nv: finalize nvdev (#11117)

* nv: finalize nvdev

* typo

tinygrad/runtime/ops_amd.py

@@ -853,8 +853,3 @@ class AMDDevice(HCQCompiled):
         self.allocator._copyout(sqtt_buf:=memoryview(bytearray(wptr)), buf0)
         Compiled.profile_events += [ProfileSQTTEvent(self.device, i, bytes(sqtt_buf), bool((self.sqtt_itrace_se_mask >> i) & 0b1))]
     super()._at_profile_finalize()
-
-  def finalize(self):
-    try: self.synchronize() # Try to finalize device in any case.
-    except RuntimeError as e: print(f"{self.device} synchronization failed before finalizing: {e}")
-    if hasattr(self.iface, 'device_fini'): self.iface.device_fini()

tinygrad/runtime/ops_nv.py

@@ -482,6 +482,8 @@ class PCIIface(PCIIfaceBase):
     res = self.dev_impl.gsp.rpc_rm_control(obj, cmd, params, self.root)
     return type(params).from_buffer_copy(res) if params is not None else None
 
+  def device_fini(self): self.dev_impl.fini()
+
 class NVDevice(HCQCompiled[NVSignal]):
   devices: ClassVar[list[HCQCompiled]] = []
   signal_pages: ClassVar[list[HCQBuffer]] = []

tinygrad/runtime/support/hcq.py

@@ -434,6 +434,13 @@ class HCQCompiled(Compiled, Generic[SignalType]):
       except Exception: errs += f"\n{iface_t.__name__}: {traceback.format_exc()}"
     raise RuntimeError(f"Cannot find a usable interface for {type(self).__name__[:-6]}:{self.device_id}:\n{errs}")
 
+  def finalize(self):
+    try: self.synchronize() # Try to finalize device in any case.
+    except RuntimeError as e: print(f"{self.device} synchronization failed before finalizing: {e}")
+
+    # If the device has an interface, call its device_fini method to clean up resources.
+    if hasattr(self, 'iface') and hasattr(self.iface, 'device_fini'): self.iface.device_fini()
+
 class HCQBuffer:
   def __init__(self, va_addr:sint, size:int, texture_info:Any=None, meta:Any=None, _base:HCQBuffer|None=None, view:MMIOInterface|None=None):
     self.va_addr, self.size, self.texture_info, self.meta, self._base, self.view = va_addr, size, texture_info, meta, _base, view

tinygrad/runtime/support/nv/nvdev.py

@@ -90,6 +90,8 @@ class NVDev(PCIDevImplBase):
     for ip in [self.flcn, self.gsp]: ip.init_sw()
     for ip in [self.flcn, self.gsp]: ip.init_hw()
 
+  def fini(self): System.pci_reset(self.devfmt) # Reset the device to clean up resources. TODO: Consider a warm start process.
+
   def reg(self, reg:str) -> NVReg: return self.__dict__[reg]
   def wreg(self, addr, value):
     self.mmio[addr // 4] = value