hidream remote-vae

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -4,15 +4,10 @@
 
 - **Models**  
   - [HiDream-I1](https://huggingface.co/HiDream-ai/HiDream-I1-Full) in fast, dev and full variants!  
-    new absolutely massive image generative foundation model with **17B** parameters  
-    and 4 (!?) text-encoders: *clip-l, clip-g, t5-1.1-xxl, llama-3.1-8b-instruct* for total of **8.3B** parameters  
+    new absolutely massive image generative foundation model with **17B** parameters and 4 text-encoders with additional **8.3B** parameters   
     simply select from *networks -> models -> reference*  
     due to size (over 25B params in 58GB), offloading and on-the-fly quantization are pretty much a necessity  
-    custom `llama` model can be set in *settings -> model options*  
-    implementation differens from reference as it bumps up default max token length from 128 to 256  
-    max token length can be further configured using env variable `HIDREAM_MAX_SEQUENCE_LENGTH`  
-    difference between variants is recommended number of steps: *fast=16, dev=28, full=50*  
-    compatible with: *flowmatching samplers*, *offloading, quantization, taesd live-preview*  
+    see [HiDream Wiki page](https://github.com/vladmandic/sdnext/wiki/HiDream) for details  
 - **Features**  
   - Custom model loader  
       can be used to load any known diffusion model with default or custom model components  

modules/sd_vae_remote.py

@@ -55,11 +55,11 @@ def remote_decode(latents: torch.Tensor, width: int = 0, height: int = 0, model_
         params = {}
         try:
             latent = latent_copy[i]
-            if model_type != 'f1' and model_type != 'h1':
+            if model_type != 'f1':
                 latent = latent.unsqueeze(0)
-            if model_type == 'h1':
-                num_channels_latents = shared.sd_model.transformer.config.in_channels
-                latent = h1_pack_latents(latent, 1, num_channels_latents, height, width) # pylint: disable=protected-access
+            # if model_type == 'h1':
+            #     num_channels_latents = shared.sd_model.transformer.config.in_channels
+            #     latent = h1_pack_latents(latent, 1, num_channels_latents, height, width) # pylint: disable=protected-access
             params = {
                 "input_tensor_type": "binary",
                 "shape": list(latent.shape),