Patch for nightly test&bench #4840
Conversation
viclafargue
commented
Jul 29, 2022
•
viclafargue
commented
- Fix for MNMG TSVD (similar issue to cudaErrorContextIsDestroyed in RandomForest)
- Update the NVTX bench helper for the new nsys utility #4826
- MNMG Kmeans testing issue : modification of accuracy threshold
- MNMG KNNRegressor testing issue : modification of input for testing
- LabelEncoder documentation test issue : modification of pandas/cuDF display configuration
- RandomForest testing issue : adjust number of estimators to the number of workers
|
is this PR going into |
Just changed it to 22.10 |
| @@ -263,6 +266,7 @@ cdef class ForestInference_impl(): | |||
| self.handle = handle | |||
| self.forest_data = forest_variant(<forest32_t> NULL) | |||
| self.shape_str = NULL | |||
| self.mr = get_current_device_resource() | |||
There was a problem hiding this comment.
Some Python models are storing rmm::device_vector and other RMM objects or structs to containing these objects. These only have a pointer to their respective memory resources. Then depending on garbage collection, the memory resource (see memory_resource.pyx) can be released before objects depending on it for CUDA deallocation. This sometimes results in a segfault (visible when benchmarking for instance). Keeping a reference to the memory resource inside the model prevents its premature release (e.g. : in DeviceBuffer).
|
rerun tests |
2 similar comments
|
rerun tests |
|
rerun tests |
|
@gpucibot merge |
- Fix for MNMG TSVD (similar issue to [cudaErrorContextIsDestroyed in RandomForest](rapidsai#2632 (comment))) - rapidsai#4826 - MNMG Kmeans testing issue : modification of accuracy threshold - MNMG KNNRegressor testing issue : modification of input for testing - LabelEncoder documentation test issue : modification of pandas/cuDF display configuration - RandomForest testing issue : adjust number of estimators to the number of workers Authors: - Victor Lafargue (https://github.com/viclafargue) Approvers: - Dante Gama Dessavre (https://github.com/dantegd) URL: rapidsai#4840
