[docker] feat: update stable image to vllm==0.17.0, sglang==0.5.9 (verl-project#5542)

Begunner · Begunner · wuxibin89 · web-flow · commit 016c1d5a7a3f · 2026-03-12T16:36:09.000+08:00
### What does this PR do? > Update dockerfiles of vllm/sglang stable images, vllm from v0.12.0 to v0.17.0 and sglang from v0.5.6 to v0.5.9. ### Checklist Before Starting - [ ] Search for similar PRs. Paste at least one query link here: ... - [ ] Format the PR title as `[{modules}] {type}: {description}` (This will be checked by the CI) - `{modules}` include `fsdp`, `megatron`, `veomni`, `sglang`, `vllm`, `rollout`, `trainer`, `ci`, `training_utils`, `recipe`, `hardware`, `deployment`, `ray`, `worker`, `single_controller`, `misc`, `perf`, `model`, `algo`, `env`, `tool`, `ckpt`, `doc`, `data`, `cfg`, `reward`, `fully_async`, `one_step_off` - 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 > For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc. ### API and Usage Example > Demonstrate how the API changes if any, and provide usage example(s) if possible. ```python # Add code snippet or script demonstrating how to use this ``` ### Design & Code Changes > Demonstrate the high-level design if this PR is complex, and list the specific changes. ### Checklist Before Submitting > [!IMPORTANT] > Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review. - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl/blob/main/CONTRIBUTING.md). - [ ] 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` - [ ] Add / Update [the documentation](https://github.com/volcengine/verl/tree/main/docs). - [ ] 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).) - [ ] If your PR is related to the `recipe` submodule, please also update the reference to the submodule commit via `git submodule update --remote` or `cd recipe && git pull origin main`. --------- Co-authored-by: Begunner <went@bytedance.com> Co-authored-by: wuxibin <wuxibin@bytedance.com>
diff --git a/.github/workflows/README.md b/.github/workflows/README.md
@@ -31,7 +31,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "your vemlp image" # e.g. "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl056.latest"
+  IMAGE: "your vemlp image" # e.g. "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl059.dev2"
   DYNAMIC_RUNNER_URL: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner" # public veFaas api
 
 jobs:
diff --git a/.github/workflows/cpu_unit_tests.yml b/.github/workflows/cpu_unit_tests.yml
@@ -56,7 +56,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/e2e_fully_async_policy.yml b/.github/workflows/e2e_fully_async_policy.yml
@@ -80,7 +80,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
@@ -119,6 +119,7 @@ jobs:
         run: |
           pip3 install -r requirements-test.txt
           pip3 install --no-deps -e .
+          pip3 install cupy-cuda12x==13.6.0
       - name: Prepare GSM8K dataset
         run: |
           python3 examples/data_preprocess/gsm8k.py --local_dataset_path ${HOME}/models/hf_data/gsm8k
@@ -147,6 +148,7 @@ jobs:
         run: |
           pip3 install -r requirements-test.txt
           pip3 install --no-deps -e .
+          pip3 install cupy-cuda12x==13.6.0
       - name: Prepare GSM8K dataset
         run: |
           python3 examples/data_preprocess/gsm8k.py --local_dataset_path ${HOME}/models/hf_data/gsm8k
diff --git a/.github/workflows/e2e_one_step_off_policy.yml b/.github/workflows/e2e_one_step_off_policy.yml
@@ -80,7 +80,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
@@ -119,6 +119,7 @@ jobs:
         run: |
           pip3 install -r requirements-test.txt
           pip3 install --no-deps -e .
+          pip3 install cupy-cuda12x==13.6.0
       - name: Prepare GSM8K dataset
         run: |
           python3 examples/data_preprocess/gsm8k.py --local_dataset_path ${HOME}/models/hf_data/gsm8k
@@ -147,6 +148,7 @@ jobs:
         run: |
           pip3 install -r requirements-test.txt
           pip3 install --no-deps -e .
+          pip3 install cupy-cuda12x==13.6.0
       - name: Prepare GSM8K dataset
         run: |
           python3 examples/data_preprocess/gsm8k.py --local_dataset_path ${HOME}/models/hf_data/gsm8k
diff --git a/.github/workflows/e2e_ppo_trainer_megatron_sglang.yml b/.github/workflows/e2e_ppo_trainer_megatron_sglang.yml
@@ -86,7 +86,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl056.latest"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl059.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/e2e_ppo_trainer_megatron_sglang_2.yml b/.github/workflows/e2e_ppo_trainer_megatron_sglang_2.yml
@@ -86,7 +86,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl056.latest"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl059.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/e2e_ppo_trainer_megatron_vllm.yml b/.github/workflows/e2e_ppo_trainer_megatron_vllm.yml
@@ -87,7 +87,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/e2e_ppo_trainer_megatron_vllm_2.yml b/.github/workflows/e2e_ppo_trainer_megatron_vllm_2.yml
@@ -87,7 +87,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/e2e_ppo_trainer_veomni_vllm.yml b/.github/workflows/e2e_ppo_trainer_veomni_vllm.yml
@@ -81,7 +81,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/e2e_sft_llm.yml b/.github/workflows/e2e_sft_llm.yml
@@ -69,7 +69,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl056.latest"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl059.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/e2e_sft_vlm.yml b/.github/workflows/e2e_sft_vlm.yml
@@ -69,7 +69,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl056.latest"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl059.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/gpu_unit_tests.yml b/.github/workflows/gpu_unit_tests.yml
@@ -70,7 +70,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl056.latest"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl059.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/model.yml b/.github/workflows/model.yml
@@ -62,7 +62,7 @@ concurrency:
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/reward_model_sglang.yml b/.github/workflows/reward_model_sglang.yml
@@ -59,7 +59,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl056.latest"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl059.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/reward_model_vllm.yml b/.github/workflows/reward_model_vllm.yml
@@ -59,7 +59,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/sgl.yml b/.github/workflows/sgl.yml
@@ -75,7 +75,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl056.latest"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:sgl059.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
diff --git a/.github/workflows/vllm.yml b/.github/workflows/vllm.yml
@@ -72,7 +72,7 @@ permissions:
   contents: read
 
