cleanup and use fp32 text/time embed if available

nodes_sampler.py

@@ -1011,7 +1011,7 @@ class WanVideoSampler:
                 lynx_ref_latent_uncond = lynx_ref_latent_uncond[0]
                 lynx_embeds["ref_feature_extractor"] = True
                 log.info(f"Lynx ref latent shape: {lynx_ref_latent.shape}")
-                log.info("Extracting Lynx ref buffer...")
+                log.info("Extracting Lynx ref cond buffer...")
                 lynx_ref_buffer = transformer(
                     [lynx_ref_latent.to(device, dtype)],
                     torch.tensor([0], device=device),
@@ -1019,7 +1019,9 @@ class WanVideoSampler:
                     seq_len=math.ceil((lynx_ref_latent.shape[2] * lynx_ref_latent.shape[3]) / 4 * lynx_ref_latent.shape[1]),
                     lynx_embeds=lynx_embeds
                 )
+                log.info(f"Extracted {len(lynx_ref_buffer)} cond ref buffers")
                 if not math.isclose(cfg[0], 1.0):
+                    log.info("Extracting Lynx ref uncond buffer...")
                     lynx_ref_buffer_uncond = transformer(
                         [lynx_ref_latent_uncond.to(device, dtype)],
                         torch.tensor([0], device=device),
@@ -1028,8 +1030,8 @@ class WanVideoSampler:
                         lynx_embeds=lynx_embeds,
                         is_uncond=True
                     )
+                    log.info(f"Extracted {len(lynx_ref_buffer_uncond)} uncond ref buffers")
 
-                log.info(f"Extracted {len(lynx_ref_buffer)} ref buffers")
                 lynx_embeds["ref_feature_extractor"] = False
                 lynx_embeds["ref_latent"] = lynx_embeds["ref_text_embed"] = None
                 lynx_embeds["ref_buffer"] = lynx_ref_buffer
@@ -1044,7 +1046,7 @@ class WanVideoSampler:
                              humo_image_cond=None, humo_image_cond_neg=None, humo_audio=None, humo_audio_neg=None, wananim_pose_latents=None,
                              wananim_face_pixels=None, uni3c_data=None,):
             nonlocal transformer
-            z = z.to(dtype)
+            #z = z.to(dtype)
             autocast_enabled = ("fp8" in model["quantization"] and not transformer.patched_linear)
             with torch.autocast(device_type=mm.get_autocast_device(device), dtype=dtype) if autocast_enabled else nullcontext():
 
@@ -2764,8 +2766,7 @@ class WanVideoSampler:
                     else:
                         noise_pred, self.cache_state = predict_with_cfg(
                             latent_model_input,
-                            cfg[idx], text_embeds["prompt_embeds"],
-                            text_embeds["negative_prompt_embeds"],
+                            cfg[idx], text_embeds["prompt_embeds"], text_embeds["negative_prompt_embeds"],
                             timestep, idx, image_cond, clip_fea, control_latents, vace_data, unianim_data, audio_proj, control_camera_latents, add_cond,
                             cache_state=self.cache_state, fantasy_portrait_input=fantasy_portrait_input, multitalk_audio_embeds=multitalk_audio_embeds, mtv_motion_tokens=mtv_motion_tokens, s2v_audio_input=s2v_audio_input,
                             humo_image_cond=humo_image_cond, humo_image_cond_neg=humo_image_cond_neg, humo_audio=humo_audio, humo_audio_neg=humo_audio_neg,
@@ -2774,8 +2775,7 @@ class WanVideoSampler:
                         if bidirectional_sampling:
                             noise_pred_flipped, self.cache_state = predict_with_cfg(
                             latent_model_input_flipped,
-                            cfg[idx], text_embeds["prompt_embeds"],
-                            text_embeds["negative_prompt_embeds"],
+                            cfg[idx], text_embeds["prompt_embeds"], text_embeds["negative_prompt_embeds"],
                             timestep, idx, image_cond, clip_fea, control_latents, vace_data, unianim_data, audio_proj, control_camera_latents, add_cond,
                             cache_state=self.cache_state, fantasy_portrait_input=fantasy_portrait_input, mtv_motion_tokens=mtv_motion_tokens,reverse_time=True)
 

wanvideo/modules/attention.py

@@ -35,9 +35,12 @@ except Exception as e:
     sageattn_func = None
 
 try:
-    from sageattn import sageattn_blackwell
+    from sageattn3 import sageattn3_blackwell as sageattn_blackwell
 except:
-    SAGE3_AVAILABLE = False
+    try:
+        from sageattn import sageattn_blackwell
+    except:
+        SAGE3_AVAILABLE = False
 
