Publish bounded Qwen3-32B six-GPU smoke 2312e1859361
Browse files
six-gpu-smoke-2312e1859361.py.txt
ADDED
|
@@ -0,0 +1,568 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Run a two-step NeMo-RL spreadsheet smoke using the existing container runtime.
|
| 2 |
+
|
| 3 |
+
Writes only a fresh run directory and an explicitly declared short IPC root.
|
| 4 |
+
The training subprocess uses existing /opt actor venvs, offline local weights,
|
| 5 |
+
an inherited smoke YAML, and redacted logs. This is not a quality benchmark.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
import argparse
|
| 9 |
+
import hashlib
|
| 10 |
+
import json
|
| 11 |
+
import os
|
| 12 |
+
import re
|
| 13 |
+
import signal
|
| 14 |
+
import socket
|
| 15 |
+
import subprocess
|
| 16 |
+
import sys
|
| 17 |
+
import threading
|
| 18 |
+
import time
|
| 19 |
+
from pathlib import Path
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
ACTORS = (
|
| 23 |
+
"nemo_rl.algorithms.async_utils.AsyncTrajectoryCollector",
|
| 24 |
+
"nemo_rl.algorithms.async_utils.ReplayBuffer",
|
| 25 |
+
"nemo_rl.environments.nemo_gym.NemoGym",
|
| 26 |
+
"nemo_rl.models.generation.vllm.vllm_worker_async.VllmAsyncGenerationWorker",
|
| 27 |
+
"nemo_rl.models.policy.workers.megatron_policy_worker.MegatronPolicyWorker",
|
| 28 |
+
)
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
def sha(path):
|
| 32 |
+
with path.open("rb") as stream:
|
| 33 |
+
return hashlib.file_digest(stream, "sha256").hexdigest()
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
def model_metadata(model):
|
| 37 |
+
config = json.loads((model / "config.json").read_text())
|
| 38 |
+
tokenizer = json.loads((model / "tokenizer_config.json").read_text())
|
| 39 |
+
template = tokenizer.get("chat_template", "")
|
| 40 |
+
if not isinstance(template, str):
|
| 41 |
+
raise ValueError("Review a non-string model chat template before running the smoke")
|
| 42 |
+
if "<function=" in template:
|
| 43 |
+
parser = "qwen3_coder"
|
| 44 |
+
elif "<tool_call>" in template and '"arguments"' in template:
|
| 45 |
+
parser = "hermes"
|
| 46 |
+
else:
|
| 47 |
+
raise ValueError("The model tool-call format is not covered by this smoke")
|
| 48 |
+
special = {value["content"]: int(key) for key, value in tokenizer.get("added_tokens_decoder", {}).items()}
|
| 49 |
+
eos = tokenizer.get("eos_token")
|
| 50 |
+
eos = eos.get("content") if isinstance(eos, dict) else eos
|
| 51 |
+
return {
|
| 52 |
+
"model_type": config.get("model_type"),
|
| 53 |
+
"tool_parser": parser,
|
| 54 |
+
"eos": special.get(eos, config.get("eos_token_id")),
|
| 55 |
+
"think_open": special.get("<think>"),
|
| 56 |
+
"think_close": special.get("</think>"),
|
| 57 |
+
"config_sha256": sha(model / "config.json"),
|
| 58 |
+
"tokenizer_config_sha256": sha(model / "tokenizer_config.json"),
|
| 59 |
+
}
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
def smoke_overrides(base, root, subset, metadata, six_gpu=False):
|
| 63 |
+
from omegaconf import OmegaConf
|
| 64 |
+
|
| 65 |
+
gym = OmegaConf.to_container(base.env.nemo_gym, resolve=False)
