Merge branch 'main' into main

Author: chaowenguo0

.github/workflows/pr_tests.yml

@@ -266,7 +266,7 @@ jobs:
         # TODO (sayakpaul, DN6): revisit `--no-deps`
         python -m pip install -U peft@git+https://github.com/huggingface/peft.git --no-deps
         python -m uv pip install -U transformers@git+https://github.com/huggingface/transformers.git --no-deps
-        python -m uv pip install -U tokenizers@git+https://github.com/huggingface/tokenizers.git --no-deps
+        python -m uv pip install -U tokenizers
         pip uninstall accelerate -y && python -m uv pip install -U accelerate@git+https://github.com/huggingface/accelerate.git --no-deps
 
     - name: Environment

examples/community/README.md

@@ -827,7 +827,9 @@ def encode_prompt(
                 )
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 prompt_embeds = prompt_embeds.hidden_states[-2]
 
                 prompt_embeds_list.append(prompt_embeds)
@@ -879,7 +881,8 @@ def encode_prompt(
                     output_hidden_states=True,
                 )
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)

examples/community/lpw_stable_diffusion_xl.py

@@ -290,7 +290,9 @@ def encode_prompt(
                 )
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 prompt_embeds = prompt_embeds.hidden_states[-2]
 
                 prompt_embeds_list.append(prompt_embeds)
@@ -342,7 +344,8 @@ def encode_prompt(
                     output_hidden_states=True,
                 )
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)

examples/community/pipeline_demofusion_sdxl.py

@@ -628,7 +628,9 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:
@@ -688,7 +690,8 @@ def encode_prompt(
                     output_hidden_states=True,
                 )
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)

examples/community/pipeline_sdxl_style_aligned.py

@@ -359,7 +359,9 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:
@@ -419,7 +421,8 @@ def encode_prompt(
                     output_hidden_states=True,
                 )
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter.py

@@ -507,7 +507,9 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:
@@ -567,7 +569,8 @@ def encode_prompt(
                     output_hidden_states=True,
                 )
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)

examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py

@@ -394,7 +394,9 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:
@@ -454,7 +456,8 @@ def encode_prompt(
                     output_hidden_states=True,
                 )
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)

examples/community/pipeline_stable_diffusion_xl_differential_img2img.py

@@ -390,7 +390,9 @@ def encode_prompt(
                 prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
 
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                pooled_prompt_embeds = prompt_embeds[0]
+                if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2:
+                    pooled_prompt_embeds = prompt_embeds[0]
+
                 if clip_skip is None:
                     prompt_embeds = prompt_embeds.hidden_states[-2]
                 else:
@@ -450,7 +452,8 @@ def encode_prompt(
                     output_hidden_states=True,
                 )
                 # We are only ALWAYS interested in the pooled output of the final text encoder
-                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2:
+                    negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                 negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
 
                 negative_prompt_embeds_list.append(negative_prompt_embeds)

examples/community/pipeline_stable_diffusion_xl_ipex.py

@@ -25,6 +25,7 @@
 CTX = init_empty_weights if is_accelerate_available else nullcontext
 
 ckpt_ids = [
+    "Efficient-Large-Model/Sana_1600M_4Kpx_BF16/checkpoints/Sana_1600M_4Kpx_BF16.pth",
     "Efficient-Large-Model/Sana_1600M_2Kpx_BF16/checkpoints/Sana_1600M_2Kpx_BF16.pth",
     "Efficient-Large-Model/Sana_1600M_1024px_MultiLing/checkpoints/Sana_1600M_1024px_MultiLing.pth",
     "Efficient-Large-Model/Sana_1600M_1024px_BF16/checkpoints/Sana_1600M_1024px_BF16.pth",
@@ -89,7 +90,10 @@ def main(args):
     converted_state_dict["caption_norm.weight"] = state_dict.pop("attention_y_norm.weight")
 
     # scheduler
-    flow_shift = 3.0
+    if args.image_size == 4096:
+        flow_shift = 6.0
+    else:
+        flow_shift = 3.0
 
     # model config
     if args.model_type == "SanaMS_1600M_P1_D20":
@@ -99,7 +103,7 @@ def main(args):
     else:
         raise ValueError(f"{args.model_type} is not supported.")
     # Positional embedding interpolation scale.
-    interpolation_scale = {512: None, 1024: None, 2048: 1.0}
+    interpolation_scale = {512: None, 1024: None, 2048: 1.0, 4096: 2.0}
 
     for depth in range(layer_num):
         # Transformer blocks.
@@ -272,9 +276,9 @@ def main(args):
         "--image_size",
         default=1024,
         type=int,
-        choices=[512, 1024, 2048],
+        choices=[512, 1024, 2048, 4096],
         required=False,
-        help="Image size of pretrained model, 512, 1024 or 2048.",
+        help="Image size of pretrained model, 512, 1024, 2048 or 4096.",
     )
     parser.add_argument(
         "--model_type", default="SanaMS_1600M_P1_D20", type=str, choices=["SanaMS_1600M_P1_D20", "SanaMS_600M_P1_D28"]