mirror of
https://github.com/Gen-Verse/OpenClaw-RL.git
synced 2026-08-14 00:47:53 +00:00
Update gui_qwen3vl_8b_rl.sh
This commit is contained in:
Yinjie Wang
+63
-89
@@ -1,6 +1,5 @@
|
||||
#!/bin/bash
|
||||
|
||||
# for rerun the task (same style as retool scripts)
|
||||
pkill -9 sglang || true
|
||||
sleep 3
|
||||
ray stop --force || true
|
||||
@@ -14,20 +13,19 @@ set -ex
|
||||
|
||||
SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
|
||||
SLIME_DIR="$(cd -- "${SCRIPT_DIR}/../slime" &>/dev/null && pwd)"
|
||||
MODEL_ARGS_ROTARY_BASE=5000000 source "${SLIME_DIR}/scripts/models/qwen3-8B.sh"
|
||||
MEGATRON_LM_PATH=${MEGATRON_LM_PATH:-"${SCRIPT_DIR}/../Megatron-LM"}
|
||||
|
||||
# keep stdout/stderr unbuffered in ray jobs
|
||||
HF_CKPT=${HF_CKPT:-/data_storage/wyj/systems/huggingface/hub/models--Qwen--Qwen3-VL-8B-Thinking/snapshots/41ea130ce6eaaf7829c72dfc0e4597d49741ed18}
|
||||
MODEL_ARGS_ROTARY_BASE=5000000 source "${SLIME_DIR}/scripts/models/qwen3-8B.sh"
|
||||
|
||||
export PYTHONUNBUFFERED=1
|
||||
export PYTHONFAULTHANDLER=1
|
||||
|
||||
# reduce false ray node death under heavy initialization
|
||||
export RAY_health_check_failure_threshold=${RAY_health_check_failure_threshold:-20}
|
||||
export RAY_health_check_period_ms=${RAY_health_check_period_ms:-5000}
|
||||
export RAY_health_check_timeout_ms=${RAY_health_check_timeout_ms:-30000}
|
||||
export RAY_num_heartbeats_timeout=${RAY_num_heartbeats_timeout:-60}
|
||||
|
||||
# default to 8 GPUs if not set by scheduler
|
||||
NUM_GPUS=${NUM_GPUS:-8}
|
||||
ACTOR_GPUS=${ACTOR_GPUS:-4}
|
||||
ROLLOUT_GPUS=${ROLLOUT_GPUS:-4}
|
||||
@@ -38,14 +36,11 @@ if (( ACTOR_GPUS + ROLLOUT_GPUS > NUM_GPUS )); then
|
||||
exit 1
|
||||
fi
|
||||
|
||||
# ---------------------------
|
||||
# GUI env pool server configs
|
||||
# ---------------------------
|
||||
export GUI_ENV_SERVER_HOST=${GUI_ENV_SERVER_HOST:-"127.0.0.1"}
|
||||
export GUI_ENV_SERVER_HOST=${GUI_ENV_SERVER_HOST:-127.0.0.1}
|
||||
export GUI_ENV_SERVER_PORT=${GUI_ENV_SERVER_PORT:-18080}
|
||||
export GUI_ENV_SERVER_URL=${GUI_ENV_SERVER_URL:-"http://${GUI_ENV_SERVER_HOST}:${GUI_ENV_SERVER_PORT}"}
|
||||
export GUI_ENV_SERVER_MAX_ENVS=${GUI_ENV_SERVER_MAX_ENVS:-32}
|
||||
export GUI_PREWARM_CONCURRENCY=${GUI_PREWARM_CONCURRENCY:-32}
|
||||
export GUI_ENV_SERVER_MAX_ENVS=${GUI_ENV_SERVER_MAX_ENVS:-64}
|
||||
export GUI_PREWARM_CONCURRENCY=${GUI_PREWARM_CONCURRENCY:-64}
|
||||
export GUI_POOL_MAX_ENVS=${GUI_POOL_MAX_ENVS:-${GUI_ENV_SERVER_MAX_ENVS}}
