fix: remove pycache from tracking

.gitignore

@@ -8,3 +8,4 @@ __pycache__/
 EOF*.jpg
 *.jpeg
 *.png
+__pycache__/

app.py

@@ -203,4 +203,4 @@ with gr.Blocks(title="ZeroWasteKitchen", theme=gr.themes.Soft()) as app:
             )
 
 if __name__ == "__main__":
-    app.launch()
+    app.launch(server_name="0.0.0.0")

modal_services.py

@@ -2,31 +2,56 @@ import modal
 from modal import Volume
 from pathlib import Path
 
-# App definition first
+# App definition
 app = modal.App("zerowastekitchen")
 
-# Volume
+# Volume for model storage
 model_volume = Volume.from_name("model-cache", create_if_missing=True)
 MODEL_DIR = Path("/models")
 
-# Images
+# Download image - lightweight, just needs huggingface_hub
 download_image = (
     modal.Image.debian_slim()
     .pip_install("huggingface_hub", "hf_transfer")
     .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
 )
 
+# Base image for OCR, expiry, dialogue, recipe
 base_image = (
     modal.Image.debian_slim()
     .pip_install(
-        "torch", "torchvision", "transformers>=5.7.0",
-        "accelerate", "pillow", "einops", "sentencepiece",
-        "timm", "open_clip_torch", "av", "numpy",
-        "huggingface_hub", "hf_transfer"
+        "torch",
+        "torchvision",
+        "transformers>=5.7.0",
+        "accelerate",
+        "pillow",
+        "einops",
+        "sentencepiece",
+        "timm",
+        "open_clip_torch",
+        "av",
+        "numpy",
+        "huggingface_hub",
+        "hf_transfer"
     )
     .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
 )
 
+tts_image = (
+    modal.Image.debian_slim()
+    .apt_install("libsndfile1", "ffmpeg")
+    .pip_install(
+        "torch==2.5.0",
+        "torchaudio==2.5.0",
+        "numpy",
+        "soundfile",
+        "transformers==4.40.0",
+        "click",
+        "coqui-tts"
+    )
+)
+
+# ── 0. DOWNLOAD MODELS ───────────────────────────────────────
 @app.function(
     image=download_image,
     volumes={str(MODEL_DIR): model_volume},
@@ -37,32 +62,46 @@ def download_models():
     from huggingface_hub import snapshot_download
     import os
     token = os.environ["HF_TOKEN"]
-    
+
+    print("Downloading MiniCPM-V 4.6...")
+    snapshot_download(
+        "openbmb/MiniCPM-V-4.6",
+        local_dir=str(MODEL_DIR / "minicpm-v"),
+        token=token
+    )
+
+    print("Downloading tiny-aya-fire...")
     snapshot_download(
         "CohereLabs/tiny-aya-fire",
         local_dir=str(MODEL_DIR / "tiny-aya-fire"),
         token=token
     )
+
+    print("Downloading Nemotron 4B...")
     snapshot_download(
         "nvidia/Nemotron-Mini-4B-Instruct",
         local_dir=str(MODEL_DIR / "nemotron-4b"),
         token=token
     )
+
+    print("Downloading Magpie TTS...")
     snapshot_download(
-        "openbmb/MiniCPM-V-4.6",
-        local_dir=str(MODEL_DIR / "minicpm-v"),
+        "nvidia/magpie_tts_multilingual_357m",
+        local_dir=str(MODEL_DIR / "magpie-tts"),
         token=token
     )
+
     model_volume.commit()
     print("All models downloaded.")
 
+
 # ── 1. RECEIPT OCR ──────────────────────────────────────────
 @app.function(
     gpu="T4",
     image=base_image,
     timeout=180,
     memory=12288,
-    volumes={"/models": model_volume},
+    volumes={str(MODEL_DIR): model_volume},
     secrets=[modal.Secret.from_name("huggingface-secret")]
 )
 def parse_receipt(image_bytes: bytes) -> dict:
@@ -72,11 +111,11 @@ def parse_receipt(image_bytes: bytes) -> dict:
     import io, json, re
 
     processor = AutoProcessor.from_pretrained(
-        "/models/minicpm-v",
+        str(MODEL_DIR / "minicpm-v"),
         trust_remote_code=True
     )
     model = AutoModelForImageTextToText.from_pretrained(
-        "/models/minicpm-v",
+        str(MODEL_DIR / "minicpm-v"),
         torch_dtype="auto",
         device_map="auto",
         trust_remote_code=True
@@ -99,14 +138,10 @@ Return as JSON only, no other text:
     ]}]
 
     text = processor.apply_chat_template(
-        messages,
-        tokenize=False,
-        add_generation_prompt=True
+        messages, tokenize=False, add_generation_prompt=True
     )
     inputs = processor(
-        text=text,
-        images=[img],
-        return_tensors="pt"
+        text=text, images=[img], return_tensors="pt"
     ).to(model.device)
 
     with torch.no_grad():
@@ -146,7 +181,7 @@ Return as JSON only, no other text:
     image=base_image,
     timeout=300,
     memory=8192,
-    volumes={"/models": model_volume},
+    volumes={str(MODEL_DIR): model_volume},
     secrets=[modal.Secret.from_name("huggingface-secret")]
 )
 def estimate_expiry_llm(item_names: list) -> dict:
@@ -154,9 +189,9 @@ def estimate_expiry_llm(item_names: list) -> dict:
     from transformers import AutoTokenizer, AutoModelForCausalLM
     from datetime import datetime, timedelta
 
-    tokenizer = AutoTokenizer.from_pretrained("/models/nemotron-4b")
+    tokenizer = AutoTokenizer.from_pretrained(str(MODEL_DIR / "nemotron-4b"))
     model = AutoModelForCausalLM.from_pretrained(
-        "/models/nemotron-4b",
+        str(MODEL_DIR / "nemotron-4b"),
         torch_dtype=torch.float16,
         device_map="auto"
     )
@@ -201,30 +236,30 @@ Reply with a single integer only. No explanation.
     image=base_image,
     timeout=180,
     memory=8192,
-    volumes={"/models": model_volume},
+    volumes={str(MODEL_DIR): model_volume},
     secrets=[modal.Secret.from_name("huggingface-secret")]
 )
 def generate_dialogue(character: str, expiring_items: list, cuisine: str) -> str:
     import torch
     from transformers import AutoTokenizer, AutoModelForCausalLM
 
-    tokenizer = AutoTokenizer.from_pretrained("/models/tiny-aya-fire")
+    tokenizer = AutoTokenizer.from_pretrained(str(MODEL_DIR / "tiny-aya-fire"))
     model = AutoModelForCausalLM.from_pretrained(
-        "/models/tiny-aya-fire",
+        str(MODEL_DIR / "tiny-aya-fire"),
         torch_dtype=torch.float16,
         device_map="auto"
     )
 
     character_prompts = {
-        "Grandma (Ammamma)": """You are Ammamma, a warm but dramatic Telugu grandma.
-Mix English with Telugu words: ayyo, nanna, babu, choodandi, waste cheyakandi.
-Scold gently about food waste but show love. 2-3 sentences only.""",
-        "Chef": "You are a sharp professional chef. Be direct and impatient but brilliant. 2-3 sentences.",
-        "Fitness Coach": "You are an enthusiastic fitness coach obsessed with gains and clean eating. 2-3 sentences.",
-        "Food Critic": "You are a pompous food critic who is secretly warm. Be dramatic. 2-3 sentences.",
+        "Grandma (Nani)": """You are Nani, a warm but dramatic Hindi-speaking grandma.
+Mix English with Hindi words: arre, beta, dekho, barbaad mat karo, achha khana, waste mat karo.
+Scold gently about food waste but show love. 2-3 sentences only. Be concise.""",
+        "Chef": "You are a sharp professional chef. Be direct and impatient but brilliant. 2-3 sentences only.",
+        "Fitness Coach": "You are an enthusiastic fitness coach obsessed with gains and clean eating. 2-3 sentences only.",
+        "Food Critic": "You are a pompous food critic who is secretly warm. Be dramatic. 2-3 sentences only.",
     }
 
-    system = character_prompts.get(character, character_prompts["Grandma (Ammamma)"])
+    system = character_prompts.get(character, character_prompts["Grandma (Nani)"])
     items_str = ", ".join(expiring_items[:5])
 
     prompt = f"""<|system|>
@@ -232,7 +267,7 @@ Scold gently about food waste but show love. 2-3 sentences only.""",
 <|user|>
 These grocery items are expiring soon: {items_str}
 We are cooking {cuisine} food today.
-Give your in-character reaction about the expiring food waste.
+Give your in-character reaction about the expiring food. 2-3 sentences maximum.
 <|assistant|>
 """
 
