ci: switch sglang to prebuilt PyPI wheels (v0.5.11)

.github/workflows/cicd-main.yml

@@ -301,7 +301,6 @@ jobs:
       build-args: |
         MAX_JOBS=4
         NEMO_RL_COMMIT=${{ needs.pre-flight.outputs.test_sha }}
-        SKIP_SGLANG_BUILD=1
 
   update-uv-cache:
     name: Update uv build cache

docker/Dockerfile

@@ -59,7 +59,7 @@ apt-get clean
 rm -rf /var/lib/apt/lists/*
 EOF
 
-# CMake (for sglang build)
+# CMake (for CUDA extension builds: transformer-engine, mamba-ssm, etc.)
 RUN GITHUB_ARTIFACTORY=github.com \
     && CMAKE_VERSION=3.31.1 \
     && ARCH=$(uname -m) \

nemo_rl/models/generation/sglang/sglang_worker.py

@@ -179,6 +179,9 @@ def __init__(
         # Get current node IP and a free port for the server
         node_ip = _get_node_ip_local()
         free_port = _get_free_port_local()
+        # sglang derives grpc_port = port + 10000, so cap at 55535 to stay within valid range
+        while free_port > 55535:
+            free_port = _get_free_port_local()
 
         # Build SGLang server arguments
         kwargs = {

pyproject.toml

@@ -143,8 +143,8 @@ vllm = [
   "nvidia-cutlass-dsl>=4.4.0.dev1",
 ]
 sglang = [
-  "sglang",
-  "sglang-kernel", # Must be a direct dep so [tool.uv.sources] VCS override applies (transitive deps don't use sources)
+  "sglang==0.5.11",
+  "sglang-kernel==0.4.2", # Direct dep for explicit version control
 ]
 mcore = [
   # also need cudnn (https://developer.nvidia.com/cudnn-downloads?target_os=Linux&target_arch=x86_64&Distribution=Ubuntu&target_version=20.04&target_type=deb_network)
@@ -249,9 +249,6 @@ causal-conv1d = { git = "https://github.com/Dao-AILab/causal-conv1d", rev = "67e
 mamba-ssm = { git = "https://github.com/state-spaces/mamba.git", rev = "d68d16ed7d5d5164eb5a57c0285f3b7eb8394ec1" }
 nv-grouped-gemm = { git = "https://github.com/fanshiqing/grouped_gemm", tag = "v1.1.4.post7" }
 emerging-optimizers = { git = "https://github.com/NVIDIA-NeMo/Emerging-Optimizers.git", rev = "v0.2.0" }
-# Official sglang v0.5.10 with transformers v5 support
-sglang = { git = "https://github.com/sgl-project/sglang.git", tag = "v0.5.10", subdirectory = "python" }
-sglang-kernel = { git = "https://github.com/sgl-project/sglang.git", tag = "v0.5.10", subdirectory = "sgl-kernel" }
 nvidia-modelopt = { git = "https://github.com/NVIDIA/Model-Optimizer", rev = "905018803414702e414a86716484ed4115b37ba6" }
 nvidia-resiliency-ext = { git = "https://github.com/NVIDIA/nvidia-resiliency-ext.git", rev = "15a851565a4ce846c04431ecb0cf09903ab4837e" }
 