|
||||
export GUI_PREWARM_ENVS=${GUI_PREWARM_ENVS:-${GUI_POOL_MAX_ENVS}}
|
||||
export GUI_FORCE_PREWARM_ALL=${GUI_FORCE_PREWARM_ALL:-1}
|
||||
@@ -54,20 +49,19 @@ if [[ "${GUI_FORCE_PREWARM_ALL}" == "1" ]]; then
|
||||
fi
|
||||
export GUI_TRAJECTORY_CONCURRENCY=${GUI_TRAJECTORY_CONCURRENCY:-${GUI_POOL_MAX_ENVS}}
|
||||
export GUI_POOL_IDLE_TTL_SECONDS=${GUI_POOL_IDLE_TTL_SECONDS:-600}
|
||||
export GUI_PROVIDER_NAME=${GUI_PROVIDER_NAME:-"volcengine"}
|
||||
export GUI_REGION=${GUI_REGION:-"cn-beijing"}
|
||||
export GUI_PROVIDER_NAME=${GUI_PROVIDER_NAME:-volcengine}
|
||||
export GUI_REGION=${GUI_REGION:-cn-beijing}
|
||||
export GUI_PATH_TO_VM=${GUI_PATH_TO_VM:-""}
|
||||
export GUI_ACTION_SPACE=${GUI_ACTION_SPACE:-"pyautogui"}
|
||||
export GUI_OBSERVATION_TYPE=${GUI_OBSERVATION_TYPE:-"screenshot"}
|
||||
export GUI_COORDINATE_TYPE=${GUI_COORDINATE_TYPE:-"relative"}
|
||||
export GUI_ACTION_SPACE=${GUI_ACTION_SPACE:-pyautogui}
|
||||
export GUI_OBSERVATION_TYPE=${GUI_OBSERVATION_TYPE:-screenshot}
|
||||
export GUI_COORDINATE_TYPE=${GUI_COORDINATE_TYPE:-relative}
|
||||
export GUI_AGENT_CLASS_PATH=${GUI_AGENT_CLASS_PATH:-"agents.qwen3vl_agent.Qwen3VLAgentLocal"}
|
||||
MULTIMODAL_KEYS=${MULTIMODAL_KEYS:-'{"image":"images"}'}
|
||||
export GUI_REUSE_VM_ON_RESET=${GUI_REUSE_VM_ON_RESET:-0}
|
||||
export GUI_RESET_ON_CLOSE=${GUI_RESET_ON_CLOSE:-1}
|
||||
export GUI_CLIENT_PASSWORD=${GUI_CLIENT_PASSWORD:-"WWbbb8b7b6314"}
|
||||
export GUI_CLIENT_PASSWORD=${GUI_CLIENT_PASSWORD:-WWbbb8b7b6314}
|
||||
export GUI_SCREEN_WIDTH=${GUI_SCREEN_WIDTH:-1920}
|
||||
export GUI_SCREEN_HEIGHT=${GUI_SCREEN_HEIGHT:-1080}
|
||||
# Keep wandb project independent from OSWORLD/result project naming.
|
||||
|
||||
WANDB_PROJECT=${WANDB_PROJECT:-slime_gui}
|
||||
GUI_PROJECT_NAME=${GUI_PROJECT_NAME:-slime_gui-8b-rl}
|
||||
export OSWORLD_PROJECT="${GUI_PROJECT_NAME}"
|
||||
@@ -75,61 +69,58 @@ export GUI_RESULT_DIR=${GUI_RESULT_DIR:-"${SCRIPT_DIR}/results"}
|
||||
export GUI_RESULT_DIR="${GUI_RESULT_DIR}/${GUI_PROJECT_NAME}"
|
||||
export GUI_TEST_CONFIG_BASE_DIR=${GUI_TEST_CONFIG_BASE_DIR:-"${SCRIPT_DIR}/evaluation_examples"}
|
||||
export GUI_TRAIN_META_PATH=${GUI_TRAIN_META_PATH:-"${GUI_TEST_CONFIG_BASE_DIR}/train_nochrome.json"}
|
||||
export GUI_EVAL_META_PATH=${GUI_EVAL_META_PATH:-"${GUI_TEST_CONFIG_BASE_DIR}/test_multinode.json"}
|
||||
export GUI_EVAL_META_PATH=${GUI_EVAL_META_PATH:-"${GUI_TEST_CONFIG_BASE_DIR}/test_nochrome.json"}
