TransferQueue
diff --git a/‎.github/workflows/e2e_ppo_grpo_trainer_trtllm.yml‎
Lines changed: 197 additions & 0 deletions b/‎.github/workflows/e2e_ppo_grpo_trainer_trtllm.yml‎
Lines changed: 197 additions & 0 deletions
diff --git a/‎docker/Dockerfile.stable.trtllm‎
Lines changed: 72 additions & 0 deletions b/‎docker/Dockerfile.stable.trtllm‎
Lines changed: 72 additions & 0 deletions
diff --git a/‎docs/index.rst‎
Lines changed: 1 addition & 0 deletions b/‎docs/index.rst‎
Lines changed: 1 addition & 0 deletions
diff --git a/‎docs/workers/trtllm_worker.rst‎
Lines changed: 65 additions & 0 deletions b/‎docs/workers/trtllm_worker.rst‎
Lines changed: 65 additions & 0 deletions
@@ -0,0 +1,197 @@
+# # 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: e2e_ppo_trainer_megatron_trtllm
+
+on:
+  # Trigger the workflow on push or pull request,
+  # but only for the main branch.
+  # For push, for now only anti-patterns are specified so it is more conservative
+  # and achieves higher coverage.
+  push:
+    branches:
+      - main
+      - v0.*
+    paths:
+      - "**/*.py"
+      # Other entrypoints
+      - "!verl/trainer/fsdp_sft_trainer.py"
+      # Recipes
+      - "!recipe/**"
+      # FSDP
+      - "!verl/workers/**/*dp_*.py"
+  pull_request:
+    branches:
+      - main
+      - v0.*
+    paths:
+      - "**/*.py"
+      # Other entrypoints
+      - "!docker/**"
+      # Docs
+      - "!**/*.md"
+      - "!docs/**"
+      - "!examples/**"
+      - "!tests/**"
+      - "!verl/trainer/main_*.py"
+      - "!verl/trainer/fsdp_sft_trainer.py"
+      # Recipes
+      - "!recipe/**"
+      # FSDP
+      - "!verl/workers/**/*dp_*.py"
+      # Entrypoints
+      - "verl/workers/rollout/trtllm_rollout/*"
+      - ".github/workflows/e2e_ppo_grpo_trainer_trtllm"
+      - "examples/data_preprocess/gsm8k.py"
+      - "examples/data_preprocess/geo3k.py"
+      # add back when ppo flow is ready
+      # - "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:trtllm1.2.0rc6"
+  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: "${{ env.IMAGE }}"
+
+  e2e_grpo_trainer_fsdp-qwen2:
+    needs: setup
+    runs-on: ["${{ 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-mirror.com"
+      HF_ENDPOINT: "https://hf-mirror.com"
+      HF_HUB_ENABLE_HF_TRANSFER: "0" # This is more stable
+      RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES: "1"
+    steps:
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          fetch-depth: 0
+      - name: Install the current repository
+        run: |
+          pip3 install -r requirements-test.txt
+          pip3 install --no-deps -e .
+      - name: Prepare GSM8K dataset
+        run: |
+          python3 examples/data_preprocess/gsm8k.py --local_dataset_path ${HOME}/models/hf_data/gsm8k --local_save_dir ${PWD}/data/gsm8k
+      - name: Running GSM8K E2E training tests with FSDP on 8 L20 GPUs (Qwen)
+        run: |
+          ray stop --force
+          DATADIR=${HOME}/data \
+            bash examples/grpo_trainer/run_qwen2-7b_math_trtllm.sh 2 \
+            trainer.total_training_steps=1 \
+            data.train_files="['${PWD}/data/gsm8k/train.parquet']" \
+            data.val_files="['${PWD}/data/gsm8k/test.parquet']" \
+            trainer.logger='["console"]' \
+            actor_rollout_ref.model.path="${HOME}/models/Qwen/Qwen2.5-0.5B-Instruct"
+      - name: clean up
+        run: |
+          rm -rf checkpoints
+
+  e2e_grpo_trainer_megatron-qwen2:
+    needs: setup
+    runs-on: ["${{ 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-mirror.com"
+      HF_ENDPOINT: "https://hf-mirror.com"
+      HF_HUB_ENABLE_HF_TRANSFER: "0" # This is more stable
+      RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES: "1"
+    steps:
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          fetch-depth: 0
+      - name: Install the current repository
+        run: |
+          pip3 install -r requirements-test.txt
+          pip3 install --no-deps -e .
