html/api/opus5-gpu-grid.php

<?php
// OPUS5 — Grid GPU v2 (doctrine 80) — MULTI-ENDPOINT PARALLEL via curl_multi
// Utilise les 3 endpoints LLM internes existants en parallèle pour ~3x throughput
// Endpoints: sovereign-proxy (SSE), weval-ia-fast-safe, opus5-autonomous-orchestrator
header('Content-Type: application/json');
$R = ['ts'=>date('c'), 'source'=>'opus5-gpu-grid-v2'];

$raw = file_get_contents('php://input');
$d = json_decode($raw, true) ?: [];
$action = $_GET['action'] ?? ($d['action'] ?? 'health');

$PROVIDERS = [
    [
        'id' => 'safe_wrapper',
        'url' => 'http://127.0.0.1/api/opus5-weval-ia-fast-safe.php',
        'type' => 'internal'
    ],
    [
        'id' => 'meta_orchestrator',
        'url' => 'http://127.0.0.1/api/opus5-autonomous-orchestrator.php',
        'type' => 'internal_orch'
    ],
    [
        'id' => 'dispatch_proxy',
        'url' => 'http://127.0.0.1/api/wevia-master-dispatch.php',
        'type' => 'internal_disp'
    ]
];

function build_ch($url, $payload, $timeout = 20) {
    $ch = curl_init($url);
    curl_setopt_array($ch, [
        CURLOPT_POST => true,
        CURLOPT_POSTFIELDS => $payload,
        CURLOPT_HTTPHEADER => ['Content-Type: application/json'],
        CURLOPT_RETURNTRANSFER => true,
        CURLOPT_TIMEOUT => $timeout,
        CURLOPT_SSL_VERIFYPEER => false,
        CURLOPT_SSL_VERIFYHOST => 0
    ]);
    return $ch;
}

function parse_response($body) {
    $d = @json_decode((string)$body, true);
    if (!$d) return ['text' => substr((string)$body, 0, 500), 'provider' => '?'];
    return [
        'text' => $d['response'] ?? $d['content'] ?? $d['final_response'] ?? substr((string)$body, 0, 300),
        'provider' => $d['provider'] ?? $d['final_provider'] ?? '?'
    ];
}

if ($action === 'health') {
    $mh = curl_multi_init();
    $handles = [];
    $t0 = microtime(true);
    foreach ($PROVIDERS as $p) {
        $ch = build_ch($p['url'], json_encode(['message'=>'ping']), 5);
        curl_multi_add_handle($mh, $ch);
        $handles[] = ['ch'=>$ch, 'provider'=>$p];
    }
    $running = null;
    do { curl_multi_exec($mh, $running); curl_multi_select($mh, 0.1); } while ($running > 0);
    
    $results = [];
    foreach ($handles as $h) {
        $http = curl_getinfo($h['ch'], CURLINFO_HTTP_CODE);
        $results[] = ['id'=>$h['provider']['id'], 'up'=>$http===200, 'http'=>$http];
        curl_multi_remove_handle($mh, $h['ch']);
        curl_close($h['ch']);
    }
    curl_multi_close($mh);
    $R['providers'] = $results;
    $R['up_count'] = count(array_filter($results, function($r){return $r['up'];}));
    $R['parallel_ms'] = round((microtime(true)-$t0)*1000);
    echo json_encode($R, JSON_PRETTY_PRINT);
    exit;
}

if ($action === 'parallel_query') {
    // Send same query to all providers in parallel, return first valid response + all for comparison
    $query = (string)($d['query'] ?? '');
    if (!$query) { http_response_code(400); echo json_encode(['err'=>'no_query']); exit; }
    
    $mh = curl_multi_init();
    $handles = [];
    $t0 = microtime(true);
    foreach ($PROVIDERS as $p) {
        $ch = build_ch($p['url'], json_encode(['message'=>$query]), 25);
        curl_multi_add_handle($mh, $ch);
        $handles[] = ['ch'=>$ch, 'provider'=>$p, 'started'=>microtime(true)];
    }
    $running = null;
    do { curl_multi_exec($mh, $running); curl_multi_select($mh, 0.1); } while ($running > 0);
    
