From b1be57803cf46b51eff2a4748c7f9a1b59159fd9 Mon Sep 17 00:00:00 2001
From: Yan Bai <bayan@nvidia.com>
Date: Mon, 15 Dec 2025 06:22:31 -0800
Subject: [PATCH 1/4] megatron engine assert mbridge=True

---
 .../open_math_reasoning/run_sft_qwen3_8b.sh   |   2 +-
 verl/models/mcore/registry.py                 |  77 ++++++----
 verl/utils/megatron_utils.py                  |   8 +
 .../engine/megatron/transformer_impl.py       | 143 +++++++-----------
 4 files changed, 112 insertions(+), 118 deletions(-)

diff --git a/recipe/open_math_reasoning/run_sft_qwen3_8b.sh b/recipe/open_math_reasoning/run_sft_qwen3_8b.sh
index 3b7e9bb5c6c..ec564a1d602 100644
--- a/recipe/open_math_reasoning/run_sft_qwen3_8b.sh
+++ b/recipe/open_math_reasoning/run_sft_qwen3_8b.sh
@@ -55,7 +55,7 @@ MEGATRON_ENGINE_CONFIG="\
     engine.pipeline_model_parallel_size=${PP_SIZE} \
     engine.virtual_pipeline_model_parallel_size=${VPP_SIZE} \
     engine.context_parallel_size=${CP_SIZE} \
-    engine.use_mbridge=False"
+    engine.use_mbridge=True"
 
 if [ "$backend" = "fsdp" ]; then
     ENGINE_CONFIG="$FSDP_ENGINE_CONFIG"
diff --git a/verl/models/mcore/registry.py b/verl/models/mcore/registry.py
index 88d90be3713..fd0f0ce1a5f 100644
--- a/verl/models/mcore/registry.py
+++ b/verl/models/mcore/registry.py
@@ -22,6 +22,54 @@
 import torch
 import torch.nn as nn
 
+from .model_forward import gptmodel_forward_no_padding, model_forward_gen
+from .model_forward_fused import fused_forward_model_gen
+
+
+class SupportedVLM(Enum):
+    QWEN2_5_VL = "Qwen2_5_VLForConditionalGeneration"
+    QWEN3_MOE_VL = "Qwen3VLMoeForConditionalGeneration"
+    QWEN3_VL = "Qwen3VLForConditionalGeneration"
+
+
+def get_mcore_forward_fn(hf_config) -> Callable:
+    """
+    Get the forward function for given model architecture.
+    """
+    assert len(hf_config.architectures) == 1, "Only one architecture is supported for now"
+    if hf_config.architectures[0] in SupportedVLM:
+        return model_forward_gen(True)
+    else:
+        # default to language model
+        return model_forward_gen(False)
+
+
+def get_mcore_forward_no_padding_fn(hf_config) -> Callable:
+    """
+    Get the forward function for given model architecture.
+    """
+    assert len(hf_config.architectures) == 1, "Only one architecture is supported for now"
+    return gptmodel_forward_no_padding
+
+
+def get_mcore_forward_fused_fn(hf_config) -> Callable:
+    """
+    Get the forward function for given model architecture.
+    """
+    assert len(hf_config.architectures) == 1, "Only one architecture is supported for now"
+    if hf_config.architectures[0] in SupportedVLM:
+        return fused_forward_model_gen(True)
+    else:
+        # default to language model
+        return fused_forward_model_gen(False)
+
+
+# ruff: noqa
+
+########################################################
+# below is the deprecated code
+########################################################
+
 from .config_converter import (
     PretrainedConfig,
     TransformerConfig,
@@ -33,8 +81,6 @@
     hf_to_mcore_config_qwen2moe,
     hf_to_mcore_config_qwen3moe,
 )
-from .model_forward import gptmodel_forward_no_padding, model_forward_gen
-from .model_forward_fused import fused_forward_model_gen
 from .model_initializer import (
     BaseModelInitializer,
     DeepseekV3Model,
@@ -239,33 +285,6 @@ def init_mcore_model(
     )
 
 
-def get_mcore_forward_fn(hf_config: PretrainedConfig) -> Callable:
-    """
-    Get the forward function for given model architecture.
-    """
-    assert len(hf_config.architectures) == 1, "Only one architecture is supported for now"
-    model = get_supported_model(hf_config.architectures[0])
-    return MODEL_FORWARD_REGISTRY[model]
-
-
-def get_mcore_forward_no_padding_fn(hf_config: PretrainedConfig) -> Callable:
-    """
-    Get the forward function for given model architecture.
-    """
-    assert len(hf_config.architectures) == 1, "Only one architecture is supported for now"
-    model = get_supported_model(hf_config.architectures[0])
-    return MODEL_FORWARD_NOPAD_REGISTRY[model]
-
-
-def get_mcore_forward_fused_fn(hf_config: PretrainedConfig) -> Callable:
-    """
-    Get the forward function for given model architecture.
-    """
-    assert len(hf_config.architectures) == 1, "Only one architecture is supported for now"
-    model = get_supported_model(hf_config.architectures[0])
-    return MODEL_FORWARD_FUSED_REGISTRY[model]
-
-
 def get_mcore_weight_converter(hf_config: PretrainedConfig, dtype: torch.dtype) -> Callable:
     """
     Get the weight converter for given model architecture.
diff --git a/verl/utils/megatron_utils.py b/verl/utils/megatron_utils.py
index c319910d855..13c63ebdce0 100644
--- a/verl/utils/megatron_utils.py
+++ b/verl/utils/megatron_utils.py
@@ -1220,3 +1220,11 @@ def register_megatron_training_hooks(model: list[torch.nn.Module], optimizer):
             config.param_sync_func = [model_chunk.start_param_sync for model_chunk in model]
             if len(model) == 1:
                 config.param_sync_func = config.param_sync_func[0]
+
+
+def mapping_string_to_attn_backend(args: dict) -> dict:
+    if "attention_backend" in args and isinstance(args["attention_backend"], str):
+        from megatron.core.transformer.enums import AttnBackend
+
+        args["attention_backend"] = AttnBackend[args["attention_backend"]]
+    return args
diff --git a/verl/workers/engine/megatron/transformer_impl.py b/verl/workers/engine/megatron/transformer_impl.py
index 00292a73c2e..1a1756012d8 100644
--- a/verl/workers/engine/megatron/transformer_impl.py
+++ b/verl/workers/engine/megatron/transformer_impl.py
@@ -41,13 +41,11 @@
     load_megatron_optimizer,
     offload_megatron_model_to_cpu,
     offload_megatron_optimizer,
-    per_tensor_generator,
     register_megatron_training_hooks,
 )
 from verl.utils.model import (
     extract_multi_modal_inputs_tensordict,
     load_mcore_dist_weights,
-    load_megatron_gptmodel_weights,
 )
 from verl.workers.config import HFModelConfig, McoreEngineConfig, McoreOptimizerConfig
 