 env:
-  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm012.dev3"
+  IMAGE: "verl-ci-cn-beijing.cr.volces.com/verlai/verl:vllm017.dev2"
   DYNAMIC_RUNNER_ENDPOINT: "https://sd10g3clalm04ug7alq90.apigateway-cn-beijing.volceapi.com/runner"
 
 jobs:
@@ -148,6 +148,7 @@ jobs:
           pip3 install -r requirements-test.txt
           pip3 install --no-deps -e .
           pip3 install --upgrade "transformers<5.0"
+          pip3 install cupy-cuda12x==13.6.0
       - name: Test vLLM ServerAdapter with Checkpoint Engine (NCCL)
         run: |
           ROLLOUT_NAME=vllm pytest -svvv tests/checkpoint_engine/test_special_server_adapter.py
diff --git a/docker/Dockerfile.stable.sglang b/docker/Dockerfile.stable.sglang
@@ -1,20 +1,22 @@
-FROM lmsysorg/sglang:v0.5.6.post2
+# sgl059
+
+FROM lmsysorg/sglang:v0.5.9
 
 ARG PIP_NO_CACHE_DIR=1
 
 RUN pip install pybind11
 
 RUN pip install nvidia-mathdx
 
-RUN MAX_JOBS=128 pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" git+https://github.com/NVIDIA/apex.git
+RUN MAX_JOBS=128 pip install -v --disable-pip-version-check --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" git+https://github.com/NVIDIA/apex.git
 
-RUN export NVTE_FRAMEWORK=pytorch && MAX_JOBS=128 NVTE_BUILD_THREADS_PER_JOB=4 pip3 install --resume-retries 999 --no-cache-dir --no-build-isolation git+https://github.com/NVIDIA/TransformerEngine.git@release_v2.10
+RUN export NVTE_FRAMEWORK=pytorch && MAX_JOBS=128 NVTE_BUILD_THREADS_PER_JOB=4 pip3 install --resume-retries 999 --no-build-isolation git+https://github.com/NVIDIA/TransformerEngine.git@release_v2.12
 
-RUN pip install --upgrade --no-cache-dir transformers tokenizers
+# RUN pip install --upgrade transformers tokenizers
 
-RUN pip install codetiming tensordict mathruler pylatexenc qwen_vl_utils
+RUN pip install codetiming mathruler pylatexenc qwen_vl_utils cachetools pytest-asyncio
 
