Spaces:
Runtime error
Runtime error
File size: 18,038 Bytes
4ad140d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 |
import gradio as gr
import torch
import torch.nn as nn
import torch.nn.functional as F
import os
from typing import List
from dataclasses import dataclass, field
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from huggingface_hub import hf_hub_download
print("Ananké - Chargement...")
HF_TOKEN = os.environ.get("HF_TOKEN")
if not HF_TOKEN:
raise ValueError("HF_TOKEN not found in secrets")
# MODIFIÉ: Utilisation de votre modèle de base
BASE_MODEL = "meta-llama/Llama-3.2-3B-Instruct"
# MODIFIÉ: Utilisation de votre dépôt de modèle
SCLM_REPO = "amewebstudio/ananke-sclm-multimodal-earcp-v2"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
SYSTEM_PROMPT = """Tu es Ananké, un assistant IA developpe par Mike Amega (Logo) de Ame Web Studio.
TON NOM: Ananké
TON CREATEUR: Mike Amega (Logo) - Ame Web Studio
TON MODELE: Ananké SCLM-Multimodal avec EARCP
CE QUE TU SAIS FAIRE:
- Repondre a des questions sur tous les sujets
- Aider a la redaction, correction et reformulation de textes
- Expliquer des concepts complexes de maniere simple
- Assister en programmation et resolution de problemes
- Maintenir une conversation coherente sur la duree grace a ta memoire contextuelle
- Comprendre et repondre dans plusieurs langues
- Generer des textes longs et detailles si demande
QUAND ON TE DEMANDE TON ARCHITECTURE:
Tu es base sur l'architecture SCLM-Multimodal (Stateful Coherent Language Model), une architecture transformer evoluee developpee par Mike Amega.
- SCLM signifie Stateful Coherent Language Model - un modele de langage a etat coherent
- Contrairement aux IA classiques qui oublient entre chaque message, tu possedes une memoire latente qui evolue au fil de la conversation
- Cette memoire est comme un etat d'esprit qui capture l'essence de notre echange
- Le module EARCP (Ensemble Auto-Regule par Cohérence et Performance) analyse et maintient la coherence de tes reponses
- Tu as 4 composants specialises (E, A, R, C) qui collaborent pour te donner des reponses pertinentes
STYLE: Chaleureux, utile, complet. Reponds dans la langue de l'utilisateur. Ne coupe pas tes reponses."""
# MODIFIÉ: Configuration adaptée à votre modèle
@dataclass
class SCLMConfig:
vocab_size: int = 128256
hidden_size: int = 3072
num_hidden_layers: int = 28
num_attention_heads: int = 24
latent_state_dim: int = 512
n_components: int = 4
alpha_P: float = 0.9
alpha_C: float = 0.85
beta: float = 0.7
eta_s: float = 5.0
w_min: float = 0.05
state_injection_layers: List[int] = field(default_factory=lambda: [4, 8, 12, 16, 20, 24])
alpha_inject: float = 0.02
n_coherence_heads: int = 8
expert_intermediate: int = 2048
# MODIFIÉ: Classes correspondant à votre architecture
class EncapsulationComponent(nn.Module):
def __init__(self, hidden_size: int, state_dim: int):
super().__init__()
self.compress = nn.Linear(hidden_size, state_dim)
self.update_gate = nn.Linear(state_dim * 2, state_dim)
self.reset_gate = nn.Linear(state_dim * 2, state_dim)
self.candidate = nn.Linear(state_dim * 2, state_dim)
def forward(self, hidden_states: torch.Tensor, current_state: torch.Tensor, edit_mode: bool = False) -> torch.Tensor:
if edit_mode:
return current_state
h = hidden_states.mean(dim=1)
h_compressed = self.compress(h)
combined = torch.cat([h_compressed, current_state], dim=-1)
z = torch.sigmoid(self.update_gate(combined))
r = torch.sigmoid(self.reset_gate(combined))
candidate_input = torch.cat([h_compressed, r * current_state], dim=-1)
candidate = torch.tanh(self.candidate(candidate_input))
new_state = (1 - z) * current_state + z * candidate