@@ -76,7 +74,7 @@ def __init__(
         self.engine_config = engine_config
         self.optimizer_config = optimizer_config
         self.checkpoint_config = checkpoint_config
-
+        assert self.engine_config.use_mbridge, "use_mbridge must be True"
         self._init_device_mesh()
 
         set_random_seed(seed=self.engine_config.seed)
@@ -110,70 +108,62 @@ def _init_device_mesh(self):
         )
 
     def _build_tf_config(self):
-        from verl.models.mcore import hf_to_mcore_config
-        from verl.models.mcore.config_converter import mapping_string_to_attn_backend
+        from verl.utils.megatron_utils import mapping_string_to_attn_backend
         from verl.utils.torch_dtypes import PrecisionType
 
         self.param_dtype = PrecisionType.to_dtype(self.engine_config.dtype)
-        if self.param_dtype == torch.float16:
-            assert self.engine_config.use_mbridge, "fp16 mode requires use_mbridge to be True"
         self.dtype = PrecisionType.to_dtype(self.param_dtype)
 
         override_transformer_config = mapping_string_to_attn_backend({**self.engine_config.override_transformer_config})
 
-        use_mbridge = self.engine_config.use_mbridge
         self.provider = None
         self.vanilla_bridge = self.engine_config.vanilla_mbridge
-        if use_mbridge:
-            if self.vanilla_bridge:
-                from verl.models.mcore.mbridge import AutoBridge
-
-                bridge = AutoBridge.from_config(self.model_config.hf_config, dtype=self.param_dtype)
-                bridge.set_extra_args(**override_transformer_config)
-                tf_config = bridge.config
-                tf_config.fp16 = self.param_dtype == torch.float16
-                tf_config.bf16 = self.param_dtype == torch.bfloat16
-            else:
-                from verl.models.mcore.bridge import AutoBridge
-
-                # Use Megatron-Bridge to convert HF config to Megatron config
-                bridge = AutoBridge.from_hf_pretrained(
-                    self.model_config.local_path, trust_remote_code=self.model_config.trust_remote_code
-                )
-                # Get Megatron provider and configure it
-                provider = bridge.to_megatron_provider(load_weights=False)
-
-                # In case of invalid overrides, we need to make sure some critical params are set correctly
-                provider.params_dtype = self.param_dtype
-
-                # Pass distributed info
-                provider.tensor_model_parallel_size = self.engine_config.tensor_model_parallel_size
-                provider.pipeline_model_parallel_size = self.engine_config.pipeline_model_parallel_size
-                provider.expert_model_parallel_size = self.engine_config.expert_model_parallel_size
-                provider.expert_tensor_parallel_size = self.engine_config.expert_tensor_parallel_size
-                provider.virtual_pipeline_model_parallel_size = self.engine_config.virtual_pipeline_model_parallel_size
-                provider.context_parallel_size = self.engine_config.context_parallel_size
-                provider.sequence_parallel = self.engine_config.sequence_parallel
-
-                # Match verl implementation (need variable_seq_lengths)
-                from megatron.core.transformer.enums import AttnBackend
-
-                provider.attention_backend = AttnBackend.flash
-                provider.variable_seq_lengths = True
-                provider.moe_token_dispatcher_type = "alltoall"
-                provider.moe_router_load_balancing_type = "none"
-
-                # Apply transformer config overrides
-                for key, value in override_transformer_config.items():
-                    setattr(provider, key, value)
-
-                provider.finalize()
-                self.provider = provider
-                tf_config = None  # Will be set after model creation
-            self.bridge = bridge
+        if self.vanilla_bridge:
+            from verl.models.mcore.mbridge import AutoBridge
+
+            bridge = AutoBridge.from_config(self.model_config.hf_config, dtype=self.param_dtype)
+            bridge.set_extra_args(**override_transformer_config)
+            tf_config = bridge.config
+            tf_config.fp16 = self.param_dtype == torch.float16
+            tf_config.bf16 = self.param_dtype == torch.bfloat16
         else:
-            self.bridge = None
-            tf_config = hf_to_mcore_config(self.model_config.hf_config, self.dtype, **override_transformer_config)
+            from verl.models.mcore.bridge import AutoBridge
+
+            # Use Megatron-Bridge to convert HF config to Megatron config
+            bridge = AutoBridge.from_hf_pretrained(
+                self.model_config.local_path, trust_remote_code=self.model_config.trust_remote_code
+            )
+            # Get Megatron provider and configure it
+            provider = bridge.to_megatron_provider(load_weights=False)
+
+            # In case of invalid overrides, we need to make sure some critical params are set correctly
+            provider.params_dtype = self.param_dtype
+
+            # Pass distributed info
+            provider.tensor_model_parallel_size = self.engine_config.tensor_model_parallel_size
+            provider.pipeline_model_parallel_size = self.engine_config.pipeline_model_parallel_size
+            provider.expert_model_parallel_size = self.engine_config.expert_model_parallel_size
+            provider.expert_tensor_parallel_size = self.engine_config.expert_tensor_parallel_size
+            provider.virtual_pipeline_model_parallel_size = self.engine_config.virtual_pipeline_model_parallel_size
+            provider.context_parallel_size = self.engine_config.context_parallel_size
+            provider.sequence_parallel = self.engine_config.sequence_parallel
+
+            # Match verl implementation (need variable_seq_lengths)
+            from megatron.core.transformer.enums import AttnBackend
+
+            provider.attention_backend = AttnBackend.flash
+            provider.variable_seq_lengths = True
+            provider.moe_token_dispatcher_type = "alltoall"
+            provider.moe_router_load_balancing_type = "none"
+
+            # Apply transformer config overrides
+            for key, value in override_transformer_config.items():
+                setattr(provider, key, value)
+
+            provider.finalize()
+            self.provider = provider
+            tf_config = None  # Will be set after model creation
+        self.bridge = bridge
 
