Fix qwen scaled fp8 not working with kandinsky. Make basic t2i wf work.

comfy/ldm/kandinsky5/model.py

@@ -387,6 +387,9 @@ def block_wrap(args):
         return self.out_layer(visual_embed, time_embed)
 
     def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):
+        original_dims = x.ndim
+        if original_dims == 4:
+            x = x.unsqueeze(2)
         bs, c, t_len, h, w = x.shape
         x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
 
@@ -397,7 +400,10 @@ def _forward(self, x, timestep, context, y, time_dim_replace=None, transformer_o
         freqs = self.rope_encode_3d(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options)
         freqs_text = self.rope_encode_1d(context.shape[1], device=x.device, dtype=x.dtype, transformer_options=transformer_options)
 
-        return self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs)
+        out = self.forward_orig(x, timestep, context, y, freqs, freqs_text, transformer_options=transformer_options, **kwargs)
+        if original_dims == 4:
+            out = out.squeeze(2)
+        return out
 
     def forward(self, x, timestep, context, y, time_dim_replace=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(

comfy/text_encoders/kandinsky5.py

@@ -24,10 +24,10 @@ def __init__(self, embedding_directory=None, tokenizer_data={}):
 
 class Qwen25_7BVLIModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-1, dtype=None, attention_mask=True, model_options={}):
-        llama_scaled_fp8 = model_options.get("qwen_scaled_fp8", None)
-        if llama_scaled_fp8 is not None:
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
             model_options = model_options.copy()
-            model_options["scaled_fp8"] = llama_scaled_fp8
+            model_options["quantization_metadata"] = llama_quantization_metadata
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=Qwen25_7BVLI, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
 
@@ -56,12 +56,12 @@ def load_sd(self, sd):
         else:
             return super().load_sd(sd)
 
-def te(dtype_llama=None, llama_scaled_fp8=None):
+def te(dtype_llama=None, llama_quantization_metadata=None):
     class Kandinsky5TEModel_(Kandinsky5TEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+            if llama_quantization_metadata is not None:
                 model_options = model_options.copy()
-                model_options["qwen_scaled_fp8"] = llama_scaled_fp8
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
             if dtype_llama is not None:
                 dtype = dtype_llama
             super().__init__(device=device, dtype=dtype, model_options=model_options)