chore(deps): bump sglang from 0.5.9/0.5.10rc0 to 0.5.10

.github/workflows/pr-test-rust.yml

@@ -832,8 +832,22 @@ jobs:
           export CGO_LDFLAGS="-L$(pwd)/bindings/golang/target/release"
           export LD_LIBRARY_PATH="$(pwd)/bindings/golang/target/release:$LD_LIBRARY_PATH"
           SHOW_WORKER_LOGS=0 SHOW_ROUTER_LOGS=1 ROUTER_LOCAL_MODEL_PATH="/models" \
+          E2E_LOG_DIR=e2e-logs \
             pytest --reruns 2 --reruns-delay 5 e2e_test/bindings_go -s -vv
 
+      - name: Worker failure diagnostics
+        if: failure() || cancelled()
+        run: bash scripts/ci_dump_worker_logs.sh e2e-logs "e2e-worker-logs-go-bindings"
+
+      - name: Upload worker logs
+        if: failure() || cancelled()
+        uses: actions/upload-artifact@v7
+        with:
+          name: e2e-worker-logs-go-bindings
+          path: e2e-logs/
+          retention-days: 7
+          if-no-files-found: ignore
+
   go-bindings-benchmark:
     name: go-bindings-benchmark
     needs: build-wheel

.github/workflows/release-sglang-docker.yml

@@ -2,7 +2,7 @@ name: Release SMG+SGLang Docker Image
 
 run-name: >-
   SMG+SGLang |
-  ${{ github.event_name == 'push' && '3-version matrix' || format('base={0}', inputs.base_image_ref || 'lmsysorg/sglang:v0.5.9') }} |
+  ${{ github.event_name == 'push' && 'base=lmsysorg/sglang:v0.5.10' || format('base={0}', inputs.base_image_ref || 'lmsysorg/sglang:v0.5.10') }} |
   engine=${{ inputs.sglang_commit || 'latest' }} |
   smg=${{ github.event_name == 'pull_request' && 'main' || inputs.smg_commit || 'v1.4.0' }} |
   ${{ github.event_name == 'pull_request' && 'dry-run |' || '' }}
@@ -23,8 +23,8 @@ on:
   workflow_dispatch:
     inputs:
       base_image_ref:
-        description: 'Base image (e.g. lmsysorg/sglang:v0.5.9)'
-        default: 'lmsysorg/sglang:v0.5.9'
+        description: 'Base image (e.g. lmsysorg/sglang:v0.5.10)'
+        default: 'lmsysorg/sglang:v0.5.10'
         required: true
         type: string
       sglang_repo:
@@ -47,7 +47,7 @@ on:
         default: 'v1.4.0'
         type: string
       tag:
-        description: 'Override image tag (e.g. v1.4.0-sglang-v0.5.9)'
+        description: 'Override image tag (e.g. v1.4.0-sglang-v0.5.10)'
         required: false
         type: string
 
@@ -60,7 +60,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        base_image: ${{ github.event_name == 'push' && fromJSON('["lmsysorg/sglang:v0.5.9","lmsysorg/sglang:v0.5.8","lmsysorg/sglang:v0.5.7"]') || fromJSON(format('["{0}"]', inputs.base_image_ref || 'lmsysorg/sglang:v0.5.9')) }}
+        base_image: ${{ github.event_name == 'push' && fromJSON('["lmsysorg/sglang:v0.5.10"]') || fromJSON(format('["{0}"]', inputs.base_image_ref || 'lmsysorg/sglang:v0.5.10')) }}
     uses: ./.github/workflows/_build-engine-image.yml
     with:
       engine: sglang

bindings/golang/internal/grpc/client_grpc.go

@@ -148,10 +148,12 @@ func (c *GrpcClient) CreateChatCompletionStream(ctx context.Context, reqJSON str
 
 	// Set sampling parameters
 	samplingParams := &proto.SamplingParams{
-		Temperature:       1.0,
-		TopP:              1.0,
-		TopK:              -1,
-		SkipSpecialTokens: true,
+		Temperature:                1.0,
+		TopP:                       1.0,
+		TopK:                       -1,
+		RepetitionPenalty:          1.0,
+		SkipSpecialTokens:          true,
+		SpacesBetweenSpecialTokens: true,
 	}
 
 	if temp, ok := reqMap["temperature"].(float64); ok {

docker/engine.Dockerfile

@@ -10,7 +10,7 @@
 #   SMG_COMMIT      - commit/ref for SMG_REPO ("latest" = HEAD)
 #
 # Usage:
-#   docker build --build-arg BASE_IMAGE_REF=lmsysorg/sglang:v0.5.9 \
+#   docker build --build-arg BASE_IMAGE_REF=lmsysorg/sglang:v0.5.10 \
 #                --build-arg ENGINE=sglang \
 #                --build-arg SMG_REPO=https://github.com/lightseekorg/smg \
 #                --build-arg SMG_COMMIT=v1.1.0 \
@@ -44,7 +44,7 @@
 COPY scripts/installation/ /tmp/scripts/
 
 # ── final stage: install SMG + conditionally install engine ──────────────────
 FROM ${BASE_IMAGE_REF}
 
 ARG ENGINE=sglang
 ARG BACKEND

e2e_test/chat_completions/test_multimodal.py

@@ -44,7 +44,7 @@ def _make_image_content(image_source: str) -> dict:
 # =============================================================================
 
