Avoid referencing the current device in LaunchConfig

This is strictly speaking not 100% safe:

cuda-python/cuda_core/cuda/core/experimental/_launch_config.py

Lines 68 to 86 in bf590e4

    
           # FIXME: Calling Device() strictly speaking is not quite right; we should instead 
        
           # look up the device from stream. We probably need to defer the checks related to 
        
           # device compute capability or attributes. 
        
           # thread block clusters are supported starting H100 
        
           if self.cluster is not None: 
        
               if not _use_ex: 
        
                   err, drvers = driver.cuDriverGetVersion() 
        
                   drvers_fmt = f" (got driver version {drvers})" if err == driver.CUresult.CUDA_SUCCESS else "" 
        
                   raise CUDAError(f"thread block clusters require cuda.bindings & driver 11.8+{drvers_fmt}") 
        
               cc = Device().compute_capability 
        
               if cc < (9, 0): 
        
                   raise CUDAError( 
        
                       f"thread block clusters are not supported on devices with compute capability < 9.0 (got {cc})" 
        
                   ) 
        
               self.cluster = cast_to_3_tuple("LaunchConfig.cluster", self.cluster) 
        
           if self.shmem_size is None: 
        
               self.shmem_size = 0 
        
           if self.cooperative_launch and not Device().properties.cooperative_launch: 
        
               raise CUDAError("cooperative kernels are not supported on this device")

because we could be changing the device after creating a launch config and before we use it to launch a kernel (on another device).

Navigation Menu

Repository navigation

Avoid referencing the current device in `LaunchConfig` #704

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	# FIXME: Calling Device() strictly speaking is not quite right; we should instead
	# look up the device from stream. We probably need to defer the checks related to
	# device compute capability or attributes.
	# thread block clusters are supported starting H100
	if self.cluster is not None:
	if not _use_ex:
	err, drvers = driver.cuDriverGetVersion()
	drvers_fmt = f" (got driver version {drvers})" if err == driver.CUresult.CUDA_SUCCESS else ""
	raise CUDAError(f"thread block clusters require cuda.bindings & driver 11.8+{drvers_fmt}")
	cc = Device().compute_capability
	if cc < (9, 0):
	raise CUDAError(
	f"thread block clusters are not supported on devices with compute capability < 9.0 (got {cc})"
	)
	self.cluster = cast_to_3_tuple("LaunchConfig.cluster", self.cluster)
	if self.shmem_size is None:
	self.shmem_size = 0
	if self.cooperative_launch and not Device().properties.cooperative_launch:
	raise CUDAError("cooperative kernels are not supported on this device")

Avoid referencing the current device in LaunchConfig #704

Description

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

Avoid referencing the current device in `LaunchConfig` #704