 
 __all__ = [

wanvideo/modules/model.py

@@ -961,39 +961,25 @@ class WanAttentionBlock(nn.Module):
 
     #region attention forward
     def forward(
-        self,
-        x,
-        e,
-        seq_lens,
-        grid_sizes,
-        freqs,
-        context,
-        current_step,
+        self, x, e, seq_lens, grid_sizes, freqs, context, current_step,
         last_step=False,
         video_attention_split_steps=[],
         clip_embed=None,
-        camera_embed=None,
-        audio_proj=None,
-        audio_scale=1.0,
+        camera_embed=None, #ReCamMaster
+        audio_proj=None, audio_scale=1.0, #fantasytalking
         num_latent_frames=21,
         original_seq_len=None,
-        enhance_enabled=False,
-        nag_params={},
-        nag_context=None,
+        enhance_enabled=False, #feta
+        nag_params={}, nag_context=None, #normalized attention guidance
         is_uncond=False,
-        multitalk_audio_embedding=None,
-        ref_target_masks=None,
-        human_num=0,
-        inner_t=None, inner_c=None,
-        cross_freqs=None,
-        x_ip=None, e_ip=None,
-        freqs_ip=None,
-        adapter_proj=None,
-        ip_scale=1.0,
+        multitalk_audio_embedding=None, ref_target_masks=None, human_num=0, #multitalk
+        inner_t=None, inner_c=None, cross_freqs=None, #echoshot
+        x_ip=None, e_ip=None, freqs_ip=None, ip_scale=1.0, #stand-in
+        adapter_proj=None, #fantasyportrait
         reverse_time=False,
-        mtv_motion_tokens=None, mtv_motion_rotary_emb=None, mtv_strength=1.0, mtv_freqs=None,
-        humo_audio_input=None, humo_audio_scale=1.0,
-        lynx_x_ip=None, lynx_ref_feature=None, lynx_ip_scale=1.0, lynx_ref_scale=1.0,
+        mtv_motion_tokens=None, mtv_motion_rotary_emb=None, mtv_strength=1.0, mtv_freqs=None, #mtv crafter
+        humo_audio_input=None, humo_audio_scale=1.0, #humo audio
+        lynx_x_ip=None, lynx_ref_feature=None, lynx_ip_scale=1.0, lynx_ref_scale=1.0, #lynx
     ):
         r"""
         Args:
@@ -1652,6 +1638,8 @@ class WanModel(torch.nn.Module):
 
         self.motion_encoder_dim = motion_encoder_dim
 
+        self.base_dtype = dtype
+
         # embeddings
         self.patch_embedding = nn.Conv3d(
             in_dim, dim, kernel_size=patch_size, stride=patch_size)
@@ -2330,7 +2318,7 @@ class WanModel(torch.nn.Module):
         if self.zero_timestep:
             t = torch.cat([t, torch.zeros([1], dtype=t.dtype, device=t.device)])
 
-        e = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(x.dtype))  # b, dim
+        e = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(self.time_embedding[0].weight.dtype))  # b, dim
         e0 = self.time_projection(e).unflatten(1, (6, self.dim))  # b, 6, dim
 
         #S2V zero timestep
@@ -2346,7 +2334,7 @@ class WanModel(torch.nn.Module):
 
         if x_ip is not None:
             timestep_ip = torch.zeros_like(t)  # [B] with 0s
-            t_ip = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, timestep_ip.flatten()).to(x.dtype))  # b, dim )
+            t_ip = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, timestep_ip.flatten()).to(self.time_embedding[0].weight.dtype))  # b, dim )
             e0_ip = self.time_projection(t_ip).unflatten(1, (6, self.dim))
 
         if fps_embeds is not None:
@@ -2387,7 +2375,7 @@ class WanModel(torch.nn.Module):
                     torch.cat(
                         [u, u.new_zeros(self.text_len - u.size(0), u.size(1))])
                     for u in context
-                ]).to(x.dtype))
+                ]).to(self.text_embedding[0].weight.dtype))
             
