[ci] chore: switch ascend ci calculation resource

.github/workflows/e2e_ascend.yml

@@ -68,27 +68,11 @@ jobs:
   test:
     if: github.repository_owner == 'volcengine'
     name: verl Ascend test (self-host)
-    runs-on: [self-hosted, npu-0]
+    runs-on: linux-aarch64-a2-8
     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:
-        - /usr/local/dcmi:/usr/local/dcmi
-        - /usr/local/bin/npu-smi:/usr/local/bin/npu-smi
-        - /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/
-        - /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info
-        - /etc/ascend_install.info:/etc/ascend_install.info
-        - /data00/dataset:/github/home/dataset
-        - /data00/models:/github/home/models
-        # Use self-host cache speed up pip and model download
-        # - /home/action/actions-runner/_work/cache:/github/home/.cache/
+      image: swr.ap-southeast-1.myhuaweicloud.com/base_image/ascend-ci/verl/verl:verl-8.3.rc1-910b-ubuntu22.04-py3.11-latest
       options: >-
-        --device /dev/davinci0
-        --device /dev/davinci_manager
-        --device /dev/devmm_svm
-        --device /dev/hisi_hdc
-        --network host
-        --privileged
         --shm-size 16g
     env:
       HTTP_PROXY: ${{ secrets.PROXY_HTTP }}
@@ -118,10 +102,10 @@ jobs:
           pip list
       - name: Preprocess gsm8k dataset
         run: |
-          python examples/data_preprocess/gsm8k.py --local_dataset_path ${HOME}/dataset/openai/gsm8k
+          python examples/data_preprocess/gsm8k.py --local_dataset_path ${HOME}/.cache/datasets/openai/gsm8k
       - name: Preprocess geo3k dataset
         run: |
-          python examples/data_preprocess/geo3k.py --local_dataset_path ${HOME}/dataset/hiyouga/geometry3k
+          python examples/data_preprocess/geo3k.py --local_dataset_path ${HOME}/.cache/datasets/hiyouga/geometry3k
       - name: Running gsm8k e2e qwen3 training tests with PPO on ASCEND NPU
         run: |
           ray stop --force

tests/special_npu/run_qwen2_5_05b_dapo.sh

@@ -2,10 +2,10 @@
 set -xeuo pipefail
 export VLLM_ASCEND_ENABLE_NZ=0
 
-NUM_GPUS=${NUM_GPUS:-16}
+NUM_GPUS=${NUM_GPUS:-8}
 
 MODEL_ID=${MODEL_ID:-Qwen/Qwen2.5-0.5B-Instruct}
-MODEL_PATH=${MODEL_PATH:-${HOME}/models/${MODEL_ID}}
+MODEL_PATH=${MODEL_PATH:-${HOME}/.cache/models/${MODEL_ID}}
 
 adv_estimator=grpo
 
@@ -25,19 +25,15 @@ overlong_penalty_factor=1.0
 
 loss_agg_mode="token-mean"
 
-enable_filter_groups=True
+enable_filter_groups=False
 filter_groups_metric=seq_reward
 max_num_gen_batches=10
 
-train_traj_micro_bsz_per_gpu=2 # b
-n_resp_per_prompt=4 # g
+train_traj_micro_bsz_per_gpu=1 # b
+n_resp_per_prompt=2 # g
 
-train_traj_micro_bsz=$((train_traj_micro_bsz_per_gpu * NUM_GPUS)) # b * n
-train_traj_mini_bsz=$((train_traj_micro_bsz * 2)) # 2 * b * n
-train_prompt_mini_bsz=$((train_traj_mini_bsz * n_resp_per_prompt)) # 2 * b * n / g
-train_prompt_bsz=$((train_prompt_mini_bsz * 2)) # 4 * b * n / g
-
-gen_prompt_bsz=$((train_prompt_bsz * 4))
+train_traj_micro_bsz=$((train_traj_micro_bsz_per_gpu * NUM_GPUS))
+train_prompt_mini_bsz=$((train_traj_micro_bsz * n_resp_per_prompt * 2))
 
 exp_name="$(basename "${MODEL_ID,,}")-dapo-minimal"
 
