experiment: remote-t5

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

modules/sd_models.py

@@ -411,7 +411,7 @@ def load_diffuser_force(model_type, checkpoint_info, diffusers_load_config, op='
         shared.log.error(f'Load {op}: path="{checkpoint_info.path}" {e}')
         if debug_load:
             errors.display(e, 'Load')
-        return None, True
+        return None
     return sd_model
 
 

modules/sd_te_remote.py

@@ -0,0 +1,40 @@
+from typing import List, Optional, Union
+import os
+import time
+import json
+import torch
+import requests
+from modules import devices, errors
+
+
+def get_t5_prompt_embeds(
+    prompt: Union[str, List[str]] = None,
+    num_images_per_prompt: int = 1, # pylint: disable=unused-argument
+    max_sequence_length: int = 512, # pylint: disable=unused-argument
+    device: Optional[torch.device] = None,
+    dtype: Optional[torch.dtype] = None,
+):
+    device = device or devices.device
+    dtype = dtype or devices.dtype
+    url = os.environ.get('SD_REMOTE_T5', None)
+    if url is None:
+        errors.log.error('Remote-TE: url is not set')
+        return None
+    try:
+        t0 = time.time()
+        response = requests.post(
+            url=url,
+            headers={ "Content-Type": "application/json" },
+            json=prompt,
+            timeout=300,
+        )
+        t1 = time.time()
+        shape = json.loads(response.headers["shape"])
+        buffer = bytearray(response.content)
+        tensor = torch.frombuffer(buffer, dtype=dtype).reshape(shape)
+        errors.log.debug(f'Remote-TE: url="{url}" prompt="{prompt}" shape={shape} time={t1-t0:.3f}')
+        return tensor.to(device=device, dtype=dtype)
+    except Exception as e:
+        errors.log.error(f'Remote-TE: {e}')
+        errors.display(e, 'remote-te')
+        return None