-RUN pip install --no-cache-dir --no-build-isolation flash_attn==2.8.1
+RUN pip install --no-build-isolation flash_attn==2.8.3
 
 RUN NSIGHT_VERSION=2025.6.1_2025.6.1.190-1_$(if [ "$(uname -m)" = "aarch64" ]; then echo "arm64"; else echo "amd64"; fi) && \
     wget https://developer.nvidia.com/downloads/assets/tools/secure/nsight-systems/2025_6/nsight-systems-${NSIGHT_VERSION}.deb && \
@@ -28,17 +30,24 @@ RUN NSIGHT_VERSION=2025.6.1_2025.6.1.190-1_$(if [ "$(uname -m)" = "aarch64" ]; t
 
 # sglang image has already installed DeepEP
 
-RUN pip3 install --no-cache-dir --no-deps trl
+RUN pip3 install --no-deps trl==0.27.0
 
 RUN pip3 install nvtx matplotlib liger_kernel
 
 RUN pip install -U git+https://github.com/ISEEKYAN/mbridge.git
 
-RUN pip install --no-deps --no-cache-dir git+https://github.com/NVIDIA/Megatron-LM.git@core_v0.15.0
+RUN pip install --no-deps git+https://github.com/NVIDIA/Megatron-LM.git@core_v0.16.0
 
-RUN pip install git+https://github.com/volcengine/verl.git@v0.6.0 && \
+RUN pip install git+https://github.com/volcengine/verl.git@v0.7.0 && \
     pip uninstall -y verl
 
-# avoid cudnn9.10&torch2.9.1 bug
-RUN sed -i '/nvidia-cudnn-cu12/d' /usr/local/lib/python3.12/dist-packages/torch-2.9.1+cu129.dist-info/METADATA
-RUN pip install --no-deps --force-reinstall nvidia-cudnn-cu12==9.16.0.29
+RUN sed -i '/nvidia-cudnn-cu12/d' /usr/local/lib/python3.12/dist-packages/torch-2.9.1+cu129.dist-info/METADATA && \
+    pip install --no-deps --force-reinstall nvidia-cudnn-cu12==9.16.0.29
+
+# for packages compiled from source code
+RUN apt-get update && \
+    apt-get install -y --allow-downgrades --allow-change-held-packages \
+    libcudnn9-cuda-12=9.16.0.29-1 \
+    libcudnn9-dev-cuda-12=9.16.0.29-1 \
+    libcudnn9-headers-cuda-12=9.16.0.29-1 && \
+    rm -rf /var/lib/apt/lists/*
diff --git a/docker/Dockerfile.stable.vllm b/docker/Dockerfile.stable.vllm
@@ -1,3 +1,5 @@
+# vllm017
+
 FROM nvidia/cuda:12.9.1-devel-ubuntu22.04
 
 ARG DEBIAN_FRONTEND=noninteractive
@@ -12,7 +14,8 @@ RUN apt-get update && apt-get install -y \
     libnuma-dev \
     librdmacm-dev \
     numactl \
-    software-properties-common && \
+    software-properties-common \
+    vim && \
     add-apt-repository ppa:deadsnakes/ppa -y && \
     apt-get update && \
     apt-get install -y \
@@ -27,20 +30,9 @@ RUN wget https://bootstrap.pypa.io/get-pip.py && \
 RUN ln -sf /usr/bin/python3.12 /usr/bin/python3 && \
     ln -sf /usr/bin/python3.12 /usr/bin/python
 
-# use torch2.9.1+cudnn9.16 to avoid conflict
-RUN pip install torch==2.9.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu129
-RUN sed -i '/nvidia-cudnn-cu12/d' /usr/local/lib/python3.12/dist-packages/torch-2.9.1+cu129.dist-info/METADATA
-RUN pip install --no-deps --force-reinstall nvidia-cudnn-cu12==9.16.0.29
-
-# NOTE: This installs the `vllm` source code in `/vllm`.
-# This might break the (based)pyright type checking. To fix it, add `/vllm` to `extraPaths` in `pyrightconfig.json`.
-# c.f. https://docs.basedpyright.com/latest/configuration/config-files/
-RUN git clone --depth 1 -b v0.12.0 https://github.com/vllm-project/vllm.git && \
-    cd vllm && \
-    find requirements -name "*.txt" -print0 | xargs -0 sed -i '/torch/d' && \
-    pip install -r requirements/build.txt && \
-    pip install -e . --no-build-isolation --no-deps && \
-    pip install -r requirements/cuda.txt
+RUN pip install torch==2.10.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu129
+
+RUN pip install vllm==0.17.0
 
