feat: add gpt oss b200 ci (#9988)

scripts/ci/ci_install_dependency.sh

@@ -47,8 +47,8 @@ $PIP_CMD install -e "python[dev]" --extra-index-url https://download.pytorch.org
 
 if [ "$IS_BLACKWELL" = "1" ]; then
     # TODO auto determine sgl-kernel version
-    SGL_KERNEL_VERSION=0.3.2
-    $PIP_CMD install https://github.com/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sgl_kernel-${SGL_KERNEL_VERSION}-cp39-abi3-manylinux2014_x86_64.whl --force-reinstall $PIP_INSTALL_SUFFIX
+    SGL_KERNEL_VERSION=0.3.8
+    $PIP_CMD install https://github.com/sgl-project/whl/releases/download/v${SGL_KERNEL_VERSION}/sgl_kernel-${SGL_KERNEL_VERSION}+cu128-cp310-abi3-manylinux2014_x86_64.whl --force-reinstall $PIP_INSTALL_SUFFIX
 fi
 
 # Show current packages

test/srt/run_suite.py

@@ -139,6 +139,7 @@ suites = {
     ],
     "per-commit-8-gpu-b200": [
         # add more here
+        TestFile("test_gpt_oss_4gpu.py", 600),
     ],
     "per-commit-4-gpu-deepep": [
         TestFile("ep/test_deepep_small.py", 531),

test/srt/test_gpt_oss_4gpu.py

@@ -9,10 +9,7 @@ class TestGptOss4Gpu(BaseTestGptOss):
             model_variant="120b",
             quantization="bf16",
             expected_score_of_reasoning_effort={
-                "low": 0.61,
-                # remove to speed up
-                # "medium": 0.61,
-                # "high": 0.61,
+                "low": 0.60,
             },
             other_args=["--tp", "4", "--cuda-graph-max-bs", "200"],
         )
@@ -22,10 +19,7 @@ class TestGptOss4Gpu(BaseTestGptOss):
             model_variant="120b",
             quantization="mxfp4",
             expected_score_of_reasoning_effort={
-                "low": 0.61,
-                # remove to speed up
-                # "medium": 0.61,
-                # "high": 0.61,
+                "low": 0.60,
             },
             other_args=[
                 "--tp",