[Clean] Remove SGlang runtime code

.claude/skills/add-dynamic-filter/SKILL.md

@@ -27,7 +27,7 @@ Use this skill when:
 
 ### Step 2: Implement the Function Signature
 
-Dynamic sampling filter (called in `slime/rollout/sglang_rollout.py`):
+Dynamic sampling filter (called in `slime/rollout/vllm_rollout.py`):
 
 ```python
 def filter_function(args, samples, **kwargs):
@@ -97,5 +97,5 @@ Example wiring:
 
 - Dynamic filter types: `slime/rollout/filter_hub/base_types.py`
 - Dynamic filter example: `slime/rollout/filter_hub/dynamic_sampling_filters.py`
-- Rollout generation hook points: `slime/rollout/sglang_rollout.py`
+- Rollout generation hook points: `slime/rollout/vllm_rollout.py`
 - Buffer filter hook point: `slime/rollout/data_source.py`

.claude/skills/add-eval-dataset-config/SKILL.md

@@ -87,5 +87,5 @@ Use a separate eval function when inference/eval behavior must differ from train
 
 - Eval config model: `slime/utils/eval_config.py`
 - Eval config resolution: `slime/utils/arguments.py`
-- Eval rollout path: `slime/rollout/sglang_rollout.py`
+- Eval rollout path: `slime/rollout/vllm_rollout.py`
 - Customization docs: `docs/en/get_started/customization.md`

.claude/skills/add-rollout-function/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: add-rollout-function
-description: Guide for adding a new rollout function in slime and wiring it through --rollout-function-path. Use when user wants to implement custom rollout data generation logic, custom train/eval rollout outputs, or migrate from the default sglang rollout path.
+description: Guide for adding a new rollout function in slime and wiring it through --rollout-function-path. Use when user wants to implement custom rollout data generation logic, custom train/eval rollout outputs, or migrate from the default vLLM rollout path.
 ---
 
 # Add Rollout Function
@@ -12,7 +12,7 @@ Implement a custom rollout function and integrate it safely with slime training/
 Use this skill when:
 
 - User asks to add a new rollout task or rollout generation function
-- User asks to replace default `slime.rollout.sglang_rollout.generate_rollout`
+- User asks to replace default `slime.rollout.vllm_rollout.generate_rollout`
 - User asks to customize train/eval data generation behavior
 
 ## Step-by-Step Guide
@@ -21,10 +21,10 @@ Use this skill when:
 
 Start from one of these references:
 
-- Async RL-style rollout: `slime/rollout/sglang_rollout.py`
+- Async RL-style rollout: `slime/rollout/vllm_rollout.py`
 - Simple SFT-style rollout: `slime/rollout/sft_rollout.py`
 
-If the task needs engine-based async generation and rewards, use the sglang path as base.
+If the task needs engine-based async generation and rewards, use the vLLM path as base.
 If the task is file/buffer-driven and simple, use sft path as base.
 
 ### Step 2: Create the New Rollout Module
@@ -100,7 +100,7 @@ The default and signature expectation are documented in:
 
 ## Reference Locations
 
-- Default rollout: `slime/rollout/sglang_rollout.py`
+- Default rollout: `slime/rollout/vllm_rollout.py`
 - Simple custom example: `slime/rollout/sft_rollout.py`
 - Output dataclasses: `slime/rollout/base_types.py`
 - Wiring/loading: `slime/ray/rollout.py`

README.md

@@ -45,7 +45,7 @@ We also provide examples for some use cases not covered in the quick start guide
 Arguments in Vime are divided into three categories:
 
 1. **Megatron arguments**: Vime reads all arguments in Megatron. You can configure Megatron by passing arguments like `--tensor-model-parallel-size 2`.
-2. **vLLM arguments**: vLLM server and engine options are exposed with a `--vllm-` prefix (for example, `--vllm-gpu-memory-utilization`). Router options live under two prefixes: vllm-router's native options are passed with `--router-` (for example, `--router-policy round_robin`), while Vime-side orchestration knobs that tell Vime *where* the router lives use `--vllm-router-` (`--vllm-router-ip`, `--vllm-router-port`, `--vllm-router-request-timeout-secs`). See [slime/backends/vllm_utils/arguments.py](slime/backends/vllm_utils/arguments.py) for the full surface.