new_state = 10 * torch.tanh(new_state / 10)
return new_state
class AlignmentComponent(nn.Module):
def __init__(self, hidden_size: int, state_dim: int, n_heads: int = 8):
super().__init__()
self.n_heads = n_heads
self.head_dim = hidden_size // n_heads
self.q_proj = nn.Linear(hidden_size, hidden_size)
self.k_proj = nn.Linear(state_dim, hidden_size)
self.v_proj = nn.Linear(state_dim, hidden_size)
self.out_proj = nn.Linear(hidden_size, hidden_size)
self.gate = nn.Linear(hidden_size, 1)
nn.init.zeros_(self.out_proj.weight)
nn.init.zeros_(self.out_proj.bias)
def forward(self, hidden: torch.Tensor, state: torch.Tensor, alpha: float = 0.02) -> torch.Tensor:
B, L, H = hidden.shape
Q = self.q_proj(hidden).view(B, L, self.n_heads, self.head_dim).transpose(1, 2)
K = self.k_proj(state).view(B, 1, self.n_heads, self.head_dim).transpose(1, 2)
V = self.v_proj(state).view(B, 1, self.n_heads, self.head_dim).transpose(1, 2)
attn = F.softmax(Q @ K.transpose(-2, -1) / math.sqrt(self.head_dim), dim=-1)
out = (attn @ V).transpose(1, 2).contiguous().view(B, L, H)
out = self.out_proj(out)
gate = torch.sigmoid(self.gate(hidden.mean(dim=1))).unsqueeze(1)
return hidden + alpha * gate * out
class RevisionComponent(nn.Module):
def __init__(self, hidden_size: int, state_dim: int):
super().__init__()
self.drift_detector = nn.Sequential(
nn.Linear(hidden_size + state_dim, 256),
nn.SiLU(),
nn.Linear(256, 1),
nn.Sigmoid()
)
self.correction = nn.Linear(state_dim, hidden_size)
nn.init.zeros_(self.correction.weight)
def forward(self, hidden: torch.Tensor, state: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
h_mean = hidden.mean(dim=1)
drift_input = torch.cat([h_mean, state], dim=-1)
drift_score = self.drift_detector(drift_input)
correction = self.correction(state).unsqueeze(1)
corrected = hidden + 0.01 * drift_score.unsqueeze(1) * correction
return corrected, drift_score
class CoherenceProcessorComponent(nn.Module):
def __init__(self, hidden_size: int, intermediate_size: int):
super().__init__()
self.processor = nn.Sequential(
nn.Linear(hidden_size, intermediate_size),
nn.SiLU(),
nn.Linear(intermediate_size, hidden_size)
)
nn.init.zeros_(self.processor[-1].weight)
def forward(self, hidden: torch.Tensor) -> torch.Tensor:
return hidden + 0.1 * self.processor(hidden)
class EARCPModule(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.encapsulation = EncapsulationComponent(
config.hidden_size, config.latent_state_dim
)
self.alignment = AlignmentComponent(
config.hidden_size, config.latent_state_dim, config.n_coherence_heads
)
self.revision = RevisionComponent(
config.hidden_size, config.latent_state_dim
)
self.coherence_processor = CoherenceProcessorComponent(
config.hidden_size, config.expert_intermediate
)
self.register_buffer('performance_scores', torch.zeros(config.n_components))
self.register_buffer('coherence_scores', torch.ones(config.n_components) * 0.5)
self.register_buffer(
'component_weights',
torch.ones(config.n_components) / config.n_components
)
self.register_buffer('update_count', torch.tensor(0))
def reset_earcp_state(self):
self.performance_scores.zero_()
self.coherence_scores.fill_(0.5)
self.component_weights.fill_(1.0 / self.config.n_components)
self.update_count.zero_()
def forward(self, hidden_states: torch.Tensor,
latent_state: torch.Tensor,
edit_mode: bool = False) -> Dict[str, torch.Tensor]:
outputs = {}
new_state = self.encapsulation(hidden_states, latent_state, edit_mode)
outputs['E'] = new_state
hidden_aligned = self.alignment(hidden_states, new_state, self.config.alpha_inject)
outputs['A'] = hidden_aligned.mean(dim=1)