         if not self.bridge:
             self.weight_converter = get_mcore_weight_converter(self.model_config.hf_config, self.dtype)
@@ -232,28 +222,14 @@ def _build_megatron_module(self):
         if self.engine_config.use_dist_checkpointing:
             load_mcore_dist_weights(module, self.engine_config.dist_checkpointing_path, is_value_model=is_value_model)
         else:
-            if self.bridge is not None:
-                if self.vanilla_bridge:
-                    self.bridge.load_weights(module, self.model_config.local_path)
-                else:
-                    allowed_mismatched_params = []
-                    if self.is_value_model:
-                        allowed_mismatched_params = ["output_layer.weight"]
-                    self.bridge.load_hf_weights(
-                        module, self.model_config.local_path, allowed_mismatched_params=allowed_mismatched_params
-                    )
+            if self.vanilla_bridge:
+                self.bridge.load_weights(module, self.model_config.local_path)
             else:
-                # (vermouth1992) this is a workaround to be compatible with the old API
-                tmp_config = OmegaConf.create(
-                    {"model": {"path": self.model_config.local_path, "use_shm": self.model_config.use_shm}}
-                )
-
-                load_megatron_gptmodel_weights(
-                    tmp_config,
-                    self.model_config.hf_config,
-                    module,
-                    params_dtype=self.dtype,
-                    is_value_model=is_value_model,
+                allowed_mismatched_params = []
+                if self.is_value_model:
+                    allowed_mismatched_params = ["output_layer.weight"]
+                self.bridge.load_hf_weights(
+                    module, self.model_config.local_path, allowed_mismatched_params=allowed_mismatched_params
                 )
 
         if torch.distributed.get_rank() == 0:
@@ -562,16 +538,7 @@ def forward_backward_batch(self, data: TensorDict, loss_function: Callable, forw
     def get_per_tensor_param(self):
         if self._is_offload_param:
             load_megatron_model_to_gpu(self.module, load_grad=False)
-        if self.bridge is not None:
-            per_tensor_param = self.bridge.export_weights(self.module)
-        else:
-            per_tensor_param = per_tensor_generator(
-                self.module,
-                self.model_config.hf_config,
-                self.weight_converter,
-                self.tf_config,
-                self.layer_name_mapping,
-            )
+        per_tensor_param = self.bridge.export_weights(self.module)
         # TODO: support megatron LoRA
         return per_tensor_param, None
 

From da5ce3db524db5a773dff94ee5d5461e586adfe9 Mon Sep 17 00:00:00 2001
From: Yan Bai <bayan@nvidia.com>
Date: Mon, 15 Dec 2025 06:40:03 -0800
Subject: [PATCH 2/4] remove legacy CI

---
 .github/workflows/checkpoint_converter.yml    | 175 ------------------
 .../e2e_ppo_trainer_megatron_sglang.yml       |  10 -
 .../e2e_ppo_trainer_megatron_sglang_2.yml     |  32 ----
 .../e2e_ppo_trainer_megatron_vllm.yml         |   5 -
 .../e2e_ppo_trainer_megatron_vllm_2.yml       |  37 ----
 .github/workflows/model.yml                   |  30 ---
 docs/advance/checkpoint.rst                   |  28 +--
 docs/perf/best_practices.rst                  |   4 +
 8 files changed, 6 insertions(+), 315 deletions(-)
 delete mode 100644 .github/workflows/checkpoint_converter.yml

