if skip_this_dim:
                         skip_this_dim = False
                         continue
@@ -270,10 +291,11 @@ class EasyCacheHolder:
                     else:
                         slicing.append(slice(None))
                 batch_slice = batch_slice + slicing
-            x[tuple(batch_slice)] += self.uuid_cache_diffs[uuid].to(x.device)
+            x[tuple(batch_slice)] += cache_diffs[uuid].to(x.device)
         return x
 
-    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]):
+    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID], is_audio: bool = False):
+        cache_diffs = self.uuid_cache_diffs_audio if is_audio else self.uuid_cache_diffs
         # if output dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
         if output.shape[1:] != x.shape[1:]:
             if not self.allow_mismatch:
@@ -293,7 +315,7 @@ class EasyCacheHolder:
         diff = output - x
         batch_offset = diff.shape[0] // len(uuids)
         for i, uuid in enumerate(uuids):
-            self.uuid_cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...]
+            cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...]
 
     def has_first_cond_uuid(self, uuids: list[UUID]) -> bool:
         return self.first_cond_uuid in uuids
@@ -324,6 +346,8 @@ class EasyCacheHolder:
         self.output_prev_norm = None
         del self.uuid_cache_diffs
         self.uuid_cache_diffs = {}
+        del self.uuid_cache_diffs_audio
+        self.uuid_cache_diffs_audio = {}
         self.total_steps_skipped = 0
         self.state_metadata = None
         return self

comfy_extras/nodes_hunyuan3d.py

@@ -618,6 +618,7 @@ class SaveGLB(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="SaveGLB",
+            display_name="Save 3D Model",
             search_aliases=["export 3d model", "save mesh"],
             category="3d",
             is_output_node=True,
@@ -626,8 +627,14 @@ class SaveGLB(IO.ComfyNode):
                     IO.Mesh.Input("mesh"),
                     types=[
                         IO.File3DGLB,
+                        IO.File3DGLTF,
+                        IO.File3DOBJ,
+                        IO.File3DFBX,
+                        IO.File3DSTL,
+                        IO.File3DUSDZ,
+                        IO.File3DAny,
                     ],
-                    tooltip="Mesh or GLB file to save",
+                    tooltip="Mesh or 3D file to save",
                 ),
                 IO.String.Input("filename_prefix", default="mesh/ComfyUI"),
             ],
@@ -649,7 +656,8 @@ class SaveGLB(IO.ComfyNode):
 
         if isinstance(mesh, Types.File3D):
             # Handle File3D input - save BytesIO data to output folder
-            f = f"{filename}_{counter:05}_.glb"
+            ext = mesh.format or "glb"
+            f = f"{filename}_{counter:05}_.{ext}"
             mesh.save_to(os.path.join(full_output_folder, f))
             results.append({
                 "filename": f,

comfy_extras/nodes_latent.py

@@ -391,8 +391,9 @@ class LatentOperationTonemapReinhard(io.ComfyNode):
             latent_vector_magnitude = (torch.linalg.vector_norm(latent, dim=(1)) + 0.0000000001)[:,None]
             normalized_latent = latent / latent_vector_magnitude
 
-            mean = torch.mean(latent_vector_magnitude, dim=(1,2,3), keepdim=True)
-            std = torch.std(latent_vector_magnitude, dim=(1,2,3), keepdim=True)
+            dims = list(range(1, latent_vector_magnitude.ndim))
+            mean = torch.mean(latent_vector_magnitude, dim=dims, keepdim=True)
+            std = torch.std(latent_vector_magnitude, dim=dims, keepdim=True)
 
             top = (std * 5 + mean) * multiplier
 

comfy_extras/nodes_load_3d.py

@@ -45,6 +45,7 @@ class Load3D(IO.ComfyNode):
                 IO.Image.Output(display_name="normal"),
                 IO.Load3DCamera.Output(display_name="camera_info"),
                 IO.Video.Output(display_name="recording_video"),
+                IO.File3DAny.Output(display_name="model_3d"),
             ],
         )
 
@@ -66,7 +67,8 @@ class Load3D(IO.ComfyNode):
 
             video = InputImpl.VideoFromFile(recording_video_path)
 
-        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video)
+        file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file))
+        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video, file_3d)
 
     process = execute  # TODO: remove
 

comfy_extras/nodes_toolkit.py

@@ -0,0 +1,47 @@
+from __future__ import annotations
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, io
+
+
+class CreateList(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        template_matchtype = io.MatchType.Template("type")
+        template_autogrow = io.Autogrow.TemplatePrefix(
+            input=io.MatchType.Input("input", template=template_matchtype),
+            prefix="input",
+        )
+        return io.Schema(
+            node_id="CreateList",
+            display_name="Create List",
+            category="logic",
+            is_input_list=True,
+            search_aliases=["Image Iterator", "Text Iterator", "Iterator"],
+            inputs=[io.Autogrow.Input("inputs", template=template_autogrow)],
+            outputs=[
+                io.MatchType.Output(
+                    template=template_matchtype,
+                    is_output_list=True,
+                    display_name="list",
+                ),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, inputs: io.Autogrow.Type) -> io.NodeOutput:
+        output_list = []
+        for input in inputs.values():
+            output_list += input
+        return io.NodeOutput(output_list)
+
+
+class ToolkitExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            CreateList,
+        ]
+
+
+async def comfy_entrypoint() -> ToolkitExtension:
+    return ToolkitExtension()

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.12.2"
+__version__ = "0.12.3"

execution.py

@@ -13,8 +13,11 @@ from contextlib import nullcontext
 
 import torch
 
+from comfy.cli_args import args
 import comfy.memory_management
 import comfy.model_management
+import comfy_aimdo.model_vbar
+
 from latent_preview import set_preview_method
 import nodes
 from comfy_execution.caching import (
@@ -527,8 +530,10 @@ async def execute(server, dynprompt, caches, current_item, extra_data, executed,
                     output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
                 finally:
                     if allocator is not None:
+                        if args.verbose == "DEBUG":
+                            comfy_aimdo.model_vbar.vbars_analyze()
                         comfy.model_management.reset_cast_buffers()
-                        torch.cuda.synchronize()
+                        comfy_aimdo.model_vbar.vbars_reset_watermark_limits()
 
             if has_pending_tasks:
                 pending_async_nodes[unique_id] = output_data

nodes.py

@@ -2433,7 +2433,8 @@ async def init_builtin_extra_nodes():
         "nodes_image_compare.py",
         "nodes_zimage.py",
         "nodes_lora_debug.py",
-        "nodes_color.py"
+        "nodes_color.py",
+        "nodes_toolkit.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.12.2"
+version = "0.12.3"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,6 +1,6 @@
-comfyui-frontend-package==1.37.11
+comfyui-frontend-package==1.38.13
 comfyui-workflow-templates==0.8.31
-comfyui-embedded-docs==0.4.0
+comfyui-embedded-docs==0.4.1
 torch
 torchsde
 torchvision
@@ -22,7 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
-comfy-aimdo>=0.1.7
+comfy-aimdo>=0.1.8
 requests
 
 #non essential dependencies: