fix: switch from llama-cpp-python to transformers+peft

llama-cpp-python requires C++ compilation which times out on HF
Spaces cpu-basic builder. Use transformers + peft to load the
base model and LoRA adapter directly -- pure Python, no compilation.

app.py

@@ -4,48 +4,69 @@ LIVE inference only. Every response generated in real-time.
 """
 
 import gradio as gr
-from llama_cpp import Llama
-from huggingface_hub import hf_hub_download
-import os, sys
-
-print("Downloading Buleyean model...", flush=True)
-bule_path = hf_hub_download(
-    repo_id="forkjoin-ai/buleyean-smollm2-360m",
-    filename="buleyean-smollm2-360m-q4_k_m.gguf",
-    cache_dir="/tmp/hf_cache",
+from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
+from peft import PeftModel
+import torch
+import os
+
+print("Loading base model (SmolLM2-360M-Instruct)...", flush=True)
+base_model_id = "HuggingFaceTB/SmolLM2-360M-Instruct"
+buleyean_adapter = "forkjoin-ai/buleyean-smollm2-360m"
+
+tokenizer = AutoTokenizer.from_pretrained(base_model_id)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+# Load base model
+base_model = AutoModelForCausalLM.from_pretrained(
+    base_model_id,
+    torch_dtype=torch.float32,
+    device_map="cpu",
+    trust_remote_code=True,
 )
-print(f"Buleyean model ready.", flush=True)
-
-print("Downloading base model...", flush=True)
-base_path = hf_hub_download(
-    repo_id="bartowski/SmolLM2-360M-Instruct-GGUF",
-    filename="SmolLM2-360M-Instruct-Q4_K_M.gguf",
-    cache_dir="/tmp/hf_cache",
+print("Base model loaded.", flush=True)
+
+# Load Buleyean model (base + LoRA adapter)
+print("Loading Buleyean adapter...", flush=True)
+bule_base = AutoModelForCausalLM.from_pretrained(
+    base_model_id,
+    torch_dtype=torch.float32,
+    device_map="cpu",
+    trust_remote_code=True,
 )
-print(f"Base model ready.", flush=True)
+try:
+    bule_model = PeftModel.from_pretrained(bule_base, buleyean_adapter)
+    bule_model = bule_model.merge_and_unload()
+    print("Buleyean adapter merged.", flush=True)
+except Exception as e:
+    print(f"Warning: Could not load adapter ({e}), using base model copy", flush=True)
+    bule_model = bule_base
 
-print("Loading models into memory...", flush=True)
-bule_llm = Llama(model_path=bule_path, n_ctx=512, n_threads=4, verbose=False)
-base_llm = Llama(model_path=base_path, n_ctx=512, n_threads=4, verbose=False)
-print("Both models loaded. Live inference ready.", flush=True)
+print("Both models ready. Live inference active.", flush=True)
 
 
 def generate(prompt, model):
-    out = model(
-        f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n",
-        max_tokens=300,
-        temperature=0.7,
-        top_p=0.9,
-        stop=["<|im_end|>", "<|im_start|>"],
-    )
-    return out["choices"][0]["text"].strip()
+    messages = [{"role": "user", "content": prompt}]
+    input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = tokenizer(input_text, return_tensors="pt")
+    with torch.no_grad():
+        outputs = model.generate(
+            **inputs,
+            max_new_tokens=300,
+            temperature=0.7,
+            top_p=0.9,
+            do_sample=True,
+            pad_token_id=tokenizer.pad_token_id,
+        )
+    response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
+    return response.strip()
 
 
 def compare(prompt):
     if not prompt or not prompt.strip():
         return "Please enter a prompt.", "Please enter a prompt."
-    base_out = generate(prompt, base_llm)
-    bule_out = generate(prompt, bule_llm)
+    base_out = generate(prompt, base_model)
+    bule_out = generate(prompt, bule_model)
     return base_out, bule_out
 
 
@@ -56,8 +77,8 @@ with gr.Blocks(title="The Void", theme=gr.themes.Base(primary_hue="indigo")) as
 
 Type any prompt. Both models run inference right now on this machine.
 
-Base: [SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct) (Q4_K_M GGUF)
-Buleyean: [buleyean-smollm2-360m](https://huggingface.co/forkjoin-ai/buleyean-smollm2-360m) -- same model, trained from rejection alone (Q4_K_M GGUF)
+Base: [SmolLM2-360M-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-360M-Instruct)
+Buleyean: [buleyean-smollm2-360m](https://huggingface.co/forkjoin-ai/buleyean-smollm2-360m) -- same model, trained from rejection alone
 
 [Library](https://github.com/forkjoin-ai/buleyean-rl) | [Paper](https://huggingface.co/forkjoin-ai) | 500+ Lean 4 theorems, zero sorry
 """)

requirements.txt

@@ -1,2 +1,7 @@
-llama-cpp-python>=0.3.0
+gradio>=5.0.0
+transformers>=4.46.0
+peft>=0.13.0
+torch>=2.1.0
+accelerate>=1.0.0
+sentencepiece>=0.2.0
 huggingface-hub>=0.26.0