diff --git a/.github/workflows/checkpoint_converter.yml b/.github/workflows/checkpoint_converter.yml
deleted file mode 100644
index 4820497f79c..00000000000
--- a/.github/workflows/checkpoint_converter.yml
+++ /dev/null
@@ -1,175 +0,0 @@
-# # Tests layout
-
-# Each folder under tests/ corresponds to a test category for a sub-namespace in verl. For instance:
-# - `tests/trainer` for testing functionality related to `verl/trainer`
-# - `tests/models` for testing functionality related to `verl/models`
-# - ...
-
-# There are a few folders with `special_` prefix, created for special purposes:
-# - `special_distributed`: unit tests that must run with multiple GPUs
-# - `special_e2e`: end-to-end tests with training/generation scripts
-# - `special_npu`: tests for NPUs
-# - `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
-# - 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.
-
-# # Workflow layout
-
-# All CI tests are configured by yaml files in `.github/workflows/`. Here's an overview of all test configs:
-# 1. A list of always triggered CPU sanity tests: `check-pr-title.yml`, `secrets_scan.yml`, `check-pr-title,yml`, `pre-commit.yml`, `doc.yml`
-# 2. Some heavy multi-GPU unit tests, such as `model.yml`, `vllm.yml`, `sgl.yml`
-# 3. End-to-end tests: `e2e_*.yml`
-# 4. Unit tests
-#   - `cpu_unit_tests.yml`, run pytest on all scripts with file name pattern `tests/**/test_*_on_cpu.py`
-#   - `gpu_unit_tests.yml`, run pytest on all scripts with file without the `on_cpu.py` suffix.
-#   - Since cpu/gpu unit tests by default runs all tests under `tests`, please make sure tests are manually excluded in them when
-#     - new workflow yaml is added to `.github/workflows`
-#     - new tests are added to workflow mentioned in 2.
-
-name: checkpoint_converter
-# latest version: Megatron-LM core_v0.14.0 https://github.com/NVIDIA/Megatron-LM/tree/core_v0.14.0
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-      - v0.*
-  pull_request:
-    branches:
-      - main
-      - v0.*
-    paths:
-      - "**/*.py"
-      # Other entrypoints
-      - "!examples/**"
-      - "!tests/**"
-      - "!verl/trainer/main_*.py"
-      - "!verl/trainer/fsdp_sft_trainer.py"
-      # Recipes
-      - "!recipe/**"
-      # FSDP
-      - "!verl/workers/**/*dp_*.py"
-      # Entrypoints
-      - ".github/workflows/checkpoint_converter.yml"
-      - ".github/workflows/e2e_ppo_trainer_megatron.yml"
-      - "examples/data_preprocess/gsm8k.py"
-      - "tests/special_e2e/run_ppo_trainer_megatron.sh"
-      - "verl/trainer/main_ppo.py"
-      - "verl/trainer/config/ppo_megatron_trainer.yaml"
-
-# Cancel jobs on the same ref if a new one is triggered
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
-
-# Declare permissions just read content.
-permissions:
-  contents: read
-
-env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl055.dev2"
-  DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
-
-jobs:
-  setup:
-    if: github.repository_owner == 'volcengine'
-    runs-on: ubuntu-latest
-    outputs:
-      runner-label: ${{ steps.create-runner.outputs.runner-label }}
-      mlp-task-id: ${{ steps.create-runner.outputs.mlp-task-id }}
-    steps:
-      - uses: actions/checkout@v4
-      - id: create-runner
-        uses: volcengine/vemlp-github-runner@v1
-        with:
-          mode: "create"
-          faas-url: "${{ env.DYNAMIC_RUNNER_ENDPOINT }}"
-          mlp-image: "${{ env.IMAGE }}"
-
-  checkpoint_converter:
-    needs: setup
-    runs-on: [ "${{ needs.setup.outputs.runner-label || 'L20x8' }}" ]
-    timeout-minutes: 20 # Increase this timeout value as needed
-    env:
-      HTTP_PROXY: ${{ secrets.PROXY_HTTP }}
-      HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }}
-      NO_PROXY: "localhost,127.0.0.1"
-      HF_HUB_ENABLE_HF_TRANSFER: "0" # This is more stable
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-      - name: Install the current repository
-        run: |
-          pip3 install -e .[test]
-#      - name: Download Model to Use
-#        run: |
-#          huggingface-cli download Qwen/Qwen2.5-0.5B --local-dir ${HOME}/models/Qwen/Qwen2.5-0.5B
-#          huggingface-cli download deepseek-ai/deepseek-coder-1.3b-instruct --local-dir ${HOME}/models/deepseek-ai/deepseek-coder-1.3b-instruct
-#          export HF_HUB_OFFLINE=1
-      - name: Running Huggingface to Megatron dist_ckpt converter (Qwen/Qwen2.5-0.5B)
-        run: |
-          ray stop --force
-          python scripts/converter_hf_to_mcore.py --hf_model_path=${HOME}/models/Qwen/Qwen2.5-0.5B --output_path checkpoints/Qwen/Qwen2.5-0.5B --test
-      - name: Running Huggingface to Megatron dist_ckpt converter (deepseek-ai/deepseek-coder-1.3b-instruct)
-        run: |
-          ray stop --force
-          python scripts/converter_hf_to_mcore.py --hf_model_path=${HOME}/models/deepseek-ai/deepseek-coder-1.3b-instruct --output_path checkpoints/deepseek-ai/deepseek-coder-1.3b-instruct --test
-      - name: Clean up
-        run: |
-          rm -rf checkpoints
-
-  checkpoint_converter_large_moe_models:
-    needs: setup
-    runs-on: [ "${{ needs.setup.outputs.runner-label || 'L20x8' }}" ]
-    timeout-minutes: 30 # Increase this timeout value as needed
-    env:
-      HTTP_PROXY: ${{ secrets.PROXY_HTTP }}
-      HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }}
-      NO_PROXY: "localhost,127.0.0.1"
-      HF_HUB_ENABLE_HF_TRANSFER: "0" # This is more stable
-      HF_ENDPOINT: "https://hf-mirror.com"
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-      - name: Install the current repository
-        run: |
-          pip3 install -e .[test]
-#      - name: Download Model to Use
-#        run: |
-#          huggingface-cli download Qwen/Qwen1.5-MoE-A2.7B-Chat --local-dir ${HOME}/models/Qwen/Qwen1.5-MoE-A2.7B-Chat
-#          export HF_HUB_OFFLINE=1
-      - name: Running Huggingface to Megatron dist_ckpt CPU converter (Qwen/Qwen1.5-MoE-A2.7B-Chat)
-        run: |
-          ray stop --force
-          python scripts/converter_hf_to_mcore.py --hf_model_path=${HOME}/models/Qwen/Qwen1.5-MoE-A2.7B-Chat --output_path checkpoints/Qwen/Qwen1.5-MoE-A2.7B-Chat --use_cpu_initialization
-      - name: Running distributed Huggingface to Megatron dist_ckpt CPU converter (Qwen/Qwen1.5-MoE-A2.7B-Chat)
-        run: |
-          ray stop --force
-          torchrun --nproc_per_node 8 --nnodes 1 scripts/converter_hf_to_mcore.py --hf_model_path=${HOME}/models/Qwen/Qwen1.5-MoE-A2.7B-Chat --output_path checkpoints/Qwen/Qwen1.5-MoE-A2.7B-Chat_dist --use_cpu_initialization
-      - name: clean up
-        run: |
-          rm -rf checkpoints
-
-  cleanup:
-    runs-on: ubuntu-latest
-    needs:
-      [
-        setup,
-        checkpoint_converter,
-        checkpoint_converter_large_moe_models
-      ]
-    if: always()
-    steps:
-      - id: destroy-runner
-        uses: volcengine/vemlp-github-runner@v1
-        with:
-          mode: "destroy"
-          faas-url: "${{ env.DYNAMIC_RUNNER_ENDPOINT }}"
-          mlp-task-id: "${{ needs.setup.outputs.mlp-task-id }}"
\ No newline at end of file
diff --git a/.github/workflows/e2e_ppo_trainer_megatron_sglang.yml b/.github/workflows/e2e_ppo_trainer_megatron_sglang.yml
index ccdc7c9c15d..df049bb0871 100644
--- a/.github/workflows/e2e_ppo_trainer_megatron_sglang.yml
+++ b/.github/workflows/e2e_ppo_trainer_megatron_sglang.yml
@@ -136,11 +136,6 @@ jobs:
           export VLLM_USE_V1=1
           ray start --head
           ENGINE=sglang MODE=async RESUME_MODE=auto MODEL_ID=deepseek-ai/deepseek-coder-1.3b-instruct TOTAL_TRAIN_STEPS=2 bash tests/special_e2e/run_ppo_trainer_megatron.sh
-      - name: Test Megatron checkpoints merging function (DeepSeek Actor and Critic)
-        run: |
-          exp_name="deepseek-coder-1.3b-instruct-megatron-gsm8k-minimal"
-          python -m verl.model_merger test --backend megatron --local_dir checkpoints/verl-test/${exp_name}/global_step_1/actor --test_hf_dir checkpoints/verl-test/${exp_name}/global_step_1/actor/huggingface
-          python -m verl.model_merger test --backend megatron --is-value-model --local_dir checkpoints/verl-test/${exp_name}/global_step_1/critic --test_hf_dir checkpoints/verl-test/${exp_name}/global_step_1/critic/huggingface
       - name: Profiling GRPO GSM8K E2E training tests with 3D parallelism on 8 L20 GPUs with Megatron (Deepseek)
         run: |
           ray stop --force
@@ -181,11 +176,6 @@ jobs:
         run: |
           ray stop --force
           ALL_OFFLOAD=True VAL_BEFORE_TRAIN=True TEST_FREQ=1 SAVE_FREQ=1 LR_WARMUP_STEPS=1 TOTAL_TRAIN_STEPS=2 MODEL_ID=Qwen/Qwen3-0.6B bash tests/special_e2e/run_ppo_trainer_megatron.sh
-      - name: Test Megatron checkpoints merging function (Qwen3 Actor and Critic)
-        run: |
-          exp_name="qwen3-0.6b-megatron-gsm8k-minimal"
-          python -m verl.model_merger test --backend megatron --tie-word-embedding --local_dir checkpoints/verl-test/${exp_name}/global_step_1/actor --test_hf_dir checkpoints/verl-test/${exp_name}/global_step_1/actor/huggingface
-          python -m verl.model_merger test --backend megatron --is-value-model --local_dir checkpoints/verl-test/${exp_name}/global_step_1/critic --test_hf_dir checkpoints/verl-test/${exp_name}/global_step_1/critic/huggingface
       - name: Running GSM8K E2E training tests with 3D parallelism on 8 L20 GPUs with FP8 rollout
         run: |
           ray stop --force
diff --git a/.github/workflows/e2e_ppo_trainer_megatron_sglang_2.yml b/.github/workflows/e2e_ppo_trainer_megatron_sglang_2.yml
index ccc503b0d58..e738fde2f8b 100644
--- a/.github/workflows/e2e_ppo_trainer_megatron_sglang_2.yml
+++ b/.github/workflows/e2e_ppo_trainer_megatron_sglang_2.yml
@@ -105,37 +105,6 @@ jobs:
           faas-url: "${{ env.DYNAMIC_RUNNER_ENDPOINT }}"
           mlp-image: "${{ env.IMAGE }}"
 