|
| 66 |
+
# The inherited template defines unrelated judges even when their config
|
| 67 |
+
# files are absent. Keep only the policy model servers used by this task.
|
| 68 |
+
for name, value in list(gym.items()):
|
| 69 |
+
if isinstance(value, dict) and any(
|
| 70 |
+
key in value for key in ("resources_servers", "responses_api_agents", "responses_api_models", "_copy")
|
| 71 |
+
):
|
| 72 |
+
if name not in ("policy_model", "policy_model_reasoning_off"):
|
| 73 |
+
del gym[name]
|
| 74 |
+
gym["nemo_gym_log_dir"] = str(root / "nemo_gym")
|
| 75 |
+
gym["policy_model"]["responses_api_models"]["vllm_model"]["num_workers"] = 2
|
| 76 |
+
if "policy_model_reasoning_off" in gym:
|
| 77 |
+
gym["policy_model_reasoning_off"]["responses_api_models"]["vllm_model"]["num_workers"] = 1
|
| 78 |
+
gym["spreadsheet_agent"] = {"responses_api_agents": {"spreadsheet_agent": {"max_steps": 6}}}
|
| 79 |
+
overrides = {
|
| 80 |
+
"checkpointing": {"enabled": False, "checkpoint_dir": str(root / "ckpt")},
|
| 81 |
+
"grpo": {
|
| 82 |
+
"num_prompts_per_step": 2,
|
| 83 |
+
"num_generations_per_prompt": 2,
|
| 84 |
+
"max_num_steps": 2,
|
| 85 |
+
"val_at_start": False,
|
| 86 |
+
"val_at_end": False,
|
| 87 |
+
"val_period": -1,
|
| 88 |
+
},
|
| 89 |
+
"policy": {
|
| 90 |
+
"generation": {
|
| 91 |
+
"max_new_tokens": 4096,
|
| 92 |
+
"vllm_cfg": {
|
| 93 |
+
"gpu_memory_utilization": 0.65,
|
| 94 |
+
"http_server_serving_chat_kwargs": {"tool_parser": metadata["tool_parser"]},
|
| 95 |
+
},
|
| 96 |
+
"vllm_kwargs": {"max_num_seqs": 8, "max_num_batched_tokens": 32768},
|
| 97 |
+
},
|
| 98 |
+
"megatron_cfg": {"sequence_parallel": False, "scheduler": {"lr_warmup_iters": 1}},
|
| 99 |
+
},
|
| 100 |
+
"data": {"shuffle": False, "train": [{"data_path": str(subset)}]},
|
| 101 |
+
"env": {
|
| 102 |
+
"_override_": True,
|
| 103 |
+
"should_use_nemo_gym": True,
|
| 104 |
+
"should_log_nemo_gym_responses": True,
|
| 105 |
+
"should_log_hashed_nemo_gym_responses": False,
|
| 106 |
+
"nemo_gym": gym,
|
| 107 |
+
},
|
| 108 |
+
"logger": {
|
| 109 |
+
"log_dir": str(root / "logs"),
|
| 110 |
+
"wandb_enabled": False,
|
| 111 |
+
"tensorboard_enabled": True,
|
| 112 |
+
"mlflow_enabled": False,
|
| 113 |
+
"swanlab_enabled": False,
|
| 114 |
+
"monitor_gpus": False,
|
| 115 |
+
},
|
| 116 |
+
"token_ids": {key: metadata[key] for key in ("eos", "think_open", "think_close") if metadata[key] is not None},
|
| 117 |
+
}
|
| 118 |
+
if six_gpu:
|
| 119 |
+
overrides["cluster"] = {"num_nodes": 1, "gpus_per_node": 6}
|
| 120 |
+
overrides["data"]["max_input_seq_length"] = 4096
|
| 121 |
+
policy = overrides["policy"]
|
| 122 |
+
policy.update(
|
| 123 |
+
precision="bfloat16",
|
| 124 |
+
train_global_batch_size=4,
|
| 125 |
+
train_micro_batch_size=1,
|
| 126 |
+
logprob_batch_size=1,
|
| 127 |
+
logprob_chunk_size=4,
|
| 128 |
+
max_total_sequence_length=8192,
|
| 129 |
+
)
|
| 130 |
+
policy["megatron_cfg"].update(
|
| 131 |
+
tensor_model_parallel_size=4,
|
| 132 |
+
pipeline_model_parallel_size=1,
|
| 133 |
+
context_parallel_size=1,
|
| 134 |
+
sequence_parallel=True,
|
| 135 |
+
activation_checkpointing=True,
|
| 136 |
+
)
|
| 137 |
+
policy["megatron_cfg"]["optimizer"] = {"optimizer_cpu_offload": False, "optimizer_offload_fraction": 0.0}
|
| 138 |
+
