Fix CI (#9012)

.github/workflows/vllm-dependency-test.yml

@@ -30,13 +30,19 @@ jobs:
       - name: Install dependencies
         run: |
           bash scripts/ci_install_dependency.sh
-          pip install "vllm==0.9.0"
+          pip install "vllm==0.10.0"
-          pip install "bitsandbytes>=0.44.0"
-          pip install torch==2.8.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu126
           pip install "openai==1.99.1"
+          pip install "bitsandbytes>=0.44.0"
+
+          # NOTE: The latest sgl-kernel depends on torch 2.8.0 but the latest vllm depends on torch 2.7.0
+          # so they are not compatible. Here we install the old sgl-kernel to make the test pass.
+          # TODO: remove this once vllm supports torch 2.8.0.
+          pip install "sgl-kernel==0.2.9"
 
       - name: Run vLLM dependency tests
         timeout-minutes: 60
         run: |
+          export SGLANG_SKIP_SGL_KERNEL_VERSION_CHECK=1
+
           cd test/srt
           python3 run_suite.py --suite vllm_dependency_test --timeout-per-file 3600

python/sglang/srt/entrypoints/engine.py

@@ -67,6 +67,7 @@ from sglang.srt.utils import (
     MultiprocessingSerializer,
     assert_pkg_version,
     configure_logger,
+    get_bool_env_var,
     get_zmq_socket,
     is_cuda,
     kill_process_tree,
@@ -627,7 +628,6 @@ def _set_envs_and_config(server_args: ServerArgs):
     os.environ["NCCL_CUMEM_ENABLE"] = str(int(server_args.enable_symm_mem))
     if not server_args.enable_symm_mem:
         os.environ["NCCL_NVLS_ENABLE"] = str(int(server_args.enable_nccl_nvls))
-    os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
     os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "4"
     os.environ["CUDA_MODULE_LOADING"] = "AUTO"
 
@@ -647,7 +647,7 @@ def _set_envs_and_config(server_args: ServerArgs):
             "reinstall the latest version by following the instructions "
             "at https://docs.flashinfer.ai/installation.html.",
         )
-    if _is_cuda:
+    if _is_cuda and not get_bool_env_var("SGLANG_SKIP_SGL_KERNEL_VERSION_CHECK"):
         assert_pkg_version(
             "sgl-kernel",
             "0.3.3",

python/sglang/srt/entrypoints/openai/tool_server.py

@@ -5,16 +5,17 @@ from abc import ABC, abstractmethod
 from contextlib import AbstractAsyncContextManager, asynccontextmanager
 from typing import Any
 
-logger = logging.getLogger(__name__)
 try:
     from mcp import ClientSession
     from mcp.client.sse import sse_client
     from mcp.types import ListToolsResult
-except ImportError:
+except ImportError as e:
-    logger.warning("Ignoring mcp import error")
+    ClientSession = sse_client = ListToolsResult = e
 
 from openai_harmony import ToolDescription, ToolNamespaceConfig
 
+logger = logging.getLogger(__name__)
+
 
 async def list_server_and_tools(server_url: str):
 

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -147,6 +147,7 @@ class FusedMoE(torch.nn.Module):
 
         self.layer_id = layer_id
         self.top_k = top_k
+        self.hidden_size = hidden_size
         self.num_experts = num_experts
         self.num_fused_shared_experts = num_fused_shared_experts
         self.expert_map_cpu = None

python/sglang/srt/layers/quantization/__init__.py

@@ -26,8 +26,9 @@ try:
     from vllm.model_executor.layers.quantization.tpu_int8 import Int8TpuConfig
 
     VLLM_AVAILABLE = True
-except ImportError:
+except ImportError as e:
     VLLM_AVAILABLE = False
+    VLLM_IMPORT_ERROR = e
 
     # Define empty classes as placeholders when vllm is not available
     class DummyConfig:
@@ -137,7 +138,8 @@ def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
     if quantization in VLLM_QUANTIZATION_METHODS and not VLLM_AVAILABLE:
         raise ValueError(
             f"{quantization} quantization requires some operators from vllm. "
-            "Please install vllm by `pip install vllm==0.9.0.1`"
+            f"Please install vllm by `pip install vllm==0.9.0.1`\n"
+            f"Import error: {VLLM_IMPORT_ERROR}"
         )
 
     return QUANTIZATION_METHODS[quantization]

python/sglang/srt/layers/quantization/modelopt_quant.py

@@ -1,9 +1,8 @@
 # Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/modelopt.py
 from __future__ import annotations
 
-import importlib.util
 import logging
-from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Union
+from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
 import torch
 from torch.nn.parameter import Parameter
@@ -42,11 +41,7 @@ if is_cuda():
 
 try:
     from flashinfer import mm_fp4 as fp4_gemm
-    from flashinfer import (
+    from flashinfer import reorder_rows_for_gated_act_gemm, shuffle_matrix_sf_a
-        reorder_rows_for_gated_act_gemm,
-        shuffle_matrix_a,
-        shuffle_matrix_sf_a,
-    )
 
     enable_flashinfer_fp4_gemm = True
 except ImportError:

python/sglang/srt/managers/multimodal_processor.py

@@ -20,7 +20,7 @@ def import_processors():
             try:
                 module = importlib.import_module(name)
             except Exception as e:
-                logger.warning(f"Ignore import error when loading {name}: " f"{e}")
+                logger.warning(f"Ignore import error when loading {name}: {e}")
                 continue
             all_members = inspect.getmembers(module, inspect.isclass)
             classes = [

python/sglang/srt/models/registry.py

@@ -83,7 +83,7 @@ def import_model_classes():
             try:
                 module = importlib.import_module(name)
             except Exception as e:
-                logger.warning(f"Ignore import error when loading {name}. " f"{e}")
+                logger.warning(f"Ignore import error when loading {name}: {e}")
                 continue
             if hasattr(module, "EntryClass"):
                 entry = module.EntryClass

test/srt/test_utils_update_weights.py

@@ -83,7 +83,6 @@ class TestUtilsUpdateWeights(unittest.TestCase):
         # Set up environment variables
         os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
         os.environ["NCCL_CUMEM_ENABLE"] = "0"
-        os.environ["TORCH_NCCL_AVOID_RECORD_STREAMS"] = "1"
         os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "4"
         os.environ["CUDA_MODULE_LOADING"] = "AUTO"