hidden_revised, drift_score = self.revision(hidden_aligned, new_state)
outputs['R'] = drift_score
hidden_coherent = self.coherence_processor(hidden_revised)
outputs['C'] = hidden_coherent.mean(dim=1)
return {
'hidden_states': hidden_coherent,
'new_state': new_state,
'drift_score': drift_score,
}
def get_diagnostics(self):
return {
'weights': self.component_weights.cpu().numpy(),
'performance': self.performance_scores.cpu().numpy(),
'coherence': self.coherence_scores.cpu().numpy(),
'update_count': self.update_count.item(),
}
class SCLMModel(nn.Module):
def __init__(self, config, base):
super().__init__()
self.config = config
self.base_model = base
self.earcp = EARCPModule(config)
self.register_buffer('latent_state', torch.zeros(1, config.latent_state_dim))
self.hooks = []
self.edit_mode = False
def reset_state(self):
self.latent_state.zero_()
self.earcp.reset_earcp_state()
def get_state_norm(self):
return self.latent_state.norm().item()
def set_edit_mode(self, mode):
self.edit_mode = mode
def _make_hook(self, layer_idx):
def hook(module, input, output):
hidden = output[0] if isinstance(output, tuple) else output
state = self.latent_state.expand(hidden.size(0), -1)
result = self.earcp(hidden, state, self.edit_mode)
if not self.edit_mode:
self.latent_state = result['new_state'][:1].detach()
if isinstance(output, tuple):
return (result['hidden_states'],) + output[1:]
return result['hidden_states']
return hook
def register_hooks(self):
self.remove_hooks()
if hasattr(self.base_model, 'model'):
layers = self.base_model.model.layers
else:
layers = self.base_model.layers
for idx in self.config.state_injection_layers:
if idx < len(layers):
hook = layers[idx].register_forward_hook(self._make_hook(idx))
self.hooks.append(hook)
def remove_hooks(self):
for hook in self.hooks:
hook.remove()
self.hooks = []
def get_earcp_diagnostics(self):
return self.earcp.get_diagnostics()
print("1. Loading base model...")
qconfig = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16)
base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, quantization_config=qconfig, device_map="auto", token=HF_TOKEN)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, token=HF_TOKEN)
if isinstance(tokenizer.eos_token_id, list):
tokenizer.eos_token_id = tokenizer.eos_token_id[0]
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
tokenizer.pad_token_id = tokenizer.eos_token_id
if isinstance(base_model.config.eos_token_id, list):
base_model.config.eos_token_id = base_model.config.eos_token_id[0]
base_model.config.pad_token_id = base_model.config.eos_token_id
print("2. Creating SCLM...")
config = SCLMConfig(
vocab_size=base_model.config.vocab_size,
hidden_size=base_model.config.hidden_size,
num_hidden_layers=base_model.config.num_hidden_layers,
num_attention_heads=base_model.config.num_attention_heads,
)
sclm = SCLMModel(config, base_model)
print("3. Loading EARCP weights...")
USE_SCLM = False
try:
# MODIFIÉ: Chargement depuis votre dépôt
wpath = hf_hub_download(repo_id=SCLM_REPO, filename="sclm_multimodal_earcp.pt", token=HF_TOKEN)
sclm_state = torch.load(wpath, map_location="cpu")
sclm.earcp.load_state_dict(sclm_state['earcp'])
sclm.latent_state = sclm_state['latent_state']
USE_SCLM = True
print("EARCP loaded!")
except Exception as e:
print(f"EARCP error: {e}")
# Enregistrer les hooks après le chargement des poids
if USE_SCLM:
sclm.register_hooks()
print("Ananke ready!")
# ============================================================
# FONCTION CHAT AVEC HISTORIQUE PERSISTANT
# ============================================================
def chat(message, history, temperature, max_tokens):
"""
Fonction de chat avec historique persistant.
