trunk/664bf32ef4ad84181302771424f7188ba4ce6a55: Graph break on CUDA manual_seed in Dynamo (#185761)

CUDA seed mutators such as torch.cuda.manual_seed and

manual_seed_all were treated as in-graph Python functions by Dynamo.

They appeared in the Dynamo FX graph, but they are not ATen operations,

so AOT/Inductor dropped the side effect before codegen. A compiled

function that reset the CUDA seed before a random operation would then

advance the generator across calls instead of replaying from the seed.

Move the CUDA manual_seed aliases onto the same SkipFunctionVariable

path already used for torch.manual_seed and RNG state setters/getters.

This forces a graph break at the side-effecting seed mutation, keeping

compiled random calls reproducible without teaching AOT/Inductor to

represent Python RNG state mutation.

The previous alternative in PR

Generated by my agent

Test Plan:

• python test/dynamo/test_functions.py FunctionTests.test_cuda_manual_seed


• python test/dynamo/test_trace_rules.py TraceRuleTests.test_cuda_manual_seed_functions_graph_break


• python test/dynamo/test_functions.py FunctionTests.test_manual_seed


• lintrunner -a

Pull Request resolved: