Create app.py

app.py

@@ -0,0 +1,108 @@
+import os
+import gradio as gr
+from huggingface_hub import hf_hub_download
+from llama_cpp import Llama
+
+# ---------------------------
+# Model settings (HF Spaces compatible)
+# ---------------------------
+REPO_ID = os.getenv("REPO_ID", "bartowski/Falcon3-1B-Instruct-GGUF")
+MODEL_FILE = os.getenv("MODEL_FILE", "Falcon3-1B-Instruct-Q4_K_M.gguf")
+
+N_CTX = int(os.getenv("N_CTX", "4096"))
+N_THREADS = int(os.getenv("N_THREADS", str(max(2, (os.cpu_count() or 4) - 1))))
+N_GPU_LAYERS = int(os.getenv("N_GPU_LAYERS", "0"))  # CPU Space -> 0
+
+_llm = None
+
+
+def get_llm() -> Llama:
+    global _llm
+    if _llm is None:
+        model_path = hf_hub_download(repo_id=REPO_ID, filename=MODEL_FILE)
+        _llm = Llama(
+            model_path=model_path,
+            n_ctx=N_CTX,
+            n_threads=N_THREADS,
+            n_gpu_layers=N_GPU_LAYERS,
+            verbose=False,
+        )
+    return _llm
+
+
+def build_prompt(topic: str, n_slides: int) -> str:
+    system = (
+        "You are a professional presentation writer. "
+        "Create a slide-by-slide PowerPoint script that is clear and easy to present. "
+        "Follow the output format exactly."
+    )
+
+    user = f"""
+Topic: {topic}
+Number of slides: {n_slides}
+
+Output format (Markdown):
+- Start with: ## Title
+- Then for each slide from 1 to {n_slides}, output exactly:
+
+### Slide X: <short slide title>
+**On-slide bullets (max 5):**
+- ...
+**Speaker notes (45–75 seconds):**
+<spoken script, conversational, with a transition to the next slide>
+
+Rules:
+- No extra sections besides the title and slides.
+- Keep bullets concise.
+- Speaker notes should sound natural when spoken aloud.
+"""
+
+    # Falcon3 chat template
+    return f"<|system|>\n{system}\n<|user|>\n{user.strip()}\n<|assistant|>\n"
+
+
+def estimate_max_tokens(n_slides: int) -> int:
+    # Rough heuristic for bullets + 45–75s notes per slide
+    # Keep it bounded to avoid long CPU runtimes.
+    return min(1800, max(600, n_slides * 180))
+
+
+def generate_ppt_script(topic: str, n_slides: int) -> str:
+    topic = (topic or "").strip()
+    if not topic:
+        return "Please enter a topic."
+
+    n_slides = int(n_slides)
+    if n_slides < 1:
+        return "Number of slides must be at least 1."
+
+    llm = get_llm()
+    prompt = build_prompt(topic, n_slides)
+
+    result = llm(
+        prompt,
+        max_tokens=estimate_max_tokens(n_slides),
+        temperature=0.6,
+        top_p=0.9,
+        repeat_penalty=1.1,
+        stop=["<|system|>", "<|user|>"],
+    )
+
+    return result["choices"][0]["text"].strip()
+
+
+# ---------------------------
+# Gradio UI
+# ---------------------------
+with gr.Blocks(title="Falcon3 PPT Script Writer (GGUF)") as demo:
+    gr.Markdown("# 🧠 Falcon3 PPT Script Writer (GGUF)\nGenerate a slide-by-slide PPT script from a topic.")
+
+    topic_in = gr.Textbox(label="Topic", placeholder="e.g., Agentic AI for Contact Centers", lines=2)
+    slides_in = gr.Slider(minimum=1, maximum=15, value=6, step=1, label="Number of slides")
+
+    generate_btn = gr.Button("Generate")
+    output_md = gr.Markdown(label="Output")
+
+    generate_btn.click(fn=generate_ppt_script, inputs=[topic_in, slides_in], outputs=output_md)
+
+demo.launch(server_name="0.0.0.0", server_port=int(os.getenv("PORT", "7860")))