remove accelerate (causes meta tensors), clean up patches

Dockerfile

@@ -72,7 +72,7 @@ RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/vae-BF16.gguf \
 # Install Python deps for Gradio UI + training
 RUN pip3 install --no-cache-dir --extra-index-url https://download.pytorch.org/whl/cpu \
     "gradio[mcp]==5.29.0" requests torch safetensors \
-    transformers>=4.51.0 peft>=0.18.0 accelerate>=1.12.0 \
+    transformers>=4.51.0 peft>=0.18.0 \
     loguru "torchaudio==2.4.0" "diffusers==0.30.3" lightning numpy tensorboard soundfile \
     einops vector_quantize_pytorch
 

app.py

@@ -339,15 +339,6 @@ try:
     torch.backends.cuda.enable_flash_sdp(False)
     os.environ["ATTN_BACKEND"] = "sdpa"
 
-    import transformers
-    _orig_from_pretrained = transformers.AutoModel.from_pretrained
-    def _cpu_from_pretrained(*args, **kwargs):
-        kwargs['low_cpu_mem_usage'] = False
-        kwargs.setdefault('torch_dtype', torch.float32)
-        return _orig_from_pretrained(*args, **kwargs)
-    transformers.AutoModel.from_pretrained = _cpu_from_pretrained
-    log("  Patched AutoModel.from_pretrained: low_cpu_mem_usage=False, dtype=float32")
-
     import torchaudio
     _orig = torchaudio.load
     def _sf(p, *a, **kw):