-  e2e_ppo_trainer_megatron-qwen2_5vl-3b:
-    needs: setup
-    runs-on: ["${{ needs.setup.outputs.runner-label || 'L20x8' }}"]
-    timeout-minutes: 60 # Increase this timeout value as needed
-    env:
-      HTTP_PROXY: ${{ secrets.PROXY_HTTP }}
-      HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }}
-      NO_PROXY: "localhost,127.0.0.1,hf-mirror.com"
-      HF_ENDPOINT: "https://hf-mirror.com"
-      HF_HUB_ENABLE_HF_TRANSFER: "0" # This is more stable
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-      - name: Install the current repository
-        run: |
-          pip3 install --no-deps -e .[test]
-      - name: Prepare Geo3k dataset
-        run: |
-          python3 examples/data_preprocess/geo3k.py --local_dataset_path ${HOME}/models/hf_data/hiyouga/geometry3k/
-      - name: Prepare dist_ckpt of Qwen2.5-VL-3B, only supports dist_ckpt
-        run: |
-          python3 scripts/converter_hf_to_mcore.py --hf_model_path ${HOME}/models/Qwen/Qwen2.5-VL-3B-Instruct --output_path checkpoints/verl-test/qwen2.5-vl-3b-megatron
-      - name: Running Geo3k E2E training tests with 3D parallelism on 8 L20 GPUs with Megatron (Qwen)
-        run: |
-          ray stop --force
-          ENGINE=sglang ROLLOUT_MODE=async TRAIN_FILES=${HOME}/data/geo3k/train.parquet VAL_FILES=${HOME}/data/geo3k/test.parquet MAX_PROMPT_LENGTH=1024 MAX_RESPONSE_LENGTH=2048  MODEL_ID=Qwen/Qwen2.5-VL-3B-Instruct ADV_ESTIMATOR=grpo USE_DYNAMIC_BSZ=False SKIP_SAVE_HF_MODEL=1 COMMON_PP=4 COMMON_VPP=null COMMON_CP=1 COMMON_TP=2 USE_DIST_CKPT=true DIST_CKPT_PATH=checkpoints/verl-test/qwen2.5-vl-3b-megatron bash tests/special_e2e/run_ppo_trainer_megatron.sh
-      - name: clean up
-        run: |
-          rm -rf checkpoints
-
   e2e_ppo_trainer_fsdp_sglang:
     needs: setup
     runs-on: [ "${{ needs.setup.outputs.runner-label || 'L20x8' }}" ]
@@ -221,7 +190,6 @@ jobs:
     needs:
       [
         setup,
-        e2e_ppo_trainer_megatron-qwen2_5vl-3b,
         e2e_ppo_trainer_fsdp-qwen2_5vl-3b,
         e2e_ppo_trainer_fsdp_sglang,
       ]
diff --git a/.github/workflows/e2e_ppo_trainer_megatron_vllm.yml b/.github/workflows/e2e_ppo_trainer_megatron_vllm.yml
index 35a50aae0ea..0fef97c8d1c 100644
--- a/.github/workflows/e2e_ppo_trainer_megatron_vllm.yml
+++ b/.github/workflows/e2e_ppo_trainer_megatron_vllm.yml
@@ -186,11 +186,6 @@ jobs:
         run: |
           ray stop --force
           ALL_OFFLOAD=True VAL_BEFORE_TRAIN=True TEST_FREQ=1 SAVE_FREQ=1 LR_WARMUP_STEPS=1 TOTAL_TRAIN_STEPS=2 MODEL_ID=Qwen/Qwen3-0.6B bash tests/special_e2e/run_ppo_trainer_megatron.sh
-      - name: Test Megatron checkpoints merging function (Qwen3 Actor and Critic)
-        run: |
-          exp_name="qwen3-0.6b-megatron-gsm8k-minimal"
-          python -m verl.model_merger test --backend megatron --tie-word-embedding --local_dir checkpoints/verl-test/${exp_name}/global_step_1/actor --test_hf_dir checkpoints/verl-test/${exp_name}/global_step_1/actor/huggingface
-          python -m verl.model_merger test --backend megatron --is-value-model --local_dir checkpoints/verl-test/${exp_name}/global_step_1/critic --test_hf_dir checkpoints/verl-test/${exp_name}/global_step_1/critic/huggingface
       - name: Running GSM8K E2E training tests with 3D parallelism on 8 L20 GPUs with FP8 rollout
         run: |
           ray stop --force
diff --git a/.github/workflows/e2e_ppo_trainer_megatron_vllm_2.yml b/.github/workflows/e2e_ppo_trainer_megatron_vllm_2.yml
index fb3e73ed02d..df99cc54e41 100644
--- a/.github/workflows/e2e_ppo_trainer_megatron_vllm_2.yml
+++ b/.github/workflows/e2e_ppo_trainer_megatron_vllm_2.yml
@@ -153,42 +153,6 @@ jobs:
         run: |
           rm -rf checkpoints
 
