upgrade versions of gradio and rest of libraries

README.md

@@ -4,7 +4,7 @@ emoji: 🎧
 colorFrom: indigo
 colorTo: gray
 sdk: gradio
-sdk_version: 4.44.0
+sdk_version: 6.19.0
 app_file: app.py
 pinned: false
 license: cc-by-sa-4.0

app.py

@@ -3,6 +3,7 @@ import os
 if os.getenv('SPACES_ZERO_GPU') == "true":
     os.environ['SPACES_ZERO_GPU'] = "1"
 
+import spaces
 import gradio as gr
 import random
 import torch
@@ -12,7 +13,6 @@ from typing import Optional, List
 import numpy as np
 from models import load_model
 import utils
-import spaces
 import huggingface_hub
 from inversion_utils import inversion_forward_process, inversion_reverse_process
 
@@ -322,7 +322,7 @@ to <code style="display:inline; background-color: lightgrey;">None</code>.
 
 """
 
-css = '.gradio-container {max-width: 1000px !important; padding-top: 1.5rem !important;}' \
+css = '.gradio-container {padding-top: 1.5rem !important;}' \
       '.audio-upload .wrap {min-height: 0px;}'
 
 # with gr.Blocks(css='style.css') as demo:
@@ -383,7 +383,7 @@ with gr.Blocks(css=css) as demo:
 
     with gr.Accordion("More Options", open=False):
         with gr.Row():
-            src_prompt = gr.Textbox(label="Source Prompt", lines=2, interactive=True,
+            src_prompt = gr.Textbox(label="Source Prompt", lines=2, interactive=True, value="",
                                     info="Optional: Describe the original audio input",
                                     placeholder="A recording of a happy upbeat classical music piece",)
 

models.py

@@ -2,7 +2,7 @@ import torch
 from diffusers import DDIMScheduler, CosineDPMSolverMultistepScheduler
 from diffusers.schedulers.scheduling_dpmsolver_sde import BrownianTreeNoiseSampler
 from diffusers import AudioLDM2Pipeline, StableAudioPipeline
-from transformers import RobertaTokenizer, RobertaTokenizerFast, VitsTokenizer
+from transformers import RobertaTokenizer, RobertaTokenizerFast, VitsTokenizer, GPT2LMHeadModel
 from diffusers.models.unets.unet_2d_condition import UNet2DConditionOutput
 from diffusers.models.embeddings import get_1d_rotary_pos_embed
 from typing import Any, Dict, List, Optional, Tuple, Union
@@ -178,14 +178,26 @@ class AudioLDM2Wrapper(PipelineWrapper):
     def __init__(self, *args, **kwargs) -> None:
         super().__init__(*args, **kwargs)
         if self.double_precision:
-            self.model = AudioLDM2Pipeline.from_pretrained(self.model_id, torch_dtype=torch.float64, token=self.token
+            language_model = GPT2LMHeadModel.from_pretrained(self.model_id, subfolder="language_model",
+                                                             local_files_only=True, token=self.token,
+                                                             torch_dtype=torch.float64)
+            self.model = AudioLDM2Pipeline.from_pretrained(self.model_id, torch_dtype=torch.float64, token=self.token,
+                                                           language_model=language_model,
                                                            ).to(self.device)
         else:
             try:
-                self.model = AudioLDM2Pipeline.from_pretrained(self.model_id, local_files_only=True, token=self.token
+                language_model = GPT2LMHeadModel.from_pretrained(self.model_id, subfolder="language_model",
+                                                                 local_files_only=True, token=self.token,
+                                                                 torch_dtype=torch.float32)
+                self.model = AudioLDM2Pipeline.from_pretrained(self.model_id, local_files_only=True, token=self.token,
+                                                               torch_dtype=torch.float32, language_model=language_model,
                                                                ).to(self.device)
             except FileNotFoundError:
-                self.model = AudioLDM2Pipeline.from_pretrained(self.model_id, local_files_only=False, token=self.token
+                language_model = GPT2LMHeadModel.from_pretrained(self.model_id, subfolder="language_model",
+                                                                 local_files_only=False, token=self.token,
+                                                                 torch_dtype=torch.float32)
+                self.model = AudioLDM2Pipeline.from_pretrained(self.model_id, local_files_only=False, token=self.token,
+                                                               torch_dtype=torch.float32, language_model=language_model,
                                                                ).to(self.device)
 
     def load_scheduler(self) -> None:
@@ -262,6 +274,10 @@ class AudioLDM2Wrapper(PipelineWrapper):
                         text_input_ids,
                         attention_mask=attention_mask,
                     )
+                    # Extract the pooler output if it's a BaseModelOutputWithPooling (Transformers v5+)
+                    # otherwise use it directly (Transformers v4)
+                    if hasattr(prompt_embeds, 'pooler_output'):
+                        prompt_embeds = prompt_embeds.pooler_output
                     # append the seq-len dim: (bs, hidden_size) -> (bs, seq_len, hidden_size)
                     prompt_embeds = prompt_embeds[:, None, :]
                     # make sure that we attend to this single hidden-state

requirements.txt

@@ -11,3 +11,5 @@ progressbar
 einops
 scipy
 librosa==0.9.2
+soxr
+torchcodec