@@ -58,8 +54,7 @@ python3 -m recipe.dapo.main_dapo \
     reward_model.overlong_buffer.len=${overlong_buffer_len} \
     reward_model.overlong_buffer.penalty_factor=${overlong_penalty_factor} \
     actor_rollout_ref.actor.loss_agg_mode=${loss_agg_mode} \
-    data.train_batch_size=${train_prompt_bsz} \
-    data.gen_batch_size=${gen_prompt_bsz} \
+    data.gen_batch_size=${train_prompt_mini_bsz} \
     algorithm.filter_groups.enable=${enable_filter_groups} \
     algorithm.filter_groups.metric=${filter_groups_metric} \
     algorithm.filter_groups.max_num_gen_batches=${max_num_gen_batches} \
@@ -96,5 +91,5 @@ python3 -m recipe.dapo.main_dapo \
     trainer.total_epochs=1 \
     trainer.resume_mode=disable \
     trainer.val_before_train=False \
-    trainer.total_training_steps=2 \
+    trainer.total_training_steps=1 \
     trainer.device=npu $@

tests/special_npu/run_qwen2_5_05b_grpo.sh

@@ -2,22 +2,22 @@ set -x
 export VLLM_ASCEND_ENABLE_NZ=0
 
 MODEL_ID=${MODEL_ID:-Qwen/Qwen2.5-0.5B-Instruct}
-MODEL_PATH=${MODEL_PATH:-${HOME}/models/${MODEL_ID}}
+MODEL_PATH=${MODEL_PATH:-${HOME}/.cache/models/${MODEL_ID}}
 
 python3 -m verl.trainer.main_ppo \
     algorithm.adv_estimator=grpo \
     data.train_files=$HOME/data/gsm8k/train.parquet \
     data.val_files=$HOME/data/gsm8k/test.parquet \
-    data.train_batch_size=128 \
+    data.train_batch_size=16 \
     data.max_prompt_length=512 \
     data.max_response_length=128 \
     data.filter_overlong_prompts=True \
     data.truncation='error' \
     actor_rollout_ref.model.path="${MODEL_PATH}" \
     actor_rollout_ref.actor.optim.lr=5e-7 \
     actor_rollout_ref.model.use_remove_padding=False \
-    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
-    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=20 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=8 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.actor.use_kl_loss=True \
     actor_rollout_ref.actor.kl_loss_coef=0.001 \
     actor_rollout_ref.actor.kl_loss_type=low_var_kl \
@@ -26,23 +26,23 @@ python3 -m verl.trainer.main_ppo \
     actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
     actor_rollout_ref.actor.use_torch_compile=False \
     actor_rollout_ref.ref.use_torch_compile=False \
-    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=40 \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
     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=5 \
-    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=40 \
+    actor_rollout_ref.rollout.n=2 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.ref.fsdp_config.param_offload=True \
     algorithm.kl_ctrl.kl_coef=0.001 \
     trainer.critic_warmup=0 \
     trainer.logger=console \
     trainer.project_name='verl_grpo_example_gsm8k' \
     trainer.experiment_name='qwen2_7b_function_rm' \
-    trainer.n_gpus_per_node=16 \
+    trainer.n_gpus_per_node=8 \
     trainer.nnodes=1 \
     trainer.save_freq=-1 \
     trainer.test_freq=-1 \
     trainer.total_epochs=1 \
-    trainer.total_training_steps=2 \
+    trainer.total_training_steps=1 \
     trainer.device=npu $@

tests/special_npu/run_qwen2_5_05b_grpo_mindspeed.sh

@@ -2,7 +2,7 @@ set -x
 export VLLM_ASCEND_ENABLE_NZ=0
 
 MODEL_ID=${MODEL_ID:-Qwen/Qwen2.5-0.5B-Instruct}
-MODEL_PATH=${MODEL_PATH:-${HOME}/models/${MODEL_ID}}
+MODEL_PATH=${MODEL_PATH:-${HOME}/.cache/models/${MODEL_ID}}
 