+      - name: Prepare GSM8K dataset
+        run: |
+          python3 examples/data_preprocess/gsm8k.py --local_dataset_path ${HOME}/models/hf_data/gsm8k --local_save_dir ${PWD}/data/gsm8k
+      - name: Running GSM8K E2E training tests with 3D parallelism on 8 L20 GPUs with Megatron (Qwen)
+        run: |
+          ray stop --force
+          DATADIR=${HOME}/data \
+          ACTOR_TP=2 \
+            bash examples/grpo_trainer/run_qwen2-7b_math_megatron_trtllm.sh 2 \
+            trainer.total_training_steps=1 \
+            data.train_files="['${PWD}/data/gsm8k/train.parquet']" \
+            data.val_files="['${PWD}/data/gsm8k/test.parquet']" \
+            trainer.logger='["console"]' \
+            actor_rollout_ref.model.path="${HOME}/models/Qwen/Qwen2.5-0.5B-Instruct"
+      - name: clean up
+        run: |
+          rm -rf checkpoints
+
+  cleanup:
+    runs-on: ubuntu-latest
+    needs:
+      [
+        setup,
+        e2e_grpo_trainer_fsdp-qwen2,
+        e2e_grpo_trainer_megatron-qwen2,
+      ]
+    if: always()
+    steps:
+      - id: destroy-runner
+        uses: volcengine/vemlp-github-runner@v1
+        with:
+          mode: "destroy"
+          faas-url: "${{ env.DYNAMIC_RUNNER_ENDPOINT }}"
+          mlp-task-id: "${{ needs.setup.outputs.mlp-task-id }}"
@@ -0,0 +1,72 @@
+# Base image from NGC TensorRT-LLM, which includes a pre-installed TensorRT-LLM.
+# For available images, visit: https://nvidia.github.io/TensorRT-LLM/installation/containers.html
+# Use TRTLLM_BASE_IMAGE to specify the base image (default: release:1.2.0rc6)
+ARG TRTLLM_BASE_IMAGE=nvcr.io/nvidia/tensorrt-llm/release:1.2.0rc6
+FROM ${TRTLLM_BASE_IMAGE}
+
+
+# ==============================================================================
+# Install Megatron dependencies
+# ==============================================================================
+# DeepEP is required for IBGDA support.
+# Clone and build gdrcopy and deepep-nvshmem dependencies.
+WORKDIR /home/dpsk_a2a
+RUN git clone -b v2.3.1 https://github.com/NVIDIA/gdrcopy.git && \
+    git clone https://github.com/deepseek-ai/DeepEP.git && cd DeepEP && git checkout a84a248 && \
+    cd /home/dpsk_a2a && \
+    wget https://developer.nvidia.com/downloads/assets/secure/nvshmem/nvshmem_src_3.2.5-1.txz && \
+    tar -xvf nvshmem_src_3.2.5-1.txz && mv nvshmem_src deepep-nvshmem && \
+    cd deepep-nvshmem && git apply /home/dpsk_a2a/DeepEP/third-party/nvshmem.patch && \
+    sed -i '16i#include <getopt.h>' /home/dpsk_a2a/deepep-nvshmem/examples/moe_shuffle.cu && \
+    sed -i 's/CUDA_STANDARD 11/CUDA_STANDARD 17/g' /home/dpsk_a2a/deepep-nvshmem/src/CMakeLists.txt && \
+    # Cleanup downloaded archive
+    rm /home/dpsk_a2a/nvshmem_src_3.2.5-1.txz
+
+# Set environment variables
+ENV CUDA_HOME=/usr/local/cuda \
+    CPATH=/usr/local/mpi/include \
+    LD_LIBRARY_PATH=/usr/local/mpi/lib:/usr/local/x86_64-linux-gnu:$LD_LIBRARY_PATH \
+    GDRCOPY_HOME=/home/dpsk_a2a/gdrcopy
+
+# Build deepep-nvshmem
+WORKDIR /home/dpsk_a2a/deepep-nvshmem
+ARG CUDA_ARCHS="80;90;100"
+RUN NVSHMEM_SHMEM_SUPPORT=0 \
+    NVSHMEM_UCX_SUPPORT=0 \
+    NVSHMEM_USE_NCCL=0 \
+    NVSHMEM_MPI_SUPPORT=0 \
+    NVSHMEM_IBGDA_SUPPORT=1 \
+    NVSHMEM_PMIX_SUPPORT=0 \
+    NVSHMEM_TIMEOUT_DEVICE_POLLING=0 \
+    NVSHMEM_USE_GDRCOPY=1 \
+    NVSHMEM_BUILD_EXAMPLES=0 \
+    NVSHMEM_BUILD_TESTS=0 \
+    cmake -S . -B build/ -DCMAKE_INSTALL_PREFIX=/home/dpsk_a2a/deepep-nvshmem/install -DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCHS}" && \
+    cd build && make install -j && \
+    # Cleanup build directory
+    rm -rf /home/dpsk_a2a/deepep-nvshmem/build
+
+# Build deepep
+WORKDIR /home/dpsk_a2a/DeepEP
+ENV NVSHMEM_DIR=/home/dpsk_a2a/deepep-nvshmem/install
+RUN NVSHMEM_DIR=/home/dpsk_a2a/deepep-nvshmem/install python setup.py install
+
+# Install Python dependencies
+RUN pip3 install --no-cache-dir --no-deps trl && \
+    pip3 install --no-cache-dir nvtx matplotlib liger_kernel && \
+    pip install --no-cache-dir -U git+https://github.com/ISEEKYAN/mbridge.git && \
+    pip install --no-deps --no-cache-dir git+https://github.com/NVIDIA/Megatron-LM.git@core_v0.14.0rc7
+
+
+# ==============================================================================
+# Install verl dependencies
+# ==============================================================================
+RUN pip install git+https://github.com/volcengine/verl.git@v0.6.0
+RUN pip uninstall -y verl
+
+
+# ==============================================================================
+# Install a specific TensorRT-LLM on demand
+# ==============================================================================
+# Note: The NGC image already includes a pre-installed TensorRT-LLM, but you can install a specific version if needed.
