Merge 1a864435f6 into 4209edf48d

app/app_settings.py

@@ -1,7 +1,6 @@
 import os
 import json
 from aiohttp import web
-import logging
 
 
 class AppSettings():
@@ -12,12 +11,8 @@ class AppSettings():
         file = self.user_manager.get_request_user_filepath(
             request, "comfy.settings.json")
         if os.path.isfile(file):
-            try:
-                with open(file) as f:
-                    return json.load(f)
-            except:
-                logging.error(f"The user settings file is corrupted: {file}")
-                return {}
+            with open(file) as f:
+                return json.load(f)
         else:
             return {}
 

comfy/ldm/lightricks/model.py

@@ -456,8 +456,9 @@ class LTXVModel(torch.nn.Module):
         x = self.patchify_proj(x)
         timestep = timestep * 1000.0
 
-        if attention_mask is not None and not torch.is_floating_point(attention_mask):
-            attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max
+        attention_mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1]))
+        attention_mask = attention_mask.masked_fill(attention_mask.to(torch.bool), float("-inf"))  # not sure about this
+        # attention_mask = (context != 0).any(dim=2).to(dtype=x.dtype)
 
         pe = precompute_freqs_cis(indices_grid, dim=self.inner_dim, out_dtype=x.dtype)
 

comfy/sd.py

@@ -111,7 +111,7 @@ class CLIP:
             model_management.load_models_gpu([self.patcher], force_full_load=True)
         self.layer_idx = None
         self.use_clip_schedule = False
-        logging.info("CLIP/text encoder model load device: {}, offload device: {}, current: {}, dtype: {}".format(load_device, offload_device, params['device'], dtype))
+        logging.info("CLIP model load device: {}, offload device: {}, current: {}, dtype: {}".format(load_device, offload_device, params['device'], dtype))
 
     def clone(self):
         n = CLIP(no_init=True)
@@ -898,7 +898,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     if output_model:
         model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
         if inital_load_device != torch.device("cpu"):
-            logging.info("loaded diffusion model directly to GPU")
+            logging.info("loaded straight to GPU")
             model_management.load_models_gpu([model_patcher], force_full_load=True)
 
     return (model_patcher, clip, vae, clipvision)

comfy/text_encoders/t5.py

@@ -227,9 +227,8 @@ class T5(torch.nn.Module):
         super().__init__()
         self.num_layers = config_dict["num_layers"]
         model_dim = config_dict["d_model"]
-        inner_dim = config_dict["d_kv"] * config_dict["num_heads"]
 
-        self.encoder = T5Stack(self.num_layers, model_dim, inner_dim, config_dict["d_ff"], config_dict["dense_act_fn"], config_dict["is_gated_act"], config_dict["num_heads"], config_dict["model_type"] != "umt5", dtype, device, operations)
+        self.encoder = T5Stack(self.num_layers, model_dim, model_dim, config_dict["d_ff"], config_dict["dense_act_fn"], config_dict["is_gated_act"], config_dict["num_heads"], config_dict["model_type"] != "umt5", dtype, device, operations)
         self.dtype = dtype
         self.shared = operations.Embedding(config_dict["vocab_size"], model_dim, device=device, dtype=dtype)
 

ruff.toml

@@ -4,8 +4,7 @@ lint.ignore = ["ALL"]
 # Enable specific rules
 lint.select = [
     "S307",  # suspicious-eval-usage
-    "S102", # exec
-    "T",  # print-usage
+    "T201",  # print-usage
     "W",
     # The "F" series in Ruff stands for "Pyflakes" rules, which catch various Python syntax errors and undefined names.
     # See all rules here: https://docs.astral.sh/ruff/rules/#pyflakes-f