[CI] Collapse the EAGLE launch matrix and the scoring engine boots on the per-commit runners (#33756)

python/sglang/multimodal_gen/runtime/layers/quantization/mxfp4.py

@@ -16,7 +16,7 @@ from sglang.multimodal_gen.runtime.models.parameter import (
     PerTensorScaleParameter,
 )
 from sglang.srt.layers.quantization.utils import is_layer_skipped
-from sglang.srt.utils import is_hip, mxfp_supported
+from sglang.srt.utils import is_gfx95_supported, is_hip
 
 logger = logging.getLogger(__name__)
 _is_hip = is_hip()
@@ -70,7 +70,7 @@ class Mxfp4Config(QuantizationConfig):
 
     @classmethod
     def get_min_capability(cls) -> int:
-        return 95  # gfx95x, Note: mxfp_supported() is a better check
+        return 95  # gfx95x, Note: is_gfx95_supported() is a better check
 
     @classmethod
     def get_config_filenames(cls) -> list[str]:
@@ -165,7 +165,7 @@ class Mxfp4LinearMethod(LinearMethodBase):
         - Packed uint8 (2 FP4 values per byte)
         - E8M0 scales (one per 32-element block)
         """
-        if not mxfp_supported():
+        if not is_gfx95_supported():
             platform = "unknown"
             if _is_hip:
                 try:
@@ -217,7 +217,7 @@ class Mxfp4LinearMethod(LinearMethodBase):
         bias: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
 
-        if not mxfp_supported():
+        if not is_gfx95_supported():
             raise RuntimeError(
                 "MXFP4 inference requires ROCm and MI350+ (gfx95x). "
                 "Current platform not supported."

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

@@ -58,13 +58,12 @@ from sglang.srt.utils import (
     cpu_has_amx_support,
     is_cpu,
     is_cuda,
-    is_hip,
+    is_gfx95_supported,
     is_mps,
     is_npu,
-    mxfp_supported,
 )
 
-_is_mxfp_supported = mxfp_supported()
+_is_gfx95_supported = is_gfx95_supported()
 
 if TYPE_CHECKING:
     from sglang.srt.layers.moe.topk import TopKOutput
@@ -102,7 +101,7 @@ BASE_QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
 }
 
 
-if is_cpu() or is_cuda() or (_is_mxfp_supported and is_hip()):
+if is_cpu() or is_cuda() or _is_gfx95_supported:
     BASE_QUANTIZATION_METHODS.update(
         {
             "mxfp4": Mxfp4Config,

python/sglang/srt/layers/quantization/mxfp4.py

@@ -59,7 +59,6 @@ from sglang.srt.utils import (
     is_sm100_supported,
     is_sm120_supported,
     is_triton_kernels_available,
-    mxfp_supported,
     next_power_of_2,
     round_up,
     set_weight_attrs,
@@ -256,7 +255,7 @@ class Mxfp4Config(QuantizationConfig):
         is_checkpoint_mxfp4_serialized = "mxfp4" in quant_method
 
         if _is_hip:
-            if mxfp_supported():
+            if is_gfx95_supported():
                 return cls(
                     is_checkpoint_mxfp4_serialized=is_checkpoint_mxfp4_serialized
                 )

python/sglang/srt/layers/quantization/quark/schemes/quark_w4a4_mxfp4.py

@@ -16,7 +16,7 @@ from sglang.srt.layers.quantization.fp8 import Fp8Config, Fp8LinearMethod
 from sglang.srt.layers.quantization.online_quantization import CopyNumelCounter
 from sglang.srt.layers.quantization.quark.schemes import QuarkLinearScheme
 from sglang.srt.utils import is_hip
-from sglang.srt.utils.common import direct_register_custom_op, mxfp_supported
+from sglang.srt.utils.common import direct_register_custom_op, is_gfx95_supported
 
 _is_hip = is_hip()
 if _is_hip:
@@ -180,7 +180,7 @@ class QuarkW4A4MXFP4(QuarkLinearScheme):
         self.dequantization_config = dequantization_config
 
         if not self.is_checkpoint_mxfp4_serialized:
-            if not mxfp_supported():
+            if not is_gfx95_supported():
                 raise NotImplementedError(
                     "Online MXFP4 quantization requires an AMD ROCm device with "
                     "FP4 hardware support (gfx95x, e.g. MI355x)."

python/sglang/srt/layers/quantization/quark/schemes/quark_w4a4_mxfp4_moe.py

@@ -24,7 +24,7 @@ from sglang.srt.utils import (
     is_hip,
     set_weight_attrs,
 )
-from sglang.srt.utils.common import mxfp_supported
+from sglang.srt.utils.common import is_gfx95_supported
 
 if TYPE_CHECKING:
     from sglang.srt.layers.moe.token_dispatcher import (
@@ -79,7 +79,7 @@ class QuarkW4A4MXFp4MoE(QuarkMoEScheme):
         self.with_bias = False
 
         if not self.is_checkpoint_mxfp4_serialized:
-            if not mxfp_supported():
+            if not is_gfx95_supported():
                 raise NotImplementedError(
                     "Online MXFP4 quantization for MoE layers requires an AMD ROCm "
                     "device with FP4 hardware support (gfx95x, e.g. MI355x)."

python/sglang/srt/utils/common.py

@@ -1023,21 +1023,11 @@ def set_cuda_arch():
         )
 
 
-def mxfp_supported():
-    """
-    Returns whether the current platform supports MX types.
-    """
-    if torch.version.hip:
-        gcn_arch = torch.cuda.get_device_properties(0).gcnArchName
-        return any(gfx in gcn_arch for gfx in ["gfx95"])
-    else:
-        return False
-
-
 @lru_cache(maxsize=1)
 def is_gfx95_supported():
-    """
-    Returns whether the current platform supports MX types.
+    """Whether the device is an AMD gfx95 GPU (the MX-capable ROCm arch).
+
+    False on every non-HIP build, so callers do not need their own is_hip().
     """
     if torch.version.hip:
         gcn_arch = torch.cuda.get_device_properties(0).gcnArchName

python/sglang/test/kits/spec_server_kits.py

@@ -386,8 +386,8 @@ class SpecLogprobKit:
         with ThreadPoolExecutor(8) as executor:
             list(executor.map(func, args))
 
-    def test_logprob_spec_v2_match(self):
-        """Verify spec v2 decode logprobs match prefill scoring logprobs."""
+    def test_logprob_decode_match_prefill(self):
+        """Decode logprobs from the spec path must match prefill scoring."""
         top_k = 5
         probe_token_ids = [1, 2, 10, 100, 1000]
         prompts = [