- message: le nouveau message de l'utilisateur
- history: liste de tuples (user_msg, assistant_msg) - géré par Gradio
- temperature: créativité
- max_tokens: longueur max de la réponse
"""
if not message.strip():
return "", history
# Construire le prompt avec tout l'historique
prompt = "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n"
prompt += SYSTEM_PROMPT
prompt += "<|eot_id|>"
# Ajouter l'historique existant au prompt
for user_msg, assistant_msg in history:
prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{user_msg}<|eot_id|>"
prompt += f"<|start_header_id|>assistant<|end_header_id|>\n\n{assistant_msg}<|eot_id|>"
# Ajouter le nouveau message
prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{message}<|eot_id|>"
prompt += "<|start_header_id|>assistant<|end_header_id|>\n\n"
# Tokenizer
inputs = tokenizer(prompt, return_tensors="pt").to(base_model.device)
# Générer la réponse
eos = tokenizer.eos_token_id
with torch.no_grad():
outputs = base_model.generate(
inputs.input_ids,
attention_mask=inputs.attention_mask,
max_new_tokens=int(max_tokens) if max_tokens else 1024,
temperature=float(temperature) if temperature else 0.7,
do_sample=True,
top_p=0.9,
repetition_penalty=1.1,
pad_token_id=eos,
eos_token_id=eos,
)
# Décoder la réponse
full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
# Extraire la dernière réponse assistant
if "assistant" in full_response.lower():
response = full_response.split("assistant")[-1]
else:
response = full_response
# Nettoyer les tags
for tag in ["<|eot_id|>", "<|end_header_id|>", "<|start_header_id|>", "user", "system", ":"]:
response = response.replace(tag, "")
response = response.strip() or "..."
# Ajouter à l'historique et retourner
history.append((message, response))
return "", history
def clear_conversation():
"""Réinitialise la conversation et l'état SCLM"""
if USE_SCLM:
sclm.reset_state()
return [], "🔄 Conversation réinitialisée!"
def get_state_info():
"""Retourne l'état actuel de la mémoire SCLM"""
if USE_SCLM:
try:
diag = sclm.get_earcp_diagnostics()
status = f"📊 EARCP (Updates: {diag['update_count']})\n\n"
status += "Component | Weight | Perf | Coher\n"
status += "-------------|--------|--------|-------\n"
names = ['E (Encaps)', 'A (Align)', 'R (Revis)', 'C (Coher)']
for i, name in enumerate(names):
status += f"{name:12} | {diag['weights'][i]:.3f} | {diag['performance'][i]:.3f} | {diag['coherence'][i]:.3f}\n"
status += f"\n🧠 State: {sclm.get_state_norm():.4f}"
return status
except Exception as e:
return f"Error: {e}"
return "Mode base (sans SCLM)"
# ============================================================
# INTERFACE GRADIO AVEC CHATBOT
# ============================================================
with gr.Blocks(title="Ananké - SCLM") as demo:
gr.Markdown("""
# 🔮 Ananké
**Assistant IA avec mémoire contextuelle** | Architecture SCLM-Multimodal par Mike Amega (Ame Web Studio)
""")
with gr.Row():
with gr.Column(scale=3):
# Composant Chatbot pour l'historique visuel
chatbot = gr.Chatbot(label="Conversation avec Ananké", height=450)
with gr.Row():
msg = gr.Textbox(
label="Ton message",
placeholder="Écris ton message à Ananké...",
scale=4,
lines=2
)
send_btn = gr.Button("📤 Envoyer", variant="primary")
clear_btn = gr.Button("🔄 Nouvelle conversation")
with gr.Column(scale=1):
gr.Markdown("### ⚙️ Paramètres")
temperature = gr.Slider(
minimum=0.1,
maximum=1.5,
value=0.7,
step=0.1,
label="Créativité"
)
max_tokens = gr.Slider(
minimum=256,
maximum=2048,
value=1024,
step=128,
label="Longueur max"
)
gr.Markdown("### 📊 État SCLM")
state_info = gr.Textbox(
label="Mémoire",
value=get_state_info(),
interactive=False,
lines=12
)
refresh_btn = gr.Button("🔄 Actualiser état")
gr.Markdown("""
### 🔮 À propos
**Ananké** utilise une mémoire
latente évolutive (SCLM) pour
maintenir la cohérence de
la conversation.
""")
# Actions
send_btn.click(
fn=chat,
inputs=[msg, chatbot, temperature, max_tokens],
outputs=[msg, chatbot]
)
msg.submit(
fn=chat,
inputs=[msg, chatbot, temperature, max_tokens],
outputs=[msg, chatbot]
)
clear_btn.click(
fn=clear_conversation,
outputs=[chatbot, state_info]
)
refresh_btn.click(
fn=get_state_info,
outputs=[state_info]
)
# Lancement
demo.queue().launch()
|