mirror of
https://github.com/R0m1k3/Loki.git
synced 2026-10-11 17:26:57 +02:00
nathaninline
6ea2e91810
- LD_LIBRARY_PATH inclut désormais les dossiers runtime CUDA (/usr/local/cuda*/lib64) : corrige le chargement du backend GPU. - Unit systemd : Nice=-10 + CPUSchedulingPolicy=other (priorité CPU). - Met à niveau le reste du projet (models, webauth, presets, web UI).
95 lines
3.2 KiB
Go
95 lines
3.2 KiB
Go
package main
|
|
|
|
import (
|
|
"fmt"
|
|
"os"
|
|
"path/filepath"
|
|
)
|
|
|
|
// cmdServe replaces the historic start.sh: read config.env, build the
|
|
// llama-server invocation, and exec it (replacing this process so systemd
|
|
// supervises llama-server directly).
|
|
func cmdServe(args []string) error {
|
|
cfg := ReadConfig()
|
|
bin := cfg["BIN"]
|
|
if bin == "" {
|
|
return fmt.Errorf("BIN non défini dans %s", confPath())
|
|
}
|
|
model := cfg["MODEL"]
|
|
if model == "" {
|
|
return fmt.Errorf("MODEL non défini dans %s", confPath())
|
|
}
|
|
|
|
// Make sure llama-server can find its bundled shared libraries (the .so/.dll
|
|
// neighbours of the binary). This is platform-specific: LD_LIBRARY_PATH on
|
|
// Linux, PATH on Windows — handled inside execServer.
|
|
setLibraryPath(filepath.Dir(bin))
|
|
|
|
// Sélection GPU (jean gpu) : on filtre les devices visibles par llama-server.
|
|
// CUDA_DEVICE_ORDER=PCI_BUS_ID garantit que les index correspondent à ceux
|
|
// affichés par nvidia-smi (sinon CUDA réordonne par "device le plus rapide").
|
|
if v := cfg["CUDA_VISIBLE_DEVICES"]; v != "" {
|
|
_ = os.Setenv("CUDA_VISIBLE_DEVICES", v)
|
|
_ = os.Setenv("CUDA_DEVICE_ORDER", "PCI_BUS_ID")
|
|
}
|
|
|
|
get := func(key, fallback string) string {
|
|
if v, ok := cfg[key]; ok && v != "" {
|
|
return v
|
|
}
|
|
return fallback
|
|
}
|
|
kv := get("KV_TYPE", "")
|
|
ktv := get("KV_TYPE_K", kv)
|
|
vtv := get("KV_TYPE_V", kv)
|
|
|
|
llmArgs := []string{bin,
|
|
"-m", model,
|
|
"-ngl", get("NGL", "999"),
|
|
"-c", get("CTX", "32768"),
|
|
"-t", get("THREADS", "0"),
|
|
"-tb", get("THREADS_BATCH", "0"),
|
|
"-b", get("BATCH", "2048"),
|
|
"-ub", get("UBATCH", "512"),
|
|
"--host", get("HOST", "0.0.0.0"),
|
|
"--port", get("PORT", "8080"),
|
|
}
|
|
if ktv != "" {
|
|
llmArgs = append(llmArgs, "-ctk", ktv)
|
|
}
|
|
if vtv != "" {
|
|
llmArgs = append(llmArgs, "-ctv", vtv)
|
|
}
|
|
if r := cfg["REASONING"]; r != "" {
|
|
// budget -1 = thinking illimité (défaut). NE PAS forcer 0 : sur llama.cpp
|
|
// vanilla, 0 = "immediate end" → coupe le raisonnement net (le fork
|
|
// ik_llama.cpp, lui, l'ignore, d'où l'ancien comportement trompeur).
|
|
// Configurable via REASONING_BUDGET (ex: 2048 pour plafonner).
|
|
llmArgs = append(llmArgs, "--reasoning", r, "--reasoning-budget", get("REASONING_BUDGET", "-1"))
|
|
}
|
|
// API_KEY protège le serveur quand il est exposé sur internet : llama-server
|
|
// exige alors l'en-tête "Authorization: Bearer <clé>". La clé est lue depuis
|
|
// $JEAN_HOME/.api_key en priorité (elle survit ainsi aux changements de preset
|
|
// qui réécrivent config.env), avec config.env comme repli rétro-compatible.
|
|
if k := readAPIKey(); k != "" {
|
|
llmArgs = append(llmArgs, "--api-key", k)
|
|
} else if k := cfg["API_KEY"]; k != "" {
|
|
llmArgs = append(llmArgs, "--api-key", k)
|
|
}
|
|
// EXTRA_ARGS is appended verbatim, split on whitespace like the shell would.
|
|
for _, a := range trimSplit(cfg["EXTRA_ARGS"], " ") {
|
|
llmArgs = append(llmArgs, a)
|
|
}
|
|
|
|
// Working dir = JEAN_HOME so relative paths in EXTRA_ARGS (e.g. --mmproj
|
|
// mmproj-F16.gguf) still resolve.
|
|
_ = os.Chdir(JeanHome())
|
|
|
|
fmt.Fprintf(os.Stderr, "[jean serve] %s model=%s port=%s\n",
|
|
bin, filepath.Base(model), get("PORT", "8080"))
|
|
|
|
// Hand off to the llama-server process. On Unix this replaces the current
|
|
// process (exec); on Windows it runs as a child and waits. See platform_*.go.
|
|
return execServer(bin, llmArgs)
|
|
}
|