generation = policy["generation"]
|
| 139 |
+
generation["colocated"] = {"enabled": False, "resources": {"gpus_per_node": 2, "num_nodes": 1}}
|
| 140 |
+
generation["vllm_cfg"].update(
|
| 141 |
+
tensor_parallel_size=2, pipeline_parallel_size=1, expert_parallel_size=1, max_model_len=8192
|
| 142 |
+
)
|
| 143 |
+
generation["vllm_kwargs"]["max_num_batched_tokens"] = 16384
|
| 144 |
+
return overrides
|
| 145 |
+
|
| 146 |
+
|
| 147 |
+
def clean_environment(root, ray_root, gym_root):
|
| 148 |
+
prefixes = ("CUDA_", "NVIDIA_", "NCCL_", "NRL_", "NEMO_", "NVTE_", "TORCH_", "OMP_", "MKL_")
|
| 149 |
+
keep = {
|
| 150 |
+
"PATH",
|
| 151 |
+
"PYTHONPATH",
|
| 152 |
+
"LD_LIBRARY_PATH",
|
| 153 |
+
"LIBRARY_PATH",
|
| 154 |
+
"HOME",
|
| 155 |
+
"USER",
|
| 156 |
+
"LOGNAME",
|
| 157 |
+
"SHELL",
|
| 158 |
+
"LANG",
|
| 159 |
+
"LC_ALL",
|
| 160 |
+
"TZ",
|
| 161 |
+
}
|
| 162 |
+
env = {
|
| 163 |
+
key: value
|
| 164 |
+
for key, value in os.environ.items()
|
| 165 |
+
if (key in keep or key.startswith(prefixes)) and not re.search(r"TOKEN|PASSWORD|SECRET|API_KEY", key)
|
| 166 |
+
}
|
| 167 |
+
paths = {
|
| 168 |
+
"NRL_MEGATRON_CHECKPOINT_DIR": "cache/nemo_rl",
|
| 169 |
+
"VLLM_CACHE_ROOT": "cache/vllm",
|
| 170 |
+
"HF_HOME": "cache/huggingface",
|
| 171 |
+
"HF_DATASETS_CACHE": "cache/huggingface/datasets",
|
| 172 |
+
"HUGGINGFACE_HUB_CACHE": "cache/huggingface/models",
|
| 173 |
+
"HF_MODULES_CACHE": "cache/hf_modules",
|
| 174 |
+
"TORCH_HOME": "cache/torch",
|
| 175 |
+
"TORCH_EXTENSIONS_DIR": "cache/torch_extensions",
|
| 176 |
+
"TORCHINDUCTOR_CACHE_DIR": "cache/torchinductor",
|
| 177 |
+
"TRITON_CACHE_DIR": "cache/triton",
|
| 178 |
+
"CUDA_CACHE_PATH": "cache/cuda",
|
| 179 |
+
"XDG_CACHE_HOME": "cache/xdg",
|
| 180 |
+
"UV_CACHE_DIR": "cache/uv",
|
| 181 |
+
}
|
| 182 |
+
for key, relative in paths.items():
|
| 183 |
+
path = root / relative
|
| 184 |
+
path.mkdir(parents=True, exist_ok=True)
|
| 185 |
+
env[key] = str(path)
|
| 186 |
+
# Include Ray's longest dashboard socket as well as its object store.
|
| 187 |
+
if len(str(ray_root / "session_9999-99-99_99-99-99_999999_9999999/sockets/dash_MetricsHead").encode()) > 107:
|
| 188 |
+
raise ValueError("Ray socket path is too long; select a shorter declared Ray root")
|
| 189 |
+
temp = ray_root / "tmp"
|
| 190 |
+
if len(str(temp / "99999999-9999-9999-9999-999999999999").encode()) > 107:
|
| 191 |
+
raise ValueError("vLLM IPC socket path is too long; select a shorter IPC root")
|
| 192 |
+
temp.mkdir(mode=0o700)
|
| 193 |
+
env["TMPDIR"] = str(temp)
|
| 194 |
+
env.update(
|
| 195 |
+
RAY_TMPDIR=str(ray_root),
|
| 196 |
+
PYTHONDONTWRITEBYTECODE="1",
|
| 197 |
+
TOKENIZERS_PARALLELISM="false",
|
| 198 |
+
HF_HUB_OFFLINE="1",
|
| 199 |
+
TRANSFORMERS_OFFLINE="1",
|
| 200 |
+
UV_OFFLINE="1",
|
| 201 |
+
PIP_NO_INDEX="1",
|
| 202 |
+
NRL_FORCE_REBUILD_VENVS="false",
|
| 203 |
+
NEMO_GYM_VENV_DIR=str(gym_root),
|
| 204 |
+
NO_PROXY="*",
|
| 205 |
+
no_proxy="*",
|
| 206 |
+
CUDA_DEVICE_MAX_CONNECTIONS="1",
|
| 207 |
+
OMP_NUM_THREADS="1",
|
| 208 |
+
MKL_NUM_THREADS="1",
|
| 209 |
+
OPENBLAS_NUM_THREADS="1",
|
| 210 |
+
WANDB_MODE="disabled",
|
| 211 |
+
NCCL_TIMEOUT="600",
|
| 212 |
+
RAY_DEFAULT_OBJECT_STORE_MAX_MEMORY_BYTES=str(8 * 1024**3),
|
| 213 |
+
)
|
| 214 |
+
return env
|
| 215 |
+
|