             # NAG
             if nag_context is not None:
@@ -2396,7 +2384,7 @@ class WanModel(torch.nn.Module):
                     torch.cat(
                         [u, u.new_zeros(self.text_len - u.size(0), u.size(1))])
                     for u in nag_context
-                ]).to(x.dtype))
+                ]).to(self.text_embedding[0].weight.dtype))
             
             if self.offload_txt_emb:
                 self.text_embedding.to(self.offload_device, non_blocking=self.use_non_blocking)
@@ -2567,6 +2555,8 @@ class WanModel(torch.nn.Module):
                 else:
                     should_calc = True
 
+        x = x.to(self.base_dtype)
+
         if self.enable_easycache:
             original_x = x.clone().to(self.cache_device)
         if should_calc:
@@ -2661,12 +2651,11 @@ class WanModel(torch.nn.Module):
             # lynx ref
             if lynx_ref_buffer is None and lynx_ref_feature_extractor:
                 lynx_ref_buffer = {}
-                print("Lynx reference feature extractor enabled.")
 
             for b, block in enumerate(self.blocks):
                 block_idx = f"{b:02d}"
                 if lynx_ref_buffer is not None and not lynx_ref_feature_extractor:
-                    print("reading from lynx ref buffer for block", block_idx)
+                    #print("reading from lynx ref buffer for block", block_idx)
                     lynx_ref_feature = lynx_ref_buffer.get(block_idx, None)
                 else:
                     lynx_ref_feature = None

wanvideo/wan_video_vae.py

@@ -998,7 +998,7 @@ class VideoVAE_(nn.Module):
         return mu
 
 
-    def encode(self, x, pbar=True):
+    def encode(self, x, pbar=True, sample=False):
         t = x.shape[2]
         iter_ = 1 + (t - 1) // 4
         if pbar:
@@ -1018,11 +1018,15 @@ class VideoVAE_(nn.Module):
             if pbar:
                 pbar.update(1)
         self.clear_cache()
-        mu = self.conv1(out).chunk(2, dim=1)[0]
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
         mu = (mu - self.mean.to(mu)) * self.inv_std.to(mu)
         if pbar:
             pbar.update_absolute(0)
-        
+
+        if sample:
+            std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+            eps = torch.randn_like(std)
+            return mu + std * eps
         return mu
 
 
@@ -1271,21 +1275,20 @@ class WanVideoVAE(nn.Module):
         return values
 
 
-    def single_encode(self, video, device, pbar=True):
+    def single_encode(self, video, device, pbar=True, sample=False):
         video = video.to(device)
-        x = self.model.encode(video, pbar=pbar)
+        x = self.model.encode(video, pbar=pbar, sample=sample)
         return x.float()
 
-
     def single_decode(self, hidden_state, device, pbar=True):
         hidden_state = hidden_state.to(device)
         video = self.model.decode(hidden_state, pbar=pbar)
         return video
 
-    def double_encode(self, video, device):
+    def double_encode(self, video, device, pbar=True, sample=False):
         print('double_encode')
         video = video.to(device)
-        x = self.model.encode_2(video)
+        x = self.model.encode_2(video, pbar=pbar, sample=sample)
         return x.float()
 
     def double_decode(self, hidden_state, device):
@@ -1294,7 +1297,7 @@ class WanVideoVAE(nn.Module):
         video = self.model.decode_2(hidden_state)
         return video
 
-    def encode(self, videos, device, tiled=False,end_=False, tile_size=None, tile_stride=None, pbar=True):
+    def encode(self, videos, device, tiled=False,end_=False, tile_size=None, tile_stride=None, pbar=True, sample=False):
         self.model.clear_cache()
         videos = [video.to("cpu") for video in videos]
         hidden_states = []
@@ -1306,7 +1309,7 @@ class WanVideoVAE(nn.Module):
                 if end_:
                     hidden_state = self.double_encode(video, device)
                 else:
-                    hidden_state = self.single_encode(video, device, pbar=pbar)
+                    hidden_state = self.single_encode(video, device, pbar=pbar, sample=sample)
             hidden_state = hidden_state.squeeze(0)
             hidden_states.append(hidden_state)
         hidden_states = torch.stack(hidden_states)
@@ -1366,7 +1369,7 @@ class VideoVAE38_(VideoVAE_):
                                     attn_scales, self.temperal_upsample, dropout)
 
 
-    def encode(self, x, pbar=True):
+    def encode(self, x, pbar=True, sample=False):
         self.clear_cache()
         x = patchify(x, patch_size=2)
         t = x.shape[2]