 
-@pytest.mark.engine("vllm")
+@pytest.mark.engine("vllm", "sglang")
 @pytest.mark.gpu(1)
 @pytest.mark.e2e
 @pytest.mark.model("Qwen/Qwen3-VL-8B-Instruct")
@@ -170,7 +170,7 @@ def test_multi_images_mixed(self, model, setup_backend):
 # =============================================================================
 
 
-@pytest.mark.engine("vllm")
+@pytest.mark.engine("vllm", "sglang")
 @pytest.mark.gpu(4)
 @pytest.mark.e2e
 @pytest.mark.model("meta-llama/Llama-4-Scout-17B-16E-Instruct")

e2e_test/infra/model_specs.py

@@ -149,6 +149,7 @@ def _resolve_model_path(hf_path: str) -> str:
             "--cuda-graph-max-bs=256",
             "--max-running-requests=300",
             "--mem-fraction-static=0.85",
+            "--enable-multimodal",
         ],
         "vllm_args": [
             "--max-model-len=196608",

grpc_servicer/pyproject.toml

@@ -31,7 +31,7 @@ classifiers = [
 
 [project.optional-dependencies]
 vllm = ["vllm>=0.19.0"]
-sglang = ["sglang>=0.5.10rc0"]
+sglang = ["sglang>=0.5.10"]
 
 [project.urls]
 Homepage = "https://github.com/lightseekorg/smg"

grpc_servicer/smg_grpc_servicer/sglang/servicer.py

@@ -37,7 +37,11 @@
     TokenizedEmbeddingReqInput,
     TokenizedGenerateReqInput,
 )
-from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
+from sglang.srt.managers.schedule_batch import (
+    Modality,
+    MultimodalDataItem,
+    MultimodalInputs,
+)
 from sglang.srt.sampling.sampling_params import SamplingParams as SGLSamplingParams
 from sglang.srt.server_args import ServerArgs
 from sglang.utils import get_exception_traceback
@@ -351,7 +355,7 @@ async def HealthCheck(
                 rid=rid,
                 input_text="",
                 input_ids=[0],
-                image_inputs={"mm_items": []},
+                image_inputs=None,
                 token_type_ids=[0],
                 sampling_params=sampling_params,
             )
@@ -768,8 +772,8 @@ def _decode_tensor_data(tensor_data):
         arr = np.frombuffer(tensor_data.data, dtype=np_dtype).reshape(shape)
         return torch.from_numpy(arr)
 
-    def _parse_mm_inputs(self, mm_proto) -> dict:
-        """Parse proto MultimodalInputs into the mm_inputs dict expected by scheduler."""
+    def _parse_mm_inputs(self, mm_proto) -> MultimodalInputs:
+        """Parse proto MultimodalInputs into the mm_inputs expected by scheduler."""
         # Decode pixel_values from typed TensorData field
         pixel_values = self._decode_tensor_data(mm_proto.pixel_values)
 
@@ -795,10 +799,10 @@ def _parse_mm_inputs(self, mm_proto) -> dict:
             offsets=offsets,
         )
 
-        result = {"mm_items": [mm_item]}
+        result = MultimodalInputs(mm_items=[mm_item])
 
         if mm_proto.HasField("im_token_id"):
-            result["im_token_id"] = mm_proto.im_token_id
+            result.im_token_id = mm_proto.im_token_id
 
         return result
 
@@ -827,7 +831,7 @@ def _convert_embed_request(
             rid=grpc_req.request_id,
             input_text=input_text,
             input_ids=input_ids,
-            image_inputs={"mm_items": []},
+            image_inputs=None,
             token_type_ids=list(grpc_req.token_type_ids),
             sampling_params=sampling_params,
         )

scripts/ci_install_sglang.sh

@@ -40,13 +40,8 @@ else
 fi
 
 # Install SGLang with all dependencies
-# TODO(changsu): Revert to `uv pip install "sglang[all]"` once sglang releases
-# a stable version (>= 0.5.10) that includes the smg_grpc_servicer migration
-# (sgl-project/sglang#20478). The rc0 is needed because the old built-in
-# grpc_server.py references proto types (EmbedError, EmbedComplete) that were
-# removed from our proto definitions.
 echo "Installing SGLang..."
-uv pip install --prerelease=allow "sglang[all]==0.5.10rc0"
+uv pip install --prerelease=allow "sglang[all]==0.5.10"
 
 # Install flashinfer-jit-cache: sglang bundles flashinfer_python but only for attention ops.
 # Multi-GPU models need trtllm_comm kernels (fused allreduce + layernorm) which FlashInfer

scripts/installation/install-sglang.sh

@@ -8,7 +8,3 @@ set -e
 SGL_SRC="${1:-/tmp/sglang-src}"
 cd "${SGL_SRC}/python"
 pip install --no-deps --force-reinstall --editable .
-
-# Pin gRPC packages to 1.78.0 — grpcio 1.78.1 was yanked from PyPI
-# TODO: remove when new SGLang 0.5.10 is released
-pip install --force-reinstall grpcio==1.78.0 grpcio-reflection==1.78.0 grpcio-tools==1.78.0 grpcio-health-checking==1.78.0