Create deploy/app.py

deploy/app.py

@@ -0,0 +1,49 @@
+import torch, gradio as gr
+from pathlib import Path
+import sys
+sys.path.insert(0, str(Path(__file__).parent.parent))
+from model.architecture import CodeLLM, CodeLLMConfig
+from model.tokenizer import get_gpt2_tokenizer_for_code
+
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+config = CodeLLMConfig()
+model  = CodeLLM(config)
+
+WEIGHTS_PATH = Path("./checkpoints/final/pytorch_model.bin")
+if WEIGHTS_PATH.exists():
+    model.load_state_dict(torch.load(WEIGHTS_PATH, map_location=DEVICE))
+    print("Loaded trained weights!")
+
+model.to(DEVICE).eval()
+tokenizer = get_gpt2_tokenizer_for_code()
+
+def generate_code(prompt, language="Python", max_new_tokens=256, temperature=0.8, top_k=50, top_p=0.95):
+    lang_map = {"Python":"<|python|>","JavaScript":"<|javascript|>",
+                "TypeScript":"<|typescript|>","Rust":"<|rust|>","Go":"<|go|>","C++":"<|cpp|>"}
+    full_prompt = f"{lang_map.get(language,'')}{prompt}"
+    input_ids = tokenizer.encode(full_prompt, return_tensors="pt").to(DEVICE)
+    with torch.no_grad():
+        out = model.generate(input_ids, max_new_tokens=max_new_tokens,
+                             temperature=temperature, top_k=top_k, top_p=top_p)
+    return tokenizer.decode(out[0][input_ids.shape[1]:], skip_special_tokens=True)
+
+with gr.Blocks(title="CodeLLM", theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# CodeLLM — Custom Coding AI\n125M param transformer built from scratch")
+    with gr.Row():
+        with gr.Column():
+            prompt = gr.Textbox(label="Code prompt", lines=5, placeholder="def fibonacci(n):")
+            lang   = gr.Dropdown(["Python","JavaScript","TypeScript","Rust","Go","C++"], value="Python", label="Language")
+            with gr.Row():
+                btn   = gr.Button("Generate ⚡", variant="primary")
+                clear = gr.Button("Clear")
+        output = gr.Code(label="Output", language="python", lines=20)
+    with gr.Accordion("Settings", open=False):
+        with gr.Row():
+            max_tok = gr.Slider(32, 512, 256, step=32, label="Max tokens")
+            temp    = gr.Slider(0.1, 2.0, 0.8, step=0.1, label="Temperature")
+            topk    = gr.Slider(1, 100, 50, step=1, label="Top-k")
+            topp    = gr.Slider(0.1, 1.0, 0.95, step=0.05, label="Top-p")
+    btn.click(generate_code, inputs=[prompt, lang, max_tok, temp, topk, topp], outputs=output)
+    clear.click(lambda: ("", ""), outputs=[prompt, output])
+
+demo.launch()