+2. **vLLM arguments**: vLLM server and engine options are exposed with a `--vllm-` prefix (for example, `--vllm-gpu-memory-utilization`). Router options live under two prefixes: vllm-router's native options are passed with `--router-` (for example, `--router-policy round_robin`, `--router-request-timeout-secs`), while Vime-side orchestration knobs that tell Vime *where* the router lives use `--vllm-router-` (`--vllm-router-ip`, `--vllm-router-port`). See [slime/backends/vllm_utils/arguments.py](slime/backends/vllm_utils/arguments.py) for the full surface.
 3. **Framework-specific arguments**: Shared slime/Vime orchestration flags (rollout GPUs, data paths, RL algorithms, etc.). Please refer to [slime/utils/arguments.py](slime/utils/arguments.py).
 
 `--rollout-num-gpus-per-engine` sets the tensor parallel size of each vLLM engine. The default rollout entry is `slime.rollout.vllm_rollout.generate_rollout`.

README_zh.md

@@ -45,7 +45,7 @@ Vime 继承 slime 的广泛模型支持，包括：
 Vime 的参数分为三类：
 
 1. **Megatron 参数**：Vime 会读取 Megatron 中的全部参数，可通过传入如 `--tensor-model-parallel-size 2` 的方式配置 Megatron；
-2. **vLLM 参数**：vLLM server 与 engine 相关选项以 `--vllm-` 为前缀（例如 `--vllm-gpu-memory-utilization`）。路由相关选项分两类前缀：vllm-router 自身的选项以 `--router-` 传入（例如 `--router-policy round_robin`），Vime 侧用于告诉 Vime *router 在哪里* 的编排参数则以 `--vllm-router-` 为前缀（`--vllm-router-ip`、`--vllm-router-port`、`--vllm-router-request-timeout-secs`）。完整参数见 [slime/backends/vllm_utils/arguments.py](slime/backends/vllm_utils/arguments.py)。
+2. **vLLM 参数**：vLLM server 与 engine 相关选项以 `--vllm-` 为前缀（例如 `--vllm-gpu-memory-utilization`）。路由相关选项分两类前缀：vllm-router 自身的选项以 `--router-` 传入（例如 `--router-policy round_robin`、`--router-request-timeout-secs`），Vime 侧用于告诉 Vime *router 在哪里* 的编排参数则以 `--vllm-router-` 为前缀（`--vllm-router-ip`、`--vllm-router-port`）。完整参数见 [slime/backends/vllm_utils/arguments.py](slime/backends/vllm_utils/arguments.py)。
 3. **框架参数**：与 slime/Vime 编排相关的开关（rollout GPU、数据路径、RL 算法等），见 [slime/utils/arguments.py](slime/utils/arguments.py)。
 
 `--rollout-num-gpus-per-engine` 对应每个 vLLM engine 的 tensor parallel size。默认 rollout 入口为 `slime.rollout.vllm_rollout.generate_rollout`。

docker/Dockerfile

@@ -1,11 +1,10 @@
-ARG BASE_IMAGE=vllm/vllm-openai:v0.21.0-cu129-ubuntu2404
+ARG BASE_IMAGE=vllm/vllm-openai:v0.22.0-cu129-ubuntu2404
 FROM ${BASE_IMAGE}
 
 # ======================================== Arguments =============================================
 
 ARG PATCH_VERSION=latest
 ARG MEGATRON_COMMIT=1dcf0dafa884ad52ffb243625717a3471643e087
-ARG SGLANG_VERSION=0.5.10.post1
 
 ARG ENABLE_CUDA_13=0
 
@@ -74,29 +73,17 @@ RUN if [ "$ENABLE_CUDA_13" = "1" ]; then \
     (cd /root && git clone -b feat/v350_plus_8045 https://github.com/fzyzcjy/triton.git && cd triton && pip install -r python/requirements.txt && pip install --verbose -e .); \
   fi
 
-# Skip sglang lines from requirements (we install --no-deps below; the full
-# transitive set conflicts with the vllm base environment).
