mirror of
https://github.com/NicolasBohn/NexQuant.git
synced 2026-07-27 15:37:44 +00:00
feat: Add start_llama and start_loop CLI commands
Added to rdagent CLI: - rdagent start_llama: Start llama.cpp server (replaces start_llama.sh) - rdagent start_loop: Start strategy generator loop (replaces start_strategy_loop.sh) Features: - start_llama: --model, --port, --gpu-layers, --ctx-size, --reasoning options - start_loop: --target, --max-wait options with auto-restart on crash - Both commands have full help (--help) and examples Moved .sh scripts to scripts/: - start_llama.sh → scripts/ (kept as reference) - start_strategy_loop.sh → scripts/ (kept as reference) Usage: rdagent start_llama # Start LLM server rdagent start_llama --gpu-layers 40 # Custom GPU layers rdagent start_loop # Start strategy loop rdagent start_loop --target 5 # Generate 5 strategies per run
This commit is contained in:
TPTBusiness
@@ -1023,3 +1023,222 @@ def _generate_single_strategy_report(strategy_file: Path, output_dir: Path) -> D
|
||||
|
||||
if __name__ == "__main__":
|
||||
app()
|
||||
|
||||
|
||||
@app.command(name="start_llama")
|
||||
def start_llama_cli(
|
||||
model: str = typer.Option(
|
||||
None, "--model", "-m", help="Path to model file"
|
||||
),
|
||||
port: int = typer.Option(8081, "--port", "-p", help="Server port"),
|
||||
gpu_layers: int = typer.Option(30, "--gpu-layers", "-g", help="GPU layers"),
|
||||
ctx_size: int = typer.Option(80000, "--ctx-size", "-c", help="Context size"),
|
||||
reasoning: bool = typer.Option(False, "--reasoning", help="Enable reasoning mode"),
|
||||
):
|
||||
"""
|
||||
Start llama.cpp server for local LLM inference.
|
||||
|
||||
Options:
|
||||
--model/-m: Path to model file (default: from .env or ~/models/qwen3.5/)
|
||||
--port/-p: Server port (default: 8081)
|
||||
--gpu-layers/-g: GPU layers (default: 30)
|
||||
--ctx-size/-c: Context size (default: 80000)
|
||||
--reasoning: Enable reasoning mode (default: off)
|
||||
|
||||
Examples:
|
||||
rdagent start_llama
|
||||
rdagent start_llama --gpu-layers 40 --ctx-size 4096
|
||||
rdagent start_llama --reasoning
|
||||
"""
|
||||
import subprocess
|
||||
import sys
|
||||
import os
|
||||
|
||||
model_path = model or os.getenv(
|
||||
"LLAMA_MODEL_PATH",
|
||||
str(Path.home() / "models" / "qwen3.5" / "Qwen3.5-35B-A3B-Q3_K_M.gguf"),
|
||||
)
|
||||
|
||||
llama_server = str(Path.home() / "llama.cpp" / "build" / "bin" / "llama-server")
|
||||
|
||||
if not Path(llama_server).exists():
|
||||
print(f"❌ llama.cpp server not found: {llama_server}")
|
||||
print("\nBuild it first:")
|
||||
print(" cd ~/llama.cpp && mkdir -p build && cd build && cmake .. && make")
|
||||
sys.exit(1)
|
||||
|
||||
if not Path(model_path).exists():
|
||||
print(f"❌ Model not found: {model_path}")
|
||||
sys.exit(1)
|
||||
|
||||
cmd = [
|
||||
llama_server,
|
||||
"--model", model_path,
|
||||
"--n-gpu-layers", str(gpu_layers),
|
||||
"--ctx-size", str(ctx_size),
|
||||
"--port", str(port),
|
||||
"--threads", "8",
|
||||
"--threads-batch", "8",
|
||||
"--parallel", "1",
|
||||
"--flash-attn",
|
||||
"--jinja",
|
||||
"--host", "0.0.0.0",
|
||||
]
|
||||
|
||||
if not reasoning:
|
||||
cmd.extend(["--reasoning", "off"])
|
||||
|
||||
print(f"🚀 Starting llama.cpp server...")
|
||||
print(f" Model: {Path(model_path).name}")
|
||||
print(f" Port: {port}")
|
||||
print(f" GPU Layers: {gpu_layers}")
|
||||
print(f" Context: {ctx_size}")
|
||||
print(f" Reasoning: {'on' if reasoning else 'off'}")
|
||||
print()
|
||||
|
||||
try:
|
||||
os.execvp(cmd[0], cmd)
|
||||
except Exception as e:
|
||||
print(f"❌ Failed to start llama.cpp server: {e}")
|
||||
sys.exit(1)
|
||||
|
||||
|
||||
@app.command(name="start_loop")
|
||||
def start_loop_cli(
|
||||
target_count: int = typer.Option(3, "--target", "-t", help="Strategies per run"),
|
||||
max_wait: int = typer.Option(1800, "--max-wait", "-w", help="Max wait per run (seconds)"),
|
||||
):
|
||||
"""
|
||||
Start PREDIX strategy generator loop.
|
||||
|
||||
Runs continuously, generating strategies with automatic restart on crash.
