app.py

import gradio as gr, requests, json, time, threading, torch, os, random, math
from transformers import AutoModelForCausalLM, AutoTokenizer
import urllib3; urllib3.disable_warnings()

sid = os.environ.get("SPACE_ID") or os.environ.get("SPACE_NAME") or ""
NUM = ""
for part in sid.split("-"):
    if part.isdigit(): NUM = part
if not NUM: NUM = "1"
TEACHER_ID = "hf-worker-" + NUM
N = int(NUM)

# === 8 MODELES du plus petit au plus gros ===
# Choisi selon le numero du worker (round-robin)
ALL_MODELS = [
    ("HuggingFaceTB/SmolLM2-360M-Instruct", 0.36),
    ("Qwen/Qwen2.5-0.5B-Instruct", 0.5),
    ("TinyLlama/TinyLlama-1.1B-Chat-v1.0", 1.1),
    ("deepseek-ai/deepseek-coder-1.3b-instruct", 1.3),
    ("Qwen/Qwen2.5-1.5B-Instruct", 1.5),
    ("Qwen/Qwen2.5-Coder-1.5B-Instruct", 1.5),
    ("stabilityai/stablelm-2-1_6b", 1.6),
    ("HuggingFaceTB/SmolLM2-1.7B-Instruct", 1.7),
]
MID, B = ALL_MODELS[N % len(ALL_MODELS)]
DELAY = max(3, int(5 + B * 10 - N % 10))  # 8-19s, petit modele = delai court
print(f"ID:{TEACHER_ID} MODEL:{MID} ({B}B) DELAY:{DELAY}s")

BIP="3.125.223.134"
BH="alfredo-agravic-saddeningly.ngrok-free.dev"
m=None; t=None; s=requests.Session(); s.verify=False
def g(p): return s.get("https://"+BIP+p,headers={"Host":BH},timeout=60)
def p(p,d): return s.post("https://"+BIP+p,headers={"Host":BH},json=d,timeout=30)
def l():
    global m,t
    m=AutoModelForCausalLM.from_pretrained(MID,torch_dtype=torch.float32,device_map="cpu",trust_remote_code=True)
    t=AutoTokenizer.from_pretrained(MID,trust_remote_code=True)
    if t.pad_token is None: t.pad_token=t.eos_token
def gen(ms, max_tokens, temp=0.7):
    x=t.apply_chat_template(ms,tokenize=False)
    i=t(x,return_tensors="pt").to("cpu")
    with torch.no_grad():
        o=m.generate(**i,max_new_tokens=max_tokens,temperature=temp,do_sample=True,
            top_p=0.9,top_k=40,repetition_penalty=1.05,pad_token_id=t.pad_token_id)
    return t.decode(o[0][i["input_ids"].shape[1]:],skip_special_tokens=True).strip()

# Prompts adaptes a la taille du modele
if B < 0.5:
    # Ultra-petit: prompts tres courts, reponses courtes
    PROMPTS = [
        lambda s: (f"Explique {s} en 1 phrase.", 64, 0.7),
        lambda s: (f"C'est quoi {s} ?", 64, 0.7),
        lambda s: (f"Donne 1 fait sur {s}.", 64, 0.8),
        lambda s: (f"Pourquoi {s} est utile ?", 96, 0.7),
        lambda s: (f"Un conseil sur {s}.", 96, 0.7),
        lambda s: (f"Decris {s} en 2 lignes.", 96, 0.7),
    ]
elif B < 1.0:
    # Petit: prompts simples, reponses courtes
    PROMPTS = [
        lambda s: (f"Explique {s} simplement.", 128, 0.7),
        lambda s: (f"C'est quoi {s} ? Donne un exemple.", 128, 0.7),
        lambda s: (f"Quels sont les points cles de {s} ?", 160, 0.6),
        lambda s: (f"Pourquoi {s} est important ?", 128, 0.8),
        lambda s: (f"Compare {s} avec son alternative.", 160, 0.7),
        lambda s: (f"Comment utiliser {s} ?", 160, 0.7),
    ]
else:
    # Moyen/Gros: prompts normaux, reponses completes
    PROMPTS = [
        lambda s: (f"Explique {s} en detail avec des exemples concrets.", 256, 0.7),
        lambda s: (f"Analyse {s} : fonctionnement, applications, limites.", 320, 0.7),
        lambda s: (f"Quels sont les concepts cles de {s} ? Liste structuree.", 256, 0.6),
        lambda s: (f"Compare {s} avec son alternative. Avantages/inconvenients.", 320, 0.7),
        lambda s: (f"Tutoriel pas a pas pour utiliser {s}.", 320, 0.7),
        lambda s: (f"L'histoire et l'avenir de {s}.", 320, 0.7),
    ]

if "coder" in MID.lower():
    # Workers code: prompts orientes code
    PROMPTS = [
        lambda s: (f"Write code example illustrating {s}.", 256, 0.6),
        lambda s: (f"Explain {s} with a code snippet.", 256, 0.7),
        lambda s: (f"How to implement {s}? Step by step.", 320, 0.7),
        lambda s: (f"Best practices for {s} in programming.", 256, 0.7),
        lambda s: (f"Compare approaches for {s} with code.", 320, 0.7),
        lambda s: (f"Common bugs with {s} and how to fix them.", 256, 0.7),
    ]

def w():
    global m,t; l()
    cycle = 0
    while True:
        try:
            p("/heartbeat",{"teacher":TEACHER_ID,"model":MID})
            r=g("/next-batch?teacher="+TEACHER_ID+"&batch_size=5")
            if r.status_code==200:
                es=r.json().get("entries",[])
                if es:
                    rs=[]
                    for e in es:
                        sj=e.get("subject","")
                        if not sj: continue
                        for idx in range(6):
                            pi = (hash(TEACHER_ID + sj + str(cycle + idx))) % len(PROMPTS)
                            fn = PROMPTS[pi]
                            try:
                                inst, mt, tmp = fn(sj)
                                resp=gen([{"role":"system","content":"Tu es Connor."},{"role":"user","content":inst}],
                                    max_tokens=mt, temp=tmp)
                                if resp and len(resp)>20:
                                    rs.append({"instruction":inst,"input":"","output":resp,
                                        "teacher":TEACHER_ID,"subject":sj,"model":MID})
                            except: pass
                    if rs:
                        p("/push-results",{"teacher":TEACHER_ID,"results":rs})
                    cycle += 1
            time.sleep(DELAY)
        except Exception as e:
            time.sleep(30)
threading.Thread(target=w,daemon=True).start()
gr.Interface(fn=lambda:json.dumps({"status":"ok","id":TEACHER_ID,"model":MID}),inputs=[],outputs="text").launch()