 USE_DIST_CKPT=${USE_DIST_CKPT:-False}
 DIST_CKPT_PATH=${DIST_CKPT_PATH:-${HOME}/dist_ckpt/qwen2_5_05b_grpo_mindspeed}
@@ -21,15 +21,15 @@ python3 -m verl.trainer.main_ppo --config-path=config \
     algorithm.adv_estimator=grpo \
     data.train_files=$HOME/data/gsm8k/train.parquet \
     data.val_files=$HOME/data/gsm8k/test.parquet \
-    data.train_batch_size=128 \
+    data.train_batch_size=16 \
     data.max_prompt_length=512 \
     data.max_response_length=128 \
     data.filter_overlong_prompts=True \
     data.truncation='error' \
-    actor_rollout_ref.model.path=${MODEL_ID} \
+    actor_rollout_ref.model.path=${MODEL_PATH} \
     actor_rollout_ref.actor.optim.lr=5e-7 \
-    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
-    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=20 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=8 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.actor.strategy=megatron \
     actor_rollout_ref.actor.megatron.pipeline_model_parallel_size=2 \
     actor_rollout_ref.actor.megatron.tensor_model_parallel_size=2 \
@@ -40,13 +40,13 @@ python3 -m verl.trainer.main_ppo --config-path=config \
     actor_rollout_ref.actor.kl_loss_coef=0.001 \
     actor_rollout_ref.actor.kl_loss_type=low_var_kl \
     actor_rollout_ref.actor.use_torch_compile=False \
-    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=40 \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
     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=5 \
-    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=40 \
+    actor_rollout_ref.rollout.n=2 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.ref.strategy=megatron \
     actor_rollout_ref.ref.megatron.pipeline_model_parallel_size=2 \
     actor_rollout_ref.ref.megatron.tensor_model_parallel_size=2 \
@@ -59,11 +59,11 @@ python3 -m verl.trainer.main_ppo --config-path=config \
     trainer.logger=console \
     trainer.project_name='verl_grpo_example_gsm8k' \
     trainer.experiment_name='qwen2_7b_function_rm' \
-    trainer.n_gpus_per_node=16 \
+    trainer.n_gpus_per_node=8 \
     trainer.nnodes=1 \
     trainer.save_freq=-1 \
     trainer.test_freq=-1 \
     trainer.total_epochs=1 \
-    trainer.total_training_steps=2 \
+    trainer.total_training_steps=1 \
     trainer.device=npu \
     +actor_rollout_ref.actor.megatron.override_transformer_config.use_flash_attn=True $@

tests/special_npu/run_qwen2_5_05b_sft_peft_sp2.sh

@@ -3,7 +3,7 @@ set -x
 mkdir -p ./save_ckpts
 
 MODEL_ID=${MODEL_ID:-Qwen/Qwen2.5-0.5B-Instruct}
-MODEL_PATH=${MODEL_PATH:-${HOME}/models/${MODEL_ID}}
+MODEL_PATH=${MODEL_PATH:-${HOME}/.cache/models/${MODEL_ID}}
 
 torchrun --standalone --nnodes=1 --nproc_per_node=8 \
      -m verl.trainer.fsdp_sft_trainer \

tests/special_npu/run_qwen2_5_vl_3b_npu.sh

@@ -8,13 +8,13 @@ ENGINE=${1:-vllm}
 export USE_OPTIMIZED_MODEL=0
 
 MODEL_ID=${MODEL_ID:-Qwen/Qwen2.5-VL-3B-Instruct}
-MODEL_PATH=${MODEL_PATH:-${HOME}/models/${MODEL_ID}}
+MODEL_PATH=${MODEL_PATH:-${HOME}/.cache/models/${MODEL_ID}}
 
 python3 -m verl.trainer.main_ppo \
     algorithm.adv_estimator=grpo \
     data.train_files=$HOME/data/geo3k/train.parquet \
     data.val_files=$HOME/data/geo3k/test.parquet \
-    data.train_batch_size=512 \
+    data.train_batch_size=16 \
     data.max_prompt_length=1024 \
     data.max_response_length=2048 \
     data.filter_overlong_prompts=True \
@@ -23,8 +23,8 @@ python3 -m verl.trainer.main_ppo \
     actor_rollout_ref.model.path="${MODEL_PATH}" \
     actor_rollout_ref.actor.optim.lr=1e-6 \
     actor_rollout_ref.model.use_remove_padding=True \
-    actor_rollout_ref.actor.ppo_mini_batch_size=32 \
-    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=8 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.actor.use_kl_loss=True \
     actor_rollout_ref.actor.kl_loss_coef=0.01 \
     actor_rollout_ref.actor.kl_loss_type=low_var_kl \
@@ -34,25 +34,25 @@ python3 -m verl.trainer.main_ppo \
     actor_rollout_ref.actor.fsdp_config.param_offload=False \
     actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
     actor_rollout_ref.ref.use_torch_compile=False \
-    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=4 \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
     actor_rollout_ref.rollout.name=$ENGINE \
     actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
     actor_rollout_ref.rollout.enable_chunked_prefill=False \
     actor_rollout_ref.rollout.enforce_eager=True \
     actor_rollout_ref.rollout.free_cache_engine=True \
-    actor_rollout_ref.rollout.n=5 \
-    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=4 \
+    actor_rollout_ref.rollout.n=2 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.ref.fsdp_config.param_offload=True \
     algorithm.use_kl_in_reward=False \
     trainer.critic_warmup=0 \
     trainer.logger=console \
     trainer.project_name='verl_grpo_example_geo3k' \
     trainer.experiment_name='qwen2_5_vl_3b_function_rm' \
-    trainer.n_gpus_per_node=16 \
+    trainer.n_gpus_per_node=8 \
     trainer.nnodes=1 \
     trainer.save_freq=-1 \
     trainer.test_freq=-1 \
     trainer.total_epochs=1 \
-    trainer.total_training_steps=2 \
+    trainer.total_training_steps=1 \
     trainer.device=npu $@