| 216 |
+
|
| 217 |
+
def redact(text, secrets):
|
| 218 |
+
for secret in secrets:
|
| 219 |
+
if secret:
|
| 220 |
+
text = text.replace(secret, "[REDACTED]")
|
| 221 |
+
return text
|
| 222 |
+
|
| 223 |
+
|
| 224 |
+
def gpu_snapshot():
|
| 225 |
+
result = subprocess.run(
|
| 226 |
+
["nvidia-smi", "--query-gpu=index,uuid,memory.used,utilization.gpu", "--format=csv,noheader,nounits"],
|
| 227 |
+
capture_output=True,
|
| 228 |
+
text=True,
|
| 229 |
+
check=True,
|
| 230 |
+
)
|
| 231 |
+
return [
|
| 232 |
+
dict(zip(("index", "uuid", "memory_mb", "utilization"), (part.strip() for part in line.split(","))))
|
| 233 |
+
for line in result.stdout.splitlines()
|
| 234 |
+
]
|
| 235 |
+
|
| 236 |
+
|
| 237 |
+
def redact_artifacts(roots, secrets):
|
| 238 |
+
changed = 0
|
| 239 |
+
for root in roots:
|
| 240 |
+
for path in root.rglob("*"):
|
| 241 |
+
if path.suffix not in {".log", ".out", ".err", ".json", ".jsonl", ".yaml"}:
|
| 242 |
+
continue
|
| 243 |
+
if (
|
| 244 |
+
path.is_symlink()
|
| 245 |
+
or not path.is_file()
|
| 246 |
+
or path.stat().st_nlink != 1
|
| 247 |
+
or path.stat().st_size > 16 * 1024**2
|
| 248 |
+
):
|
| 249 |
+
continue
|
| 250 |
+
try:
|
| 251 |
+
text = path.read_text()
|
| 252 |
+
except UnicodeDecodeError:
|
| 253 |
+
continue
|
| 254 |
+
safe = redact(text, secrets)
|
| 255 |
+
if safe != text:
|
| 256 |
+
path.write_text(safe)
|
| 257 |
+
changed += 1
|
| 258 |
+
return changed
|
| 259 |
+
|
| 260 |
+
|
| 261 |
+
def existing_actor_root(trainer_root):
|
| 262 |
+
selected = os.environ.get("NEMO_RL_VENV_DIR")
|
| 263 |
+
candidates = [Path(selected)] if selected else [trainer_root / "venvs", Path("/opt/ray_venvs")]
|
| 264 |
+
for candidate in candidates:
|
| 265 |
+
if all((candidate / actor / "bin/python").is_file() for actor in ACTORS):
|
| 266 |
+
return candidate.resolve()
|
| 267 |
+
raise ValueError(
|
| 268 |
+
"Missing prebuilt actor venvs; refusing to bootstrap. Checked: " + ", ".join(map(str, candidates))
|
| 269 |
+
)
|
| 270 |
+
|
| 271 |
+
|
| 272 |
+
def main():
|
| 273 |
+
parser = argparse.ArgumentParser(description=__doc__)
|
| 274 |
+
for name in ("trainer-root", "base-config", "run-dir", "ipc-root", "private-config"):
|
| 275 |
+
parser.add_argument("--" + name, type=Path, required=True)
|
| 276 |
+
parser.add_argument("--expected-hostname", required=True)
|
| 277 |
+
parser.add_argument("--max-wall-seconds", type=int, default=1800)
|
| 278 |
+
parser.add_argument("--gpu-indices", help="Qwen3-32B six-GPU smoke profile, e.g. 0,1,2,3,4,5")
|
| 279 |
+
parser.add_argument("--helpers-root", type=Path, default=Path(__file__).resolve().parent)
|
| 280 |
+
parser.add_argument("--reuse-cache-root", type=Path, help="Existing launcher directory containing cache/")
|
| 281 |
+
args = parser.parse_args()
|
| 282 |
+
if socket.gethostname() != args.expected_hostname:
|
| 283 |
+
parser.error("Wrong training instance")
|
| 284 |
+
root, ipc_root = args.run_dir.resolve(), args.ipc_root.resolve()
|
| 285 |
+
ray_root = ipc_root
|
| 286 |
+
if root.exists() or ipc_root.exists():
|
| 287 |
+
parser.error("Run and IPC directories must both be fresh")
|
| 288 |
+
root.mkdir(mode=0o700)
|
| 289 |
+
ipc_root.mkdir(mode=0o700)
|
| 290 |
+
from nemo_rl.utils.config import load_config, register_omegaconf_resolvers
