Downloading LLM in HF

Dockerfile

@@ -23,7 +23,6 @@ RUN pip install --no-cache-dir -r requirements.txt
 COPY main.py .
 COPY yolo_deart.onnx .
 COPY style_classifier.onnx .
-COPY tinyllama.gguf .
 
 # ─────────────────────────────────────────────
 # HF / LLM performance config

main.py

@@ -12,6 +12,7 @@ from fastapi.responses import JSONResponse
 
 from ultralytics import YOLO
 from torchvision import transforms
+from huggingface_hub import hf_hub_download
 from llama_cpp import Llama
 
 from ctransformers import AutoModelForCausalLM
@@ -26,7 +27,11 @@ import re
 
 YOLO_PATH = Path("yolo_deart.onnx")
 STYLE_MODEL_PATH = Path("style_classifier.onnx")
-LLM_PATH = "phi3_mini_4k_q4.gguf"
+
+LLM_PATH = hf_hub_download(
+    repo_id="Qwen/Qwen2.5-1.5B-Instruct-GGUF", 
+    filename="qwen2.5-1.5b-instruct-fp16.gguf"
+)
 
 FRONTEND_ORIGIN = os.environ.get(
     "FRONTEND_ORIGIN",
@@ -75,13 +80,11 @@ STYLES = [
 # LLM (phi3 GGUF)
 # ─────────────────────────────────────────────
 
+# Inicializando o modelo
 llm = Llama(
     model_path=LLM_PATH,
-    n_ctx=4096,
-    n_threads=8,
-    n_batch=256,
-    verbose=False,
-    use_mlock=False
+    n_ctx=2048, # Tamanho da janela de contexto
+    n_threads=2 # O Space free tem 2 vCPUs
 )
 
 def generate_poem(detections, style):
@@ -92,40 +95,31 @@ def generate_poem(detections, style):
         for k, v in counts.items()
     ) or "silence"
 
-    prompt = f"""
-You are a world-class art poet.
-
-Write ONLY a poem based on an artwork with the following characteristics:
-
-Style: {style["style_name"]}
-Objects: {objects}
+    prompt = f"""<|im_start|>system
+            You are a master of classical poetry. You write only the poem requested, following strict metrical and rhyme rules. No titles, no explanations, no markdown.<|im_end|>
+            <|im_start|>user
+            Write a Shakespearean sonnet (14 lines, ABAB CDCD EFEF GG) about: {objects}.
+            Style of the artwork: {style["style_name"]}.
 
-Rules:
-- only poem
-- Shakespearean English Sonnet (14 lines, iambic pentameter, ABAB CDCD EFEF GG)
-- no explanation
-- no repetition of prompt
-- metaphorical, emotional, artistic
-
-Ouput expected: 
-
-- Response: formatted poem text only, no markdown, no code blocks, no extra text.
-
-"""
-    
-    print("LLM PROMPT:\n", prompt)
+            Directives:
+            - Strictly 14 lines.
+            - No markdown (no bold, no italics).
+            - No title.
+            - Output only the poem text.<|im_end|>
+            <|im_start|>assistant
+            """
 
     output = llm(
         prompt,
-        max_tokens=200,
-        temperature=0.8,
+        max_tokens=300,
+        temperature=0.7, # Equilíbrio entre criatividade e ordem
         top_p=0.9,
-        repeat_penalty=1.2
+        repeat_penalty=1.1, # Reduzido, pois o FP16 é naturalmente mais coerente
+        stop=["<|im_end|>", "User:", "\n\n\n\n"] 
     )
 
-    print("LLM RAW OUTPUT:", output)
-
-    return output["choices"][0]["text"]
+    poem_text = output["choices"][0]["text"].strip()
+    return poem_text
 
 def format_poem(text: str, max_lines: int = 17):
     # ── 1. Remove prefixos tipo [response]: ou "response:"

requirements.txt

@@ -10,4 +10,6 @@ python-multipart==0.0.9
 numpy==1.26.4
 
 # ── LLM (TinyLlama GGUF) ─────────────────────
-ctransformers==0.2.27
+ctransformers==0.2.27
+huggingface-hub==0.23.0
+llama-cpp-python==0.2.90