-  e2e_ppo_trainer_megatron-qwen2_5vl-3b:
-    needs: setup
-    runs-on: ["${{ needs.setup.outputs.runner-label || 'L20x8' }}"]
-    timeout-minutes: 60 # Increase this timeout value as needed
-    env:
-      HTTP_PROXY: ${{ secrets.PROXY_HTTP }}
-      HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }}
-      NO_PROXY: "localhost,127.0.0.1,hf-mirror.com"
-      HF_ENDPOINT: "https://hf-mirror.com"
-      HF_HUB_ENABLE_HF_TRANSFER: "0" # This is more stable
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-      - name: Install the current repository
-        run: |
-          pip3 install --no-deps -e .[test]
-          pip3 install transformers==$TRANSFORMERS_VERSION
-      - name: Prepare Geo3k dataset
-        run: |
-          python3 examples/data_preprocess/geo3k.py --local_dataset_path ${HOME}/models/hf_data/hiyouga/geometry3k/
-      - name: Prepare dist_ckpt of Qwen2.5-VL-3B, only supports dist_ckpt
-        run: |
-          python3 scripts/converter_hf_to_mcore.py --hf_model_path ${HOME}/models/Qwen/Qwen2.5-VL-3B-Instruct --output_path checkpoints/verl-test/qwen2.5-vl-3b-megatron
-      - name: Running Geo3k E2E training tests with 3D parallelism on 8 L20 GPUs with Megatron (Qwen)
-        run: |
-          ray stop --force
-          TRAIN_FILES=${HOME}/data/geo3k/train.parquet VAL_FILES=${HOME}/data/geo3k/test.parquet \
-          MAX_PROMPT_LENGTH=1024 MAX_RESPONSE_LENGTH=2048  MODEL_ID=Qwen/Qwen2.5-VL-3B-Instruct ADV_ESTIMATOR=grpo \
-          USE_DYNAMIC_BSZ=False USE_FUSED_KERNELS=True SKIP_SAVE_HF_MODEL=1 \
-          COMMON_PP=4 COMMON_VPP=null COMMON_CP=1 COMMON_TP=2 USE_DIST_CKPT=true \
-          DIST_CKPT_PATH=checkpoints/verl-test/qwen2.5-vl-3b-megatron bash tests/special_e2e/run_ppo_trainer_megatron.sh
-      - name: clean up
-        run: |
-          rm -rf checkpoints
-
   e2e_ppo_trainer_fsdp_vllm:
     needs: setup
     runs-on: [ "${{ needs.setup.outputs.runner-label || 'L20x8' }}" ]
@@ -330,7 +294,6 @@ jobs:
       [
         setup,
         e2e_ppo_trainer_megatron-moe-expert-parallel,
-        e2e_ppo_trainer_megatron-qwen2_5vl-3b,
         e2e_ppo_trainer_fsdp-qwen2_5vl-3b,
         e2e_ppo_trainer_fsdp_vllm,
       ]
diff --git a/.github/workflows/model.yml b/.github/workflows/model.yml
index cab35a68d96..c9f1f2deac2 100644
--- a/.github/workflows/model.yml
+++ b/.github/workflows/model.yml
@@ -48,7 +48,6 @@ on:
       # Entrypoints
       - ".github/workflows/model.yml"
       - "tests/special_distributed/test_fsdp_ckpt.py"
-      - "tests/special_distributed/test_mcore_config_converter.py"
       - "tests/special_distributed/test_tensor_dict.py"
       - "tests/models/**"
       - "tests/special_distributed/run_all.sh"
@@ -144,34 +143,6 @@ jobs:
         run: |
           STRATEGY=fsdp2 torchrun --nproc_per_node=8 tests/special_distributed/test_fsdp_ckpt.py
 