|
| 291 |
+
from omegaconf import OmegaConf
|
| 292 |
+
|
| 293 |
+
register_omegaconf_resolvers()
|
| 294 |
+
base = load_config(args.base_config)
|
| 295 |
+
if args.gpu_indices and Path(load_config.__code__.co_filename).resolve().parents[2] != args.trainer_root.resolve():
|
| 296 |
+
raise ValueError("Activation selected a different trainer checkout")
|
| 297 |
+
metadata = model_metadata(Path(base.policy.model_name))
|
| 298 |
+
six_gpu = args.gpu_indices is not None
|
| 299 |
+
scope = None
|
| 300 |
+
if six_gpu:
|
| 301 |
+
sys.path.insert(0, str(args.helpers_root))
|
| 302 |
+
from gpu_scope import inspect_selected, parse_indices
|
| 303 |
+
|
| 304 |
+
indices = parse_indices(args.gpu_indices)
|
| 305 |
+
if indices != list(range(6)) or int(base.cluster.num_nodes) != 1 or int(base.cluster.gpus_per_node) != 6:
|
| 306 |
+
raise ValueError("Require the assigned single-node GPU 0-5 layout")
|
| 307 |
+
if Path(base.policy.model_name).name != "Qwen3-32B" or metadata["model_type"] != "qwen3":
|
| 308 |
+
raise ValueError("This six-GPU profile requires the selected dense Qwen3-32B model")
|
| 309 |
+
if not OmegaConf.select(base, "policy.megatron_cfg.enabled") or base.policy.generation.backend != "vllm":
|
| 310 |
+
raise ValueError("The six-GPU profile requires the existing Megatron trainer and vLLM backend")
|
| 311 |
+
if (
|
| 312 |
+
base.get("on_policy_distillation", {}).get("enabled", False)
|
| 313 |
+
or base.grpo.get("adv_estimator", {}).get("name") == "opd"
|
| 314 |
+
):
|
| 315 |
+
raise ValueError("This is an RLVR smoke, not an OPD run")
|
| 316 |
+
scope = inspect_selected(indices, args.expected_hostname)
|
| 317 |
+
capacity = subprocess.check_output(
|
| 318 |
+
["nvidia-smi", "-i", args.gpu_indices, "--query-gpu=memory.total", "--format=csv,noheader,nounits"],
|
| 319 |
+
text=True,
|
| 320 |
+
)
|
| 321 |
+
if len(capacity.splitlines()) != 6 or any(int(line.strip()) < 160000 for line in capacity.splitlines()):
|
| 322 |
+
raise ValueError("The six-GPU profile requires the assigned large-memory GPUs")
|
| 323 |
+
elif int(base.cluster.num_nodes) != 1 or int(base.cluster.gpus_per_node) != 2:
|
| 324 |
+
raise ValueError("Legacy smoke requires two GPUs; select --gpu-indices for the six-GPU profile")
|
| 325 |
+
actor_root = (
|
| 326 |
+
existing_actor_root(args.trainer_root)
|
| 327 |
+
if six_gpu
|
| 328 |
+
else Path(os.environ.get("NEMO_RL_VENV_DIR", "/opt/ray_venvs"))
|
| 329 |
+
)
|
| 330 |
+
for actor in ACTORS:
|
| 331 |
+
if not (actor_root / actor / "bin/python").is_file():
|
| 332 |
+
raise ValueError(f"Missing prebuilt actor venv; refusing to bootstrap: {actor}")
|
| 333 |
+
source = Path(base.data.train[0].data_path)
|
| 334 |
+
with source.open() as stream:
|
| 335 |
+
rows = [json.loads(next(stream)) for _ in range(4)]
|
| 336 |
+
if any(row.get("agent_ref", {}).get("name") != "spreadsheet_agent" for row in rows):
|
| 337 |
+
raise ValueError("Smoke rows must route to spreadsheet_agent")
|
| 338 |
+
subset = root / "train.jsonl"
|
| 339 |
+
subset.write_text("".join(json.dumps(row, ensure_ascii=False) + "\n" for row in rows))
|
| 340 |
+
overrides = smoke_overrides(base, root, subset, metadata, six_gpu=six_gpu)
|
| 341 |
+
from nemo_gym import PARENT_DIR as gym_source
|
| 342 |
+
|
| 343 |
+
if six_gpu:
|