@@ -278,7 +275,6 @@ explicit = true
 
 [tool.uv]
 preview = true # Enable preview features like extra-build-dependencies
-extra-build-variables = { sglang-kernel = { CMAKE_BUILD_PARALLEL_LEVEL = "8", FLASHINFER_CUDA_ARCH_LIST = "9.0a 10.0a", CMAKE_ARGS = "-DCMAKE_POLICY_VERSION_MINIMUM=3.5" } }
 no-build-isolation-package = [
   "transformer-engine-torch",
   "transformer-engine",
@@ -288,7 +284,6 @@ no-build-isolation-package = [
   "deep_gemm",
   "deep_ep",
   "nv-grouped-gemm",          # from mlm (added here to make sure it's built no isolation since mlm workspace uses setup.py)
-  "sglang-kernel",
 ]
 # Always apply the build group since dependencies like TE/mcore/nemo-run require build dependencies
 # and this lets us assume they are implicitly installed with a simply `uv sync`. Ideally, we'd
@@ -318,20 +313,17 @@ override-dependencies = [
   "setuptools>=80.10.2",
   "deep_ep @ git+https://github.com/deepseek-ai/DeepEP.git@bfded34800dfec415b71503f8205181de90b2480 ; platform_machine == 'x86_64'",
   "deep_ep @ git+https://github.com/deepseek-ai/DeepEP.git@a48493600c4886c1b297aaa78db0e1ebc2d8dd6c ; platform_machine == 'aarch64'",
-  # Note: flashinfer versions are pinned per-extra (vllm uses 0.6.8.post1, sglang uses 0.6.7.post2)
-  # since vllm and sglang extras are mutually exclusive and have different requirements.
-  # Override megatron-core's flashinfer~=0.5.0 constraint to allow both vllm (0.6.8.post1) and sglang (0.6.7.post2)
+  # Override megatron-core's flashinfer~=0.5.0 constraint to allow both vllm and sglang (both use 0.6.8.post1)
   "flashinfer-python>=0.5.0",
   "flashinfer-cubin>=0.5.0",
-  # sglang 0.5.10 requires nvidia-cutlass-dsl>=4.4.1 (via flashinfer 0.6.7.post2 which uses CUTLASS 4.4.2).
   # Override to >=4.4.1 so uv can resolve to a version satisfying both vllm and sglang.
   "nvidia-cutlass-dsl>=4.4.1",
   # Relax megatron-core workspace member's opentelemetry-api ceiling (<1.34) for protobuf 6.x compat with ray
   "opentelemetry-api>=1.33.1",
   # vLLM 0.20.0 excludes transformers 5.0-5.5.0 but our codebase needs 5.3.0.
   # Override so uv can resolve transformers==5.3.0 for non-vllm extras.
   "transformers==5.3.0",
-  #Override till we can upgrade sglang version to address CVE GHSA-7rgv-gqhr-fxg3
+  # Override sglang's xgrammar==0.1.32 to address CVE GHSA-7rgv-gqhr-fxg3
   "xgrammar==0.1.33",
   # Override dependencies to address CVEs
   "mlflow>=3.11.1",
@@ -411,7 +403,6 @@ transformer-engine-torch = [{ requirement = "torch", match-runtime = true }]
 mamba-ssm = [{ requirement = "torch", match-runtime = true }]
 causal-conv1d = [{ requirement = "torch", match-runtime = true }]
 nv-grouped-gemm = [{ requirement = "torch", match-runtime = true }]
-sglang-kernel = [{ requirement = "torch", match-runtime = true }]
 
 # Needed when building from source
 [[tool.uv.dependency-metadata]]
@@ -458,83 +449,6 @@ name = "nv-grouped-gemm"
 version = "v1.1.4.post7"
 requires-dist = ["setuptools", "wheel", "torch", "numpy"]
 
-[[tool.uv.dependency-metadata]]
-name = "sglang-kernel"
-# This version has to match the version in the commit/rev/tag used
-version = "0.4.1"
-requires-dist = ["torch", "scikit-build-core", "wheel"]
-
-[[tool.uv.dependency-metadata]]
-name = "sglang"
-# VCS install from official sgl-project/sglang v0.5.10
-# Version is dynamic (setuptools-scm), so uv cannot resolve deps from the VCS source automatically.
-# This requires-dist list must be kept in sync with the official python/pyproject.toml [project].dependencies.
-# Source: https://github.com/sgl-project/sglang/blob/v0.5.10/python/pyproject.toml
-version = "0.5.10"
-requires-dist = [
-  "IPython",
-  "aiohttp",
-  "apache-tvm-ffi>=0.1.5,<0.2",
-  "anthropic>=0.20.0",
-  "blobfile==3.0.0",
-  "build",
-  "compressed-tensors",
-  "cuda-python==13.0",
-  "decord2",
-  "datasets",
-  "einops",
-  "fastapi",
-  "flashinfer_python==0.6.7.post2",
-  "flashinfer_cubin==0.6.7.post2",
-  "gguf",
-  "interegular",
-  "llguidance>=0.7.11,<0.8.0",
-  "modelscope",
-  "msgspec",
-  "ninja",
-  "numpy",
-  "nvidia-cutlass-dsl>=4.4.1",
-  "nvidia-ml-py",
-  "openai-harmony==0.0.4",
-  "openai==2.6.1",
-  "orjson",
-  "outlines==0.1.11",
-  "packaging",
-  "partial_json_parser",
-  "pillow",
-  "prometheus-client>=0.20.0",
-  "psutil",
-  "py-spy",
-  "pybase64",
-  "pydantic",
-  "python-multipart",
-  "pyzmq>=25.1.2",
-  "quack-kernels>=0.3.0",
-  "requests",
-  "scipy",
-  "sentencepiece",
-  "setproctitle",
-  "flash-attn-4>=4.0.0b4",
-  "sglang-kernel==0.4.1",
-  "soundfile==0.13.1",
-  "tiktoken",
-  "timm==1.0.16",
-  "torch_memory_saver==0.0.9",
-  "torch==2.9.1",
-  "torchao==0.9.0",
-  "torchaudio==2.9.1",
-  "torchcodec==0.9.1 ; sys_platform != 'linux' or (sys_platform == 'linux' and platform_machine != 'aarch64' and platform_machine != 'arm64' and platform_machine != 'armv7l')",
-  "torchvision",
-  "tqdm",
-  "mistral_common>=1.9.0",
-  "transformers==5.3.0",
-  "uvicorn",
-  "uvloop",
-  "watchfiles",
-  "xgrammar==0.1.32",
-  "smg-grpc-servicer>=0.5.0",
-]
-
 [[tool.uv.dependency-metadata]]
 name = "megatron-bridge"
 # Static metadata so the root resolver handles megatron-bridge's deps

