Merge 7b5e2b5662 into b08debceca

The TAESD preview decoders (TAESDPreviewerImpl and TAEHVPreviewerImpl) were
passing tensor views (slices) of x0 directly to the decoder. This allowed the
decoder's internal operations to potentially modify the original latent in-place,
corrupting the midsampling latent.

The fix clones the sliced tensor before passing it to the decoder, ensuring the
original x0 is never affected by any in-place operations within the preview decode
path.

This specifically addresses the case where lighttaew2_1.safetensors (WanVAE) is
used as the TAESD preview decoder for models using the Wan21 latent format (e.g.,
QwenImage), where the full VAE decode pipeline could write back to the input slice.

comfy/text_encoders/llama.py

@@ -937,22 +937,41 @@ class BaseGenerate:
             return torch.argmax(logits, dim=-1, keepdim=True)
 
         # Sampling mode
-        if repetition_penalty != 1.0:
-            for i in range(logits.shape[0]):
-                for token_id in set(token_history):
-                    logits[i, token_id] *= repetition_penalty if logits[i, token_id] < 0 else 1/repetition_penalty
-
-        if presence_penalty is not None and presence_penalty != 0.0:
-            for i in range(logits.shape[0]):
-                for token_id in set(token_history):
-                    logits[i, token_id] -= presence_penalty
+        if len(token_history) > 0 and (repetition_penalty != 1.0 or (presence_penalty is not None and presence_penalty != 0.0)):
+            token_ids = torch.tensor(list(set(token_history)), device=logits.device)
+            token_logits = logits[:, token_ids]
+            if repetition_penalty != 1.0:
+                token_logits = torch.where(token_logits < 0, token_logits * repetition_penalty, token_logits / repetition_penalty)
+            if presence_penalty is not None and presence_penalty != 0.0:
+                token_logits = token_logits - presence_penalty
+            logits[:, token_ids] = token_logits
 
         if temperature != 1.0:
             logits = logits / temperature
 
         if top_k > 0:
-            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
-            logits[indices_to_remove] = torch.finfo(logits.dtype).min
+            top_k = min(top_k, logits.shape[-1])
+            logits, top_indices = torch.topk(logits, top_k)
+
+            if min_p > 0.0:
+                probs_before_filter = torch.nn.functional.softmax(logits, dim=-1)
+                top_probs, _ = probs_before_filter.max(dim=-1, keepdim=True)
+                min_threshold = min_p * top_probs
+                indices_to_remove = probs_before_filter < min_threshold
+                logits[indices_to_remove] = torch.finfo(logits.dtype).min
+
+            if top_p < 1.0:
+                sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+                cumulative_probs = torch.cumsum(torch.nn.functional.softmax(sorted_logits, dim=-1), dim=-1)
+                sorted_indices_to_remove = cumulative_probs > top_p
+                sorted_indices_to_remove[..., 0] = False
+                indices_to_remove = torch.zeros_like(logits, dtype=torch.bool)
+                indices_to_remove.scatter_(1, sorted_indices, sorted_indices_to_remove)
+                logits[indices_to_remove] = torch.finfo(logits.dtype).min
+
+            probs = torch.nn.functional.softmax(logits, dim=-1)
+            next_token = torch.multinomial(probs, num_samples=1, generator=generator)
+            return top_indices.gather(1, next_token)
 
         if min_p > 0.0:
             probs_before_filter = torch.nn.functional.softmax(logits, dim=-1)

latent_preview.py

@@ -41,12 +41,14 @@ class TAESDPreviewerImpl(LatentPreviewer):
         self.taesd = taesd
 
     def decode_latent_to_preview(self, x0):
-        x_sample = self.taesd.decode(x0[:1])[0].movedim(0, 2)
+        # Clone to prevent the decoder from modifying the latent in-place
+        x_sample = self.taesd.decode(x0[:1].clone())[0].movedim(0, 2)
         return preview_to_image(x_sample)
 
 class TAEHVPreviewerImpl(TAESDPreviewerImpl):
     def decode_latent_to_preview(self, x0):
-        x_sample = self.taesd.decode(x0[:1, :, :1])[0][0]
+        # Clone to prevent the decoder from modifying the latent in-place
+        x_sample = self.taesd.decode(x0[:1, :, :1].clone())[0][0]
         return preview_to_image(x_sample, do_scale=False)
 
 class Latent2RGBPreviewer(LatentPreviewer):

requirements.txt

@@ -1,6 +1,6 @@
 comfyui-frontend-package==1.45.20
 comfyui-workflow-templates==0.11.2
-comfyui-embedded-docs==0.5.6
+comfyui-embedded-docs==0.5.7
 torch
 torchsde
 torchvision