diff --git a/training/a4x/qwen3-235b/README.md b/training/a4x/qwen3-235b/README.md
new file mode 100644
index 00000000..e184f139
--- /dev/null
+++ b/training/a4x/qwen3-235b/README.md
@@ -0,0 +1,9 @@
+# To run this recipe:
+
+1. Insert your Hugging Face token into `launch_script.sh` by replacing `<YOUR_HF_TOKEN>`.
+2. SSH into the Slurm login node
+3. Copy this recipe folder to `/home/$USER/recipe`
+4. From your home directory (`/home/$USER`), run:
+   ```
+   sbatch recipe/sbatch_script.sh
+   ```
diff --git a/training/a4x/qwen3-235b/custom_setup_experiment.py b/training/a4x/qwen3-235b/custom_setup_experiment.py
new file mode 100644
index 00000000..32173cbc
--- /dev/null
+++ b/training/a4x/qwen3-235b/custom_setup_experiment.py
@@ -0,0 +1,233 @@
+import glob
+import logging
+import os
+from pathlib import Path
+import sys
+import time
+from typing import Any, Dict, List, Optional
+
+import nemo_run as run
+from nemo_run.config import get_nemorun_home
+
+
+try:
+  from argument_parser import parse_cli_args
+  from utils.evaluate import calc_convergence_and_performance
+  from utils.executors import dgxc_executor, slurm_executor
+except (ImportError, ModuleNotFoundError):
+  from .argument_parser import parse_cli_args
+  from .utils.evaluate import calc_convergence_and_performance
+  from .utils.executors import dgxc_executor, slurm_executor
+
+try:
+  import wandb
+
+  HAVE_WANDB = True
+except (ImportError, ModuleNotFoundError):
+  HAVE_WANDB = False
+
+try:
+  from perf_plugins import NsysPlugin, PerfEnvPlugin
+  from resiliency_plugins import FaultTolerancePlugin
+except (ImportError, ModuleNotFoundError):
+  from .perf_plugins import NsysPlugin, PerfEnvPlugin
+  from .resiliency_plugins import FaultTolerancePlugin
+
+import logging
+
+
+SCRIPT_DIR = Path(__file__).parent.resolve()
+ENTRYPOINT_PEFORMANCE = "run_script.py"
+ENTRYPOINT_RECIPE = "run_recipe.py"
+
+logging.basicConfig(level=logging.DEBUG)
+logger = logging.getLogger(__name__)
+
+
+def main(
+    use_recipes: bool,
+    model_family_name: str,
+    model_recipe_name: str,
+    task: str,
+    compute_dtype: str,
+    gpu: str,
+    hf_token: str,
+    detach: bool,
+    dryrun: bool,
+    enable_vboost: bool,
+    enable_nsys: bool,
+    moe_a2a_overlap: bool,
+    tp_size: Optional[int],
+    pp_size: Optional[int],
+    cp_size: Optional[int],
+    wandb_key: str,
+    wandb_project_name: str,
+    wandb_experiment_name: str,
+    wandb_entity_name: str,
+    profiling_start_step: int,
+    profiling_stop_step: int,
+    profiling_gpu_metrics: bool,
+    profiling_ranks: Optional[List[int]],
+    nemo_home: str,
+    account: str,
+    partition: str,
+    log_dir: str,
+    gpus_per_node: int,
+    time_limit: str,
+    container_image: str,
+    custom_mounts: List[str],
+    custom_env_vars: List[str],
+    custom_srun_args: List[str],
+    pretrained_checkpoint: Optional[str],
+    num_gpus: int,
+    is_long_convergence_run: bool,
+    additional_slurm_params: Optional[Dict[str, Any]],
+    golden_values_path: str,
+    convergence_params: Dict[str, Any],
+    performance_params: Dict[str, Any],
+    max_retries: int,
+    dgxc_base_url: str,
+    dgxc_cluster: str,
+    dgxc_kube_apiserver_url: str,
+    dgxc_app_id: str,
+    dgxc_app_secret: str,