-  mcore_config_converter:
-    needs: setup
-    runs-on: [ "${{ needs.setup.outputs.runner-label || 'L20x8' }}" ]
-    timeout-minutes: 20 # Increase this timeout value as needed
-    env:
-      HTTP_PROXY: ${{ secrets.PROXY_HTTP }}
-      HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }}
-      NO_PROXY: "localhost,127.0.0.1,hf-mirror.com"
-      HF_ENDPOINT: "https://hf-mirror.com"
-      HF_HUB_ENABLE_HF_TRANSFER: "0" # This is more stable
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-      - name: Install the current repository
-        run: |
-          pip3 install -e .[test]
-#      - name: Download model config files
-#        run: |
-#          hf download Qwen/Qwen2.5-7B config.json --local-dir $HOME/configs/Qwen/Qwen2.5-7B
-#          hf download Qwen/Qwen3-8B config.json --local-dir $HOME/configs/Qwen/Qwen3-8B
-#          hf download deepseek-ai/deepseek-coder-1.3b-instruct config.json --local-dir $HOME/configs/deepseek-ai/deepseek-coder-1.3b-instruct
-#          hf download Qwen/Qwen2-57B-A14B config.json --local-dir $HOME/configs/Qwen/Qwen2-57B-A14B
-#          hf download Qwen/Qwen3-30B-A3B config.json --local-dir $HOME/configs/Qwen/Qwen3-30B-A3B
-#          hf download deepseek-ai/DeepSeek-V3-Base config.json --local-dir $HOME/configs/deepseek-ai/DeepSeek-V3-Base
-      - name: Running mcore config converter tests on 8 L20 GPUs
-        run: |
-          torchrun --nproc_per_node=8 tests/special_distributed/test_mcore_config_converter.py
 
   model_engine:
     needs: setup
@@ -206,7 +177,6 @@ jobs:
         setup,
         model_rmpad,
         model_rmpad_fsdp2_unstable,
-        mcore_config_converter,
         model_engine
       ]
     if: always()
diff --git a/docs/advance/checkpoint.rst b/docs/advance/checkpoint.rst
index 56bec4a75c3..9782af951d9 100644
--- a/docs/advance/checkpoint.rst
+++ b/docs/advance/checkpoint.rst
@@ -137,32 +137,8 @@ Current implementation use solution 2.
 HuggingFace to Megatron DistCheckpoint details
 ----------------------------------------------
 
-If your model is quite huge, we recommend you to use Megatron dist-checkpoint to load the model.
-Megatron dist-checkpoint supports loading with different kinds of model parallelism,
-and it is much faster than the original checkpoint loading.
-
-To convert original HuggingFace model to Megatron dist-checkpoint,
-you can use the ``scripts/converter_hf_to_mcore.py`` script. Large MoE models are temporarily supported with CPU initialization,
-which is a little slower. While we are working on a better solution to support large models.
-
-Example command to convert the model is as follows:
-
-.. code:: bash
-
-    python scripts/converter_hf_to_mcore.py \
-        --hf_model_path Qwen/Qwen1.5-MoE-A2.7B-Chat \
-        --output_path /mnt/disk/Qwen/Qwen1.5-MoE-A2.7B-Chat \
-        --use_cpu_initialization    # Only work for MoE models
-
-
-Example command to distributed convert the huge model like deepseekv3 671B is as follows:
-
-.. code:: bash
-
-    torchrun --nproc_per_node 1 --nnodes 8 --node_rank ${RANK} scripts/converter_hf_to_mcore.py \
-        --hf_model_path deepseek-ai/DeepSeek-V3 \
-        --output_path /mnt/disk/deepseek-ai/DeepSeek-V3 \
-        --use_cpu_initialization    # Only work for MoE models
+Through ``mbridge``, we can directly save the mcore model to huggingface format during training.
+No need to convert the model to Megatron dist-checkpoint format.
 