| 344 |
+
if Path(gym_source).resolve() != (args.trainer_root / "3rdparty/Gym-workspace/Gym").resolve():
|
| 345 |
+
raise ValueError("Activation selected a different Gym checkout")
|
| 346 |
+
from prepare_training_node import probe_venv, service_probe
|
| 347 |
+
import importlib.metadata as package_metadata
|
| 348 |
+
|
| 349 |
+
expected_versions = {name: package_metadata.version(name) for name in ("ray", "openai")}
|
| 350 |
+
cache = Path(base.env.nemo_gym.uv_venv_dir)
|
| 351 |
+
for relative, module in (
|
| 352 |
+
("responses_api_models/vllm_model", None),
|
| 353 |
+
("resources_servers/spreadsheet", "resources_servers.spreadsheet.app"),
|
| 354 |
+
("responses_api_agents/spreadsheet_agent", "responses_api_agents.spreadsheet_agent.app"),
|
| 355 |
+
):
|
| 356 |
+
result = probe_venv(cache / relative / ".venv", Path(gym_source), module)
|
| 357 |
+
if (
|
| 358 |
+
Path(result["prefix"]).resolve() != Path(result["venv_path"]).resolve()
|
| 359 |
+
or result["versions"] != expected_versions
|
| 360 |
+
):
|
| 361 |
+
raise ValueError("Existing cache interpreter/version mismatch")
|
| 362 |
+
service_probe(args.private_config, "LibreOffice 24.2.7.2 420(Build:2)")
|
| 363 |
+
|
| 364 |
+
audit_agent = args.helpers_root / "audit_agent.py"
|
| 365 |
+
agent = overrides["env"]["nemo_gym"]["spreadsheet_agent"]["responses_api_agents"]["spreadsheet_agent"]
|
| 366 |
+
agent.update(
|
| 367 |
+
entrypoint=os.path.relpath(audit_agent, gym_source / "responses_api_agents/spreadsheet_agent"),
|
| 368 |
+
num_workers=1,
|
| 369 |
+
audit_path=str(root / "model_rollouts.jsonl"),
|
| 370 |
+
)
|
| 371 |
+
overrides = {"defaults": str(args.base_config.resolve()), **overrides}
|
| 372 |
+
config = root / "smoke.yaml"
|
| 373 |
+
config.write_text(OmegaConf.to_yaml(OmegaConf.create(overrides), resolve=False))
|
| 374 |
+
resolved = OmegaConf.to_container(load_config(config), resolve=True)
|
| 375 |
+
assert resolved["grpo"]["max_num_steps"] == 2
|
| 376 |
+
assert resolved["policy"]["train_global_batch_size"] == 4
|
| 377 |
+
if six_gpu:
|
| 378 |
+
assert resolved["policy"]["megatron_cfg"]["tensor_model_parallel_size"] == 4
|
| 379 |
+
assert resolved["policy"]["generation"]["vllm_cfg"]["tensor_parallel_size"] == 2
|
| 380 |
+
# Freeze the effective smoke config, avoiding a later reread of the
|
| 381 |
+
# user's YAML/defaults while this run is in flight.
|
| 382 |
+
config.write_text(OmegaConf.to_yaml(OmegaConf.create(resolved), resolve=True))
|
| 383 |
+
print(
|
| 384 |
+
json.dumps(
|
| 385 |
+
{
|
| 386 |
+
"event": "SIX_GPU_CONFIG_CHECKED",
|
| 387 |
+
"model": resolved["policy"]["model_name"],
|
| 388 |
+
"train_gpus": 4,
|
| 389 |
+
"train_tp": 4,
|
| 390 |
+
"train_dp": 1,
|
| 391 |
+
"rollout_gpus": 2,
|
| 392 |
+
"rollout_tp": 2,
|
| 393 |
+
"context": 8192,
|
| 394 |
+
"precision": "bfloat16",
|
| 395 |
+
"optimizer_cpu_offload": False,
|
| 396 |
+
"max_new_tokens": 4096,
|
| 397 |
+
"optimizer_steps": 2,
|
| 398 |
+
"prompts_per_step": 2,
|
| 399 |
+
"rollouts_per_prompt": 2,
|
| 400 |
+
"config": str(config),
|
| 401 |
+
}
|
| 402 |
+
),
|
| 403 |
+
flush=True,
|
| 404 |
+
)
|
| 405 |
+
env = clean_environment(
|
| 406 |
+
args.reuse_cache_root.resolve() if args.reuse_cache_root else root,
|
| 407 |
+
ray_root,
|
| 408 |
+
