Avoid using non-polymorphic CUDA graph APIs internally #2778

Activity

1. leofang

   added

   enhancementAny code-related improvementsAny code-related improvements

   cuda.coreEverything related to the cuda.core moduleEverything related to the cuda.core module

   on Sep 8, 2026
2. leofang

   added this to the cuda.core 1.3.0 milestone on Sep 8, 2026
3. leofang commented on Sep 8, 2026

   

   leofang

   on Sep 8, 2026

   MemberAuthor

   More actions

   Deprecated driver-side surface, by node family. Replacement is the polymorphic API (cuGraphAddNode, cuGraphNodeGetParams, cuGraphNodeSetParams, cuGraphExecNodeSetParams on CUgraphNodeParams).

   Node-type-specific APIs

   • Kernel: cuGraphAddKernelNode, cuGraphKernelNodeGetParams, cuGraphKernelNodeSetParams, cuGraphExecKernelNodeSetParams
   • Memcpy: cuGraphAddMemcpyNode, cuGraphMemcpyNodeGetParams, cuGraphMemcpyNodeSetParams, cuGraphExecMemcpyNodeSetParams
   • Memset: cuGraphAddMemsetNode, cuGraphMemsetNodeGetParams, cuGraphMemsetNodeSetParams, cuGraphExecMemsetNodeSetParams
   • Host: cuGraphAddHostNode, cuGraphHostNodeGetParams, cuGraphHostNodeSetParams, cuGraphExecHostNodeSetParams
   • Child graph: cuGraphAddChildGraphNode, cuGraphChildGraphNodeGetGraph, cuGraphExecChildGraphNodeSetParams
   • Empty: cuGraphAddEmptyNode
   • Event record: cuGraphAddEventRecordNode, cuGraphEventRecordNodeGetEvent, cuGraphEventRecordNodeSetEvent, cuGraphExecEventRecordNodeSetEvent
   • Event wait: cuGraphAddEventWaitNode, cuGraphEventWaitNodeGetEvent, cuGraphEventWaitNodeSetEvent, cuGraphExecEventWaitNodeSetEvent
   • External-semaphore signal: cuGraphAddExternalSemaphoresSignalNode, cuGraphExternalSemaphoresSignalNodeGetParams, cuGraphExternalSemaphoresSignalNodeSetParams, cuGraphExecExternalSemaphoresSignalNodeSetParams
   • External-semaphore wait: cuGraphAddExternalSemaphoresWaitNode, cuGraphExternalSemaphoresWaitNodeGetParams, cuGraphExternalSemaphoresWaitNodeSetParams, cuGraphExecExternalSemaphoresWaitNodeSetParams
   • Batch memop: cuGraphAddBatchMemOpNode, cuGraphBatchMemOpNodeGetParams, cuGraphBatchMemOpNodeSetParams, cuGraphExecBatchMemOpNodeSetParams
   • Mem alloc/free: cuGraphAddMemAllocNode, cuGraphMemAllocNodeGetParams, cuGraphAddMemFreeNode, cuGraphMemFreeNodeGetParams

   Legacy parameter type aliases — use the versioned ..._v2 types that CUgraphNodeParams already references:
   CUDA_KERNEL_NODE_PARAMS, CUDA_MEMSET_NODE_PARAMS, CUDA_HOST_NODE_PARAMS, CUDA_EXT_SEM_SIGNAL_NODE_PARAMS, CUDA_EXT_SEM_WAIT_NODE_PARAMS, CUDA_BATCH_MEM_OP_NODE_PARAMS, CUDA_MEM_ALLOC_NODE_PARAMS

   Kernel-node attribute aliases — CUkernelNodeAttrID and the CU_KERNEL_NODE_ATTRIBUTE_* enumerators are deprecated in favor of CUlaunchAttributeID and the corresponding CU_LAUNCH_ATTRIBUTE_*.

   -- Leo's bot
4. Andy-Jost commented on Sep 8, 2026

   

   Andy-Jost

   on Sep 8, 2026

   Contributor

   More actions

   The polymorphic APIs were added in CUDA 12.2 and 12.3 and cuda.core 1.x still supports drivers all the way back to 12.0. There are two viable migration paths for updating a non-polymorphic call, A(), to a polymorphic one, B():

   1. Wait for CUDA 14 / cuda.core 2.0, then directly replace A() with B() throughout.
   2. Migrate early. The code goes through an intermediate state with extra runtime checks, e.g., A() if version() < (12, 2, 0) else B().

   If there is no compelling reason to update early, the first option should be preferred: fewer changes, simpler code, marginally more performant.

   👍React with 👍1Reacted by Leo Fang
5. leofang commented on Sep 9, 2026

   

   leofang

   on Sep 9, 2026

   MemberAuthor

   More actions

   Agreed to move this to cuda.core v2.0.0.
6. leofang

   added

   breakingBreaking changes are introducedBreaking changes are introduced

   on Sep 9, 2026
7. leofang

   modified the milestones: cuda.core 1.3.0, cuda.core 2.0.0 on Sep 9, 2026