[megatron,recipe] feat: support Qwen3-30B (MoE) DAPO training on ASCEND NPU (#3203)

### What does this PR do? Fix of megatron config, and example shell of Qwen3-30B-Dapo with megatron. ### Checklist Before Starting - [x] Search for similar PRs. Paste at least one query link here: ... - [x] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data` - If this PR involves multiple modules, separate them with `,` like `[megatron, fsdp, doc]` - `{type}` is in `feat`, `fix`, `refactor`, `chore`, `test` - If this PR breaks any API (CLI arguments, config, function signature, etc.), add `[BREAKING]` to the beginning of the title. - Example: `[BREAKING][fsdp, megatron] feat: dynamic batching` ### Test critic/reward/mean: <img width="1304" height="704" alt="dapo_30b_megatron" src="https://github.com/user-attachments/assets/f2062e24-b37d-4d54-8dd6-e9da25f8c69b" /> response_length/mean: <img width="815" height="407" alt="image" src="https://github.com/user-attachments/assets/f59b6c7b-4f24-4aa7-9b9e-bb8184dac5d3" /> ### Checklist Before Submitting > [!IMPORTANT] > Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review. - [x] Read the [Contribute Guide](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [x] Apply [pre-commit checks](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md#code-linting-and-formatting): `pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always` - [x] Add / Update [the documentation](https://github.com/volcengine/verl/tree/main/docs). - [x] Add unit or end-to-end test(s) to [the CI workflow](https://github.com/volcengine/verl/tree/main/.github/workflows) to cover all the code. If not feasible, explain why: ... - [ ] Once your PR is ready for CI, send a message in [the `ci-request` channel](https://verl-project.slack.com/archives/C091TCESWB1) in [the `verl` Slack workspace](https://join.slack.com/t/verl-project/shared_invite/zt-3855yhg8g-CTkqXu~hKojPCmo7k_yXTQ). (If not accessible, please try [the Feishu group (飞书群)](https://applink.larkoffice.com/client/chat/chatter/add_by_link?link_token=772jd4f1-cd91-441e-a820-498c6614126a).)
2025-10-20 13:43:50 +08:00 · 2025-09-13 19:08:23 +08:00
parent 6e6fafdc74
commit b00b090149
6 changed files with 202 additions and 1 deletions
--- a/recipe/dapo/config/dapo_megatron_trainer.yaml
+++ b/recipe/dapo/config/dapo_megatron_trainer.yaml
@ -0,0 +1,28 @@
+hydra:
+  searchpath:
+    - file://verl/trainer/config
+
+defaults:
+  - ppo_megatron_trainer
+  - _self_
+
+data:
+  gen_batch_size: ${data.train_batch_size}
+
+reward_model:
+  reward_manager: dapo
+  overlong_buffer: 
+    enable: False # We try to avoid forgetting to set enable
+    len: 0
+    penalty_factor: 0.0
+    log: False
+
+algorithm:
+  filter_groups:
+    _target_: verl.trainer.config.FilterGroupsConfig
+    enable: False # We try to avoid forgetting to set enable
+    metric: null # acc / score / seq_reward / seq_final_reward / ...
+    max_num_gen_batches: 0 # Non-positive values mean no upper limit
+
+trainer:
+  project_name: verl-dapo
--- a/recipe/dapo/run_dapo_qwen3_moe_30b_base_fsdp_npu.sh
+++ b/recipe/dapo/run_dapo_qwen3_moe_30b_base_fsdp_npu.sh
--- a/recipe/dapo/run_dapo_qwen3_moe_30b_megatron_npu.sh
+++ b/recipe/dapo/run_dapo_qwen3_moe_30b_megatron_npu.sh
@ -0,0 +1,169 @@
+#!/bin/bash
+
+project_name='DAPO'
+exp_name='DAPO-Qwen3-30B-megatron'
+
+adv_estimator=grpo
+
+use_kl_in_reward=False
+kl_coef=0.0
+use_kl_loss=False
+kl_loss_coef=0.0
+
+clip_ratio_low=0.2
+clip_ratio_high=0.28
+
+max_prompt_length=$((1024 * 2))
+max_response_length=$((1024 * 20))
+enable_overlong_buffer=True
+overlong_buffer_len=$((1024 * 4))
+overlong_penalty_factor=1.0
+
+loss_agg_mode="token-mean"
+
+enable_filter_groups=True
+filter_groups_metric=acc
+max_num_gen_batches=10
+train_prompt_bsz=16
+gen_prompt_bsz=$((train_prompt_bsz * 2))
+n_resp_per_prompt=16
+train_prompt_mini_bsz=2
+
+# Ray
+RAY_ADDRESS=${RAY_ADDRESS:-"http://localhost:8265"}
+WORKING_DIR=${WORKING_DIR:-"${PWD}"}
