diff --git a/docs/advance/megatron_lite_backend.rst b/docs/advance/megatron_lite_backend.rst
new file mode 100644
index 00000000000..b465754ce01
--- /dev/null
+++ b/docs/advance/megatron_lite_backend.rst
@@ -0,0 +1,81 @@
+Megatron Lite backend
+=====================
+
+Last updated: 06/17/2026.
+
+Megatron Lite (``mlite``) is Megatron's experimental, agent-friendly training
+path for work that needs to move quickly. It is optimized for fast iteration,
+small reviewable changes, and agentic development: model/runtime code can be
+changed without touching unrelated Megatron subsystems, and new experiments can
+live in their own source checkout instead of being copied into the verl tree.
+
+The verl integration intentionally keeps the backend glue outside this
+repository. The ``mlite`` checkout provides ``megatron.lite`` and the
+``verl_mlite`` launcher/config package used by the example scripts here. Put
+custom extensions in your own code path, add that path through ``MLITE_ROOT`` or
+``PYTHONPATH``, and keep verl focused on orchestration. See the upstream
+Megatron Lite path at
+`NVIDIA/Megatron-LM experimental/lite <https://github.com/NVIDIA/Megatron-LM/tree/dev/experimental/lite>`_.
+
+For the ``dist_opt`` optimizer path, Megatron Lite is intended to preserve
+Megatron-Core behavior rather than trade correctness for flexibility. In
+deterministic runs, the ``mlite`` path has been validated against the
+Megatron-Core distributed optimizer path with bitwise-aligned loss and gradient
+norms, and its step time / throughput are also aligned with the Core path.
+
+Install the backend
+-------------------
+
+Clone Megatron-LM's upstream ``dev`` branch and install its Megatron Lite verl
+integration:
+
+.. code-block:: bash
+
+   git clone -b dev https://github.com/NVIDIA/Megatron-LM.git
+   pip install -e Megatron-LM/experimental/lite/examples/verl
+
+Alternatively, keep the checkout outside the Python environment and set
+``MLITE_ROOT`` when running a launcher. The scripts add both
+``$MLITE_ROOT/experimental/lite`` and
+``$MLITE_ROOT/experimental/lite/examples/verl`` to ``PYTHONPATH``.
+
+Run an example
+--------------
+
+The DeepSeek-V4 examples use the ``mlite`` engine for training and vLLM for
+rollout where applicable:
+
+.. code-block:: bash
+
+   MODEL_PATH=/path/to/deepseek-v4 \
+   MLITE_ROOT=/path/to/mlite \
+   OPTIMIZER=fsdp2 \
+   bash examples/sft/gsm8k/run_deepseek_v4_megatron_lite.sh
+
+.. code-block:: bash
+
+   MODEL_PATH=/path/to/deepseek-v4 \
+   MLITE_ROOT=/path/to/mlite \
+   OPTIMIZER=fsdp2 \
+   bash examples/grpo_trainer/run_deepseek_v4_megatron_lite.sh
+
+``OPTIMIZER`` accepts ``dist_opt`` for the vanilla Megatron distributed
+optimizer and ``fsdp2`` for the Megatron Lite FSDP2 wrapper. The DeepSeek-V4
+launchers default to a 128-GPU mesh with PP4, EP8, CP4, full activation
+recompute, and ``fsdp2``.
+
+Further reading
+---------------
+
+For a practical discussion of long-sequence MoE RL tuning with Megatron Lite,
+including memory, recompute, communication overlap, and FSDP2 trade-offs, see
+`Making Long-Context MoE RL Training Easier to Tune <https://iseekyan.github.io/posts/qwen35-long-sequence-moe-rl/>`_.
+
+DeepSeek-V4 DSA note
+--------------------
+
+DeepSeek-V4 uses fused DSA kernels on Hopper and Blackwell GPUs. In addition to
+the normal verl runtime, the critical DSA-only dependencies are
+``nvidia-cutlass-dsl==4.5.2`` and ``nvidia-cudnn-frontend``. The
+``nvidia-cudnn-frontend`` 1.24.1 release is sufficient for Blackwell, while
+Hopper still needs a develop-branch build with ``IndexerForwardSm90`` support.
diff --git a/docs/index.rst b/docs/index.rst
index 4d6e304953b..d98d874a2cf 100644
--- a/docs/index.rst
+++ b/docs/index.rst
@@ -115,6 +115,7 @@ verl is fast with:
 
    advance/fsdp_extension
    advance/megatron_extension
+   advance/megatron_lite_backend
 
 .. toctree::
    :maxdepth: 1
diff --git a/examples/README.md b/examples/README.md
index a2d675a5d19..f38ca231f8e 100644
--- a/examples/README.md
+++ b/examples/README.md
@@ -19,9 +19,9 @@ All run scripts follow the same shape:
    - `<model>`: a single canonical size per model family. E.g.
      `qwen3_8b`, `qwen3_30b_a3b`, `qwen3_235b_a22b`, `qwen3_vl_8b`,
      `deepseek_v3`, `mimo_7b`, `nemotron_nano_v3`.
-   - `<train-backend>`: one of `fsdp`, `fsdp2`, `megatron`, `mindspeed`,
-     `automodel`, or `veomni`. **Must be the last underscore-separated
-     token before `.sh`**.
+   - `<train-backend>`: one of `fsdp`, `fsdp2`, `megatron`,
+     `megatron_lite`, `mindspeed`, `automodel`, or `veomni`. **Must be the
+     final suffix before `.sh`**.
 
    Nothing follows `<train-backend>`. Per-example *features* — including
    the inference backend (`vllm`/`sglang`/`trtllm`), the platform
diff --git a/examples/grpo_trainer/run_deepseek_v4_megatron_lite.sh b/examples/grpo_trainer/run_deepseek_v4_megatron_lite.sh
new file mode 100755
index 00000000000..24f2384f51f
--- /dev/null
+++ b/examples/grpo_trainer/run_deepseek_v4_megatron_lite.sh
@@ -0,0 +1,191 @@
+#!/usr/bin/env bash
+# GRPO scale demo | DeepSeek-V4 | vLLM rollout | Megatron Lite training | GPU
+#
+# Megatron Lite is Megatron's agentic experimental path. Its upstream home is
+# Megatron-LM's dev branch:
+# https://github.com/NVIDIA/Megatron-LM/tree/dev/experimental/lite
+#
+# This launcher currently tracks the submitter's active branch until the latest
+# mlite changes merge upstream. That checkout provides both megatron.lite and
+# the verl_mlite backend glue:
+#
+#   git clone https://github.com/ISEEKYAN/mlite
+#   pip install -e mlite/experimental/lite/examples/verl
+#
+# DeepSeek-V4 uses fused DSA kernels on Hopper and Blackwell GPUs. The critical
+# DSA-only dependencies are nvidia-cutlass-dsl==4.5.2 and nvidia-cudnn-frontend.