    $results = [];
    foreach ($handles as $h) {
        $body = curl_multi_getcontent($h['ch']);
        $http = curl_getinfo($h['ch'], CURLINFO_HTTP_CODE);
        $parsed = parse_response($body);
        $results[] = [
            'provider_id' => $h['provider']['id'],
            'http' => $http,
            'ms' => round((microtime(true) - $h['started']) * 1000),
            'response_preview' => substr($parsed['text'] ?? '', 0, 300),
            'response_len' => strlen($parsed['text'] ?? ''),
            'engine' => $parsed['provider'] ?? '?'
        ];
        curl_multi_remove_handle($mh, $h['ch']);
        curl_close($h['ch']);
    }
    curl_multi_close($mh);
    
    // Pick winner: fastest with non-empty response
    $valid = array_filter($results, function($r){ return $r['http'] === 200 && $r['response_len'] > 10; });
    usort($valid, function($a,$b){ return $a['ms'] - $b['ms']; });
    $R['winner'] = !empty($valid) ? $valid[0] : null;
    $R['all_results'] = $results;
    $R['total_wall_ms'] = round((microtime(true) - $t0) * 1000);
    $R['speedup'] = count($results) > 0 ? round(array_sum(array_column($results, 'ms')) / $R['total_wall_ms'], 1) . 'x' : 'N/A';
    $R['doctrine'] = '77 — grid GPU parallel (curl_multi multi-endpoint + fastest-wins)';
    
    echo json_encode($R, JSON_PRETTY_PRINT|JSON_UNESCAPED_UNICODE);
    exit;
}

if ($action === 'shard_text') {
    // Split large text into chunks, process each on different provider in parallel
    $text = (string)($d['text'] ?? '');
    $instruction = (string)($d['instruction'] ?? 'Résume');
    if (!$text) { http_response_code(400); echo json_encode(['err'=>'no_text']); exit; }
    
    $chunk_size = 600;
    $chunks = [];
    for ($i = 0; $i < strlen($text); $i += $chunk_size) {
        $chunks[] = substr($text, $i, $chunk_size);
        if (count($chunks) >= 6) break; // cap 6 shards
    }
    
    $t0 = microtime(true);
    $mh = curl_multi_init();
    $handles = [];
    foreach ($chunks as $idx => $chunk) {
        $p = $PROVIDERS[$idx % count($PROVIDERS)];
        $full = "$instruction\n\n$chunk";
        $ch = build_ch($p['url'], json_encode(['message'=>$full]), 20);
        curl_multi_add_handle($mh, $ch);
        $handles[] = ['ch'=>$ch, 'provider'=>$p, 'idx'=>$idx];
    }
    $running = null;
    do { curl_multi_exec($mh, $running); curl_multi_select($mh, 0.1); } while ($running > 0);
    
    $shards = [];
    foreach ($handles as $h) {
        $body = curl_multi_getcontent($h['ch']);
        $parsed = parse_response($body);
        $shards[] = [
            'idx' => $h['idx'],
            'provider' => $h['provider']['id'],
            'response_preview' => substr($parsed['text'] ?? '', 0, 200),
            'response_len' => strlen($parsed['text'] ?? '')
        ];
        curl_multi_remove_handle($mh, $h['ch']);
        curl_close($h['ch']);
    }
    curl_multi_close($mh);
    
    usort($shards, function($a,$b){ return $a['idx'] - $b['idx']; });
    
    $R['chunks_count'] = count($chunks);
    $R['shards'] = $shards;
    $R['total_ms'] = round((microtime(true) - $t0) * 1000);
    $R['merged'] = implode("\n---\n", array_column($shards, 'response_preview'));
    $R['providers_used'] = array_values(array_unique(array_column($shards, 'provider')));
    
    echo json_encode($R, JSON_PRETTY_PRINT|JSON_UNESCAPED_UNICODE);
    exit;
}

http_response_code(400);
echo json_encode(['err'=>'unknown_action', 'available'=>['health','parallel_query','shard_text']]);