+    dgxc_project_name: str,
+    dgxc_pvc_claim_name: str,
+    dgxc_pvc_mount_path: str,
+):
+    logger.info("Hello World")
+
+    rank = os.environ['RANK']
+
+    exp_name = f"{model_recipe_name}_{model_family_name}"
+    exp_name += f'_worker{rank}'
+    if use_recipes:
+        script_name = ENTRYPOINT_RECIPE
+
+    else:
+        script_name = ENTRYPOINT_PEFORMANCE
+
+    run_script_path = SCRIPT_DIR / script_name
+    logger.info(f"Run script path: {run_script_path}")
+    if not run_script_path.is_file():
+        logger.error(f"Specified run script not found: {run_script_path}")
+        sys.exit(1)
+
+    nemorun_script = run.Script(
+        path=str(run_script_path),
+        entrypoint="python",
+        env={"PYTHONPATH": f"{SCRIPT_DIR}:$PYTHONPATH"},
+        args=list(sys.argv[1:]),
+    )
+
+    plugins = []
+
+    if not use_recipes:
+        plugins.append(
+            PerfEnvPlugin(
+                enable_vboost=enable_vboost,
+                moe_a2a_overlap=moe_a2a_overlap,
+                tp_size=tp_size,
+                pp_size=pp_size,
+                cp_size=cp_size,
+                model_family_name=model_family_name,
+                model_recipe_name=model_recipe_name,
+                gpu=gpu,
+                compute_dtype=compute_dtype,
+                train_task=task,
+            )
+        )
+
+    if enable_nsys:
+        plugins.append(
+            NsysPlugin(
+                profile_step_start=profiling_start_step,
+                profile_step_end=profiling_stop_step,
+                nsys_gpu_metrics=profiling_gpu_metrics,
+                profile_ranks=profiling_ranks,
+            )
+        )
+
+    executor = run.LocalExecutor()
+    run.run(
+        nemorun_script,
+        executor=executor,
+        plugins=plugins,
+        dryrun=False,
+        detach=False,
+        name=exp_name,
+    )
+
+
+if __name__ == "__main__":
+    parser = parse_cli_args()
+    args, unknown_args = parser.parse_known_args()
+
+    # probably better to use parser.parse_args() and make unknowns an error,
+    # but for now we'll just issue a warning.
+    if unknown_args:
+        logger.warning(f"Ignoring unrecognized arguments: {' '.join(unknown_args)}")
+
+    main(
+        use_recipes=args.use_recipes,
+        model_family_name=args.model_family_name,
+        model_recipe_name=args.model_recipe_name,
+        task=args.task,
+        compute_dtype=args.compute_dtype,
+        gpu=args.gpu,
+        hf_token=args.hf_token,
+        detach=args.detach,
+        dryrun=args.dryrun,
+        enable_vboost=args.enable_vboost,
+        enable_nsys=args.enable_nsys,
+        moe_a2a_overlap=args.moe_a2a_overlap,
+        tp_size=args.tensor_model_parallel_size,
+        pp_size=args.pipeline_model_parallel_size,
+        cp_size=args.context_parallel_size,
+        wandb_key=args.wandb_key,
+        wandb_project_name=args.wandb_project_name,
+        wandb_experiment_name=args.wandb_experiment_name,
+        wandb_entity_name=args.wandb_entity_name,
+        profiling_start_step=args.profiling_start_step,
+        profiling_stop_step=args.profiling_stop_step,
+        profiling_gpu_metrics=args.profiling_gpu_metrics,
+        profiling_ranks=args.profiling_ranks,