+# cudnn-frontend release 1.24.1 is sufficient for Blackwell, while Hopper still
+# needs a develop-branch build with IndexerForwardSm90 support.
+#
+# MODEL_VARIANT selects the DeepSeek-V4 target and its default mlite mesh:
+#   - flash: 16 nodes, PP4 EP8  CP4
+#   - pro:   64 nodes, PP8 EP16 CP4
+#
+# DS4 is fixed to TP1/ETP1. The architecture does not support TP/ETP
+# sharding, and there is no plan to support it.
+#
+# OPTIMIZER selects the Megatron Lite optimizer path:
+#   - fsdp2:    Megatron Lite FSDP2 wrapper, lower memory pressure, default
+#   - dist_opt: original Megatron distributed optimizer
+# When using dist_opt, prefer a larger PP*EP mesh to reduce per-rank model and
+# optimizer memory pressure and avoid OOM.
+
+set -xeuo pipefail
+
+########################### mlite backend knobs ###########################
+MODEL_VARIANT=${MODEL_VARIANT:-flash}
+MLITE_ROOT=${MLITE_ROOT:-$HOME/mlite}
+MLITE_VERL_ROOT=${MLITE_VERL_ROOT:-${MLITE_ROOT}/experimental/lite/examples/verl}
+MLITE_LITE_ROOT=${MLITE_LITE_ROOT:-${MLITE_ROOT}/experimental/lite}
+
+OPTIMIZER=${OPTIMIZER:-fsdp2} # dist_opt
+ALL_OFFLOAD=${ALL_OFFLOAD:-True}
+
+case "${MODEL_VARIANT}" in
+    flash)
+        MODEL_PATH=${MODEL_PATH:-${FLASH_MODEL_PATH:-}}
+        NNODES=${NNODES:-16}
+        PP=${PP:-4}
+        EP=${EP:-8}
+        CP=${CP:-4}
+        ;;
+    pro)
+        MODEL_PATH=${MODEL_PATH:-${PRO_MODEL_PATH:-}}
+        NNODES=${NNODES:-64}
+        PP=${PP:-8}
+        EP=${EP:-16}
+        CP=${CP:-4}
+        ;;
+    *)
+        echo "Unsupported MODEL_VARIANT=${MODEL_VARIANT}. Expected flash or pro." >&2
+        exit 1
+        ;;
+esac
+
+: "${MODEL_PATH:?set MODEL_PATH, or set FLASH_MODEL_PATH/PRO_MODEL_PATH for MODEL_VARIANT=${MODEL_VARIANT}}"
+########################### end mlite backend knobs ###########################
+
+########################### user-adjustable ###########################
+NDEVICES_PER_NODE=${NDEVICES_PER_NODE:-8}
+
+TRAIN_FILE=${TRAIN_FILE:-$HOME/data/gsm8k/train.parquet}
+TEST_FILE=${TEST_FILE:-$HOME/data/gsm8k/test.parquet}
+TRAIN_BATCH_SIZE=${TRAIN_BATCH_SIZE:-128}
+PPO_MINI_BATCH_SIZE=${PPO_MINI_BATCH_SIZE:-32}
+PPO_MICRO_BATCH_SIZE_PER_GPU=${PPO_MICRO_BATCH_SIZE_PER_GPU:-1}
+MAX_PROMPT_LENGTH=${MAX_PROMPT_LENGTH:-1024}
+MAX_RESPONSE_LENGTH=${MAX_RESPONSE_LENGTH:-2048}
+
+ACTOR_LR=${ACTOR_LR:-1e-6}
+CLIP_RATIO_LOW=${CLIP_RATIO_LOW:-0.2}
+CLIP_RATIO_HIGH=${CLIP_RATIO_HIGH:-0.28}
+CLIP_RATIO_C=${CLIP_RATIO_C:-10.0}
+ENTROPY_COEFF=${ENTROPY_COEFF:-0}
+
+ROLLOUT_TP=${ROLLOUT_TP:-2}
+ROLLOUT_GPU_MEM_UTIL=${ROLLOUT_GPU_MEM_UTIL:-0.8}
+ROLLOUT_N=${ROLLOUT_N:-16}
+
+TOTAL_EPOCHS=${TOTAL_EPOCHS:-1}
+PROJECT_NAME=${PROJECT_NAME:-verl-mlite-deepseek_v4_${MODEL_VARIANT}-grpo}
+EXPERIMENT_NAME=${EXPERIMENT_NAME:-deepseek_v4_${MODEL_VARIANT}_grpo_${OPTIMIZER}}
+########################### end user-adjustable ###########################
+
+########################### derived defaults ###########################
+export PYTHONPATH="${MLITE_VERL_ROOT}:${MLITE_LITE_ROOT}:${MLITE_ROOT}:${VERL_ROOT:-}:${MEGATRON_ROOT:-}:${PYTHONPATH:-}"
+export CUDA_DEVICE_MAX_CONNECTIONS="${CUDA_DEVICE_MAX_CONNECTIONS:-1}"
+
+if [[ -n "${CUDA_VISIBLE_DEVICES:-}" ]]; then
+    unset ROCR_VISIBLE_DEVICES
+    unset HIP_VISIBLE_DEVICES
+fi
+
+########################### parameter arrays ###########################
+ALGORITHM=(
+    algorithm.adv_estimator=grpo
+    algorithm.use_kl_in_reward=False
+    algorithm.kl_ctrl.kl_coef=0.0
+)
+
+DATA=(
+    data.train_files="${TRAIN_FILE}"
+    data.val_files="${TEST_FILE}"
+    data.train_batch_size=${TRAIN_BATCH_SIZE}
+    data.prompt_key=prompt
+    data.return_raw_chat=True
+    data.filter_overlong_prompts=True
+    data.truncation=error
+    data.max_prompt_length=${MAX_PROMPT_LENGTH}
+    data.max_response_length=${MAX_RESPONSE_LENGTH}
+)
+
+MODEL=(
+    actor_rollout_ref.model.path="${MODEL_PATH}"
+    actor_rollout_ref.model.trust_remote_code=True
+    actor_rollout_ref.model.use_fused_kernels=False
+)
+
+ACTOR=(
+    actor@actor_rollout_ref.actor=mlite_actor
+    actor_rollout_ref.actor.optim.lr=${ACTOR_LR}
+    actor_rollout_ref.actor.ppo_mini_batch_size=${PPO_MINI_BATCH_SIZE}
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=${PPO_MICRO_BATCH_SIZE_PER_GPU}