 Original Checkpoint Utils
 -------------------------
diff --git a/docs/perf/best_practices.rst b/docs/perf/best_practices.rst
index d7ff382c250..69d8286710a 100644
--- a/docs/perf/best_practices.rst
+++ b/docs/perf/best_practices.rst
@@ -110,6 +110,10 @@ Parameter Reference
     Path to the actor checkpoint in HuggingFace-compatible format.
   - ``actor_rollout_ref.actor.megatron.use_mbridge``:
     Enable mbridge format conversion when the model was trained with Megatron. Use the latest mbridge release: https://github.com/ISEEKYAN/mbridge.
+    Now it must be True.
+  - ``actor_rollout_ref.actor.megatron.vanilla_mbridge``:
+    If set to True, use mbridge, else use Megatron-Bridge https://github.com/NVIDIA-NeMo/Megatron-Bridge.
+    Now it is True by default. and it will defaultly be set to False in the future(v0.8).
 
 :math:`\pi`
   - ``actor_rollout_ref.rollout.name``:

From 9cab72bccc27d7d31ea4343ff584a37e9f137d8e Mon Sep 17 00:00:00 2001
From: Yan Bai <bayan@nvidia.com>
Date: Mon, 15 Dec 2025 21:33:06 -0800
Subject: [PATCH 3/4] fix CI

---
 verl/trainer/config/_generated_ppo_megatron_trainer.yaml | 4 ++--
 verl/trainer/config/engine/megatron.yaml                 | 2 +-
 verl/workers/config/engine.py                            | 2 +-
 3 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/verl/trainer/config/_generated_ppo_megatron_trainer.yaml b/verl/trainer/config/_generated_ppo_megatron_trainer.yaml
index 9e8e0e813bd..a117c0f332f 100644
--- a/verl/trainer/config/_generated_ppo_megatron_trainer.yaml
+++ b/verl/trainer/config/_generated_ppo_megatron_trainer.yaml
@@ -52,7 +52,7 @@ actor_rollout_ref:
         recompute_num_layers: null
         attention_backend: flash
       override_mcore_model_config: {}
-      use_mbridge: false
+      use_mbridge: true
       vanilla_mbridge: true
       use_remove_padding: true
       forward_only: false
@@ -433,7 +433,7 @@ critic:
       recompute_num_layers: null
       attention_backend: flash
     override_mcore_model_config: {}
-    use_mbridge: false
+    use_mbridge: true
     vanilla_mbridge: true
     use_remove_padding: true
     forward_only: false
diff --git a/verl/trainer/config/engine/megatron.yaml b/verl/trainer/config/engine/megatron.yaml
index 84601f5a3f5..b588a96c1b3 100644
--- a/verl/trainer/config/engine/megatron.yaml
+++ b/verl/trainer/config/engine/megatron.yaml
@@ -75,7 +75,7 @@ override_transformer_config:
 override_mcore_model_config: {}
 
 # oc.select: default val for ref.megatron.use_mbridge
-use_mbridge: False
+use_mbridge: True
 
 # oc.select: default val for ref.megatron.vanilla_mbridge
 vanilla_mbridge: True
diff --git a/verl/workers/config/engine.py b/verl/workers/config/engine.py
index a8799c35691..b645715489e 100644
--- a/verl/workers/config/engine.py
+++ b/verl/workers/config/engine.py
@@ -120,7 +120,7 @@ class McoreEngineConfig(EngineConfig):
     override_ddp_config: dict[str, Any] = field(default_factory=dict)
     override_transformer_config: dict[str, Any] = field(default_factory=dict)
     override_mcore_model_config: dict[str, Any] = field(default_factory=dict)
-    use_mbridge: bool = False
+    use_mbridge: bool = True
     vanilla_mbridge: bool = True
     strategy: str = "megatron"
 

From d9a02cc011e6dcf1bba360ea1154f4c00c25b7c7 Mon Sep 17 00:00:00 2001
From: Yan Bai <bayan@nvidia.com>
Date: Mon, 15 Dec 2025 22:31:10 -0800
Subject: [PATCH 4/4] fix CI, defaultl use_mbridge

---
 tests/trainer/config/legacy_ppo_megatron_trainer.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tests/trainer/config/legacy_ppo_megatron_trainer.yaml b/tests/trainer/config/legacy_ppo_megatron_trainer.yaml
index ea2a15d685e..06e2e94a662 100644
--- a/tests/trainer/config/legacy_ppo_megatron_trainer.yaml
+++ b/tests/trainer/config/legacy_ppo_megatron_trainer.yaml
@@ -111,7 +111,7 @@ actor_rollout_ref:
       dist_checkpointing_path: null
       seed: 42
       override_transformer_config: {} # additional transformer config like: num_layers_in_first(/last)_pipeline_stage
-      use_mbridge: False
+      use_mbridge: True
       vanilla_mbridge: True
     profile: # profile the actor model in `update_policy`
       use_profile: False # open it when you want to profile the actor model