app.py

# Modelos
import torch
import numpy as np
import networkx as nx

from transformers import AutoTokenizer, BertForPreTraining, AutoModelForCausalLM

# API
from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel, Field

def compute_attention_rollout(attn_mean):
    rollout_list = []

    L, S, _ = attn_mean.shape

    I = torch.eye(S, device=attn_mean.device)
    acummulated = I.clone()

    for layer_idx in range(L):
        A = attn_mean[layer_idx]

        # Se le suma la identidad para la residual connection que indica el paper
        A = A + I

        # Se normaliza
        A = A / A.sum(dim=-1, keepdim=True).clamp_min(1e-12)

        acummulated = A @ acummulated

        rollout_list.append(acummulated.clone())

    return torch.stack(rollout_list, dim=0)

def residual_and_normalize(attention_layers):
    L, seq_len, _ = attention_layers.shape
    augmented_attention = attention_layers.copy()
    identity_matrix = np.eye(seq_len)

    for layer_idx in range(L):
        # Conexión residual
        augmented_attention[layer_idx] += identity_matrix

        # Normalización
        row_sums = augmented_attention[layer_idx].sum(axis=-1, keepdims=True)
        augmented_attention[layer_idx] /= row_sums

    return augmented_attention

def get_node_index(layer_idx, token_position, seq_len):
    # El índice del nodo se calcula como el número de capa por la secuencia y
    # la posición del token en esa capa
    return layer_idx * seq_len + token_position

def build_attention_graph(augmented_attentions):
    L, T, _ = augmented_attentions.shape
    G = nx.DiGraph()
    total_nodes = (L + 1) * T              # Nodos: todas las capas + capa de entrada
    super_sink = total_nodes               # Añadimos super nodo
    G.add_nodes_from(range(total_nodes + 1))   # Añadir todos los nodos del grafo

    # Crear aristas con capacidad según las matrices de atención
    for layer_idx in range(1, L + 1):
        for token_from in range(T):
            # Se obtiene el índice del token que observa al otro
            u = get_node_index(layer_idx, token_from, T)
            for token_to in range(T):
                # Se obtiene el índice del token que es observado
                v = get_node_index(layer_idx - 1, token_to, T)
                # Se obtiene su atención (capacidad de flujo del que observa hacia el que es observado)
                capacity = float(augmented_attentions[layer_idx - 1, token_from, token_to])
                if capacity > 0:
                    G.add_edge(u, v, capacity=capacity)

    for token_to in range(T):
        v = get_node_index(0, token_to, T)
        G.add_edge(v, super_sink, capacity=float(1e3))

    return G, super_sink

def compute_attention_flow_matrices(layers_mean):
    A = np.asarray(layers_mean)  # (L, T, T)

    L, T, _ = A.shape

    # Agrega residual y normaliza las matrices de atención
    aug = residual_and_normalize(A)

    # Construye el grafo de flujo (edges: capa i → capa i-1)
    G, super_sink = build_attention_graph(aug)

    # Índices de los nodos de la capa 0 (tokens de entrada)
    input_nodes = [get_node_index(0, v, T) for v in range(T)]

    flow_layers = []

    for layer_idx in range(1, L + 1):
        layer_flow = np.zeros((T, T), dtype=np.float64)

        for u in range(T):
            src = get_node_index(layer_idx, u, T)

            flow_val, flow_dict = nx.maximum_flow(G, src, super_sink, flow_func=nx.algorithms.flow.preflow_push)

            row = np.zeros(T)

            for v, node_in in enumerate(input_nodes):
                row[v] = float(flow_dict.get(node_in, {}).get(super_sink, 0))

            # Normalización
            s = row.sum()
            row /= s

            layer_flow[u, :] = row

        flow_layers.append(layer_flow)

    return flow_layers

def process_prompt(prompt):
    inputs = tokenizer(prompt, return_tensors="pt", return_offsets_mapping=True).to(model.device)
    offsets = inputs.pop("offset_mapping")[0].tolist()
    tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])

    with torch.no_grad():
        outputs = model(**inputs, output_attentions=True, return_dict=True)

    attn = torch.stack(outputs.attentions, dim=0).squeeze(1)
    att_mean = attn.mean(dim=1)
    rollout = compute_attention_rollout(att_mean)

    flow = compute_attention_flow_matrices(att_mean.detach().cpu().numpy())

    layers_mean = [att_mean[l].detach().cpu().numpy().tolist() for l in range(att_mean.shape[0])]
    attention_rollout = [rollout[l].detach().cpu().numpy().tolist() for l in range(rollout.shape[0])]
    attention_flow = [flow[l].tolist() for l in range(len(flow))]

    return {
        "model": model_name,
        "prompt": prompt,
        "tokens": tokens,
        "offsets": offsets,
        "layers_mean": layers_mean,
        "attention_rollout": attention_rollout,
        "attention_flow": attention_flow
    }

print(torch.__version__)
print(torch.cuda.is_available())

name = "gpt2"

if name == "gpt2":
    model_name = "gpt2"
elif name == "bert":
    model_name = "bert-base-uncased"
elif name == "qwen":
    model_name = "Qwen/Qwen3-1.7B"

device = "cuda" if torch.cuda.is_available() else "cpu"

if name == "bert":
    model = BertForPreTraining.from_pretrained(model_name, attn_implementation="eager")
else:
    model = AutoModelForCausalLM.from_pretrained(model_name, attn_implementation="eager")

model.eval().to(device)

tokenizer = AutoTokenizer.from_pretrained(model_name)

# API

app = FastAPI(title="Attention Server", version="1.0")

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

class AttnIn(BaseModel):
    prompt: str = Field(..., description="Texto de entrada")

@app.get("/health")
def health():
    return {"status": "ok", "model": model_name, "device": device}

@app.post("/attentions")
def attentions(payload: AttnIn):
    return process_prompt(payload.prompt)