+    actor_rollout_ref.actor.use_dynamic_bsz=True
+    actor_rollout_ref.actor.use_kl_loss=False
+    actor_rollout_ref.actor.entropy_coeff=${ENTROPY_COEFF}
+    actor_rollout_ref.actor.clip_ratio_low=${CLIP_RATIO_LOW}
+    actor_rollout_ref.actor.clip_ratio_high=${CLIP_RATIO_HIGH}
+    actor_rollout_ref.actor.clip_ratio_c=${CLIP_RATIO_C}
+    actor_rollout_ref.actor.loss_agg_mode=token-mean
+    actor_rollout_ref.actor.engine.tp=1
+    actor_rollout_ref.actor.engine.pp=${PP}
+    actor_rollout_ref.actor.engine.vpp=1
+    actor_rollout_ref.actor.engine.ep=${EP}
+    actor_rollout_ref.actor.engine.cp=${CP}
+    actor_rollout_ref.actor.engine.etp=1
+    actor_rollout_ref.actor.engine.param_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.optimizer_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.grad_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.attention_backend_override=flash
+    actor_rollout_ref.actor.engine.impl_cfg.use_thd=True
+    +actor_rollout_ref.actor.engine.impl_cfg.optimizer=${OPTIMIZER}
+    +actor_rollout_ref.actor.engine.impl_cfg.recompute=[full]
+    +actor_rollout_ref.actor.optim.override_optimizer_config.offload_fraction=1.0
+)
+
+ROLLOUT=(
+    actor_rollout_ref.rollout.name=vllm
+    actor_rollout_ref.rollout.mode=async
+    actor_rollout_ref.rollout.tensor_model_parallel_size=${ROLLOUT_TP}
+    actor_rollout_ref.rollout.gpu_memory_utilization=${ROLLOUT_GPU_MEM_UTIL}
+    actor_rollout_ref.rollout.n=${ROLLOUT_N}
+    actor_rollout_ref.rollout.prompt_length=${MAX_PROMPT_LENGTH}
+    actor_rollout_ref.rollout.response_length=${MAX_RESPONSE_LENGTH}
+    actor_rollout_ref.rollout.free_cache_engine=True
+)
+
+TRAINER=(
+    critic.enable=False
+    trainer.logger=[console]
+    trainer.project_name=${PROJECT_NAME}
+    trainer.experiment_name=${EXPERIMENT_NAME}
+    trainer.val_before_train=False
+    trainer.nnodes=${NNODES}
+    trainer.n_gpus_per_node=${NDEVICES_PER_NODE}
+    trainer.total_epochs=${TOTAL_EPOCHS}
+)
+
+EXTRA=(
+    hydra.searchpath=[pkg://verl_mlite.config]
+)
+
+########################### launch ###########################
+python3 -m verl.trainer.main_ppo \
+    "${EXTRA[@]}" \
+    "${ALGORITHM[@]}" \
+    "${DATA[@]}" \
+    "${MODEL[@]}" \
+    "${ACTOR[@]}" \
+    "${ROLLOUT[@]}" \
+    "${TRAINER[@]}" \
+    "$@"
diff --git a/examples/grpo_trainer/run_kimi_k2_6_glm5_1_megatron_lite.sh b/examples/grpo_trainer/run_kimi_k2_6_glm5_1_megatron_lite.sh
new file mode 100755
index 00000000000..d6a5400e4f4
--- /dev/null
+++ b/examples/grpo_trainer/run_kimi_k2_6_glm5_1_megatron_lite.sh
@@ -0,0 +1,209 @@
+#!/usr/bin/env bash
+# GRPO scale demo | Kimi K2.6 / GLM 5.1 | vLLM rollout | Megatron Lite training | GPU
+#
+# Megatron Lite is Megatron's agentic experimental path. Its upstream home is
+# Megatron-LM's dev branch:
+# https://github.com/NVIDIA/Megatron-LM/tree/dev/experimental/lite
+#
+# This launcher currently tracks the submitter's active branch until the latest
+# mlite changes merge upstream. That checkout provides both megatron.lite and
+# the verl_mlite backend glue:
+#
+#   git clone https://github.com/ISEEKYAN/mlite
+#   pip install -e mlite/experimental/lite/examples/verl
+#
+# MODEL_VARIANT selects the target model. Both defaults are 256-GPU mlite runs:
+#   - kimi_k2_6: 32 nodes, PP8 EP8 CP8, fsdp2
+#   - glm5_1:    32 nodes, PP8 EP8 CP8, fsdp2
+#
+# GLM 5.1 uses fused DSA kernels on Hopper and Blackwell GPUs. The critical
+# DSA-only dependencies are nvidia-cutlass-dsl==4.5.2 and nvidia-cudnn-frontend.
+# cudnn-frontend release 1.24.1 is sufficient for Blackwell, while Hopper still
+# needs a develop-branch build with IndexerForwardSm90 support.
+#
+# Mesh accounting follows Megatron Lite's per-pipeline-stage layout:
+#   ngpu / pp = tp * ep * dp = etp * ep * edp
+# With the default 256 GPUs, PP8, EP8, TP1, and ETP1, this gives DP=4 and EDP=4.
+#
+# OPTIMIZER selects the Megatron Lite optimizer path:
+#   - fsdp2:    Megatron Lite FSDP2 wrapper, lower memory pressure, default
+#   - dist_opt: original Megatron distributed optimizer
+# When using dist_opt, prefer a larger PP*EP mesh to reduce per-rank model and
+# optimizer memory pressure and avoid OOM.