tests/functional/L1_Functional_Tests_GPU.sh

@@ -68,10 +68,7 @@ run_test      uv run --no-sync bash ./tests/functional/grpo_multiple_dataloaders
 run_test      uv run --no-sync bash ./tests/functional/grpo_multiturn.sh
 run_test      uv run --no-sync bash ./tests/functional/grpo_non_colocated.sh
 run_test      uv run --no-sync bash ./tests/functional/grpo_rm_env.sh
-# Skip sglang functional test if sglang was not built into the container
-if python -c "import sglang" 2>/dev/null; then
-    run_test      uv run --no-sync bash ./tests/functional/grpo_sglang.sh
-fi
+run_test      uv run --no-sync bash ./tests/functional/grpo_sglang.sh
 run_test fast uv run --no-sync bash ./tests/functional/grpo_topp_topk.sh
 run_test      uv run --no-sync bash ./tests/functional/prorlv2.sh
 run_test      uv run --no-sync bash ./tests/functional/qa_distillation_megatron.sh

tests/unit/L0_Unit_Tests_Generation.sh

@@ -57,14 +57,10 @@ else
     uv run --extra vllm bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --vllm-only
 fi
 
-# Check and run sglang tests (skip if sglang was not built into the container)
-if python -c "import sglang" 2>/dev/null; then
-    exit_code=$(cd ${PROJECT_ROOT}/tests && uv run --extra sglang pytest "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --collect-only --hf-gated --sglang-only -q >/dev/null 2>&1; echo $?)
-    if [[ $exit_code -eq 5 ]]; then
-        echo "No sglang tests to run"
-    else
-        uv run --extra sglang bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --sglang-only
-    fi
+# Check and run sglang tests
+exit_code=$(cd ${PROJECT_ROOT}/tests && uv run --extra sglang pytest "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --collect-only --hf-gated --sglang-only -q >/dev/null 2>&1; echo $?)
+if [[ $exit_code -eq 5 ]]; then
+    echo "No sglang tests to run"
 else
-    echo "sglang not installed, skipping sglang tests"
+    uv run --extra sglang bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --sglang-only
 fi

tests/unit/L0_Unit_Tests_Other.sh

@@ -57,16 +57,12 @@ else
     uv run --extra vllm bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --vllm-only
 fi
 
-# Check and run sglang tests (skip if sglang was not built into the container)
-if python -c "import sglang" 2>/dev/null; then
-    exit_code=$(cd ${PROJECT_ROOT}/tests && uv run --extra sglang pytest "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --collect-only --hf-gated --sglang-only -q >/dev/null 2>&1; echo $?)
-    if [[ $exit_code -eq 5 ]]; then
-        echo "No sglang tests to run"
-    else
-        uv run --extra sglang bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --sglang-only
-    fi
+# Check and run sglang tests
+exit_code=$(cd ${PROJECT_ROOT}/tests && uv run --extra sglang pytest "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --collect-only --hf-gated --sglang-only -q >/dev/null 2>&1; echo $?)
+if [[ $exit_code -eq 5 ]]; then
+    echo "No sglang tests to run"
 else
-    echo "sglang not installed, skipping sglang tests"
+    uv run --extra sglang bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --sglang-only
 fi
 
 # Check and run nemo_gym tests

tests/unit/L0_Unit_Tests_Policy.sh

@@ -57,14 +57,10 @@ else
     uv run --extra vllm bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --vllm-only
 fi
 
-# Check and run sglang tests (skip if sglang was not built into the container)
-if python -c "import sglang" 2>/dev/null; then
-    exit_code=$(cd ${PROJECT_ROOT}/tests && uv run --extra sglang pytest "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --collect-only --hf-gated --sglang-only -q >/dev/null 2>&1; echo $?)
-    if [[ $exit_code -eq 5 ]]; then
-        echo "No sglang tests to run"
-    else
-        uv run --extra sglang bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --sglang-only
-    fi
+# Check and run sglang tests
+exit_code=$(cd ${PROJECT_ROOT}/tests && uv run --extra sglang pytest "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --collect-only --hf-gated --sglang-only -q >/dev/null 2>&1; echo $?)
+if [[ $exit_code -eq 5 ]]; then
+    echo "No sglang tests to run"
 else
-    echo "sglang not installed, skipping sglang tests"
+    uv run --extra sglang bash -x ./tests/run_unit.sh "${TEST_PATHS[@]}" "${IGNORE[@]}" "${EXCLUDED_UNIT_TESTS[@]}" --cov=nemo_rl --cov-append --cov-report=term-missing --cov-report=json --hf-gated --sglang-only
 fi