[ci, megatron] test: Add Qwen3 Megatron+Mindspeed Ascend NPU CI

.github/workflows/e2e_ascend.yml

@@ -12,7 +12,7 @@
 # - `special_sanity`: a suite of quick sanity tests
 # - `special_standalone`: a set of test that are designed to run in dedicated environments
 
-# Accelerators for tests 
+# Accelerators for tests
 # - By default tests are run with GPU available, except for the ones under `special_npu`, and any test script whose name ends with `on_cpu.py`.
 # - For test scripts with `on_cpu.py` name suffix would be tested on CPU resources in linux environment.
 
@@ -67,7 +67,7 @@ jobs:
     if: github.repository_owner == 'volcengine'
     name: verl Ascend test (self-host)
     runs-on: [self-hosted, npu-0]
-    timeout-minutes: 45 # Increase this timeout value as needed
+    timeout-minutes: 60 # Increase this timeout value as needed
     container:
       image: quay.io/ascend/verl:verl-8.3.rc1-910b-ubuntu22.04-py3.11-latest
       volumes:
@@ -88,7 +88,7 @@ jobs:
         --network host
         --privileged
         --shm-size 16g
-    env: 
+    env:
       HTTP_PROXY: ${{ secrets.PROXY_HTTP }}
       HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }}
       NO_PROXY: "localhost,127.0.0.1,hf-mirror.com"
@@ -140,6 +140,11 @@ jobs:
           ray stop --force
           bash tests/special_npu/run_qwen2_5_05b_dapo.sh
           rm -rf $HOME/ckpts
+      - name: Running gsm8k e2e qwen3 MoE training tests with DAPO MindSpeed on ASCEND NPU
+        run: |
+          ray stop --force
+          export PYTHONPATH=$PYTHONPATH:/Megatron-LM
+          USE_DIST_CKPT=True USE_DUMMY_MODEL=True DUMMY_MODEL_CONFIG_PATH=tests/special_e2e/ppo_trainer/expert_parallel/qwen3moe_minimal.json DUMMY_MODEL_PATH=$HOME/dist_ckpt/qwen3_30b_dapo_mindspeed bash tests/special_npu/run_qwen3_30b_dapo_mindspeed.sh
       - name: Running gsm8k e2e training tests with GRPO MindSpeed on ASCEND NPU
         run: |
           ray stop --force

recipe/dapo/run_dapo_qwen3_moe_30b_megatron_npu.sh

@@ -58,6 +58,9 @@ offload=True
 
 max_num_batched_tokens=$((max_prompt_length + max_response_length))
 