tests/special_npu/run_qwen3_06b_ppo.sh

@@ -1,33 +1,30 @@
 set -x
 export VLLM_ASCEND_ENABLE_NZ=0
 
-# TODO (FightingZhen) Env VLLM_USE_V1=1 is not supported in vllm==0.7.3
-# export VLLM_USE_V1=1
-
 MODEL_ID=${MODEL_ID:-Qwen/Qwen2.5-0.5B-Instruct}  # TODO: change to Qwen3-0.6B when CI server is ready
-MODEL_PATH=${MODEL_PATH:-${HOME}/models/${MODEL_ID}}
+MODEL_PATH=${MODEL_PATH:-${HOME}/.cache/models/${MODEL_ID}}
 
 python3 -m verl.trainer.main_ppo \
     algorithm.adv_estimator=gae \
     data.train_files=$HOME/data/gsm8k/train.parquet \
     data.val_files=$HOME/data/gsm8k/test.parquet \
-    data.train_batch_size=128 \
+    data.train_batch_size=16 \
     data.max_prompt_length=512 \
     data.max_response_length=128 \
     data.shuffle=False \
     actor_rollout_ref.model.path="${MODEL_PATH}" \
     actor_rollout_ref.model.use_remove_padding=True \
     actor_rollout_ref.model.enable_gradient_checkpointing=True \
     actor_rollout_ref.actor.optim.lr=1e-6 \
-    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
-    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=8 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.actor.fsdp_config.param_offload=True \
     actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
     actor_rollout_ref.actor.use_kl_loss=False \
     actor_rollout_ref.actor.ulysses_sequence_parallel_size=2 \
     actor_rollout_ref.actor.use_dynamic_bsz=True \
     actor_rollout_ref.actor.use_torch_compile=False \
-    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
     actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
     actor_rollout_ref.rollout.name=vllm \
     actor_rollout_ref.rollout.gpu_memory_utilization=0.8 \
@@ -38,7 +35,7 @@ python3 -m verl.trainer.main_ppo \
     critic.model.use_remove_padding=True \
     critic.model.path="${MODEL_PATH}" \
     critic.model.enable_gradient_checkpointing=True \
-    critic.ppo_micro_batch_size_per_gpu=8 \
+    critic.ppo_micro_batch_size_per_gpu=1 \
     critic.ulysses_sequence_parallel_size=2 \
     critic.model.fsdp_config.param_offload=True \
     critic.model.fsdp_config.optimizer_offload=True \
@@ -52,5 +49,5 @@ python3 -m verl.trainer.main_ppo \
     trainer.save_freq=-1 \
     trainer.test_freq=-1 \
     trainer.total_epochs=1 \
-    trainer.total_training_steps=2 \
+    trainer.total_training_steps=1 \
     trainer.device=npu $@

tests/special_npu/run_qwen3_30b_dapo_mindspeed.sh

@@ -4,7 +4,7 @@ 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}}
+MODEL_PATH=${MODEL_PATH:-${HOME}/.cache/models/${MODEL_ID}}
 USE_DIST_CKPT=${USE_DIST_CKPT:-False}
 DIST_CKPT_PATH=${DIST_CKPT_PATH:-${HOME}/dist_ckpt/qwen3_30b_dapo_mindspeed}
 
@@ -73,12 +73,12 @@ python3 -m recipe.dapo.main_dapo \
     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.log_prob_micro_batch_size_per_gpu=1 \
     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.n=2 \
     actor_rollout_ref.rollout.temperature=1.0 \
     actor_rollout_ref.rollout.top_p=1.0 \
     actor_rollout_ref.rollout.top_k=-1 \
@@ -96,8 +96,8 @@ python3 -m recipe.dapo.main_dapo \
     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.actor.ppo_mini_batch_size=8 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
     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 \
@@ -108,7 +108,7 @@ python3 -m recipe.dapo.main_dapo \
     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.log_prob_micro_batch_size_per_gpu=1 \
     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 \
@@ -119,12 +119,12 @@ python3 -m recipe.dapo.main_dapo \
     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.n_gpus_per_node=8 \
     trainer.nnodes=1 \
     trainer.save_freq=-1 \
     trainer.test_freq=-1 \
     trainer.total_epochs=1 \
-    trainer.total_training_steps=2 \
+    trainer.total_training_steps=1 \
     trainer.device=npu \
     actor_rollout_ref.actor.use_torch_compile=False \
     actor_rollout_ref.ref.use_torch_compile=False \