restoring pyharp version

app.py

@@ -1,6 +1,8 @@
 import sys
 sys.stdout.reconfigure(line_buffering=True)
 
+import traceback
+import contextlib
 import torch
 import os
 import threading
@@ -10,6 +12,7 @@ import soundfile as sf
 import gradio as gr
 import time
 from huggingface_hub import snapshot_download
+from pyharp import ModelCard, build_endpoint
 
 from anyaccomp.inference_utils import Sing2SongInferencePipeline
 
@@ -63,12 +66,21 @@ def load_model():
         model_loading = False
 
 
+# Create a ModelCard describing this endpoint for HARP.
+model_card = ModelCard(
+    name="AnyAccomp",
+    description="Upload a 3-30 second vocal or instrument track and the model will generate an accompaniment.",
+    author="Amphion",
+    tags=["audio", "music", "accompaniment"],
+)
+
 # Start model loading in the background so the server can start immediately.
 threading.Thread(target=load_model, daemon=True).start()
 
 
 # @spaces.GPU  # uncomment when running on ZeroGPU
-def sing2song_inference(vocal_filepath, n_timesteps, cfg_scale):
+@torch.inference_mode()
+def process_fn(vocal_filepath, n_timesteps, cfg_scale):
     if model_loading:
         raise gr.Error("Model is still loading, please wait a moment and try again.")
     if inference_pipeline is None:
@@ -90,8 +102,7 @@ def sing2song_inference(vocal_filepath, n_timesteps, cfg_scale):
 
         vocal_mel = inference_pipeline.encode_vocal(vocal_tensor)
 
-        # on GPU: torch.amp.autocast("cuda", dtype=torch.bfloat16)
-        autocast_ctx = torch.amp.autocast("cuda", dtype=torch.bfloat16) if torch.cuda.is_available() else torch.inference_mode()
+        autocast_ctx = torch.amp.autocast("cuda", dtype=torch.bfloat16) if torch.cuda.is_available() else contextlib.nullcontext()
         with autocast_ctx:
             mel = inference_pipeline.model.reverse_diffusion(
                 vocal_mel=vocal_mel,
@@ -116,56 +127,47 @@ def sing2song_inference(vocal_filepath, n_timesteps, cfg_scale):
         sf.write(accompaniment_path, wav, 24000)
         sf.write(mixture_path, mixture_wav, 24000)
 
-        return accompaniment_path, mixture_path, "Status: Complete!"
+        return accompaniment_path, mixture_path
 
     except Exception as e:
-        import traceback
         traceback.print_exc()
         raise gr.Error(f"An error occurred during processing: {e}")
 
 
-
-with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown(
-        """
-        # AnyAccomp: GENERALIZABLE ACCOMPANIMENT GENERATION
-        Upload a 3-30 second vocal or instrument track (.wav or .mp3) and the model will generate an accompaniment for it.
-        """
+with gr.Blocks() as demo:
+    input_components = [
+        gr.Audio(
+            type="filepath",
+            label="Input Vocal or Instrument Audio",
+        ).harp_required(True),
+        gr.Slider(
+            minimum=10, maximum=100, value=50, step=1,
+            label="Inference Steps (n_timesteps)",
+            info="Number of diffusion steps. More steps = slower but higher quality",
+        ),
+        gr.Slider(
+            minimum=1.0, maximum=10.0, value=3.0, step=0.1,
+            label="CFG Scale",
+            info="Classifier-free guidance scale",
+        ),
+    ]
+
+    output_components = [
+        gr.Audio(type="filepath", label="Generated Accompaniment").set_info(
+            "The generated instrumental accompaniment."
+        ),
+        gr.Audio(type="filepath", label="Mixture (Vocal + Accompaniment)").set_info(
+            "Vocal mixed with the generated accompaniment."
+        ),
+    ]
+
+    build_endpoint(
+        model_card=model_card,
+        input_components=input_components,
+        output_components=output_components,
+        process_fn=process_fn,
     )
 
-    with gr.Row():
-        with gr.Column(scale=1):
-            gr.Markdown("### 1. Upload Audio")
-            vocal_input = gr.Audio(
-                type="filepath",
-                label="Upload Vocal or Instrument Audio",
-                sources=["upload", "microphone"],
-            )
-
-            gr.Markdown("### 2. Adjust Parameters (Optional)")
-            with gr.Accordion("Advanced Settings", open=True):
-                n_timesteps_slider = gr.Slider(
-                    minimum=10, maximum=100, value=50, step=1,
-                    label="Inference Steps (n_timesteps)",
-                )
-                cfg_slider = gr.Slider(
-                    minimum=1.0, maximum=10.0, value=3.0, step=0.1,
-                    label="CFG Scale",
-                )
-
-        with gr.Column(scale=1):
-            gr.Markdown("### 3. Generate and Listen")
-            status_text = gr.Markdown("Status: Ready")
-            accompaniment_output = gr.Audio(label="Generated Accompaniment", type="filepath")
-            mixture_output = gr.Audio(label="Mixture (Vocal + Accompaniment)", type="filepath")
-            submit_btn = gr.Button("Generate Accompaniment", variant="primary")
-
-    submit_btn.click(
-        fn=sing2song_inference,
-        inputs=[vocal_input, n_timesteps_slider, cfg_slider],
-        outputs=[accompaniment_output, mixture_output, status_text],
-    )
-
-
 
-demo.launch(server_name="0.0.0.0", ssr_mode=False)
+import time; time.sleep(60)  # TEST: simulate blocking startup
+demo.queue().launch()

requirements.txt

@@ -1,4 +1,5 @@
 # spaces  # uncomment when running on ZeroGPU
+git+https://github.com/TEAMuP-dev/pyharp.git@v0.3.0
 torch==2.6.0
 torchaudio==2.6.0
 numpy==1.26.4