Update ngpus_per_node to 16 for all e2e_ascend CI test cases

Author: FightingZhen

.github/workflows/e2e_ascend.yml

@@ -57,6 +57,7 @@ on:
       - "examples/data_preprocess/gsm8k.py"
       - "examples/data_preprocess/geo3k.py"
       - "tests/special_e2e/ppo_trainer"
+      - "tests/special_npu"
       - "verl/trainer/main_ppo.py"
       - "verl/trainer/config/ppo_trainer.yaml"
 
@@ -123,6 +124,11 @@ jobs:
         run: |
           ray stop --force
           python3 examples/data_preprocess/geo3k.py
+      - name: Running gsm8k e2e qwen3 training tests with PPO on ASCEND NPU
+        run: |
+          ray stop --force
+          bash tests/special_npu/run_qwen3_06b_ppo.sh
+          rm -rf $HOME/ckpts
       - name: Running gsm8k e2e training tests with peft sft on ASCEND NPU
         run: |
           ray stop --force
@@ -143,16 +149,6 @@ jobs:
           ray stop --force
           bash tests/special_npu/run_qwen2_5_05b_dapo.sh
           rm -rf $HOME/ckpts
-      - name: Running gsm8k e2e qwen3 training tests with GRPO on ASCEND NPU
-        run: |
-          ray stop --force
-          bash tests/special_npu/run_qwen3_06b_grpo.sh
-          rm -rf $HOME/ckpts
-      - name: Running gsm8k e2e qwen3 training tests with PPO on ASCEND NPU
-        run: |
-          ray stop --force
-          bash tests/special_npu/run_qwen3_06b_ppo.sh
-          rm -rf $HOME/ckpts
       - name: Running gsm8k e2e training tests with GRPO MindSpeed on ASCEND NPU
         run: |
           ray stop --force

tests/special_npu/run_qwen2_5_05b_dapo.sh

@@ -1,7 +1,7 @@
 #!/usr/bin/env bash
 set -xeuo pipefail
 
-NUM_GPUS=${NUM_GPUS:-8}
+NUM_GPUS=${NUM_GPUS:-16}
 
 MODEL_ID=${MODEL_ID:-Qwen/Qwen2.5-0.5B-Instruct}
 MODEL_PATH=${MODEL_PATH:-${HOME}/models/${MODEL_ID}}

tests/special_npu/run_qwen2_5_05b_grpo.sh

@@ -36,7 +36,7 @@ python3 -m verl.trainer.main_ppo \
     trainer.logger=console \
     trainer.project_name='verl_grpo_example_gsm8k' \
     trainer.experiment_name='qwen2_7b_function_rm' \
-    trainer.n_gpus_per_node=8 \
+    trainer.n_gpus_per_node=16 \
     trainer.nnodes=1 \
     trainer.save_freq=-1 \
     trainer.test_freq=5 \

tests/special_npu/run_qwen2_5_05b_grpo_mindspeed.sh

@@ -59,7 +59,7 @@ 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=8 \
+    trainer.n_gpus_per_node=16 \
     trainer.nnodes=1 \
     trainer.save_freq=-1 \
     trainer.test_freq=5 \

tests/special_npu/run_qwen2_5_vl_3b_npu.sh

@@ -18,7 +18,7 @@ python3 -m verl.trainer.main_ppo \
     actor_rollout_ref.model.path=Qwen/Qwen2.5-VL-3B-Instruct \
     actor_rollout_ref.actor.optim.lr=1e-6 \
     actor_rollout_ref.model.use_remove_padding=True \
-    actor_rollout_ref.actor.ppo_mini_batch_size=16 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=32 \
     actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=2 \
     actor_rollout_ref.actor.use_kl_loss=True \
     actor_rollout_ref.actor.kl_loss_coef=0.01 \
@@ -44,7 +44,7 @@ python3 -m verl.trainer.main_ppo \
     trainer.logger=console \
     trainer.project_name='verl_grpo_example_geo3k' \
     trainer.experiment_name='qwen2_5_vl_3b_function_rm' \
-    trainer.n_gpus_per_node=8 \
+    trainer.n_gpus_per_node=16 \
     trainer.nnodes=1 \
     trainer.save_freq=-1 \
     trainer.test_freq=-1 \

tests/special_npu/run_qwen3_06b_grpo.sh

@@ -36,6 +36,7 @@
     get_ulysses_sequence_parallel_world_size,
     validate_ulysses_config,
 )
+from verl.utils.device import is_npu_available
 
 logger = logging.getLogger(__file__)
 logger.setLevel(os.getenv("VERL_LOGGING_LEVEL", "WARN"))
@@ -46,9 +47,19 @@
 
     _flash_supports_window_size = "window_size" in inspect.signature(flash_attn_func).parameters
     _flash_supports_deterministic = "deterministic" in inspect.signature(flash_attn_func).parameters
-    _flash_deterministic_enabled = os.getenv("FLASH_ATTENTION_DETERMINISTIC", "0") == "1"
     _flash_use_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
 
+if is_npu_available:
+    from transformers.integrations.npu_flash_attention import npu_flash_attn_func as flash_attn_func
+    from transformers.integrations.npu_flash_attention import npu_flash_attn_varlen_func as flash_attn_varlen_func
+    from transformers.modeling_flash_attention_utils import flash_attn_supports_top_left_mask
+
+    _flash_supports_window_size = "window_size" in inspect.signature(flash_attn_func).parameters
+    _flash_supports_deterministic = "deterministic" in inspect.signature(flash_attn_func).parameters
+    _flash_use_top_left_mask = flash_attn_supports_top_left_mask()
+
+_flash_deterministic_enabled = os.getenv("FLASH_ATTENTION_DETERMINISTIC", "0") == "1"
+
 
 def get_rope_index(
     processor,