Path(base.env.nemo_gym.uv_venv_dir),
|
| 409 |
+
)
|
| 410 |
+
before = scope["devices"] if scope else gpu_snapshot()
|
| 411 |
+
if not six_gpu and (len(before) != 2 or any(int(gpu["memory_mb"]) > 1024 for gpu in before)):
|
| 412 |
+
raise ValueError("The two allocated GPUs are no longer idle")
|
| 413 |
+
env["CUDA_VISIBLE_DEVICES"] = ",".join(scope["uuids"] if scope else [gpu["index"] for gpu in before])
|
| 414 |
+
env["NEMO_RL_VENV_DIR"] = str(actor_root)
|
| 415 |
+
private = OmegaConf.load(args.private_config).spreadsheet_resources.resources_servers.spreadsheet
|
| 416 |
+
secrets = [str(private.api_key), str(private.session_secret_key)]
|
| 417 |
+
entry = args.trainer_root / "examples/nemo_gym/run_grpo_nemo_gym.py"
|
| 418 |
+
driver = args.helpers_root / "smoke_driver.py"
|
| 419 |
+
if not driver.is_file() or not audit_agent.is_file():
|
| 420 |
+
raise ValueError("Pinned driver/audit helper is missing")
|
| 421 |
+
report = {
|
| 422 |
+
"status": "running",
|
| 423 |
+
"hostname": socket.gethostname(),
|
| 424 |
+
"model": metadata,
|
| 425 |
+
"base_config": str(args.base_config),
|
| 426 |
+
"base_config_sha256": sha(args.base_config),
|
| 427 |
+
"config": str(config),
|
| 428 |
+
"config_sha256": sha(config),
|
| 429 |
+
"data_sha256": sha(subset),
|
| 430 |
+
"source_data_sha256": sha(source),
|
| 431 |
+
"task_ids": [row["task_id"] for row in rows],
|
| 432 |
+
"entrypoint_sha256": sha(entry),
|
| 433 |
+
"driver_sha256": sha(driver),
|
| 434 |
+
"audit_agent_sha256": sha(audit_agent),
|
| 435 |
+
"gpus_before": before,
|
| 436 |
+
"actor_venv_root": str(actor_root),
|
| 437 |
+
"gpu_profile": "train_tp4_rollout_tp2" if six_gpu else "legacy_two_gpu",
|
| 438 |
+
"requested_optimizer_steps": 2,
|
| 439 |
+
"max_wall_seconds": args.max_wall_seconds,
|
| 440 |
+
"log": str(root / "train.log"),
|
| 441 |
+
"ipc_root": str(ipc_root),
|
| 442 |
+
}
|
| 443 |
+
(root / "run.json").write_text(json.dumps(report, indent=2) + "\n")
|
| 444 |
+
print(
|
| 445 |
+
json.dumps(
|
| 446 |
+
{
|
| 447 |
+
"event": "SPREADSHEET_TRAIN_SMOKE_START",
|
| 448 |
+
"run_dir": str(root),
|
| 449 |
+
"model": metadata,
|
| 450 |
+
"steps": 2,
|
| 451 |
+
"prompts_per_step": 2,
|
| 452 |
+
"rollouts_per_prompt": 2,
|
| 453 |
+
}
|
| 454 |
+
),
|
| 455 |
+
flush=True,
|
| 456 |
+
)
|
| 457 |
+
start = time.monotonic()
|
| 458 |
+
command = [
|
| 459 |
+
sys.executable,
|
| 460 |
+
"-B",
|
| 461 |
+
str(driver),
|
| 462 |
+
"--entrypoint",
|
| 463 |
+
str(entry),
|
| 464 |
+
"--config",
|
| 465 |
+
str(config),
|
| 466 |
+
"--ray-root",
|
| 467 |
+
str(ray_root),
|
| 468 |
+
]
|
| 469 |
+
if scope:
|
| 470 |
+
command.extend(
|
| 471 |
+
[
|
| 472 |
+
"--gpu-indices",
|
| 473 |
+
args.gpu_indices,
|
| 474 |
+
"--gpu-uuids",
|
| 475 |
+
",".join(scope["uuids"]),
|
| 476 |
+
"--expected-hostname",
|
| 477 |
+
args.expected_hostname,
|
| 478 |
+
]
|
| 479 |
+
)
|
| 480 |
+
process = subprocess.Popen(
|
| 481 |
+
command,
|
| 482 |
+
cwd=root,
|
| 483 |
+
env=env,
|
| 484 |
+
stdout=subprocess.PIPE,
|
| 485 |
+
stderr=subprocess.STDOUT,
|
| 486 |
+
stdin=subprocess.DEVNULL,
|
| 487 |
+
start_new_session=True,
|
| 488 |
+
)
|
| 489 |
+
report.update(driver_pid=process.pid, driver_pgid=process.pid)
|
| 490 |
+
(root / "run.json").write_text(json.dumps(report, indent=2) + "\n")