+# vllm
+gen_tp=4
+
 # Megatron backen
 train_tp=4
 train_ep=2

tests/special_e2e/ppo_trainer/expert_parallel/qwen3moe_minimal.json

@@ -0,0 +1,4 @@
+{
+    "num_hidden_layers": 2,
+    "max_window_layers": 2
+}

tests/special_npu/run_qwen3_30b_dapo_mindspeed.sh

@@ -0,0 +1,139 @@
+#!/usr/bin/env bash
+set -xeuo pipefail
+
+export VLLM_ASCEND_ENABLE_NZ=0
+
+MODEL_ID=${MODEL_ID:-Qwen/Qwen3-30B-A3B-Instruct-2507}
+MODEL_PATH=${MODEL_PATH:-${HOME}/models/${MODEL_ID}}
+USE_DIST_CKPT=${USE_DIST_CKPT:-False}
+DIST_CKPT_PATH=${DIST_CKPT_PATH:-${HOME}/dist_ckpt/qwen3_30b_dapo_mindspeed}
+
+# use dummy model
+if [[ "$USE_DUMMY_MODEL" == "True" ]]; then
+    DUMMY_MODEL_PATH=${DUMMY_MODEL_PATH:-${HOME}/models_dummy/${MODEL_ID}}
+    if [ -z "${DUMMY_MODEL_CONFIG_PATH}" ]; then
+        echo "[ERROR] DUMMY_MODEL_CONFIG_PATH not set"
+        exit 1
+    fi
+
+    # make sure the path is empty
+    if [[ -d $DUMMY_MODEL_PATH && $DUMMY_MODEL_PATH != "/" ]]; then
+        rm -rf $DUMMY_MODEL_PATH
+    fi
+
+    # init model
+    python scripts/init_random_model.py \
+        --hf_model_path "${MODEL_PATH}" \
+        --new_config_path "${DUMMY_MODEL_CONFIG_PATH}" \
+        --output_path "${DUMMY_MODEL_PATH}"
+
+    # replace model path
+    MODEL_PATH=$DUMMY_MODEL_PATH
+fi
+
+# convert to megatron
+if [[ "$USE_DIST_CKPT" == "True" ]]; then
+
+    if [[ "$USE_DUMMY_MODEL" == "True" ]]; then
+        DIST_CKPT_PATH=${HOME}/dist_ckpt/qwen3_30b_dapo_mindspeed_dummy
+
+        if [[ -d $DIST_CKPT_PATH && $DIST_CKPT_PATH != "/" ]];then
+            rm -rf $DIST_CKPT_PATH
+        fi
+    fi
+
+    torchrun --nproc_per_node 2 --nnodes 1 scripts/converter_hf_to_mcore.py \
+        --hf_model_path "${MODEL_PATH}" \
+        --output_path "${DIST_CKPT_PATH}"
+fi
+
+exp_name='Qwen3-30B-A3B-DAPO-MindSpeed'
+
+max_prompt_length=512
+max_response_length=1024
+
+train_prompt_bsz=16
+
+actor_ppo_max_token_len=$(((max_prompt_length + max_response_length)))
+infer_ppo_max_token_len=$(((max_prompt_length + max_response_length)))
+
+python3 -m recipe.dapo.main_dapo \
+    --config-name="dapo_megatron_trainer" \
+    data.train_files=${HOME}/data/gsm8k/train.parquet \
+    data.val_files=${HOME}/data/gsm8k/test.parquet \
+    data.train_batch_size=${train_prompt_bsz} \
+    data.max_prompt_length=${max_prompt_length} \
+    data.max_response_length=${max_response_length} \
+    data.filter_overlong_prompts=True \
+    data.shuffle=False \
+    data.truncation='left' \
+    data.gen_batch_size=${train_prompt_bsz} \
+    algorithm.adv_estimator=grpo \
+    algorithm.use_kl_in_reward=False \
+    algorithm.filter_groups.enable=False \
+    algorithm.filter_groups.max_num_gen_batches=10 \
+    algorithm.filter_groups.metric=acc \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.rollout.enable_chunked_prefill=False \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
+    actor_rollout_ref.rollout.name=vllm \
+    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
+    actor_rollout_ref.rollout.n=8 \
+    actor_rollout_ref.rollout.temperature=1.0 \
+    actor_rollout_ref.rollout.top_p=1.0 \
+    actor_rollout_ref.rollout.top_k=-1 \
+    actor_rollout_ref.rollout.enforce_eager=True \
+    actor_rollout_ref.rollout.free_cache_engine=True \
+    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=${actor_ppo_max_token_len} \
+    actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=${infer_ppo_max_token_len} \
+    actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=${infer_ppo_max_token_len} \
+    actor_rollout_ref.actor.strategy=megatron \
+    actor_rollout_ref.actor.kl_loss_coef=0.0 \
+    actor_rollout_ref.actor.clip_ratio_low=0.2 \
+    actor_rollout_ref.actor.clip_ratio_high=0.28 \
+    actor_rollout_ref.actor.clip_ratio_c=10.0 \
+    actor_rollout_ref.actor.ppo_epochs=1 \
+    actor_rollout_ref.actor.use_dynamic_bsz=True \
+    actor_rollout_ref.model.path="${MODEL_PATH}" \
+    actor_rollout_ref.actor.optim.lr=1e-6 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=16 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=True \
+    actor_rollout_ref.ref.log_prob_use_dynamic_bsz=True \
+    actor_rollout_ref.actor.megatron.pipeline_model_parallel_size=2 \
+    actor_rollout_ref.actor.megatron.tensor_model_parallel_size=2 \
+    actor_rollout_ref.actor.megatron.expert_model_parallel_size=2 \
+    actor_rollout_ref.actor.megatron.use_dist_checkpointing=${USE_DIST_CKPT} \
+    actor_rollout_ref.actor.megatron.dist_checkpointing_path=${DIST_CKPT_PATH} \
+    actor_rollout_ref.actor.use_kl_loss=False \
+    actor_rollout_ref.actor.loss_agg_mode="token-mean" \
+    actor_rollout_ref.ref.strategy=megatron \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.ref.megatron.pipeline_model_parallel_size=2 \
+    actor_rollout_ref.ref.megatron.tensor_model_parallel_size=2 \
+    actor_rollout_ref.ref.megatron.expert_model_parallel_size=2 \
+    actor_rollout_ref.ref.megatron.use_dist_checkpointing=${USE_DIST_CKPT} \
+    actor_rollout_ref.ref.megatron.dist_checkpointing_path=${DIST_CKPT_PATH} \
+    reward_model.reward_manager=dapo \
+    algorithm.kl_ctrl.kl_coef=0.0 \
+    trainer.logger=['console'] \
+    trainer.project_name='verl_gsm8k_example' \
+    trainer.experiment_name='qwen3_30b_a3b_cut_gsm8k_mindspeed' \
+    trainer.n_gpus_per_node=16 \
+    trainer.nnodes=1 \
+    trainer.save_freq=-1 \
+    trainer.test_freq=-1 \
+    trainer.total_epochs=1 \
+    trainer.total_training_steps=2 \
+    trainer.device=npu \
+    actor_rollout_ref.actor.use_torch_compile=False \
+    actor_rollout_ref.ref.use_torch_compile=False \
+    +actor_rollout_ref.actor.megatron.override_transformer_config.use_flash_attn=True $@
+
+# clean up
+if [[ "$USE_DUMMY_MODEL" == "True" ]]; then
+    rm -rf $DUMMY_MODEL_PATH
+    if [[ "$USE_DIST_CKPT" == "True" ]]; then
+        rm -rf $DIST_CKPT_PATH
+    fi
+fi