[Perf] Add opt-in SM100 Oink RMSNorm custom-op path (#31828)

Signed-off-by: Laura Wang <3700467+Laurawly@users.noreply.github.com>
Co-authored-by: Lu Fang <30275821+houseroad@users.noreply.github.com>

vllm/envs.py

@@ -97,6 +97,7 @@ if TYPE_CHECKING:
     VLLM_SKIP_P2P_CHECK: bool = False
     VLLM_DISABLED_KERNELS: list[str] = []
     VLLM_DISABLE_PYNCCL: bool = False
+    VLLM_USE_OINK_OPS: bool = False
     VLLM_ROCM_USE_AITER: bool = False
     VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
     VLLM_ROCM_USE_AITER_LINEAR: bool = True
@@ -896,6 +897,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_DISABLE_PYNCCL": lambda: (
         os.getenv("VLLM_DISABLE_PYNCCL", "False").lower() in ("true", "1")
     ),
+    # Optional: enable external Oink custom ops (e.g., Blackwell RMSNorm).
+    # Disabled by default.
+    "VLLM_USE_OINK_OPS": lambda: (
+        os.getenv("VLLM_USE_OINK_OPS", "False").lower() in ("true", "1")
+    ),
     # Disable aiter ops unless specifically enabled.
     # Acts as a parent switch to enable the rest of the other operations.
     "VLLM_ROCM_USE_AITER": lambda: (