|
| 491 |
+
timed_out = threading.Event()
|
| 492 |
+
|
| 493 |
+
def stop_owned_run():
|
| 494 |
+
if process.poll() is None:
|
| 495 |
+
timed_out.set()
|
| 496 |
+
try:
|
| 497 |
+
os.killpg(process.pid, signal.SIGTERM)
|
| 498 |
+
except ProcessLookupError:
|
| 499 |
+
pass
|
| 500 |
+
|
| 501 |
+
timer = threading.Timer(args.max_wall_seconds, stop_owned_run)
|
| 502 |
+
timer.daemon = True
|
| 503 |
+
timer.start()
|
| 504 |
+
with (root / "train.log").open("w") as log:
|
| 505 |
+
for raw in iter(process.stdout.readline, b""):
|
| 506 |
+
line = redact(raw.decode("utf-8", errors="replace"), secrets)
|
| 507 |
+
log.write(line)
|
| 508 |
+
log.flush()
|
| 509 |
+
if re.search(r"Traceback|ERROR|Exception|\bstep\b|reward|training|RL INIT|local Ray", line, re.IGNORECASE):
|
| 510 |
+
print(line, end="", flush=True)
|
| 511 |
+
code = process.wait()
|
| 512 |
+
timer.cancel()
|
| 513 |
+
report.update(
|
| 514 |
+
status="exited" if code == 0 else "failed",
|
| 515 |
+
exit_code=code,
|
| 516 |
+
timed_out=timed_out.is_set(),
|
| 517 |
+
elapsed_seconds=time.monotonic() - start,
|
| 518 |
+
gpus_after=gpu_snapshot(),
|
| 519 |
+
redacted_artifacts=redact_artifacts((root, ipc_root), secrets),
|
| 520 |
+
base_config_unchanged=sha(args.base_config) == report["base_config_sha256"],
|
| 521 |
+
)
|
| 522 |
+
(root / "run.json").write_text(json.dumps(report, indent=2) + "\n")
|
| 523 |
+
print(
|
| 524 |
+
json.dumps(
|
| 525 |
+
{
|
| 526 |
+
"event": "SPREADSHEET_TRAIN_SMOKE_EXIT",
|
| 527 |
+
"exit_code": code,
|
| 528 |
+
"timed_out": timed_out.is_set(),
|
| 529 |
+
"receipt": str(root / "run.json"),
|
| 530 |
+
}
|
| 531 |
+
),
|
| 532 |
+
flush=True,
|
| 533 |
+
)
|
| 534 |
+
if six_gpu:
|
| 535 |
+
if timed_out.is_set() and code == 0:
|
| 536 |
+
code = 124
|
| 537 |
+
if not report["base_config_unchanged"] and code == 0:
|
| 538 |
+
code = 1
|
| 539 |
+
try:
|
| 540 |
+
sys.path.insert(0, str(args.helpers_root))
|
| 541 |
+
from analyze_training_smoke import analyze
|
| 542 |
+
|
| 543 |
+
analysis = analyze(root)
|
| 544 |
+
(root / "analysis.json").write_text(json.dumps(analysis, indent=2) + "\n")
|
| 545 |
+
summary = {
|
| 546 |
+
key: analysis[key]
|
| 547 |
+
for key in (
|
| 548 |
+
"logged_steps",
|
| 549 |
+
"trajectory_count",
|
| 550 |
+
"model_tool_calls",
|
| 551 |
+
"reward_counts",
|
| 552 |
+
"advantages",
|
| 553 |
+
"environment_error_count",
|
| 554 |
+
"problems",
|
| 555 |
+
)
|
| 556 |
+
}
|
| 557 |
+
print(json.dumps({"event": "SIX_GPU_SMOKE_ANALYSIS", **summary}), flush=True)
|
| 558 |
+
if code == 0 and analysis["problems"]:
|
| 559 |
+
code = 1
|
| 560 |
+
except Exception as error:
|
| 561 |
+
print(json.dumps({"event": "SIX_GPU_ANALYSIS_FAILED", "error_type": type(error).__name__}), flush=True)
|
| 562 |
+
if code == 0:
|
| 563 |
+
code = 1
|
| 564 |
+
raise SystemExit(code if code >= 0 else 128 - code)
|
| 565 |
+
|
| 566 |
+
|
| 567 |
+
if __name__ == "__main__":
|
| 568 |
+
main()
|
six-gpu-smoke-2312e1859361.py.txt.json
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"source_commit": "2312e1859361d037a689afbe3d19f99eedf6a0ab",
|
| 3 |
+
"file": "six-gpu-smoke-2312e1859361.py.txt",
|
| 4 |
+
"sha256": "58dcdd57ffdafe13ecb8d513c5dc31b42cb3de0ce3b19e5a58ab3c43f6bdbb7d",
|
| 5 |
+
"bytes": 23400
|
| 6 |
+
}
|