Merge branch 'master' into asset-management

Author: Kosinkadink

comfy/ldm/wan/model.py

@@ -1355,7 +1355,7 @@ def forward(self, x, context, transformer_options={}, **kwargs):
 
         x = optimized_attention(q, k, v, heads=self.num_heads, skip_reshape=True, skip_output_reshape=True, transformer_options=transformer_options)
 
-        x = x.transpose(1, 2).view(b, -1, n, d).flatten(2)
+        x = x.transpose(1, 2).reshape(b, -1, n * d)
         x = self.o(x)
         return x
 
@@ -1551,6 +1551,9 @@ def forward_orig(
         context_img_len = None
 
         if audio_embed is not None:
+            if reference_latent is not None:
+                zero_audio_pad = torch.zeros(audio_embed.shape[0], reference_latent.shape[-3], *audio_embed.shape[2:], device=audio_embed.device, dtype=audio_embed.dtype)
+                audio_embed = torch.cat([audio_embed, zero_audio_pad], dim=1)
             audio = self.audio_proj(audio_embed).permute(0, 3, 1, 2).flatten(2).transpose(1, 2)
         else:
             audio = None

comfy/ldm/wan/model_animate.py

@@ -39,6 +39,7 @@
 import comfy.ldm.cosmos.predict2
 import comfy.ldm.lumina.model
 import comfy.ldm.wan.model
+import comfy.ldm.wan.model_animate
 import comfy.ldm.hunyuan3d.model
 import comfy.ldm.hidream.model
 import comfy.ldm.chroma.model
@@ -1253,6 +1254,23 @@ def extra_conds(self, **kwargs):
 
         return out
 
+class WAN22_Animate(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model_animate.AnimateWanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        face_video_pixels = kwargs.get("face_video_pixels", None)
+        if face_video_pixels is not None:
+            out['face_pixel_values'] = comfy.conds.CONDRegular(face_video_pixels)
+
+        pose_latents = kwargs.get("pose_video_latent", None)
+        if pose_latents is not None:
+            out['pose_latents'] = comfy.conds.CONDRegular(self.process_latent_in(pose_latents))
+        return out
+
 class WAN22_S2V(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel_S2V)

comfy/model_base.py

@@ -404,6 +404,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["model_type"] = "s2v"
         elif '{}audio_proj.audio_proj_glob_1.layer.bias'.format(key_prefix) in state_dict_keys:
             dit_config["model_type"] = "humo"
+        elif '{}face_adapter.fuser_blocks.0.k_norm.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "animate"
         else:
             if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
                 dit_config["model_type"] = "i2v"

comfy/model_detection.py

@@ -348,7 +348,7 @@ def amd_min_version(device=None, min_rdna_version=0):
 #                    if any((a in arch) for a in ["gfx1201"]):
 #                        ENABLE_PYTORCH_ATTENTION = True
         if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4):
-            if any((a in arch) for a in ["gfx1201", "gfx942", "gfx950"]):  # TODO: more arches
+            if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx942", "gfx950"]):  # TODO: more arches
                 SUPPORT_FP8_OPS = True
 
 except:
@@ -645,7 +645,9 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             if loaded_model.model.is_clone(current_loaded_models[i].model):
                 to_unload = [i] + to_unload
         for i in to_unload:
-            current_loaded_models.pop(i).model.detach(unpatch_all=False)
+            model_to_unload = current_loaded_models.pop(i)
+            model_to_unload.model.detach(unpatch_all=False)
+            model_to_unload.model_finalizer.detach()
 
     total_memory_required = {}
     for loaded_model in models_to_load:

comfy/model_management.py

@@ -365,12 +365,13 @@ def reset_parameters(self):
             return None
 
         def forward_comfy_cast_weights(self, input):
-            try:
-                out = fp8_linear(self, input)
-                if out is not None:
-                    return out
-            except Exception as e:
-                logging.info("Exception during fp8 op: {}".format(e))
+            if not self.training:
+                try:
+                    out = fp8_linear(self, input)
+                    if out is not None:
+                        return out
+                except Exception as e:
+                    logging.info("Exception during fp8 op: {}".format(e))
 
             weight, bias = cast_bias_weight(self, input)
             return torch.nn.functional.linear(input, weight, bias)

comfy/ops.py

@@ -995,7 +995,7 @@ class WAN21_T2V(supported_models_base.BASE):
     unet_extra_config = {}
     latent_format = latent_formats.Wan21
 
-    memory_usage_factor = 1.0
+    memory_usage_factor = 0.9
 
     supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
 
@@ -1004,7 +1004,7 @@ class WAN21_T2V(supported_models_base.BASE):
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
-        self.memory_usage_factor = self.unet_config.get("dim", 2000) / 2000
+        self.memory_usage_factor = self.unet_config.get("dim", 2000) / 2222
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN21(self, device=device)
@@ -1096,6 +1096,19 @@ def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN22_S2V(self, device=device)
         return out
 