|
||||
MULTIMODAL_KEYS=${MULTIMODAL_KEYS:-'{"image":"images"}'}
|
||||
|
||||
|
||||
|
||||
# Start each run with a clean project result directory.
|
||||
if [[ -n "${GUI_RESULT_DIR}" && "${GUI_RESULT_DIR}" != "/" ]]; then
|
||||
rm -rf "${GUI_RESULT_DIR}"
|
||||
fi
|
||||
mkdir -p "${GUI_RESULT_DIR}"
|
||||
|
||||
|
||||
|
||||
# ---------------------------
|
||||
# Volcengine non-secret configs
|
||||
# Secrets (AK/SK) should be exported in shell before running this script.
|
||||
# ---------------------------
|
||||
export VOLCENGINE_REGION=${VOLCENGINE_REGION:-"cn-beijing"}
|
||||
export VOLCENGINE_IMAGE_ID=${VOLCENGINE_IMAGE_ID:-"image-id"}
|
||||
export VOLCENGINE_SUBNET_ID=${VOLCENGINE_SUBNET_ID:-"subnet-id"}
|
||||
export VOLCENGINE_SECURITY_GROUP_ID=${VOLCENGINE_SECURITY_GROUP_ID:-"sg-id"}
|
||||
export VOLCENGINE_ZONE_ID=${VOLCENGINE_ZONE_ID:-"cn-beijing-a"}
|
||||
export VOLCENGINE_DEFAULT_PASSWORD=${VOLCENGINE_DEFAULT_PASSWORD:-"WWbbb180314"}
|
||||
# Lower interval speeds up create/delete under moderate throttling.
|
||||
export VOLCENGINE_REGION=${VOLCENGINE_REGION:-cn-beijing}
|
||||
export VOLCENGINE_IMAGE_ID=${VOLCENGINE_IMAGE_ID:-image-xxx}
|
||||
export VOLCENGINE_SUBNET_ID=${VOLCENGINE_SUBNET_ID:-subnet-xxx}
|
||||
export VOLCENGINE_SECURITY_GROUP_ID=${VOLCENGINE_SECURITY_GROUP_ID:-sg-xxx}
|
||||
export VOLCENGINE_ZONE_ID=${VOLCENGINE_ZONE_ID:-cn-beijing-a}
|
||||
export VOLCENGINE_DEFAULT_PASSWORD=${VOLCENGINE_DEFAULT_PASSWORD:-WWbbb180314}
|
||||
export VOLCENGINE_RUNINST_MIN_INTERVAL=${VOLCENGINE_RUNINST_MIN_INTERVAL:-0.1}
|
||||
export VOLCENGINE_DELINST_MIN_INTERVAL=${VOLCENGINE_DELINST_MIN_INTERVAL:-0.1}
|
||||
|
||||
# Instance type fallback list (comma-separated).
|
||||
# desktop_env/providers/volcengine/manager.py already supports this format.
|
||||
export VOLCENGINE_INSTANCE_TYPE=${VOLCENGINE_INSTANCE_TYPE:-"ecs.e-c1m2.large,ecs.e-c1m4.large,ecs.e-c1m8.large,ecs.e-c1m1.large,ecs.c3al.large,ecs.c3a.large,ecs.c3il.large,ecs.g3il.large,ecs.r3il.large,ecs.c3a.large,ecs.g3a.large,ecs.r3a.large,ecs.c3i.large,ecs.g3i.large,ecs.r3i.large,ecs.g3al.large,ecs.r3al.large,ecs.r1ie.large,ecs.g1ie.large,ecs.c1ie.large,ecs.g3ine.large"}
|
||||
|
||||
# Optional download proxy used by setup/evaluator code.
|
||||
# Do not put a default value here; export it in terminal when needed (same as keys).
|
||||
export download_proxy=${download_proxy:-}
|
||||
|
||||
HF_CKPT=${HF_CKPT:-/absolute/path/to/models--Qwen--Qwen3-VL-8B-Thinking/snapshots/41ea130ce6eaaf7829c72dfc0e4597d49741ed18}
|
||||
# For bridge+multimodal path, prefer HF checkpoint path for ref model loading.
|
||||
if [[ -z "${VOLCENGINE_ACCESS_KEY_ID:-}" ]]; then
|
||||
echo "VOLCENGINE_ACCESS_KEY_ID must be set before starting GUI RL."
|
||||
exit 1
|
||||
fi
|
||||
if [[ -z "${VOLCENGINE_SECRET_ACCESS_KEY:-}" ]]; then
|
||||
echo "VOLCENGINE_SECRET_ACCESS_KEY must be set before starting GUI RL."