+        nemo_home=args.nemo_home,
+        account=args.account,
+        partition=args.partition,
+        log_dir=args.log_dir,
+        gpus_per_node=args.gpus_per_node,
+        time_limit=args.time_limit,
+        container_image=args.container_image,
+        custom_mounts=args.custom_mounts,
+        custom_env_vars=args.custom_env_vars,
+        custom_srun_args=args.custom_srun_args,
+        pretrained_checkpoint=args.pretrained_checkpoint,
+        num_gpus=args.num_gpus,
+        is_long_convergence_run=args.is_long_convergence_run,
+        additional_slurm_params=args.additional_slurm_params,
+        golden_values_path=args.golden_values_path,
+        convergence_params={
+            "correlation_threshold": args.correlation_threshold,
+            "high_loss_tolerance": args.high_loss_tolerance,
+            "medium_loss_tolerance": args.medium_loss_tolerance,
+            "low_loss_tolerance": args.low_loss_tolerance,
+            "final_loss_tolerance": args.final_loss_tolerance,
+            "max_outlier_ratio": args.max_outlier_ratio,
+            "outlier_threshold": args.outlier_threshold,
+            "skip_first_percent_loss": args.skip_first_percent_loss,
+        },
+        performance_params={
+            "timing_threshold": args.timing_threshold,
+            "skip_first_percent_time": args.skip_first_percent_time,
+        },
+        max_retries=args.max_retries,
+        dgxc_base_url=args.dgxc_base_url,
+        dgxc_cluster=args.dgxc_cluster,
+        dgxc_kube_apiserver_url=args.dgxc_kube_apiserver_url,
+        dgxc_app_id=args.dgxc_app_id,
+        dgxc_app_secret=args.dgxc_app_secret,
+        dgxc_project_name=args.dgxc_project_name,
+        dgxc_pvc_claim_name=args.dgxc_pvc_claim_name,
+        dgxc_pvc_mount_path=args.dgxc_pvc_mount_path,
+    )
diff --git a/training/a4x/qwen3-235b/launch_script.sh b/training/a4x/qwen3-235b/launch_script.sh
new file mode 100644
index 00000000..dfc0f282
--- /dev/null
+++ b/training/a4x/qwen3-235b/launch_script.sh
@@ -0,0 +1,151 @@
+usage()
+{
+cat << EOF
+usage: bash ./launcher.sh [config-override  [config-override ...]]
+config-override  (Optional) A  NeMo configuration override. E.g. trainer.max_steps=10000.
+EOF
+}
+
+parse_args() {
+  while [[ "$1" != "" ]]; do
+    case $(grep -o "=" <<< "$1" | wc -l) in
+        1  )
+        config_overrides+=("$1")
+        ;;
+        * )
+            echo "Invalid config override: $1"
+            usage
+            exit 1
+    esac
+    shift
+  done
+  config_overrides="${config_overrides[*]}"
+}
+
+config_overrides=()
+parse_args "$@"
+
+if [[ -z "${config_overrides[*]}" ]]; then
+  echo "No NeMo config overrides specified"
+else
+  echo "NeMo config overrides:"
+  echo "  ${config_overrides}"
+fi
+
+export LD_LIBRARY_PATH="/usr/local/cuda/compat/lib:$NCCL_PLUGIN_PATH:$LD_LIBRARY_PATH"
+ldconfig "$LD_LIBRARY_PATH"
+echo "Added $LD_LIBRARY_PATH to ldconfig:"
+ldconfig -p | grep libcuda | sed 's/^/  /'
+echo ""
+
+if [[ -n "${EXPLICIT_LOG_DIR}" ]]; then
+  explicit_log_dir=${EXPLICIT_LOG_DIR}
+else
+  explicit_log_dir=workload_logs
+fi
+echo "Logging to ${explicit_log_dir}"
+
+if [[ -n "${TOKENIZER_PATH}" ]]; then
+  echo "Getting tokenizer files"
+  cp "${TOKENIZER_PATH}"/* .