 RUN pip install pybind11
 
@@ -54,13 +46,16 @@ RUN pip install nvidia-mathdx
 
 RUN MAX_JOBS=128 pip install -v --disable-pip-version-check --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" git+https://github.com/NVIDIA/apex.git
 
-RUN export NVTE_FRAMEWORK=pytorch && MAX_JOBS=128 NVTE_BUILD_THREADS_PER_JOB=4 pip3 install --resume-retries 999 --no-build-isolation git+https://github.com/NVIDIA/TransformerEngine.git@release_v2.10
+RUN export NVTE_FRAMEWORK=pytorch && \
+    MAX_JOBS=128 \
+    NVTE_BUILD_THREADS_PER_JOB=4 \
+    pip3 install --resume-retries 999 --no-build-isolation git+https://github.com/NVIDIA/TransformerEngine.git@release_v2.12
 
-RUN pip install --upgrade transformers tokenizers
+# RUN pip install --upgrade transformers tokenizers
 
-RUN pip install codetiming tensordict mathruler pylatexenc qwen_vl_utils
+RUN pip install codetiming mathruler pylatexenc qwen_vl_utils cachetools pytest-asyncio
 
-RUN pip install flash_attn==2.8.1 --no-build-isolation
+RUN export FLASH_ATTENTION_FORCE_BUILD="TRUE" && MAX_JOBS=16 pip install --no-build-isolation flash_attn==2.8.3
 
 RUN NSIGHT_VERSION=2025.6.1_2025.6.1.190-1_$(if [ "$(uname -m)" = "aarch64" ]; then echo "arm64"; else echo "amd64"; fi) && \
     wget https://developer.nvidia.com/downloads/assets/tools/secure/nsight-systems/2025_6/nsight-systems-${NSIGHT_VERSION}.deb && \
@@ -85,25 +80,33 @@ RUN git clone -b v2.5.1 https://github.com/NVIDIA/gdrcopy.git && \
 
 ENV GDRCOPY_HOME=/usr/local
 
-RUN git clone -b v1.2.1 https://github.com/deepseek-ai/DeepEP.git && \
+RUN git clone -b hybrid-ep https://github.com/deepseek-ai/DeepEP.git && \
     export NVSHMEM_DIR=/usr/local/lib/python3.12/dist-packages/nvidia/nvshmem && \
     export LD_LIBRARY_PATH="${NVSHMEM_DIR}/lib:$LD_LIBRARY_PATH" && \
     export PATH="${NVSHMEM_DIR}/bin:$PATH" && \
     cd ${NVSHMEM_DIR}/lib && \
-    ln -s libnvshmem_host.so.3 libnvshmem_host.so && \
+    ln -sf libnvshmem_host.so.3 libnvshmem_host.so && \
     cd /home/dpsk_a2a/DeepEP && \
+    export CPATH=/usr/local/cuda/targets/x86_64-linux/include/cccl:$CPATH && \
     python setup.py install
 
-RUN pip3 install --no-deps trl
+RUN pip3 install --no-deps trl==0.27.0
 
 RUN pip3 install nvtx matplotlib liger_kernel
 
 RUN pip install -U git+https://github.com/ISEEKYAN/mbridge.git
 
-RUN pip install --no-deps git+https://github.com/NVIDIA/Megatron-LM.git@core_v0.15.0
+RUN pip install --no-deps git+https://github.com/NVIDIA/Megatron-LM.git@core_v0.16.0
 
-RUN pip install git+https://github.com/volcengine/verl.git@v0.6.0 && \
+RUN pip install git+https://github.com/volcengine/verl.git@v0.7.0 && \
     pip uninstall -y verl
 