+
+set -xeuo pipefail
+
+########################### mlite backend knobs ###########################
+MODEL_VARIANT=${MODEL_VARIANT:-kimi_k2_6}
+MLITE_ROOT=${MLITE_ROOT:-$HOME/mlite}
+MLITE_VERL_ROOT=${MLITE_VERL_ROOT:-${MLITE_ROOT}/experimental/lite/examples/verl}
+MLITE_LITE_ROOT=${MLITE_LITE_ROOT:-${MLITE_ROOT}/experimental/lite}
+
+NNODES=${NNODES:-32}
+NDEVICES_PER_NODE=${NDEVICES_PER_NODE:-8}
+TP=${TP:-1}
+PP=${PP:-8}
+EP=${EP:-8}
+CP=${CP:-8}
+ETP=${ETP:-1}
+OPTIMIZER=${OPTIMIZER:-fsdp2} # dist_opt
+ALL_OFFLOAD=${ALL_OFFLOAD:-True}
+
+case "${MODEL_VARIANT}" in
+    kimi_k2_6)
+        MODEL_PATH=${MODEL_PATH:-${KIMI_K2_6_MODEL_PATH:-}}
+        ;;
+    glm5_1)
+        MODEL_PATH=${MODEL_PATH:-${GLM5_1_MODEL_PATH:-}}
+        ;;
+    *)
+        echo "Unsupported MODEL_VARIANT=${MODEL_VARIANT}. Expected kimi_k2_6 or glm5_1." >&2
+        exit 1
+        ;;
+esac
+
+: "${MODEL_PATH:?set MODEL_PATH, or set KIMI_K2_6_MODEL_PATH/GLM5_1_MODEL_PATH for MODEL_VARIANT=${MODEL_VARIANT}}"
+
+NGPU=$((NNODES * NDEVICES_PER_NODE))
+if (( NGPU % PP != 0 )); then
+    echo "Invalid mesh: NGPU=${NGPU} must be divisible by PP=${PP}." >&2
+    exit 1
+fi
+
+NGPU_PER_PP=$((NGPU / PP))
+if (( NGPU_PER_PP % (TP * EP) != 0 )); then
+    echo "Invalid mesh: NGPU/PP=${NGPU_PER_PP} must be divisible by TP*EP=$((TP * EP))." >&2
+    exit 1
+fi
+if (( NGPU_PER_PP % (ETP * EP) != 0 )); then
+    echo "Invalid mesh: NGPU/PP=${NGPU_PER_PP} must be divisible by ETP*EP=$((ETP * EP))." >&2
+    exit 1
+fi
+
+DP=$((NGPU_PER_PP / (TP * EP)))
+EDP=$((NGPU_PER_PP / (ETP * EP)))
+echo "MLITE_MESH model=${MODEL_VARIANT} ngpu=${NGPU} pp=${PP} tp=${TP} ep=${EP} etp=${ETP} cp=${CP} dp=${DP} edp=${EDP} optimizer=${OPTIMIZER}"
+########################### end mlite backend knobs ###########################
+
+########################### user-adjustable ###########################
+TRAIN_FILE=${TRAIN_FILE:-$HOME/data/gsm8k/train.parquet}
+TEST_FILE=${TEST_FILE:-$HOME/data/gsm8k/test.parquet}
+TRAIN_BATCH_SIZE=${TRAIN_BATCH_SIZE:-128}
+PPO_MINI_BATCH_SIZE=${PPO_MINI_BATCH_SIZE:-32}
+PPO_MICRO_BATCH_SIZE_PER_GPU=${PPO_MICRO_BATCH_SIZE_PER_GPU:-1}
+MAX_PROMPT_LENGTH=${MAX_PROMPT_LENGTH:-1024}
+MAX_RESPONSE_LENGTH=${MAX_RESPONSE_LENGTH:-2048}
+
+ACTOR_LR=${ACTOR_LR:-1e-6}
+CLIP_RATIO_LOW=${CLIP_RATIO_LOW:-0.2}
+CLIP_RATIO_HIGH=${CLIP_RATIO_HIGH:-0.28}
+CLIP_RATIO_C=${CLIP_RATIO_C:-10.0}
+ENTROPY_COEFF=${ENTROPY_COEFF:-0}
+
+ROLLOUT_TP=${ROLLOUT_TP:-2}
+ROLLOUT_GPU_MEM_UTIL=${ROLLOUT_GPU_MEM_UTIL:-0.8}
+ROLLOUT_N=${ROLLOUT_N:-16}
+
+TOTAL_EPOCHS=${TOTAL_EPOCHS:-1}
+PROJECT_NAME=${PROJECT_NAME:-verl-mlite-${MODEL_VARIANT}-grpo}
+EXPERIMENT_NAME=${EXPERIMENT_NAME:-${MODEL_VARIANT}_grpo_${OPTIMIZER}}
+########################### end user-adjustable ###########################
+
+########################### derived defaults ###########################
+export PYTHONPATH="${MLITE_VERL_ROOT}:${MLITE_LITE_ROOT}:${MLITE_ROOT}:${VERL_ROOT:-}:${MEGATRON_ROOT:-}:${PYTHONPATH:-}"
+export CUDA_DEVICE_MAX_CONNECTIONS="${CUDA_DEVICE_MAX_CONNECTIONS:-1}"
+
+if [[ -n "${CUDA_VISIBLE_DEVICES:-}" ]]; then
+    unset ROCR_VISIBLE_DEVICES
+    unset HIP_VISIBLE_DEVICES
+fi
+
+########################### parameter arrays ###########################
+ALGORITHM=(
+    algorithm.adv_estimator=grpo
+    algorithm.use_kl_in_reward=False
+    algorithm.kl_ctrl.kl_coef=0.0
+)
+
+DATA=(
+    data.train_files="${TRAIN_FILE}"
+    data.val_files="${TEST_FILE}"
+    data.train_batch_size=${TRAIN_BATCH_SIZE}
+    data.prompt_key=prompt
+    data.return_raw_chat=True
+    data.filter_overlong_prompts=True
+    data.truncation=error
+    data.max_prompt_length=${MAX_PROMPT_LENGTH}
+    data.max_response_length=${MAX_RESPONSE_LENGTH}
+)
+
+MODEL=(
+    actor_rollout_ref.model.path="${MODEL_PATH}"
+    actor_rollout_ref.model.trust_remote_code=True
+    actor_rollout_ref.model.use_fused_kernels=False
+)
+
+ACTOR=(
+    actor@actor_rollout_ref.actor=mlite_actor
+    actor_rollout_ref.actor.optim.lr=${ACTOR_LR}
+    actor_rollout_ref.actor.ppo_mini_batch_size=${PPO_MINI_BATCH_SIZE}
