Tortoise and MBD improvements (rsxdalv#118)

Author: rsxdalv

README.md

@@ -34,6 +34,16 @@ Google Colab demo: [![Open In Colab](https://colab.research.google.com/assets/co
 https://rsxdalv.github.io/bark-speaker-directory/
 
 ## Changelog
+Aug 8:
+* Update AudioCraft, improving MultiBandDiffusion performance
+* Fix Tortoise parameter 'cond_free' mismatch with 'ultra_fast' preset
+
+Aug 7:
+* add tortoise deepspeed fix to colab
+
+Aug 6:
+* Fix audiogen + mbd error, add tortoise fix for colab 
+
 Aug 4:
 * Add MultiBandDiffusion option to MusicGen https://github.com/rsxdalv/tts-generation-webui/pull/109
 * MusicGen/AudioGen save tokens on generation as .npz files. 

requirements_audiocraft.txt

@@ -1,2 +1,2 @@
 # audiocraft==0.0.1 # MIT License
-audiocraft @ git+https://git@github.com/facebookresearch/audiocraft@c5157b5bf14bf83449c17ea1eeb66c19fb4bc7f0#egg=audiocraft # MIT License
+audiocraft @ git+https://git@github.com/facebookresearch/audiocraft@d6df4f0fff69a69299c5883fae4a4d6d7b8a7948#egg=audiocraft # MIT License

server.py

@@ -35,7 +35,7 @@ def reload_config_and_restart_ui():
     title="TTS Generation WebUI",
     analytics_enabled=False,  # it broke too many times
 ) as demo:
-    gr.Markdown("# TTS Generation WebUI (Bark, MusicGen, Tortoise, RVC)")
+    gr.Markdown("# TTS Generation WebUI (Bark, MusicGen + AudioGen, Tortoise, RVC)")
     with Joutai.singleton.tabs:
         from src.tortoise.generation_tab_tortoise import generation_tab_tortoise
         from src.settings_tab_gradio import settings_tab_gradio

src/tortoise/diffusion_params.py

@@ -6,7 +6,7 @@ def diffusion_params():
         iterations = gr.Slider(
             label="Diffusion Iterations", value=30, minimum=4, maximum=400, step=1
         )
-        cond_free = gr.Checkbox(label="Cond Free", value=True)
+        cond_free = gr.Checkbox(label="Cond Free", value=False)
         cond_free_k = gr.Slider(
             label="Cond Free K", value=2, minimum=0, maximum=10, step=1
         )

src/tortoise/gen_tortoise.py

@@ -70,6 +70,7 @@ def get_tts(
     kv_cache=False,
     use_deepspeed=False,
     half=False,
+    device=None,
 ):
     global MODEL
     if MODEL is None or force_reload:
@@ -79,6 +80,7 @@ def get_tts(
             kv_cache=kv_cache,
             use_deepspeed=use_deepspeed,
             half=half,
+            device=device,
         )
     return MODEL