+  echo ""
+fi
+
+echo "Launching Torch distributed on the node rank $JOB_COMPLETION_INDEX out of $NNODES nodes"
+
+pip install git+https://github.com/NVIDIA/dllogger#egg=dllogger
+
+# Create the nsys directory.
+mkdir -p "${explicit_log_dir}/nsys"
+
+# Collect diagnostics to a single line
+kv="\"kernel_version\": \"$(uname --kernel-release)\""
+if command -v nvidia-smi &> /dev/null; then
+  cuda_v=$(nvidia-smi -q -x | grep -Po '(?<=<cuda_version>).*(?=</cuda_version>)' || true)
+  driver_v=$(nvidia-smi -q -x | grep -Po '(?<=<driver_version>).*(?=</driver_version>)' || true)
+  vbios_v=$(nvidia-smi -q -x | grep -Po '(?<=<vbios_version>).*(?=</vbios_version>)' | head -n1 || true)
+  kv="${kv}, \"cuda_version\": \"${cuda_v}\""
+  kv="${kv}, \"driver_version\": \"${driver_v}\""
+  kv="${kv}, \"vbios_version\": \"${vbios_v}\""
+fi
+echo "VERSION_DIAGNOSTICS: {${kv}}"
+
+
+export HF_TOKEN=<YOUR_HF_TOKEN>
+
+cd /opt
+rm -rf Megatron-Bridge
+git clone https://github.com/NVIDIA-NeMo/Megatron-Bridge.git
+cd Megatron-Bridge
+git checkout 7695d4acbfac19353d20e456509117efe4733d6b
+sed -i -e '/pretrain(config=recipe/i \    recipe.dist.distributed_timeout_minutes = 10' scripts/performance/run_script.py
+ls
+
+cp $CUSTOM_SETUP_EXPERIMENT_SCRIPT_PATH scripts/performance/
+
+worker_command=$(cat <<- EOM
+  if [ "\$RANK" -eq "0" ]; then
+    echo "Worker 0 is stalling for a few seconds.." ;
+    sleep 3 ;
+    echo "The detected environment within worker rank 0 is:" ;
+    env | sed 's/^/  /' ;
+  fi ;
+
+  cd /opt/Megatron-Bridge ;
+
+  numactl \
+    --cpunodebind=\$((LOCAL_RANK/2)) \
+    --membind=\$((LOCAL_RANK/2))           nsys profile \
+    -t nvtx,cuda \
+    --cuda-event-trace=false \
+    --sample=none \
+    --capture-range=cudaProfilerApi \
+    --capture-range-end=stop \
+    --kill none \
+    -o "/${explicit_log_dir}/$JOB_IDENTIFIER/rank-\$RANK" \
+    --force-overwrite true \
+    --session-new "nsys-\$RANDOM-\$RANK" \
+  nice -10 \
+  python scripts/performance/custom_setup_experiment.py \
+    --gpu gb200 \
+    --model_family_name qwen \
+    --model_recipe_name qwen3_235b_a22b \
+    --gpus_per_node 4 \
+    --num_gpus 64 \
+    --seq_length 4096 \
+    --compute_dtype bf16 \
+    --global_batch_size 1024 \
+    --tensor_model_parallel_size 1 \
+    --pipeline_model_parallel_size 8 \
+    --context_parallel_size 1 \
+    --expert_model_parallel_size 8 \
+    --virtual_pipeline_model_parallel_size 3 \
+    --micro_batch_size 1 \
+    --cuda_graph_impl transformer_engine \
+    --cuda_graph_scope moe_router,moe_preprocess,attn \
+    --max_steps 30
+
+EOM
+)
+
+echo "$worker_command" > worker_command.sh
+chmod 777 worker_command.sh
+
+torchrun \