 COPY requirements.txt /tmp/requirements.txt
-RUN grep -vE '^[[:space:]]*(sglang|sglang-router)([[:space:]]|[<>=!]|$)' /tmp/requirements.txt > /tmp/requirements-vllm.txt && \
-    pip install --ignore-installed PyJWT && \
-    pip install -r /tmp/requirements-vllm.txt
-
-# Temporarily install another sgl-kernel version for GB300 without rebuilding the whole image
-RUN if [ "$ENABLE_CUDA_13" = "1" ]; then \
-    SGL_KERNEL_VERSION=0.3.17.post2 && \
-    python3 -m pip install https://github.com/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sgl_kernel-${SGL_KERNEL_VERSION}+cu130-cp310-abi3-manylinux2014_$(uname -m).whl --force-reinstall --no-deps; \
-  fi
+RUN pip install --ignore-installed PyJWT && \
+    pip install -r /tmp/requirements.txt
 
 # https://github.com/pytorch/pytorch/issues/168167
 RUN pip install nvidia-cudnn-cu12==9.16.0.29
 
 # reinstall numpy 1.x for megatron
 RUN pip install "numpy<2"
 
-# vime's slime/utils/arguments.py + slime/ray/rollout.py top-level import sglang_router;
-# the cu129 vllm base does not ship sglang, so install --no-deps stubs here.
-RUN pip install --no-deps "sglang==${SGLANG_VERSION}" sglang-router==0.3.2 && \
-    pip install IPython
+RUN pip install IPython
 
 # Pin vllm-router explicitly so the vllm rollout routing layer is a visible build step
 # (also in requirements.txt; pulling it here makes the layer cache-able and fail-fast).
@@ -135,10 +122,9 @@ RUN cd /root/slime/slime/backends/megatron_utils/kernels/int4_qat && \
 # Fail-fast import smoke + flashinfer version pin check. Catches ABI / version
 # regressions at build time instead of first GPU run.
 RUN python3 -c "\
-import vllm, sglang, slime, flashinfer; \
+import vllm, slime, flashinfer; \
 from slime.backends.vllm_utils.vllm_engine import VLLMEngine; \
 print('vllm', vllm.__version__); \
-print('sglang', sglang.__version__); \
 print('flashinfer', flashinfer.__version__); \
 print('VLLMEngine import ok')"
 

docker/justfile

@@ -1,13 +1,17 @@
 release-primary:
     ARG_TAG_POSTFIX="" ARG_BUILD_EXTRA_ARGS="" just _release-raw
 
-# Should be executed on ARM machines
+# Should be executed on ARM machines.
+# Inherits the Dockerfile's default cu129 BASE_IMAGE; that tag is a multi-arch
+# manifest, so docker selects the arm64 image automatically on an ARM host.
 release-cu129-arm64:
-    ARG_TAG_POSTFIX="-cu129-arm64" ARG_BUILD_EXTRA_ARGS='--build-arg SGLANG_IMAGE_TAG=v0.5.5.post3-cu129-arm64 --build-arg ENABLE_SGLANG_PATCH=0' just _release-raw
+    ARG_TAG_POSTFIX="-cu129-arm64" ARG_BUILD_EXTRA_ARGS="" just _release-raw
 
-# Should be executed on ARM machines
+# Should be executed on ARM machines.
+# The default-CUDA vLLM tag (no cuXXX suffix) already ships CUDA 13.0;
+# ENABLE_CUDA_13 then builds the CUDA-13 TransformerEngine/Triton on top.
 release-cu13-arm64:
-    ARG_TAG_POSTFIX="-cu13-arm64" ARG_BUILD_EXTRA_ARGS='--build-arg SGLANG_IMAGE_TAG=dev-arm64-cu13-20251122 --build-arg ENABLE_CUDA_13=1 --build-arg ENABLE_SGLANG_PATCH=0' just _release-raw
+    ARG_TAG_POSTFIX="-cu13-arm64" ARG_BUILD_EXTRA_ARGS='--build-arg BASE_IMAGE=vllm/vllm-openai:v0.22.0-ubuntu2404 --build-arg ENABLE_CUDA_13=1' just _release-raw
 
 _release-raw:
     #!/bin/bash