+# Refer to https://nvidia.github.io/TensorRT-LLM/installation/index.html for more details.
@@ -90,6 +90,7 @@ verl is fast with:
    workers/fsdp_workers
    workers/megatron_workers
    workers/sglang_worker
+   workers/trtllm_worker
    workers/model_engine
 
 .. toctree::
 
@@ -0,0 +1,65 @@
+TensorRT-LLM Backend
+====================
+
+Last updated: 12/31/2025.
+
+**Authored By TensorRT-LLM Team**
+
+Introduction
+------------
+`TensorRT-LLM <https://github.com/NVIDIA/TensorRT-LLM>`_ is a high-performance LLM inference engine with state-of-the-art optimizations for NVIDIA GPUs.
+The verl integration of TensorRT-LLM is based on TensorRT-LLM's `Ray orchestrator <https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/ray_orchestrator>`_. This integration is in its early stage, with more features and optimizations to come.
+
+The TensorRT-LLM rollout engine primarily targets the colocated mode. Instead of relying purely on standard colocated mode, we adopted a mixed design combining aspects of the hybrid engine and colocated mode.
+
+Installation
+------------
+We provide ``docker/Dockerfile.stable.trtllm`` for building a docker image with TensorRT-LLM pre-installed. The verl integration is supported from ``nvcr.io/nvidia/tensorrt-llm/release:1.2.0rc6``, and you can choose other TensorRT-LLM versions via ``TRTLLM_BASE_IMAGE`` from the `NGC Catalog <https://catalog.ngc.nvidia.com/orgs/nvidia/teams/tensorrt-llm/containers/release>`_.
+
+Alternatively, refer to the `TensorRT-LLM installation guide <https://nvidia.github.io/TensorRT-LLM/installation/index.html>`_ for compatible environments if you want to build your own.
+
+Install verl with TensorRT-LLM:
+
+.. code-block:: bash
+
+    pip install --upgrade pip
+    pip install -e ".[trtllm]"
+
+.. note::
+
+    Using the TensorRT-LLM rollout requires setting the following environment variables before launching the Ray cluster. These have been included in all the example scripts:
+
+    .. code-block:: bash
+
+        # Clean all SLURM/MPI/PMIx env to avoid PMIx mismatch error.
+        for v in $(env | awk -F= '/^(PMI|PMIX|MPI|OMPI|SLURM)_/{print $1}'); do
+            unset "$v"
+        done
+
+        # Required for IPC UUID detection
+        export RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES=1
+
+Using TensorRT-LLM as the Rollout Engine for GRPO
+-------------------------------------------------
+
+We provide the following GRPO recipe scripts for you to test the performance and accuracy curve of TensorRT-LLM as the rollout engine:
+
+.. code-block:: bash
+
+    ## For FSDP training engine
+    bash examples/grpo_trainer/run_qwen2-7b_math_trtllm.sh
+    ## For Megatron-Core training engine
+    bash examples/grpo_trainer/run_qwen2-7b_math_megatron_trtllm.sh
+
+Using TensorRT-LLM as the Rollout Engine for DAPO
+-------------------------------------------------
+
+We provide a DAPO recipe script ``recipe/dapo/test_dapo_7b_math_trtllm.sh``.
+
+.. code-block:: bash
+
+    ## For FSDP training engine
+    bash recipe/dapo/test_dapo_7b_math_trtllm.sh
+    ## For Megatron-Core training engine
+    TRAIN_ENGINE=megatron bash recipe/dapo/test_dapo_7b_math_trtllm.sh
+