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=${PPO_MICRO_BATCH_SIZE_PER_GPU}
+    actor_rollout_ref.actor.use_dynamic_bsz=True
+    actor_rollout_ref.actor.use_kl_loss=False
+    actor_rollout_ref.actor.entropy_coeff=${ENTROPY_COEFF}
+    actor_rollout_ref.actor.clip_ratio_low=${CLIP_RATIO_LOW}
+    actor_rollout_ref.actor.clip_ratio_high=${CLIP_RATIO_HIGH}
+    actor_rollout_ref.actor.clip_ratio_c=${CLIP_RATIO_C}
+    actor_rollout_ref.actor.loss_agg_mode=token-mean
+    actor_rollout_ref.actor.engine.tp=${TP}
+    actor_rollout_ref.actor.engine.pp=${PP}
+    actor_rollout_ref.actor.engine.vpp=1
+    actor_rollout_ref.actor.engine.ep=${EP}
+    actor_rollout_ref.actor.engine.cp=${CP}
+    actor_rollout_ref.actor.engine.etp=${ETP}
+    actor_rollout_ref.actor.engine.param_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.optimizer_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.grad_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.attention_backend_override=flash
+    actor_rollout_ref.actor.engine.impl_cfg.use_thd=True
+    +actor_rollout_ref.actor.engine.impl_cfg.optimizer=${OPTIMIZER}
+    +actor_rollout_ref.actor.engine.impl_cfg.recompute=[full]
+    +actor_rollout_ref.actor.optim.override_optimizer_config.offload_fraction=1.0
+)
+
+ROLLOUT=(
+    actor_rollout_ref.rollout.name=vllm
+    actor_rollout_ref.rollout.mode=async
+    actor_rollout_ref.rollout.tensor_model_parallel_size=${ROLLOUT_TP}
+    actor_rollout_ref.rollout.gpu_memory_utilization=${ROLLOUT_GPU_MEM_UTIL}
+    actor_rollout_ref.rollout.n=${ROLLOUT_N}
+    actor_rollout_ref.rollout.prompt_length=${MAX_PROMPT_LENGTH}
+    actor_rollout_ref.rollout.response_length=${MAX_RESPONSE_LENGTH}
+    actor_rollout_ref.rollout.free_cache_engine=True
+)
+
+TRAINER=(
+    critic.enable=False
+    trainer.logger=[console]
+    trainer.project_name=${PROJECT_NAME}
+    trainer.experiment_name=${EXPERIMENT_NAME}
+    trainer.val_before_train=False
+    trainer.nnodes=${NNODES}
+    trainer.n_gpus_per_node=${NDEVICES_PER_NODE}
+    trainer.total_epochs=${TOTAL_EPOCHS}
+)
+
+EXTRA=(
+    hydra.searchpath=[pkg://verl_mlite.config]
+)
+
+########################### launch ###########################
+python3 -m verl.trainer.main_ppo \
+    "${EXTRA[@]}" \
+    "${ALGORITHM[@]}" \
+    "${DATA[@]}" \
+    "${MODEL[@]}" \
+    "${ACTOR[@]}" \
+    "${ROLLOUT[@]}" \
+    "${TRAINER[@]}" \
+    "$@"
diff --git a/examples/grpo_trainer/run_qwen3_5_35b_megatron_lite.sh b/examples/grpo_trainer/run_qwen3_5_35b_megatron_lite.sh
new file mode 100755
index 00000000000..984aa2c706b
--- /dev/null
+++ b/examples/grpo_trainer/run_qwen3_5_35b_megatron_lite.sh
@@ -0,0 +1,192 @@
+#!/usr/bin/env bash
+# GRPO scale demo | Qwen3.5-35B-A3B | vLLM rollout | Megatron Lite training | GPU
+#
+# Megatron Lite is Megatron's agentic experimental path. Its upstream home is
+# Megatron-LM's dev branch:
+# https://github.com/NVIDIA/Megatron-LM/tree/dev/experimental/lite
+#
+# This launcher currently tracks the submitter's active branch until the latest
+# mlite changes merge upstream. That checkout provides both megatron.lite and
+# the verl_mlite backend glue:
+#
+#   git clone https://github.com/ISEEKYAN/mlite
+#   pip install -e mlite/experimental/lite/examples/verl
+#
+# Qwen3.5 uses the Megatron Lite allgather CP path. CP is intentionally kept as
+# an explicit mlite knob below; the default single-node run uses CP8. This path
+# depends on FLA (flash-linear-attention) 5.0 in the runtime environment:
+# https://github.com/fla-org/flash-linear-attention
+#
+# Default mlite mesh:
+#   - 8 GPUs, PP1 EP8 CP8, TP1 ETP1, fsdp2
+#
+# Mesh accounting follows Megatron Lite's per-pipeline-stage layout:
+#   ngpu / pp = tp * ep * dp = etp * ep * edp
+# With the default 8 GPUs, PP1, EP8, TP1, and ETP1, this gives DP=1 and EDP=1.
+#
+# OPTIMIZER selects the Megatron Lite optimizer path:
+#   - fsdp2:    Megatron Lite FSDP2 wrapper, lower memory pressure, default
+#   - dist_opt: original Megatron distributed optimizer
+# When using dist_opt, prefer a larger PP*EP mesh to reduce per-rank model and
+# optimizer memory pressure and avoid OOM.