+RUNTIME_ENV=${RUNTIME_ENV:-"${WORKING_DIR}/verl/trainer/runtime_env.yaml"}
+NNODES=${NNODES:-1}
+# Paths
+RAY_DATA_HOME=${RAY_DATA_HOME:-"${HOME}/verl"}
+MODEL_PATH=${MODEL_PATH:-"${RAY_DATA_HOME}/models/Qwen3-30B-A3B"}
+# MCORE_MODEL_PATH points to the converted checkpoint.
+# To avoid loading these weights, set actor_rollout_ref.actor.megatron.use_dist_checkpointing=False.
+MCORE_MODEL_PATH=${MCORE_MODEL_PATH:-"${RAY_DATA_HOME}/models/Qwen3-30B-A3B-dist_ckpt"}
+CKPTS_DIR=${CKPTS_DIR:-"${RAY_DATA_HOME}/ckpts/${project_name}/${exp_name}"}
+TRAIN_FILE=${TRAIN_FILE:-"${RAY_DATA_HOME}/data/dapo-math-17k.parquet"}
+TEST_FILE=${TEST_FILE:-"${RAY_DATA_HOME}/data/aime-2024.parquet"}
+
+# Algorithm
+temperature=1.0
+top_p=1.0
+top_k=-1 # 0 for HF rollout, -1 for vLLM rollout
+
+# Performance Related Parameter
+sp_size=8
+use_dynamic_bsz=True
+actor_ppo_max_token_len=$(((max_prompt_length + max_response_length)))
+infer_ppo_max_token_len=$(((max_prompt_length + max_response_length)))
+offload=True
+
+max_num_batched_tokens=$((max_prompt_length + max_response_length))
+
+# Megatron backen
+train_tp=4
+train_ep=2
+train_pp=2
+train_cp=1
+
+ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
+    --address "${RAY_ADDRESS}" \
+    -- python3 -m recipe.dapo.main_dapo \
+    --config-name="dapo_megatron_trainer" \
+    data.filter_overlong_prompts=False \
+    data.train_files="${TRAIN_FILE}" \
+    data.val_files="${TEST_FILE}" \
+    data.shuffle=False \
+    data.prompt_key=prompt \
+    data.truncation='left' \
+    data.max_prompt_length=${max_prompt_length} \
+    data.max_response_length=${max_response_length} \
+    data.gen_batch_size=${gen_prompt_bsz} \
+    data.train_batch_size=${train_prompt_bsz} \
+    algorithm.adv_estimator=${adv_estimator} \
+    algorithm.use_kl_in_reward=${use_kl_in_reward} \
+    algorithm.kl_ctrl.kl_coef=${kl_coef} \
+    actor_rollout_ref.rollout.n=${n_resp_per_prompt} \
+    actor_rollout_ref.rollout.name=vllm \
+    actor_rollout_ref.actor.use_kl_loss=${use_kl_loss} \
+    actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef} \
+    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=10.0 \
+    actor_rollout_ref.actor.ppo_epochs=1 \
+    algorithm.filter_groups.enable=${enable_filter_groups} \
+    algorithm.filter_groups.max_num_gen_batches=${max_num_gen_batches} \
+    algorithm.filter_groups.metric=${filter_groups_metric} \
+    actor_rollout_ref.model.use_remove_padding=True \
+    actor_rollout_ref.actor.use_dynamic_bsz=${use_dynamic_bsz} \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=1 \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
+    actor_rollout_ref.ref.log_prob_use_dynamic_bsz=${use_dynamic_bsz} \
+    actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=${use_dynamic_bsz} \
+    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=${infer_ppo_max_token_len} \
+    actor_rollout_ref.model.path="${MODEL_PATH}" \
+    +actor_rollout_ref.model.override_config.attention_dropout=0. \
+    +actor_rollout_ref.model.override_config.embd_pdrop=0. \
+    +actor_rollout_ref.model.override_config.resid_pdrop=0. \
+    actor_rollout_ref.actor.optim.lr=1e-6 \
+    +actor_rollout_ref.critic.optim.lr=5e-8 \
+    actor_rollout_ref.actor.optim.lr_warmup_steps=10 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=${train_prompt_mini_bsz} \
+    actor_rollout_ref.actor.megatron.param_offload=${offload} \
+    actor_rollout_ref.actor.megatron.optimizer_offload=${offload} \
+    actor_rollout_ref.actor.megatron.grad_offload=${offload} \
+    actor_rollout_ref.actor.megatron.pipeline_model_parallel_size=${train_pp} \
+    actor_rollout_ref.actor.megatron.tensor_model_parallel_size=${train_tp} \
+    actor_rollout_ref.actor.megatron.expert_model_parallel_size=${train_ep} \
+    actor_rollout_ref.actor.megatron.context_parallel_size=${train_cp} \
+    actor_rollout_ref.actor.megatron.dist_checkpointing_path=${MCORE_MODEL_PATH} \