|
||||
exit 1
|
||||
fi
|
||||
|
||||
if [[ -z "${HF_CKPT}" ]]; then
|
||||
echo "Set HF_CKPT to your Qwen3-VL-8B checkpoint path"
|
||||
exit 1
|
||||
fi
|
||||
if [[ ! -e "${HF_CKPT}" ]]; then
|
||||
echo "HF_CKPT does not exist: ${HF_CKPT}"
|
||||
exit 1
|
||||
fi
|
||||
REF_LOAD=${REF_LOAD:-${HF_CKPT}}
|
||||
|
||||
CKPT_ARGS=(
|
||||
--hf-checkpoint ${HF_CKPT}
|
||||
--ref-load ${REF_LOAD}
|
||||
--hf-checkpoint "${HF_CKPT}"
|
||||
--ref-load "${REF_LOAD}"
|
||||
--save "${SAVE_CKPT:-${SCRIPT_DIR}/../ckpt/gui-qwen3vl-8b-rl}"
|
||||
--save-interval 20
|
||||
)
|
||||
|
||||
ENABLE_RESUME_LOAD=${ENABLE_RESUME_LOAD:-0}
|
||||
RESUME_LOAD=${RESUME_LOAD:-/absolute/path/to/OpenClaw-RL/ckpt/gui-qwen3vl-8b-rl}
|
||||
RESUME_LOAD=${RESUME_LOAD:-"${SCRIPT_DIR}/../ckpt/gui-qwen3vl-8b-rl"}
|
||||
if [[ "${ENABLE_RESUME_LOAD}" == "1" ]]; then
|
||||
CKPT_ARGS+=(--load "${RESUME_LOAD}")
|
||||
echo "Resume load enabled: ${RESUME_LOAD}"
|
||||
else
|
||||
echo "Resume load disabled (ENABLE_RESUME_LOAD=${ENABLE_RESUME_LOAD})"
|
||||
fi
|
||||
|
||||
ROLLOUT_BATCH_SIZE=${ROLLOUT_BATCH_SIZE:-4}
|
||||
ROLLOUT_BATCH_SIZE=${ROLLOUT_BATCH_SIZE:-8}
|
||||
N_SAMPLES_PER_PROMPT=${N_SAMPLES_PER_PROMPT:-8}
|
||||
|
||||
ROLLOUT_ARGS=(
|
||||
@@ -138,14 +129,14 @@ ROLLOUT_ARGS=(
|
||||
--num-rollout 1000
|
||||
--rollout-batch-size ${ROLLOUT_BATCH_SIZE}
|
||||
--n-samples-per-prompt ${N_SAMPLES_PER_PROMPT}
|
||||
--rollout-max-response-len 512
|
||||
--rollout-max-response-len 1024
|
||||
--rollout-temperature 1.0
|
||||
--gui-max-steps 30
|
||||
--gui-wait-after-reset 60
|
||||
--gui-sleep-after-execution 0.0
|
||||
--gui-sleep-after-execution 0.5
|
||||
--gui-max-image-history-length 3
|
||||
--gui-max-reward-image-history-length 2
|
||||
--num-steps-per-rollout 2
|
||||
--gui-max-reward-image-history-length 1
|
||||
--num-steps-per-rollout 1
|
||||
)
|
||||
|
||||
ROLLOUT_NUM_GPUS_PER_ENGINE=${ROLLOUT_NUM_GPUS_PER_ENGINE:-1}
|
||||
@@ -157,9 +148,12 @@ echo "GUI_TRAJECTORY_CONCURRENCY=${GUI_TRAJECTORY_CONCURRENCY} (independent from
|
||||
EVAL_ARGS=(
|
||||
--eval-temperature 0.0
|
||||
--gui-eval-max-steps 30
|
||||
--gui-eval-sleep-after-execution 0.0
|
||||
--gui-eval-sleep-after-execution 5.0
|
||||
--gui-eval-wait-after-reset 60
|
||||
--n-samples-per-eval-prompt 1
|
||||
--eval-interval 20
|
||||
--eval-reward-key acc
|
||||
--eval-function-path generate_with_gui.gui_generate_rollout
|
||||
)
|
||||
if [ -n "${GUI_EVAL_INTERVAL}" ]; then
|
||||
EVAL_ARGS+=(--eval-interval "${GUI_EVAL_INTERVAL}")
