session_pool_util.h

Go to the documentation of this file.

// SPDX-License-Identifier: Apache-2.0
#pragma once
 
#include <entropic/types/config.h>
 
#include <algorithm>
#include <string>
 
namespace entropic {
 
struct PoolGeometry {
    int n_seq_max = 1;        
    bool kv_unified = false;  
    int n_ctx = 0;            
    int temp_seq_base = 1;    
};
 
inline PoolGeometry derive_pool_geometry(const ModelConfig& cfg) {
    const int sessions = std::max(1, cfg.max_sessions);
    const int parallel = std::max(1, cfg.n_parallel);
 
    PoolGeometry g;
    g.n_seq_max = std::max(parallel, sessions);
    // gh#98's same-prefix batch fan-out REQUIRES a unified buffer (seq_cp
    // asserts on per-sequence buffers), so preserve that exactly — but a
    // session pool must NOT be unified, per the header. The two are made
    // mutually exclusive at configure time, so this can never be asked to
    // satisfy both.
    g.kv_unified = (parallel > 1) && (sessions <= 1);
    // context_length is PER SESSION.
    g.n_ctx = cfg.context_length * sessions;
    // Sessions own seq ids [0, sessions). The batch fan-out's temp ids must
    // start above them: allocate_temp_seq_id hands out 1, 2, 3... which
    // would otherwise collide with session slots. The configure-time
    // exclusion already makes this unreachable; raising the base is defence
    // in depth against a future change that removes the exclusion silently.
    g.temp_seq_base = std::max(1, sessions);
    return g;
}
 
inline std::string session_pool_conflict_reason(const ModelConfig& cfg) {
    std::string reason;
    if (cfg.max_sessions > 1) {
        if (cfg.n_parallel > 1) {
            reason = "max_sessions > 1 conflicts with n_parallel > 1: a "
                     "session pool needs private per-session KV streams, "
                     "while the same-prefix batch fan-out "
                     "(entropic_run_batch) needs a unified buffer. Set one "
                     "or the other, not both.";
        } else if (cfg.context_length <= 0) {
            reason = "max_sessions > 1 requires a positive context_length: "
                     "the pool allocates context_length * max_sessions "
                     "cells.";
        }
    }
    return reason;
}
 
}  // namespace entropic