ask user only for the edit part, use non-deprecated autocast api

app.py

@@ -104,10 +104,14 @@ def process_fn(input_audio_path: str, instruction: str) -> str:
     if not instruction or not instruction.strip():
         raise gr.Error("Instruction cannot be empty.")
 
+    # model was trained on "Music piece. Instruct: <desired_edit>." 
+    # only <desired_edit> is variable, so rest is hardcoded
+    instruction = f"Music piece. Instruct: {instruction.strip().rstrip('.')}."
+
     input_audio, sample_rate = sf.read(input_audio_path)
     input_audio_tensor = torch.tensor(input_audio).float()
     if input_audio_tensor.ndim == 2:
-        # downmix to mono -- the original inference.py assumes a mono waveform
+        # downmix to mono -- the original inference.py assumes a mono waveform!!
         input_audio_tensor = input_audio_tensor.mean(dim=-1)
     input_audio_tensor = input_audio_tensor.unsqueeze(0).unsqueeze(0)
     if sample_rate != SAMPLE_RATE:
@@ -118,7 +122,7 @@ def process_fn(input_audio_path: str, instruction: str) -> str:
 
     instruction_list = [instruction]
 
-    with model.model.musicgen.autocast:
+    with torch.autocast("cuda", dtype=torch.float16):
         description, cond_code = model.model.musicgen._prepare_tokens_and_attributes(
             instruction_list, input_audio_tensor
         )
@@ -127,7 +131,7 @@ def process_fn(input_audio_path: str, instruction: str) -> str:
         [cond_code, torch.ones_like(cond_code[:, :, 0:1]) * CODEBOOK_SIZE], dim=-1
     )
 
-    with model.model.musicgen.autocast:
+    with torch.autocast("cuda", dtype=torch.float16):
         audio_values = model.model.generate(
             text_description=instruction_list,
             condition_audio_code=cond_code,
@@ -154,9 +158,8 @@ with gr.Blocks() as demo:
         gr.Audio(type="filepath", label="Input Audio").harp_required(True),
         gr.Textbox(
             label="Instruction",
-            value="Music piece. Instruct: Only Drums.",
-            info="Editing instruction, in the form 'Music piece. Instruct: <edit>.' "
-                 "<edit> is add/only/no + an instrument, e.g. 'Only Drums', 'No Bass', "
+            value="Only Drums",
+            info="What to change: add/only/no + an instrument, e.g. 'Only Drums', 'No Bass', "
                  "'Add Piano' (per the model's demo examples; piano/bass/drums/guitar work best)",
         ),
     ]

src/models/instructmusicgenadapter_module.py

@@ -1,8 +1,7 @@
 """PyTorch Lightning module for InstructMusicGen model with adapter-based instruction tuning.
 
-Trimmed to the inference path only: the original also had training_step/
-validation_step/test_step and their torchmetrics/wandb logging, none of
-which load_from_checkpoint() needs to reconstruct the model for inference.
+Trimmed to the inference path only i.e. removed training_step/ validation_step/test_step, torchmetrics/wandb 
+logging bcs not needed for inference; only kept what load_from_checkpoint() needs.
 """
 import torch
 from lightning import LightningModule