Spaces:
Sleeping
Sleeping
File size: 4,027 Bytes
509d302 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 | """
Modal app for running Sakha GRPO training.
Usage:
modal run scripts/modal_train.py --mode demo --task hard --episodes 50
modal run scripts/modal_train.py --mode smoke
"""
import modal
image = (
modal.Image.debian_slim(python_version="3.12")
.apt_install("git")
.pip_install("uv")
.run_commands(
"uv pip install --system 'transformers>=5.2.0' 'trl[quantization]' peft accelerate bitsandbytes jmespath wandb",
"uv pip install --system 'git+https://github.com/meta-pytorch/OpenEnv.git' pydantic fastapi uvicorn openai python-dotenv tenacity",
)
.add_local_dir(
"/home/verma/projects/sakha",
remote_path="/sakha",
ignore=[
".git",
".venv",
"__pycache__",
".pytest_cache",
".ruff_cache",
"artifacts",
".sisyphus",
"*.pyc",
"*.pyo",
],
)
)
volume = modal.Volume.from_name("sakha-training", create_if_missing=True)
app = modal.App("sakha-grpo-training", image=image)
@app.function(
gpu="T4",
timeout=3600,
volumes={"/artifacts": volume},
)
def run_training(
mode: str = "demo",
task: str = "hard",
episodes: int | None = None,
max_steps: int = 96,
model: str = "Qwen/Qwen3-0.6B",
seed: int = 42,
) -> dict:
"""Run GRPO training on Modal."""
import os
import subprocess
import json
import sys
from pathlib import Path
os.environ["PYTHONPATH"] = "/sakha/src"
sys.path.insert(0, "/sakha/src")
install = subprocess.run(
["pip", "install", "--no-deps", "-e", "/sakha"],
capture_output=True,
text=True,
)
if install.returncode != 0:
subprocess.run(["pip", "install", "-e", "/sakha"], capture_output=True, text=True)
output_dir = "/artifacts/grpo"
cmd = [
"python",
"/sakha/scripts/train_grpo.py",
"--mode",
mode,
"--task",
task,
"--model",
model,
"--max-steps",
str(max_steps),
"--seed",
str(seed),
"--output-dir",
output_dir,
]
if episodes is not None:
cmd.extend(["--episodes", str(episodes)])
env = os.environ.copy()
env["PYTHONPATH"] = "/sakha/src"
env["TRL_EXPERIMENTAL_SILENCE"] = "1"
env["WANDB_MODE"] = "disabled"
result = subprocess.run(
cmd,
capture_output=True,
text=True,
cwd="/sakha",
env=env,
)
output = {
"stdout": result.stdout,
"stderr": result.stderr,
"returncode": result.returncode,
"success": result.returncode == 0,
}
results_files = list(Path(output_dir).rglob("results.json"))
if results_files:
latest = max(results_files, key=lambda p: p.stat().st_mtime)
output["results_file"] = str(latest)
output["results"] = json.loads(latest.read_text())
checkpoints = list(Path(output_dir).rglob("checkpoint-*"))
if checkpoints:
output["checkpoints"] = [str(c) for c in checkpoints]
return output
@app.local_entrypoint()
def main(
mode: str = "demo",
task: str = "hard",
episodes: int | None = None,
max_steps: int = 96,
model: str = "Qwen/Qwen3-0.6B",
seed: int = 42,
):
print(f"Running Sakha GRPO training: mode={mode}, task={task}, model={model}")
result = run_training.remote(
mode=mode,
task=task,
episodes=episodes,
max_steps=max_steps,
model=model,
seed=seed,
)
print(f"\nExit code: {result['returncode']}")
print(f"Success: {result['success']}")
print("\n--- STDOUT ---")
print(result["stdout"])
if result["stderr"]:
print("\n--- STDERR ---")
print(result["stderr"])
if "results" in result:
print("\n--- RESULTS ---")
print(result["results"])
if not result["success"]:
raise RuntimeError("Training failed!")
print("\nTraining completed!")
|