[Platform] Route pin memory availability through current_platform (#28113)

Co-authored-by: N3u0ns <N3u0ns@users.noreply.github.com>
Co-authored-by: Alex Nails <alex.nails@radixark.ai>

python/sglang/srt/platforms/cpu.py

@@ -125,9 +125,9 @@ class CpuSRTPlatform(CpuDeviceMixin, SRTPlatform):
 
     supports_fp8 / support_cuda_graph / support_piecewise_cuda_graph keep the
     conservative SRTPlatform defaults (all False), so they are not repeated
-    here. Only is_pin_memory_available is overridden: the base defaults to
-    True, but CPU has no GPU to pin host memory to.
+    here. is_pin_memory_available is repeated for explicitness: CPU has no GPU
+    to pin host memory to.
     """
 
-    def is_pin_memory_available(self) -> bool:
+    def is_pin_memory_available(self, device=None) -> bool:
         return False

python/sglang/srt/platforms/cuda.py

@@ -52,6 +52,11 @@ class CudaDeviceMixin(DeviceMixin):
     def get_available_memory(self, device_id: int = 0) -> tuple[int, int]:
         return torch.cuda.mem_get_info(device_id)
 
+    def is_pin_memory_available(self, device=None) -> bool:
+        if device is not None and str(device) == "cpu":
+            return False
+        return True
+
     def get_torch_distributed_backend_str(self) -> str:
         return "nccl"
 

python/sglang/srt/platforms/device_mixin.py

@@ -159,6 +159,10 @@ class DeviceMixin:
         """[Active] Get current peak memory usage in bytes."""
         raise NotImplementedError
 
+    def is_pin_memory_available(self, device=None) -> bool:
+        """[Active] Whether pinned host memory is available for a target device."""
+        return False
+
     # ------------------------------------------------------------------
     # Planned methods — reserved interface.  Core still uses hardcoded
     # calls (e.g. torch.cuda.*).  OOT implementations will NOT take

python/sglang/srt/platforms/interface.py

@@ -103,10 +103,6 @@ class SRTPlatform(DeviceMixin):
         """Whether this platform supports FP8 quantization."""
         return False
 
-    def is_pin_memory_available(self) -> bool:
-        """Whether pinned memory is available on this platform."""
-        return True
-
     def support_cuda_graph(self) -> bool:
         """Whether this platform supports device graph capture and replay.
         Controls CUDA graph (CudaGraphRunner) for the decode path.

python/sglang/srt/utils/common.py

@@ -539,11 +539,7 @@ def get_available_gpu_memory(
 
 
 def is_pin_memory_available(device=None) -> bool:
-    if not torch.cuda.is_available():
-        return False
-    if device is not None and str(device) == "cpu":
-        return False
-    return True
+    return current_platform.is_pin_memory_available(device)
 
 
 def get_dispatch_device_backend():