+
+set -xeuo pipefail
+
+########################### mlite backend knobs ###########################
+MLITE_ROOT=${MLITE_ROOT:-$HOME/mlite}
+MLITE_VERL_ROOT=${MLITE_VERL_ROOT:-${MLITE_ROOT}/experimental/lite/examples/verl}
+MLITE_LITE_ROOT=${MLITE_LITE_ROOT:-${MLITE_ROOT}/experimental/lite}
+
+MODEL_PATH=${MODEL_PATH:-${QWEN3_5_35B_MODEL_PATH:-Qwen3.5-35B-A3B}}
+NNODES=${NNODES:-1}
+NDEVICES_PER_NODE=${NDEVICES_PER_NODE:-8}
+TP=${TP:-1}
+PP=${PP:-1}
+EP=${EP:-8}
+CP=${CP:-8}
+ETP=${ETP:-1}
+OPTIMIZER=${OPTIMIZER:-fsdp2} # dist_opt
+ALL_OFFLOAD=${ALL_OFFLOAD:-True}
+
+NGPU=$((NNODES * NDEVICES_PER_NODE))
+if (( NGPU % PP != 0 )); then
+    echo "Invalid mesh: NGPU=${NGPU} must be divisible by PP=${PP}." >&2
+    exit 1
+fi
+
+NGPU_PER_PP=$((NGPU / PP))
+if (( NGPU_PER_PP % (TP * EP) != 0 )); then
+    echo "Invalid mesh: NGPU/PP=${NGPU_PER_PP} must be divisible by TP*EP=$((TP * EP))." >&2
+    exit 1
+fi
+if (( NGPU_PER_PP % (ETP * EP) != 0 )); then
+    echo "Invalid mesh: NGPU/PP=${NGPU_PER_PP} must be divisible by ETP*EP=$((ETP * EP))." >&2
+    exit 1
+fi
+
+DP=$((NGPU_PER_PP / (TP * EP)))
+EDP=$((NGPU_PER_PP / (ETP * EP)))
+echo "MLITE_MESH model=qwen3_5_35b ngpu=${NGPU} pp=${PP} tp=${TP} ep=${EP} etp=${ETP} cp=${CP} dp=${DP} edp=${EDP} optimizer=${OPTIMIZER}"
+########################### end mlite backend knobs ###########################
+
+########################### user-adjustable ###########################
+TRAIN_FILE=${TRAIN_FILE:-$HOME/data/geo3k/train.parquet}
+TEST_FILE=${TEST_FILE:-$HOME/data/geo3k/test.parquet}
+TRAIN_BATCH_SIZE=${TRAIN_BATCH_SIZE:-32}
+PPO_MINI_BATCH_SIZE=${PPO_MINI_BATCH_SIZE:-32}
+PPO_MICRO_BATCH_SIZE_PER_GPU=${PPO_MICRO_BATCH_SIZE_PER_GPU:-1}
+MAX_PROMPT_LENGTH=${MAX_PROMPT_LENGTH:-1024}
+MAX_RESPONSE_LENGTH=${MAX_RESPONSE_LENGTH:-2048}
+PPO_MAX_TOKEN_LEN_PER_GPU=${PPO_MAX_TOKEN_LEN_PER_GPU:-4096}
+
+ACTOR_LR=${ACTOR_LR:-1e-6}
+ENTROPY_COEFF=${ENTROPY_COEFF:-0}
+
+ROLLOUT_TP=${ROLLOUT_TP:-8}
+ROLLOUT_GPU_MEM_UTIL=${ROLLOUT_GPU_MEM_UTIL:-0.6}
+ROLLOUT_N=${ROLLOUT_N:-5}
+
+SAVE_FREQ=${SAVE_FREQ:-20}
+TEST_FREQ=${TEST_FREQ:-5}
+TOTAL_EPOCHS=${TOTAL_EPOCHS:-15}
+PROJECT_NAME=${PROJECT_NAME:-verl-mlite-qwen3_5_35b-geo3k-grpo}
+EXPERIMENT_NAME=${EXPERIMENT_NAME:-qwen3_5_35b_grpo_${OPTIMIZER}}
+########################### end user-adjustable ###########################
+
+########################### derived defaults ###########################
+export PYTHONPATH="${MLITE_VERL_ROOT}:${MLITE_LITE_ROOT}:${MLITE_ROOT}:${VERL_ROOT:-}:${MEGATRON_ROOT:-}:${PYTHONPATH:-}"
+export CUDA_DEVICE_MAX_CONNECTIONS="${CUDA_DEVICE_MAX_CONNECTIONS:-1}"
+export VLLM_USE_V1="${VLLM_USE_V1:-1}"
+export VLLM_ALLREDUCE_USE_SYMM_MEM="${VLLM_ALLREDUCE_USE_SYMM_MEM:-0}"
+
+if [[ -n "${CUDA_VISIBLE_DEVICES:-}" ]]; then
+    unset ROCR_VISIBLE_DEVICES
+    unset HIP_VISIBLE_DEVICES
+fi
+
+########################### parameter arrays ###########################
+ALGORITHM=(
+    algorithm.adv_estimator=grpo
+    algorithm.use_kl_in_reward=False
+)
+
+DATA=(
+    data.train_files="${TRAIN_FILE}"
+    data.val_files="${TEST_FILE}"
+    data.train_batch_size=${TRAIN_BATCH_SIZE}
+    data.max_prompt_length=${MAX_PROMPT_LENGTH}
+    data.max_response_length=${MAX_RESPONSE_LENGTH}
+    data.filter_overlong_prompts=True
+    data.truncation=error
+)
+
+MODEL=(
+    actor_rollout_ref.model.path="${MODEL_PATH}"
+    actor_rollout_ref.model.trust_remote_code=True
+    actor_rollout_ref.model.use_fused_kernels=False
+)
+
+ACTOR=(
+    actor@actor_rollout_ref.actor=mlite_actor
+    actor_rollout_ref.actor.optim.lr=${ACTOR_LR}
+    actor_rollout_ref.actor.ppo_mini_batch_size=${PPO_MINI_BATCH_SIZE}
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=${PPO_MICRO_BATCH_SIZE_PER_GPU}
+    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=${PPO_MAX_TOKEN_LEN_PER_GPU}
+    actor_rollout_ref.actor.use_dynamic_bsz=True
+    actor_rollout_ref.actor.use_kl_loss=False
+    actor_rollout_ref.actor.entropy_coeff=${ENTROPY_COEFF}
+    actor_rollout_ref.actor.loss_agg_mode=token-mean
+    actor_rollout_ref.actor.engine.tp=${TP}
+    actor_rollout_ref.actor.engine.pp=${PP}
+    actor_rollout_ref.actor.engine.vpp=1
+    actor_rollout_ref.actor.engine.ep=${EP}
+    actor_rollout_ref.actor.engine.cp=${CP}
+    actor_rollout_ref.actor.engine.etp=${ETP}
+    actor_rollout_ref.actor.engine.param_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.optimizer_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.grad_offload=${ALL_OFFLOAD}
+    actor_rollout_ref.actor.engine.attention_backend_override=flash
+    actor_rollout_ref.actor.engine.impl_cfg.use_thd=True
+    +actor_rollout_ref.actor.engine.impl_cfg.optimizer=${OPTIMIZER}
+    +actor_rollout_ref.actor.engine.impl_cfg.recompute=[full]
+    +actor_rollout_ref.actor.optim.override_optimizer_config.offload_fraction=1.0
+)
+
+ROLLOUT=(
+    actor_rollout_ref.rollout.name=vllm
+    actor_rollout_ref.rollout.mode=async
+    actor_rollout_ref.rollout.tensor_model_parallel_size=${ROLLOUT_TP}
+    actor_rollout_ref.rollout.gpu_memory_utilization=${ROLLOUT_GPU_MEM_UTIL}
+    actor_rollout_ref.rollout.n=${ROLLOUT_N}
+    actor_rollout_ref.rollout.prompt_length=${MAX_PROMPT_LENGTH}
+    actor_rollout_ref.rollout.response_length=${MAX_RESPONSE_LENGTH}
+    actor_rollout_ref.rollout.free_cache_engine=True
+)
+
+TRAINER=(
+    critic.enable=False
+    trainer.logger=[console]
+    trainer.project_name=${PROJECT_NAME}
+    trainer.experiment_name=${EXPERIMENT_NAME}
+    trainer.val_before_train=False
+    trainer.nnodes=${NNODES}
+    trainer.n_gpus_per_node=${NDEVICES_PER_NODE}
+    trainer.save_freq=${SAVE_FREQ}
+    trainer.test_freq=${TEST_FREQ}
+    trainer.total_epochs=${TOTAL_EPOCHS}
+)
+
+EXTRA=(
+    hydra.searchpath=[pkg://verl_mlite.config]
+)
+
+########################### launch ###########################
+python3 -m verl.trainer.main_ppo \
+    "${EXTRA[@]}" \
+    "${ALGORITHM[@]}" \
+    "${DATA[@]}" \
+    "${MODEL[@]}" \
+    "${ACTOR[@]}" \
+    "${ROLLOUT[@]}" \
+    "${TRAINER[@]}" \
+    "$@"
diff --git a/examples/sft/gsm8k/run_deepseek_v4_megatron_lite.sh b/examples/sft/gsm8k/run_deepseek_v4_megatron_lite.sh
new file mode 100755
index 00000000000..6387429b98e
--- /dev/null
+++ b/examples/sft/gsm8k/run_deepseek_v4_megatron_lite.sh
@@ -0,0 +1,170 @@
+#!/usr/bin/env bash
+# GSM8K SFT scale demo | DeepSeek-V4 | Megatron Lite training | GPU
+#
+# Megatron Lite is Megatron's agentic experimental path. Its upstream home is
+# Megatron-LM's dev branch:
+# https://github.com/NVIDIA/Megatron-LM/tree/dev/experimental/lite
+#
+# This launcher currently tracks the submitter's active branch until the latest
+# mlite changes merge upstream. That checkout provides both megatron.lite and
+# the verl_mlite backend glue:
+#
+#   git clone https://github.com/ISEEKYAN/mlite
+#   pip install -e mlite/experimental/lite/examples/verl
+#
+# DeepSeek-V4 uses fused DSA kernels on Hopper and Blackwell GPUs. The critical
+# DSA-only dependencies are nvidia-cutlass-dsl==4.5.2 and nvidia-cudnn-frontend.
