mike code

app.py

@@ -7,7 +7,9 @@ For information on how to customize the ChatInterface, peruse the gradio docs: h
 demo = gr.Interface(fn=gpt.get_response, inputs=["textbox",
                                                 gr.Slider(0, 100, value=50, step=1),
                                                 gr.Slider(0.1, 2.0, value=1.0),
-                                                 ], outputs="textbox", title="Mike Chat", article="""Mike is the greatest AI ever created. It was trained for about 8 hrs on my pc using fineweb-edu and open orca datasets. While it hallucinates a lot, it seems to be about on par with other lms of its size (about 160M params). Model details:
+                                                gr.Dropdown(
+            ["mike-chat", "mike-code"], value="mike-chat"),
+                                                 ], outputs=gr.Markdown(line_breaks=True), title="Mike Chat", article="""Mike is the greatest AI ever created. It was trained for about 8 hrs on my pc using fineweb-edu and open orca datasets. While it hallucinates a lot, it seems to be about on par with other lms of its size (about 160M params). Model details:
                                                  block_size: 512
                                                  n_layers: 12
                                                  n_heads: 12

gpt.py

@@ -134,9 +134,21 @@ my_GPT = load_compiled_model_state_dict(my_GPT, 'latest_model_finetune.pth')
 #my_GPT.load_state_dict(torch.load('latest_model_finetune.pth', map_location=torch.device('cpu')))
 my_GPT.eval()
 
+my_GPT_code = GPT(enc.n_vocab, block_size, n_layers, n_heads, d_model, dropout=0.1) #enc.n_vocab
+my_GPT_code = my_GPT_code.to(device)
+#my_GPT = torch.compile(my_GPT, mode='reduce-overhead')
+my_GPT_code = load_compiled_model_state_dict(my_GPT_code, 'mike-code-15k.pth')
+#my_GPT.load_state_dict(torch.load('latest_model_finetune.pth', map_location=torch.device('cpu')))
+my_GPT_code.eval()
+
+models = {
+   "mike-chat": my_GPT,
+   "mike-code": my_GPT_code
+}
+
 eot = enc._special_tokens['<|endoftext|>']
 
-def get_response(in_text, top_k=50, temperature=1):
+def get_response(in_text, top_k=50, temperature=1, model="mike-chat"):
   with torch.inference_mode():
     prompt = "USER: " + in_text + "\nASSISTANT: "
     input_tokens = enc.encode(prompt)
@@ -146,7 +158,7 @@ def get_response(in_text, top_k=50, temperature=1):
         input_tokens = input_tokens[1:]
       context_tensor = torch.tensor(input_tokens).view(1, -1).to(device)
 
-      logits, loss = my_GPT(context_tensor)
+      logits, loss = models[model](context_tensor)
       logits = logits[:, -1, :] / temperature
       if top_k > 0:
             # Remove all tokens with a probability less than the last token of the top-k

mike-code-15k.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a3f92898af4acb8f41a571e6a4f2b597b7cca1316120787f51cb8d11ba84977
+size 650652710