|
||||
@@ -181,18 +175,15 @@ PERF_ARGS=(
|
||||
--tensor-model-parallel-size 4
|
||||
--sequence-parallel
|
||||
--pipeline-model-parallel-size 1
|
||||
# VLM train forward in current Megatron bridge expects unsliced vision inputs.
|
||||
# CP>1 slices text tokens per rank but does not slice multimodal tensors here,
|
||||
# which can trigger "assert video_embeds is None" with image-only data.
|
||||
--context-parallel-size 1
|
||||
--expert-model-parallel-size 1
|
||||
--expert-tensor-parallel-size 1
|
||||
--recompute-granularity full
|
||||
--recompute-method uniform
|
||||
--recompute-num-layers 36
|
||||
--recompute-num-layers 1
|
||||
--megatron-to-hf-mode bridge
|
||||
--use-dynamic-batch-size
|
||||
--max-tokens-per-gpu 512
|
||||
--max-tokens-per-gpu 1024
|
||||
)
|
||||
|
||||
GRPO_ARGS=(
|
||||
@@ -203,12 +194,6 @@ GRPO_ARGS=(
|
||||
--kl-loss-coef 0.01
|
||||
)
|
||||
|
||||
#ROUTER_ARGS=(
|
||||
# Use SlimeRouter's least-active-worker routing to avoid over-pinning
|
||||
# GUI requests with similar prefixes to a single SGLang worker.
|
||||
#--use-slime-router
|
||||
#)
|
||||
|
||||
SGLANG_ARGS=(
|
||||
--rollout-num-gpus-per-engine ${ROLLOUT_NUM_GPUS_PER_ENGINE}
|
||||
--sglang-mem-fraction-static 0.85
|
||||
@@ -231,9 +216,8 @@ else
|
||||
WANDB_ARGS=()
|
||||
fi
|
||||
|
||||
# Start GUI env pool server
|
||||
mkdir -p logs
|
||||
ENV_SERVER_LOG=${ENV_SERVER_LOG:-"./logs/gui_env_pool_server.log"}
|
||||
ENV_SERVER_LOG=${ENV_SERVER_LOG:-"./logs/gui_env_pool_server_qwen3vl_8b.log"}
|
||||
PYTHONPATH="${SLIME_DIR}:${SCRIPT_DIR}:${PYTHONPATH}" \
|
||||
python3 -m env_pool_server \
|
||||
--host "${GUI_ENV_SERVER_HOST}" \
|
||||
@@ -252,7 +236,6 @@ PYTHONPATH="${SLIME_DIR}:${SCRIPT_DIR}:${PYTHONPATH}" \
|
||||
--screen-height "${GUI_SCREEN_HEIGHT}" \
|
||||
> "${ENV_SERVER_LOG}" 2>&1 &
|
||||
GUI_ENV_SERVER_PID=$!
|
||||
echo "GUI env pool server PID=${GUI_ENV_SERVER_PID}, log=${ENV_SERVER_LOG}"
|
||||
|
||||
cleanup() {
|
||||
set +e
|
||||
@@ -262,23 +245,19 @@ cleanup() {
|
||||
}
|
||||
trap cleanup EXIT INT TERM
|
||||
|
||||
# Wait until env server is healthy
|
||||
for i in {1..60}; do
|
||||
if curl -fsS "${GUI_ENV_SERVER_URL}/healthz" >/dev/null 2>&1; then
|
||||
echo "GUI env pool server is ready: ${GUI_ENV_SERVER_URL}"