+    actor_rollout_ref.actor.megatron.use_dist_checkpointing=True \
+    actor_rollout_ref.ref.megatron.pipeline_model_parallel_size=${train_pp} \
+    actor_rollout_ref.ref.megatron.tensor_model_parallel_size=${train_tp} \
+    actor_rollout_ref.ref.megatron.expert_model_parallel_size=${train_ep} \
+    actor_rollout_ref.ref.megatron.context_parallel_size=${train_cp} \
+    actor_rollout_ref.ref.megatron.param_offload=${offload} \
+    actor_rollout_ref.ref.megatron.dist_checkpointing_path=${MCORE_MODEL_PATH} \
+    actor_rollout_ref.actor.entropy_coeff=0 \
+    actor_rollout_ref.actor.loss_agg_mode=${loss_agg_mode} \
+    actor_rollout_ref.rollout.gpu_memory_utilization=0.7 \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=${gen_tp} \
+    actor_rollout_ref.rollout.enable_chunked_prefill=True \
+    actor_rollout_ref.rollout.enable_prefix_caching=False \
+    actor_rollout_ref.rollout.max_num_batched_tokens=${max_num_batched_tokens} \
+    actor_rollout_ref.rollout.max_model_len=$((max_prompt_length + max_response_length)) \
+    actor_rollout_ref.rollout.temperature=${temperature} \
+    actor_rollout_ref.rollout.top_p=${top_p} \
+    actor_rollout_ref.rollout.top_k=${top_k} \
+    actor_rollout_ref.rollout.val_kwargs.temperature=${temperature} \
+    actor_rollout_ref.rollout.val_kwargs.top_p=${top_p} \
+    actor_rollout_ref.rollout.val_kwargs.top_k=${top_k} \
+    actor_rollout_ref.rollout.val_kwargs.do_sample=True \
+    actor_rollout_ref.rollout.val_kwargs.n=1 \
+    actor_rollout_ref.rollout.enforce_eager=True \
+    actor_rollout_ref.rollout.free_cache_engine=True \
+    actor_rollout_ref.ref.megatron.use_dist_checkpointing=True \
+    reward_model.reward_manager=dapo \
+    reward_model.overlong_buffer.enable=${enable_overlong_buffer} \
+    reward_model.overlong_buffer.len=${overlong_buffer_len} \
+    reward_model.overlong_buffer.penalty_factor=${overlong_penalty_factor} \
+    trainer.logger=['console'] \
+    trainer.project_name="${project_name}" \
+    trainer.experiment_name="${exp_name}" \
+    trainer.n_gpus_per_node=16 \
+    trainer.nnodes="${NNODES}" \
+    trainer.val_before_train=False \
+    trainer.test_freq=-1 \
+    trainer.save_freq=-1 \
+    trainer.total_epochs=1 \
+    trainer.default_local_dir="${CKPTS_DIR}" \
+    trainer.device="npu" \
+    actor_rollout_ref.nccl_timeout=14400 \
+    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 \
+    +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_method=uniform \
+    +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_granularity=full \
+    +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_num_layers=1
+
--- a/verl/trainer/config/_generated_ppo_megatron_trainer.yaml
+++ b/verl/trainer/config/_generated_ppo_megatron_trainer.yaml
@ -264,6 +264,7 @@ actor_rollout_ref:
        freeze_moe_router: false
    use_fused_kernels: false
    trust_remote_code: false
+    use_remove_padding: false
 data:
  tokenizer: null
  use_shm: false
--- a/verl/trainer/config/ppo_megatron_trainer.yaml
+++ b/verl/trainer/config/ppo_megatron_trainer.yaml
@ -43,6 +43,9 @@ actor_rollout_ref:

    trust_remote_code: False

+    # Whether to remove padding tokens in inputs during training
+    use_remove_padding: false
+
  rollout:
    # may get higher throughput when set to True. When activated, Please increase max_num_batched_tokens or decrease max_model_len.
    enable_chunked_prefill: False
--- a/verl/utils/reward_score/init.py
+++ b/verl/utils/reward_score/init.py
@ -55,7 +55,7 @@ def default_compute_score(

        # from . import math_verify
        # res = math_verify.compute_score(solution_str, ground_truth)
-    elif data_source == "math_dapo" or data_source.startswith("aime"):
+    elif data_source in ["math_dapo", "math", "math_dapo_reasoning"] or data_source.startswith("aime"):
        from . import math_dapo

        res = math_dapo.compute_score(solution_str, ground_truth)