revert sageattn back to optional dependency because it's not on windows, force resize_modules on by default because I broke something

setup.py

@@ -25,6 +25,13 @@ def write_version(version_core, pre_release=True):
 with open("README.md", "r") as f:
 	long_description = f.read()
 
+platform_dependencies = []
+
+if sys.platform.startswith("win"):
+	platform_dependencies += ["psutil"]
+else:
+	platform_dependencies += ["deepspeed>=0.7.7"]
+
 setup(
 	name="vall-e",
 	python_requires=">=3.10.0",
@@ -35,10 +42,8 @@ setup(
 	long_description=long_description,
 	long_description_content_type="text/markdown",
 	packages=find_packages(),
-    install_requires=(
-        # training backends
-        ["deepspeed>=0.7.7"] if not sys.platform.startswith("win") else ["psutil"])
-        + [
+	install_requires=
+		platform_dependencies + [
 		# logging niceties
 		"coloredlogs>=15.0.1", # barely required
 		"humanize>=4.4.0", # not really required
@@ -74,7 +79,6 @@ setup(
 		# for the web UI
 		"gradio",
 		"nltk",
-        "sageattention==1.0.6",
 	],
 	extras_require = {
 		"all": [
@@ -88,6 +92,7 @@ setup(
 
 			# attention helpers
 			"xformers",
+			"sageattention==1.0.6",
 			# "flash-attn" --no-build-isolation # commented out right now because I want to query this for Volta freaks like me who can't use it
 			
 			# other audio backend that doesn't prove fruitful

vall_e/config.py

@@ -698,7 +698,7 @@ class Trainer:
 	strict_loading: bool = False # sets strict_loading=True when loading the state dict
 	load_module_only: bool = False # 
 	restart_step_count: bool = False # clears the training stats when loading a checkpoint
-	resize_modules: bool = False # automatically resizes 
+	resize_modules: bool = True # automatically resizes 
 
 	activation_checkpointing: bool | None = None # deprecated, should technically be used for only on activations and not the entire gradients, but HF only has gradient checkpointing
 	gradient_checkpointing: bool = True # enables gradient checkpointing to save VRAM at the cost of slightly reduced performance when training