|
||||
break
|
||||
fi
|
||||
sleep 2
|
||||
done
|
||||
|
||||
# Block until all prewarm envs are visible in the pool.
|
||||
if (( GUI_PREWARM_ENVS > 0 )); then
|
||||
for i in {1..600}; do
|
||||
if python3 - "${GUI_ENV_SERVER_URL}" "${GUI_PREWARM_ENVS}" <<'PY'
|
||||
import json
|
||||
import sys
|
||||
import urllib.request
|
||||
|
||||
status_url = sys.argv[1].rstrip("/") + "/status"
|
||||
target = int(sys.argv[2])
|
||||
with urllib.request.urlopen(status_url, timeout=5) as resp:
|
||||
@@ -286,32 +265,24 @@ with urllib.request.urlopen(status_url, timeout=5) as resp:
|
||||
pool = data.get("pool", {})
|
||||
total_envs = int(pool.get("total_envs", 0))
|
||||
ok = bool(data.get("ok", False))
|
||||
print(f"pool total_envs={total_envs}, target={target}, ok={ok}")
|
||||
raise SystemExit(0 if ok and total_envs >= target else 1)
|
||||
PY
|
||||
then
|
||||
echo "GUI prewarm complete: ${GUI_PREWARM_ENVS}/${GUI_POOL_MAX_ENVS}"
|
||||
break
|
||||
fi
|
||||
sleep 2
|
||||
if (( i == 600 )); then
|
||||
echo "Timed out waiting prewarm completion: target=${GUI_PREWARM_ENVS}"
|
||||
exit 1
|
||||
fi
|
||||
done
|
||||
fi
|
||||
|
||||
NVLINK_COUNT=$(nvidia-smi topo -m 2>/dev/null | grep -o 'NV[0-9][0-9]*' | wc -l)
|
||||
if [ "$NVLINK_COUNT" -gt 0 ]; then
|
||||
HAS_NVLINK=0
|
||||
if [[ "${NVLINK_COUNT}" -gt 0 ]]; then
|
||||
HAS_NVLINK=1
|
||||
else
|
||||
HAS_NVLINK=0
|
||||
fi
|
||||
echo "HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)"
|
||||
|
||||
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:2048
|
||||
|
||||
ray start --head --node-ip-address 127.0.0.1 --num-gpus ${NUM_GPUS} --disable-usage-stats --dashboard-host=0.0.0.0 --dashboard-port=8265
|
||||
ray start --head --node-ip-address 127.0.0.1 --num-gpus "${NUM_GPUS}" --disable-usage-stats --dashboard-host=0.0.0.0 --dashboard-port=8265
|
||||
|
||||
RUNTIME_ENV_JSON="{
|
||||
\"env_vars\": {
|
||||
@@ -333,17 +304,21 @@ RUNTIME_ENV_JSON="{
|
||||
\"GUI_TRAIN_META_PATH\": \"${GUI_TRAIN_META_PATH}\",
|
||||
\"GUI_EVAL_META_PATH\": \"${GUI_EVAL_META_PATH}\",
|
||||
\"OSWORLD_PROJECT\": \"${OSWORLD_PROJECT}\",
|
||||
\"download_proxy\": \"${download_proxy}\"
|
||||
\"download_proxy\": \"${download_proxy}\",
|
||||
\"MEGATRON_LM_PATH\": \"${MEGATRON_LM_PATH}\",
|
||||
\"HF_CKPT\": \"${HF_CKPT}\",
|
||||
\"GUI_AGENT_CLASS_PATH\": \"${GUI_AGENT_CLASS_PATH}\"
|
||||
}
|
||||
}"
|
||||
|
||||
RAY_JOB_SUBMISSION_ID=${RAY_JOB_SUBMISSION_ID:-"gui_qwen3vl_$(date +%Y%m%d_%H%M%S)"}
|
||||
TRAIN_ENTRY=${TRAIN_ENTRY:-train_async.py}
|
||||
RAY_JOB_SUBMISSION_ID=${RAY_JOB_SUBMISSION_ID:-"gui_qwen3vl_8b_rl_$(date +%Y%m%d_%H%M%S)"}
|
||||
|
||||
ray job submit --address="http://127.0.0.1:8265" \
|
||||
--submission-id "${RAY_JOB_SUBMISSION_ID}" \
|
||||
--no-wait \
|
||||
--runtime-env-json="${RUNTIME_ENV_JSON}" \
|
||||
-- python3 -u train.py \
|
||||
-- python3 -u "${TRAIN_ENTRY}" \
|
||||
--actor-num-nodes 1 \
|
||||
--actor-num-gpus-per-node ${ACTOR_GPUS} \
|
||||
--rollout-num-gpus ${ROLLOUT_GPUS} \
|
||||
@@ -374,4 +349,3 @@ fi
|
||||
|
||||
echo "Ray job failed (submission id: ${RAY_JOB_SUBMISSION_ID}, logs exit: ${RAY_LOG_EXIT})"
|
||||
exit 1
|
||||
|
||||
|
||||
Reference in New Issue
Block a user