+--nproc-per-node="4" \
+--nnodes="16" \
+--node_rank="${JOB_COMPLETION_INDEX}" \
+--rdzv_id="${JOB_IDENTIFIER}" \
+--master_addr="${MASTER_ADDR}" \
+--master_port="${MASTER_PORT}" \
+--no-python bash worker_command.sh
+
+
+if [[ "$JOB_COMPLETION_INDEX" == "0" ]]; then
+  mkdir -p "${ARTIFACT_DIR}"
+  cp -r "${explicit_log_dir}"/* "${ARTIFACT_DIR}/"
+  env > "${ARTIFACT_DIR}/environ.txt"
+  ls "${ARTIFACT_DIR}"
+fi
+echo "Training completed"
+echo "Pod on $(hostname --fqdn) is exiting"
\ No newline at end of file
diff --git a/training/a4x/qwen3-235b/sbatch_script.sh b/training/a4x/qwen3-235b/sbatch_script.sh
new file mode 100644
index 00000000..6fa2dadc
--- /dev/null
+++ b/training/a4x/qwen3-235b/sbatch_script.sh
@@ -0,0 +1,54 @@
+#!/bin/bash
+#SBATCH --job-name=rishabhbaghel-ubench-6uxs
+#SBATCH --nodes=16
+#SBATCH --ntasks-per-node=1
+#SBATCH --gres=gpu:4
+#SBATCH --mem=0
+
+# Exit early on failures
+set -e
+
+# Validate that the recipe location is setup correctly.
+# Recipe is expected to be in "recipe" folder inside current working directory
+RECIPE_DIR="$(pwd)/recipe"
+LAUNCH_SCRIPT="${RECIPE_DIR}/launch_script.sh"
+if [[ ! -f "${LAUNCH_SCRIPT}" ]]; then
+    echo "Error: Recipe is not located correctly. The recipe is expected to be in "recipe" folder inside current working directory. We could not find the launch script there." >&2
+    exit 1
+fi
+chmod +x "${LAUNCH_SCRIPT}"
+
+# Enroot the image if it is not already enrooted.
+export ENROOT_CONFIG_PATH=${HOME}/.config/enroot
+ORIG_IMAGE=nvcr.io#nvidia/nemo:25.11
+SQSH_IMAGE_PATH=${RECIPE_DIR}/sqsh/nvcr.io_nvidia_nemo:25.11
+if [[ ! -f "${SQSH_IMAGE_PATH}" ]]; then
+  mkdir -p "$(dirname "${SQSH_IMAGE_PATH}")"
+  echo "enrooting $ORIG_IMAGE to ${SQSH_IMAGE_PATH}"
+  enroot import --output "${SQSH_IMAGE_PATH}" -- "docker://${ORIG_IMAGE}"
+fi
+
+# get the master node
+master_addr=$(scontrol show hostnames "$SLURM_JOB_NODELIST" | head -n 1)
+master_port=29500
+
+ARTIFACT_DIR_HOME="/home/$USER/job_artifacts/${SLURM_JOB_ID}"
+mkdir -p "$ARTIFACT_DIR_HOME"
+
+export NNODES=$SLURM_NNODES
+export MASTER_ADDR=$master_addr
+export MASTER_PORT=$master_port
+export ARTIFACT_DIR=/artifacts
+export JOB_NAME=rishabhbaghel-ubench-6uxs
+export JOB_IDENTIFIER=rishabhbaghel-ubench-6uxs
+export CUSTOM_SETUP_EXPERIMENT_SCRIPT_PATH=/recipe/custom_setup_experiment.py 
+
+
+export PMIX_MCA_gds="^ds12"
+export GLOO_SOCKET_IFNAME=enp0s3
+
+srun --container-image="$SQSH_IMAGE_PATH" \
+  --container-mounts="${RECIPE_DIR}:/recipe:mkdir,${ARTIFACT_DIR_HOME}:${ARTIFACT_DIR}:mkdir,/usr/local/gib:/usr/local/gib" \
+  --container-workdir=/recipe \
+  --container-writable \
+  bash -c 'export JOB_COMPLETION_INDEX=$SLURM_NODEID; ./launch_script.sh'