+# cudnn-frontend release 1.24.1 is sufficient for Blackwell, while Hopper still
+# needs a develop-branch build with IndexerForwardSm90 support.
+#
+# MODEL_VARIANT selects the DeepSeek-V4 target and its default mlite mesh:
+#   - flash: 16 nodes, PP4 EP8  CP4
+#   - pro:   64 nodes, PP8 EP16 CP4
+#
+# DS4 is fixed to TP1/ETP1. The architecture does not support TP/ETP
+# sharding, and there is no plan to support it.
+#
+# OPTIMIZER selects the Megatron Lite optimizer path:
+#   - fsdp2:    Megatron Lite FSDP2 wrapper, lower memory pressure, default
+#   - dist_opt: original Megatron distributed optimizer
+# When using dist_opt, prefer a larger PP*EP mesh to reduce per-rank model and
+# optimizer memory pressure and avoid OOM.
+
+set -xeuo pipefail
+
+########################### mlite backend knobs ###########################
+MODEL_VARIANT=${MODEL_VARIANT:-flash}
+MLITE_ROOT=${MLITE_ROOT:-$HOME/mlite}
+MLITE_VERL_ROOT=${MLITE_VERL_ROOT:-${MLITE_ROOT}/experimental/lite/examples/verl}
+MLITE_LITE_ROOT=${MLITE_LITE_ROOT:-${MLITE_ROOT}/experimental/lite}
+
+OPTIMIZER=${OPTIMIZER:-fsdp2} # dist_opt
+ALL_OFFLOAD=${ALL_OFFLOAD:-True}
+
+case "${MODEL_VARIANT}" in
+    flash)
+        MODEL_PATH=${MODEL_PATH:-${FLASH_MODEL_PATH:-}}
+        NNODES=${NNODES:-16}
+        PP=${PP:-4}
+        EP=${EP:-8}
+        CP=${CP:-4}
+        ;;
+    pro)
+        MODEL_PATH=${MODEL_PATH:-${PRO_MODEL_PATH:-}}
+        NNODES=${NNODES:-64}
+        PP=${PP:-8}
+        EP=${EP:-16}
+        CP=${CP:-4}
+        ;;
+    *)
+        echo "Unsupported MODEL_VARIANT=${MODEL_VARIANT}. Expected flash or pro." >&2
+        exit 1
+        ;;
+esac
+
+: "${MODEL_PATH:?set MODEL_PATH, or set FLASH_MODEL_PATH/PRO_MODEL_PATH for MODEL_VARIANT=${MODEL_VARIANT}}"
+########################### end mlite backend knobs ###########################
+
+########################### user-adjustable ###########################
+NDEVICES_PER_NODE=${NDEVICES_PER_NODE:-8}
+NODE_RANK=${NODE_RANK:-0}
+MASTER_ADDR=${MASTER_ADDR:-127.0.0.1}
+MASTER_PORT=${MASTER_PORT:-29500}
+
+TRAIN_FILE=${TRAIN_FILE:-$HOME/data/gsm8k/train.parquet}
+TRAIN_BATCH_SIZE=${TRAIN_BATCH_SIZE:-32}
+MICRO_BATCH_SIZE_PER_GPU=${MICRO_BATCH_SIZE_PER_GPU:-1}
+MAX_LENGTH=${MAX_LENGTH:-2048}
+
+LR=${LR:-1e-5}
+MIN_LR=${MIN_LR:-1e-5}
+WEIGHT_DECAY=${WEIGHT_DECAY:-0.1}
+CLIP_GRAD=${CLIP_GRAD:-1.0}
+
+TOTAL_EPOCHS=${TOTAL_EPOCHS:-1}
+PROJECT_NAME=${PROJECT_NAME:-verl-mlite-deepseek_v4_${MODEL_VARIANT}-gsm8k-sft}
+EXPERIMENT_NAME=${EXPERIMENT_NAME:-deepseek_v4_${MODEL_VARIANT}_gsm8k_sft_${OPTIMIZER}}
+########################### end user-adjustable ###########################
+
+########################### derived defaults ###########################
+export PYTHONPATH="${MLITE_VERL_ROOT}:${MLITE_LITE_ROOT}:${MLITE_ROOT}:${VERL_ROOT:-}:${MEGATRON_ROOT:-}:${PYTHONPATH:-}"
+export CUDA_DEVICE_MAX_CONNECTIONS="${CUDA_DEVICE_MAX_CONNECTIONS:-1}"
+
+if [[ -n "${CUDA_VISIBLE_DEVICES:-}" ]]; then
+    unset ROCR_VISIBLE_DEVICES
+    unset HIP_VISIBLE_DEVICES
+fi
+
+########################### parameter arrays ###########################
+DATA=(
+    data.train_files="${TRAIN_FILE}"
+    data.train_batch_size=${TRAIN_BATCH_SIZE}
+    data.micro_batch_size_per_gpu=${MICRO_BATCH_SIZE_PER_GPU}
+    data.use_dynamic_bsz=True
+    data.max_token_len_per_gpu=${MAX_LENGTH}
+    data.max_length=${MAX_LENGTH}
+    data.pad_mode=no_padding
+    data.truncation=error
+    data.messages_key=messages
+)
+
+MODEL=(
+    model=hf_model
+    model.path="${MODEL_PATH}"
+    model.trust_remote_code=True
+)
+
+OPTIM=(
+    optim=megatron
+    optim.lr=${LR}
+    optim.min_lr=${MIN_LR}
+    optim.weight_decay=${WEIGHT_DECAY}
+    optim.clip_grad=${CLIP_GRAD}
+    optim.lr_warmup_steps=0
+    optim.lr_decay_style=constant
+    +optim.override_optimizer_config.offload_fraction=1.0
+    +optim.override_optimizer_config.use_precision_aware_optimizer=True
+    +optim.override_optimizer_config.decoupled_weight_decay=True
+)
+
+ENGINE=(
+    hydra.searchpath=[pkg://verl_mlite.config]
+    engine=mlite
+    engine.tp=1
+    engine.pp=${PP}
+    engine.vpp=1
+    engine.ep=${EP}
+    engine.cp=${CP}
+    engine.etp=1
