perf: improve memory usage

add warning for corrupt weights files

imaginairy/api/generate_refiners.py

@@ -310,7 +310,9 @@ def latent_logger(latents):
                 condition_scale=prompt.prompt_strength,
                 **text_conditioning_kwargs,
             )
+        # trying to clear memory. not sure if this helps
         sd.unet.set_context(context="self_attention_map", value={})
+        sd.unet._reset_context()
         clear_gpu_cache()
 
         logger.debug("Decoding image")

imaginairy/enhancers/upscale_realesrgan.py

@@ -12,7 +12,7 @@
 
 
 @memory_managed_model("realesrgan_upsampler", memory_usage_mb=70)
-def realesrgan_upsampler(tile=1024, tile_pad=50, ultrasharp=False):
+def realesrgan_upsampler(tile=512, tile_pad=50, ultrasharp=False):
     model = RRDBNet(
         num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4
     )

imaginairy/utils/model_manager.py

@@ -674,7 +674,7 @@ def load_sdxl_diffusers_weights(base_url: str, device=None, dtype=torch.float16)
     vae_weights_path = download_diffusers_weights(
         base_url=base_url, sub="vae", prefer_fp16=False
     )
-    print(vae_weights_path)
+    logger.debug(f"vae: {vae_weights_path}")
     vae_weights = translator.load_and_translate_weights(
         source_path=vae_weights_path,
         device="cpu",
@@ -684,8 +684,10 @@ def load_sdxl_diffusers_weights(base_url: str, device=None, dtype=torch.float16)
     del vae_weights
 
     translator = translators.diffusers_unet_sdxl_to_refiners_translator()
-    unet_weights_path = download_diffusers_weights(base_url=base_url, sub="unet")
-    print(unet_weights_path)
+    unet_weights_path = download_diffusers_weights(
+        base_url=base_url, sub="unet", prefer_fp16=True
+    )
+    logger.debug(f"unet: {unet_weights_path}")
     unet_weights = translator.load_and_translate_weights(
         source_path=unet_weights_path,
         device="cpu",
@@ -700,8 +702,8 @@ def load_sdxl_diffusers_weights(base_url: str, device=None, dtype=torch.float16)
     text_encoder_2_path = download_diffusers_weights(
         base_url=base_url, sub="text_encoder_2"
     )
-    print(text_encoder_1_path)
-    print(text_encoder_2_path)
+    logger.debug(f"text encoder 1: {text_encoder_1_path}")
+    logger.debug(f"text encoder 2: {text_encoder_2_path}")
     text_encoder_weights = (
         translators.DoubleTextEncoderTranslator().load_and_translate_weights(
             text_encoder_l_weights_path=text_encoder_1_path,

imaginairy/vendored/realesrgan.py

@@ -1,3 +1,4 @@
+import logging
 import math
 import os
 import queue
@@ -12,6 +13,7 @@
 
 ROOT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
 
+logger = logging.getLogger(__name__)
 
 class RealESRGANer:
     """A helper class for upsampling images with RealESRGAN.
@@ -146,7 +148,7 @@ def tile_process(self):
         self.output = self.img.new_zeros(output_shape)
         tiles_x = math.ceil(width / self.tile_size)
         tiles_y = math.ceil(height / self.tile_size)
-
+        logger.debug(f"Tiling with {tiles_x}x{tiles_y} ({tiles_x*tiles_y}) tiles")
         # loop over all tiles
         for y in range(tiles_y):
             for x in range(tiles_x):

imaginairy/weight_management/translation.py

@@ -4,7 +4,7 @@
 from typing import Dict
 
 import torch
-from refiners.fluxion import load_from_safetensors
+from safetensors import safe_open
 from torch import device as Device
 
 logger = logging.getLogger(__name__)
@@ -29,7 +29,8 @@ def load_and_translate_weights(
             source_weights = torch.load(source_path, map_location="cpu")
 
         elif extension in ["safetensors"]:
-            source_weights = load_from_safetensors(source_path, device=device)
+            with safe_open(source_path, framework="pt", device=device) as f:  # type: ignore
+                source_weights = {k: f.get_tensor(k) for k in f.keys()}  # noqa
         else:
             msg = f"Unsupported extension {extension}"
             raise ValueError(msg)
@@ -79,10 +80,30 @@ def load(cls, path):
         return cls(**d)
 
 
+def check_nan_path(path: str, device):
+    from safetensors import safe_open
+
+    with safe_open(path, framework="pt", device=device) as f:  # type: ignore
+        for k in f.keys():  # noqa
+            if torch.any(torch.isnan(f.get_tensor(k))):
+                print(f"Found nan values in {k} of {path}")
+
+
 def translate_weights(
     source_weights: TensorDict, weight_map: WeightTranslationMap
 ) -> TensorDict:
     new_state_dict: TensorDict = {}
+    # check source weights for nan
+    for k, v in source_weights.items():
+        nan_count = torch.sum(torch.isnan(v)).item()
+        if nan_count:
+            msg = (
+                f"Found {nan_count} nan values in {k} of source state dict."
+                " This could indicate the source weights are corrupted and "
+                "need to be re-downloaded. "
+            )
+            logger.warning(msg)
+
     # print(f"Translating {len(source_weights)} weights")
     # print(f"Using {len(weight_map.name_map)} name mappings")
     # print(source_weights.keys())
@@ -142,7 +163,7 @@ def translate_weights(
 
     if source_weights:
         msg = f"Unmapped keys: {list(source_weights.keys())}"
-        print(msg)
+        logger.info(msg)
         for k in source_weights:
             if isinstance(source_weights[k], torch.Tensor):
                 print(f"  {k}: {source_weights[k].shape}")
@@ -154,6 +175,15 @@ def translate_weights(
             if key in new_state_dict:
                 new_state_dict[key] = new_state_dict[key].reshape(new_shape)
 
+    # check for nan values
+    for k in list(new_state_dict.keys()):
+        v = new_state_dict[k]
+        nan_count = torch.sum(torch.isnan(v)).item()
+        if nan_count:
+            logger.warning(
+                f"Found {nan_count} nan values in {k} of converted state dict."
+            )
+
     return new_state_dict