|
||||
|
||||
Options:
|
||||
--target/-t: Strategies to generate per run (default: 3)
|
||||
--max-wait/-w: Max wait time per run in seconds (default: 1800 = 30min)
|
||||
|
||||
Examples:
|
||||
rdagent start_loop
|
||||
rdagent start_loop --target 5 --max-wait 3600
|
||||
"""
|
||||
import subprocess
|
||||
import signal
|
||||
import sys
|
||||
import os
|
||||
from datetime import datetime
|
||||
import time
|
||||
|
||||
script_dir = str(Path(__file__).parent.parent.parent.parent)
|
||||
generator = f"python {script_dir}/scripts/predix_smart_strategy_gen.py"
|
||||
logfile = f"{script_dir}/results/logs/generator_loop.log"
|
||||
pidfile = "/tmp/predix_loop.pid"
|
||||
|
||||
os.makedirs(f"{script_dir}/results/logs", exist_ok=True)
|
||||
|
||||
def log(msg: str):
|
||||
ts = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
|
||||
line = f"{ts} - {msg}"
|
||||
print(line)
|
||||
with open(logfile, "a") as f:
|
||||
f.write(line + "\n")
|
||||
|
||||
def cleanup(signum=None, frame=None):
|
||||
log("Received termination signal. Cleaning up...")
|
||||
try:
|
||||
subprocess.run(["pkill", "-f", "predix_smart_strategy_gen.py"], capture_output=True)
|
||||
except Exception:
|
||||
pass
|
||||
try:
|
||||
os.remove(pidfile)
|
||||
except FileNotFoundError:
|
||||
pass
|
||||
log("Cleanup complete. Exiting.")
|
||||
sys.exit(0)
|
||||
|
||||
signal.signal(signal.SIGTERM, cleanup)
|
||||
signal.signal(signal.SIGINT, cleanup)
|
||||
|
||||
with open(pidfile, "w") as f:
|
||||
f.write(str(os.getpid()))
|
||||
|
||||
log("=========================================")
|
||||
log("🚀 PREDIX Generator Loop Starting")
|
||||
log("=========================================")
|
||||
log(f"Target: {target_count} strategies per run")
|
||||
log(f"Max wait: {max_wait}s per run")
|
||||
log(f"Log: {logfile}")
|
||||
|
||||
attempt = 0
|
||||
|
||||
while True:
|
||||
attempt += 1
|
||||
log("")
|
||||
log(f"=== Attempt #{attempt} ===================================")
|
||||
|
||||
# Check disk space
|
||||
try:
|
||||
usage = subprocess.run(["df", "-h", script_dir], capture_output=True, text=True)
|
||||
disk_line = usage.stdout.strip().split("\n")[-1]
|
||||
pct = int(disk_line.split()[4].replace("%", ""))
|
||||
if pct > 90:
|
||||
log(f"⚠️ Disk usage at {pct}%. Pausing...")
|
||||
time.sleep(300)
|
||||
continue
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
# Count existing strategies
|
||||
from pathlib import Path as P
|
||||
strat_dir = P(f"{script_dir}/results/strategies_new")
|
||||
strat_count = len(list(strat_dir.glob("*.json"))) if strat_dir.exists() else 0
|
||||
log(f"📁 Existing strategies: {strat_count}")
|
||||
|
||||
# Kill stale processes
|
||||
try:
|
||||
subprocess.run(["pkill", "-9", "-f", "predix_smart_strategy_gen.py"], capture_output=True)
|
||||
except Exception:
|
||||
pass
|
||||
time.sleep(2)
|
||||
|
||||
# Start generator
|
||||
log("🤖 Starting generator...")
|
||||
proc = subprocess.Popen(
|
||||
generator.split(),
|
||||
cwd=script_dir,
|
||||
stdout=subprocess.DEVNULL,
|
||||
stderr=subprocess.DEVNULL,
|
||||
)
|
||||
log(f" PID: {proc.pid}")
|
||||
|
||||
# Monitor progress
|
||||
elapsed = 0
|
||||
while proc.poll() is None:
|
||||
time.sleep(30)
|
||||
elapsed += 30
|
||||
|
||||
if elapsed % 120 == 0:
|
||||
log(f" ⏱️ {elapsed}s elapsed")
|
||||
|
||||
if elapsed >= max_wait:
|
||||
log(f" ⏰ Timeout after {elapsed}s. Killing...")
|
||||
proc.kill()
|
||||
break
|
||||
|
||||
# Check results
|
||||
exit_code = proc.wait()
|
||||
if exit_code == 0:
|
||||
log("✅ Generator completed successfully")
|
||||
elif exit_code == -9:
|
||||
log("❌ Generator killed (OOM? Exit 137)")
|
||||
else:
|
||||
log(f"⚠️ Generator exited with code {exit_code}")
|
||||
|
||||
# Count new strategies
|
||||
new_strats = sorted(strat_dir.glob("*.json"), key=lambda x: x.stat().st_mtime, reverse=True)[:3]
|
||||
if new_strats:
|
||||
log("📊 Latest strategies:")
|
||||
for s in new_strats:
|
||||
log(f" - {s.name}")
|
||||
|
||||
log("⏳ Waiting 60s before next attempt...")
|
||||
time.sleep(60)
|
||||
|
||||
Reference in New Issue
Block a user