 RUN apt-get update && apt-get install -y curl \
     && rm -rf /var/lib/apt/lists/*
+
+RUN apt-get update && \
+    apt-get install -y --allow-downgrades --allow-change-held-packages \
+    libcudnn9-cuda-12=9.16.0.29-1 \
+    libcudnn9-dev-cuda-12=9.16.0.29-1 \
+    libcudnn9-headers-cuda-12=9.16.0.29-1 && \
+    rm -rf /var/lib/apt/lists/*
diff --git a/docker/README.md b/docker/README.md
@@ -24,7 +24,7 @@ Latest docker file:
 - [Dockerfile.stable.vllm](https://github.com/volcengine/verl/blob/main/docker/Dockerfile.stable.vllm)
 - [Dockerfile.stable.sglang](https://github.com/volcengine/verl/blob/main/docker/Dockerfile.stable.sglang)
 
-All pre-built images are available in dockerhub: https://hub.docker.com/r/verlai/verl. For example, `verlai/verl:sgl056.latest`, `verlai/verl:vllm012.dev3`.
+All pre-built images are available in dockerhub: https://hub.docker.com/r/verlai/verl. For example, `verlai/verl:sgl059.latest`, `verlai/verl:vllm017.latest`.
 
 You can find the latest images used for development and ci in our github workflows:
 - [.github/workflows/vllm.yml](https://github.com/volcengine/verl/blob/main/.github/workflows/vllm.yml)
@@ -62,6 +62,7 @@ pip3 install -e .[sglang]
 
 ## Release History
 
+- 2026/03/10: update vllm stable image to vllm==0.17.0; update sglang stable image to sglang==0.5.9
 - 2026/01/17: update vllm stable image to torch==2.9.1, cudnn==9.16, deepep==1.2.1
 - 2025/12/23: update vllm stable image to vllm==0.12.0; update sglang stable image to sglang==0.5.6
 - 2025/11/18: update vllm stable image to vllm==0.11.1; update sglang stable image to sglang==0.5.5
diff --git a/tests/experimental/agent_loop/test_multi_modal.py b/tests/experimental/agent_loop/test_multi_modal.py
@@ -230,10 +230,14 @@ def test_multimodal_tool_agent(init_config):
             assert "pixel_values_videos" in multi_modal_inputs[i], f"Sample {i} should have pixel_values_videos"
             assert "video_grid_thw" in multi_modal_inputs[i], f"Sample {i} should have video_grid_thw"
 
-        if i // n <= 1:
-            # TODO: prompt with video not generate tool call as expected
+        if i // n == 0:
             # First prompt: "How are you?" - should have 2 turns [user, assistant]
             assert num_turns[i] == 2, f"Expected 2 turns but got {num_turns[i]} for sample {i}"
+        elif i // n == 1:
+            # TODO: prompt with video not generate tool call as expected
+            assert num_turns[i] == 2 or num_turns[i] == 4, (
+                f"Expected 2 or 4 turns but got {num_turns[i]} for sample {i}"
+            )
         else:
             # Tool-calling prompts should have 4 turns [user, assistant, tool, assistant]
             assert num_turns[i] == 4, f"Expected 4 turns but got {num_turns[i]} for sample {i}"
diff --git a/tests/experimental/agent_loop/test_standalone_rollout.py b/tests/experimental/agent_loop/test_standalone_rollout.py
@@ -68,7 +68,10 @@ async def test_standalone_rollout(init_config, tp_size):
     rollout_server_class = get_rollout_replica_class(init_config.actor_rollout_ref.rollout.name)
     rollout_servers = [
         rollout_server_class(
-            replica_rank=replica_rank, config=rollout_config, model_config=model_config, gpus_per_node=2
+            replica_rank=replica_rank,
+            config=rollout_config,
+            model_config=model_config,
+            gpus_per_node=init_config.trainer.n_gpus_per_node,
         )
         for replica_rank in range(num_replicas)
     ]
diff --git a/tests/special_e2e/run_ppo_trainer_veomni.sh b/tests/special_e2e/run_ppo_trainer_veomni.sh
diff --git a/verl/utils/vllm/vllm_fp8_utils.py b/verl/utils/vllm/vllm_fp8_utils.py
diff --git a/verl/workers/engine/base.py b/verl/workers/engine/base.py
diff --git a/verl/workers/rollout/vllm_rollout/vllm_async_server.py b/verl/workers/rollout/vllm_rollout/vllm_async_server.py
