Switch from llama-cpp to ctransformers (no compilation)

app.py

@@ -1,29 +1,32 @@
 import gradio as gr
 from huggingface_hub import hf_hub_download
-import llama_cpp
-import os
+from ctransformers import AutoModelForCausalLM
 
-MODEL_REPO = "bartowski/Mistral-7B-Instruct-v0.3-GGUF"
-MODEL_FILE = "Mistral-7B-Instruct-v0.3-Q4_K_M.gguf"
+MODEL_REPO = "TheBloke/Mistral-7B-Instruct-v0.3-GGUF"
+MODEL_FILE = "mistral-7b-instruct-v0.3.Q4_K_M.gguf"
 
 print("📥 Téléchargement du modèle...")
 model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE, resume=True)
 
 print("⚙️ Chargement du modèle...")
-llm = llama_cpp.Llama(
-    model_path=model_path,
-    n_ctx=2048,
-    n_threads=2,
-    n_batch=256,
-    n_gpu_layers=0,
-    verbose=False
+llm = AutoModelForCausalLM.from_pretrained(
+    model_path,
+    model_type="mistral",
+    threads=2,
+    context_length=2048,
+    max_new_tokens=256
 )
 print("✅ Modèle chargé !")
 
 def respond(message, history):
     prompt = f"<s>[INST] {message} [/INST]"
-    response = llm(prompt, max_tokens=256, temperature=0.7, top_p=0.95, stop=["</s>", "[INST]", "[/INST]"], echo=False)
-    return response["choices"][0]["text"].strip()
+    response = llm.generate(prompt, temperature=0.7)
+    return response.strip()
+
+demo = gr.ChatInterface(
+    fn=respond,
+    title="🤖 Assistant Mistral 7B",
+    description="Version GGUF optimisée - réponses ~30-60 secondes"
+)
 
-demo = gr.ChatInterface(fn=respond, title="🤖 Assistant Mistral 7B (GGUF optimisé)")
 demo.launch(server_name="0.0.0.0", server_port=7860)

requirements.txt

@@ -1,4 +1,3 @@
 gradio
---find-links https://github.com/abetlen/llama-cpp-python/releases/tag/v0.2.90
-llama-cpp-python
+ctransformers
 huggingface-hub