+class WAN22_Animate(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "animate",
+    }
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN22_Animate(self, device=device)
+        return out
+
 class WAN22_T2V(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
@@ -1361,6 +1374,6 @@ def get_model(self, state_dict, prefix="", device=None):
         out = model_base.HunyuanImage21Refiner(self, device=device)
         return out
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage]
 
 models += [SVD_img2vid]

comfy/supported_models.py

@@ -400,21 +400,25 @@ def preprocess_embed(self, embed, device):
 
     def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
         grid = None
+        position_ids = None
+        offset = 0
         for e in embeds_info:
             if e.get("type") == "image":
                 grid = e.get("extra", None)
-                position_ids = torch.zeros((3, embeds.shape[1]), device=embeds.device)
                 start = e.get("index")
-                position_ids[:, :start] = torch.arange(0, start, device=embeds.device)
+                if position_ids is None:
+                    position_ids = torch.zeros((3, embeds.shape[1]), device=embeds.device)
+                    position_ids[:, :start] = torch.arange(0, start, device=embeds.device)
                 end = e.get("size") + start
                 len_max = int(grid.max()) // 2
                 start_next = len_max + start
-                position_ids[:, end:] = torch.arange(start_next, start_next + (embeds.shape[1] - end), device=embeds.device)
-                position_ids[0, start:end] = start
+                position_ids[:, end:] = torch.arange(start_next + offset, start_next + (embeds.shape[1] - end) + offset, device=embeds.device)
+                position_ids[0, start:end] = start + offset
                 max_d = int(grid[0][1]) // 2
-                position_ids[1, start:end] = torch.arange(start, start + max_d, device=embeds.device).unsqueeze(1).repeat(1, math.ceil((end - start) / max_d)).flatten(0)[:end - start]
+                position_ids[1, start:end] = torch.arange(start + offset, start + max_d + offset, device=embeds.device).unsqueeze(1).repeat(1, math.ceil((end - start) / max_d)).flatten(0)[:end - start]
                 max_d = int(grid[0][2]) // 2
-                position_ids[2, start:end] = torch.arange(start, start + max_d, device=embeds.device).unsqueeze(0).repeat(math.ceil((end - start) / max_d), 1).flatten(0)[:end - start]
+                position_ids[2, start:end] = torch.arange(start + offset, start + max_d + offset, device=embeds.device).unsqueeze(0).repeat(math.ceil((end - start) / max_d), 1).flatten(0)[:end - start]
+                offset += len_max - (end - start)
 
         if grid is None:
             position_ids = None

comfy/text_encoders/llama.py

@@ -130,12 +130,12 @@ def __init__(self, loaded_keys, weights):
     def create_train(cls, weight, rank=1, alpha=1.0):
         out_dim = weight.shape[0]
         in_dim = weight.shape[1:].numel()
-        mat1 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
-        mat2 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        mat1 = torch.empty(out_dim, rank, device=weight.device, dtype=torch.float32)
+        mat2 = torch.empty(rank, in_dim, device=weight.device, dtype=torch.float32)
         torch.nn.init.normal_(mat1, 0.1)
         torch.nn.init.constant_(mat2, 0.0)
-        mat3 = torch.empty(out_dim, rank, device=weight.device, dtype=weight.dtype)
-        mat4 = torch.empty(rank, in_dim, device=weight.device, dtype=weight.dtype)
+        mat3 = torch.empty(out_dim, rank, device=weight.device, dtype=torch.float32)
+        mat4 = torch.empty(rank, in_dim, device=weight.device, dtype=torch.float32)
         torch.nn.init.normal_(mat3, 0.1)
         torch.nn.init.normal_(mat4, 0.01)
         return LohaDiff(

comfy/weight_adapter/loha.py

@@ -89,8 +89,8 @@ def create_train(cls, weight, rank=1, alpha=1.0):
         in_dim = weight.shape[1:].numel()
         out1, out2 = factorization(out_dim, rank)
         in1, in2 = factorization(in_dim, rank)
-        mat1 = torch.empty(out1, in1, device=weight.device, dtype=weight.dtype)
-        mat2 = torch.empty(out2, in2, device=weight.device, dtype=weight.dtype)
+        mat1 = torch.empty(out1, in1, device=weight.device, dtype=torch.float32)
+        mat2 = torch.empty(out2, in2, device=weight.device, dtype=torch.float32)
         torch.nn.init.kaiming_uniform_(mat2, a=5**0.5)
         torch.nn.init.constant_(mat1, 0.0)
         return LokrDiff(