@@ -240,7 +275,7 @@ Give your in-character reaction about the expiring food waste.
     with torch.no_grad():
         outputs = model.generate(
             **inputs,
-            max_new_tokens=150,
+            max_new_tokens=100,
             temperature=0.8,
             do_sample=True,
             pad_token_id=tokenizer.eos_token_id
@@ -251,7 +286,6 @@ Give your in-character reaction about the expiring food waste.
         skip_special_tokens=True
     ).strip()
 
-    # Clean up any leaked instructions
     response = response.split("##")[0].strip()
     response = response.split("Instruction")[0].strip()
     return response
@@ -263,7 +297,7 @@ Give your in-character reaction about the expiring food waste.
     image=base_image,
     timeout=240,
     memory=8192,
-    volumes={"/models": model_volume},
+    volumes={str(MODEL_DIR): model_volume},
     secrets=[modal.Secret.from_name("huggingface-secret")]
 )
 def generate_recipe_llm(
@@ -274,9 +308,9 @@ def generate_recipe_llm(
     import torch
     from transformers import AutoTokenizer, AutoModelForCausalLM
 
-    tokenizer = AutoTokenizer.from_pretrained("/models/nemotron-4b")
+    tokenizer = AutoTokenizer.from_pretrained(str(MODEL_DIR / "nemotron-4b"))
     model = AutoModelForCausalLM.from_pretrained(
-        "/models/nemotron-4b",
+        str(MODEL_DIR / "nemotron-4b"),
         torch_dtype=torch.float16,
         device_map="auto"
     )
@@ -328,11 +362,54 @@ IMPORTANT: Maximum 5 steps. Stop after step 5.
     return response
 
 
+# ── 5. TEXT TO SPEECH ────────────────────────────────────────
+@app.function(
+    gpu="T4",
+    image=tts_image,
+    timeout=600,
+    memory=8192,
+    volumes={str(MODEL_DIR): model_volume},
+    secrets=[modal.Secret.from_name("huggingface-secret")]
+)
+def text_to_speech(text: str, character: str) -> bytes:
+    import torch
+    import io
+    import soundfile as sf
+    import os
+
+    print("Downloading XTTS-v2...")
+    os.environ["COQUI_TOS_AGREED"] = "1"
+    from TTS.api import TTS
+    tts  = TTS("tts_models/multilingual/multi-dataset/xtts_v2")    
+
+    speaker_map = {
+        "Grandma (Ammamma)": "Claribel Dervla",
+        "Chef": "Damien Black",
+        "Fitness Coach": "Abrahan Mack",
+        "Food Critic": "Annmarie Nele",
+    }
+    speaker = speaker_map.get(character, "Claribel Dervla")
+
+    # Keep short
+    sentences = text.replace("!", ".").replace("?", ".").split(".")
+    short_text = ". ".join([s.strip() for s in sentences[:2] if s.strip()]) + "."
+
+    wav = tts.tts(
+        text=short_text,
+        speaker=speaker,
+        language="hi"
+    )
+
+    buf = io.BytesIO()
+    sf.write(buf, wav, samplerate=24000, format="WAV")
+    return buf.getvalue()
+
+
 @app.local_entrypoint()
 def main():
-    print("Downloading models...")
+    print("Downloading all models to volume...")
     download_models.remote()
-    print("Done. Running tests...")
+    print("Done.")
     # Test expiry estimation
     test_items = ["INDIA WALA PANEER", "CURRIS LEAVES PACKET", "BASMATI RICE"]
     print("Testing expiry estimation...")
@@ -340,9 +417,9 @@ def main():
     print(expiry_map)
 
     # Test dialogue
-    print("\nTesting Ammamma dialogue...")
+    print("\nTesting Nani dialogue...")
     dialogue = generate_dialogue.remote(
-        "Grandma (Ammamma)",
+        "Grandma (Nani)",
         ["INDIA WALA PANEER", "CURRIS LEAVES PACKET"],
         "South Indian"
     )
@@ -353,4 +430,11 @@ def main():
     all_items = [{"name": i} for i in test_items]
     expiring = [{"name": i} for i in test_items[:2]]
     recipe = generate_recipe_llm.remote(all_items, expiring, "South Indian")
-    print(recipe)
+    print(recipe)
+
+    # Test TTS
+    print("\nTesting TTS...")
+    audio_bytes = text_to_speech.remote(dialogue, "Grandma (Nani)")
+    with open("test_audio.wav", "wb") as f:
+        f.write(audio_bytes)
+    print("Audio saved to test_audio.wav")

requirements.txt

@@ -0,0 +1,5 @@
+gradio
+modal
+pillow
+soundfile
+numpy