+    engine.param_offload=${ALL_OFFLOAD}
+    engine.optimizer_offload=${ALL_OFFLOAD}
+    engine.grad_offload=${ALL_OFFLOAD}
+    engine.attention_backend_override=flash
+    engine.impl_cfg.use_thd=True
+    +engine.impl_cfg.optimizer=${OPTIMIZER}
+    +engine.impl_cfg.recompute=[full]
+)
+
+TRAINER=(
+    trainer.logger=[console]
+    trainer.project_name=${PROJECT_NAME}
+    trainer.experiment_name=${EXPERIMENT_NAME}
+    trainer.total_epochs=${TOTAL_EPOCHS}
+    trainer.nnodes=${NNODES}
+    trainer.n_gpus_per_node=${NDEVICES_PER_NODE}
+)
+
+########################### launch ###########################
+torchrun \
+    --nnodes="${NNODES}" \
+    --nproc_per_node="${NDEVICES_PER_NODE}" \
+    --node_rank="${NODE_RANK}" \
+    --master_addr="${MASTER_ADDR}" \
+    --master_port="${MASTER_PORT}" \
+    -m verl_mlite.launch verl.trainer.sft_trainer \
+    "${DATA[@]}" \
+    "${MODEL[@]}" \
+    "${OPTIM[@]}" \
+    "${ENGINE[@]}" \
+    "${TRAINER[@]}" \
+    "$@"
diff --git a/tests/special_sanity/check_example_naming.py b/tests/special_sanity/check_example_naming.py
index 8f1f92798a0..da145007087 100644
--- a/tests/special_sanity/check_example_naming.py
+++ b/tests/special_sanity/check_example_naming.py
@@ -19,8 +19,8 @@
     run_<model>_<train-backend>.sh
 
 Where ``<train-backend>`` is one of ``fsdp``, ``fsdp2``, ``megatron``,
-``mindspeed``, ``automodel`` or ``veomni``, and **must be the last
-underscore-separated token before** ``.sh`` — nothing follows it. The legacy
+``megatron_lite``, ``mindspeed``, ``automodel`` or ``veomni``, and **must be
+the final suffix before** ``.sh`` — nothing follows it. The legacy
 convention used to embed the inference backend (``vllm``/``sglang``/``trtllm``),
 platform tokens (``_npu``/``_amd``), machine-type tokens (``_gb200``,
 ``_blackwell``), quantization variants (``_fp8``), and ad-hoc trailing
@@ -66,13 +66,15 @@
 )
 
 # Recognised train-backend / engine markers. The filename must end with
-# ``_<one of these>.sh`` (i.e. the train-backend is the LAST underscore-
-# separated token). Generation-only scripts that do not run a trainer are
-# listed in ``DEFAULT_IGNORE_FILES`` instead.
+# ``_<one of these>.sh``. Some backends, such as ``megatron_lite``, contain an
+# underscore; validate the full suffix rather than only the last token.
+# Generation-only scripts that do not run a trainer are listed in
+# ``DEFAULT_IGNORE_FILES`` instead.
 ALLOWED_BACKENDS = (
     "fsdp",
     "fsdp2",
     "megatron",
+    "megatron_lite",
     "mindspeed",
     "automodel",
     "veomni",
@@ -114,6 +116,11 @@ def _split_tokens(stem: str) -> list[str]:
     return parts
 
 
+def _has_allowed_backend(stem: str) -> bool:
+    """Return whether ``stem`` ends with one of the allowed backend suffixes."""
+    return any(stem.endswith(f"_{backend}") for backend in ALLOWED_BACKENDS)
+
+
 def _is_ignored(path: Path, repo_root: Path, ignore_dirs: tuple[str, ...], ignore_files: tuple[str, ...]) -> bool:
     rel = path.relative_to(repo_root).as_posix()
     if rel in ignore_files:
@@ -153,7 +160,7 @@ def check_filename(path: Path, display: str | None = None) -> list[str]:
             f"the script, not embedded in the filename."
         )
 
-    if not tokens or tokens[-1] not in ALLOWED_BACKENDS:
+    if not tokens or not _has_allowed_backend(path.stem):
         errors.append(
             f"{shown}: filename must end with '_<train-backend>.sh' where "
             f"train-backend ∈ {list(ALLOWED_BACKENDS)} "
diff --git a/tests/special_sanity/test_check_example_naming.py b/tests/special_sanity/test_check_example_naming.py
index 41b6814df28..09a5542b3b4 100644
--- a/tests/special_sanity/test_check_example_naming.py
+++ b/tests/special_sanity/test_check_example_naming.py
@@ -46,6 +46,10 @@ def test_all_train_backends_accepted():
         assert _violations(f"run_qwen3_8b_{backend}.sh") == [], backend
 
 
+def test_multi_token_train_backend_accepted():
+    assert _violations("run_deepseek_v4_megatron_lite.sh") == []
+
+
 def test_forbidden_engine_token_rejected():
     errs = _violations("run_qwen3_8b_vllm_fsdp.sh")
     # `vllm` is both a forbidden token AND occupies the last-token slot