chore: drop P100 from the CI pipeline

The P100 Jenkins agent (label `ephemeral-linux-gpu`) is being deleted, so
move everything that ran on it over to T4x2:

- Remove the dedicated `Test on P100` stage. T4x2 remains the GPU test bed.
- Repoint the `Build GPU Image` / `Diff GPU Image` stages at
  `ephemeral-linux-gpu-t4x2`.
- Drop the now-dead `p100_exempt` decorator and its stale cuDNN/sm_60
  comments. Those tests were only ever skipped on Pascal hardware, so they
  now run unconditionally on T4.

Co-Authored-By: Claude <noreply@anthropic.com>

Author: calderjo

Jenkinsfile

@@ -51,8 +51,8 @@ pipeline {
           }
         }
         stage('GPU') {
-          agent { label 'ephemeral-linux-gpu' }
-          stages {  
+          agent { label 'ephemeral-linux-gpu-t4x2' }
+          stages {
             stage('Build GPU Image') {
               options {
                 timeout(time: 4324, unit: 'MINUTES')
@@ -135,23 +135,6 @@ pipeline {
             }
           }
         }
-        stage('Test on P100') {
-          agent { label 'ephemeral-linux-gpu' }
-          options {
-            timeout(time: 40, unit: 'MINUTES')
-          }
-          steps {
-            retry(2) {
-              sh '''#!/bin/bash
-                set -exo pipefail
-
-                date
-                docker pull gcr.io/kaggle-private-byod/python:${PRETEST_TAG}
-                ./test --gpu --image gcr.io/kaggle-private-byod/python:${PRETEST_TAG}
-              '''
-            }
-          }
-        }
         stage('Test on T4x2') {
           agent { label 'ephemeral-linux-gpu-t4x2' }
           options {

tests/common.py

@@ -15,6 +15,4 @@ def isGPU():
     return os.path.isfile('/proc/driver/nvidia/version')
 
 gpu_test = unittest.skipIf(not isGPU(), 'Not running GPU tests')
-# b/342143152 P100s are slowly being unsupported in new release of popular ml tools such as RAPIDS. 
-p100_exempt = unittest.skipIf(getAcceleratorName() == "Tesla P100-PCIE-16GB", 'Not running p100 exempt tests')
 tpu_test = unittest.skipIf(len(os.environ.get('ISTPUVM', '')) == 0, 'Not running TPU tests')

tests/test_cudf.py

@@ -1,11 +1,10 @@
 import unittest
 
-from common import gpu_test, p100_exempt
+from common import gpu_test
 
 
 class TestCudf(unittest.TestCase):
     @gpu_test
-    @p100_exempt # b/342143152: cuDL(>=24.4v) is inompatible with p100 GPUs.
     def test_cudf_dataframe_operations(self):
         import cudf
 

tests/test_cuml.py

@@ -1,11 +1,10 @@
 import unittest
 
-from common import gpu_test, p100_exempt
+from common import gpu_test
 
 
 class TestCuml(unittest.TestCase):
     @gpu_test
-    @p100_exempt # b/342143152: cuML(>=24.4v) is inompatible with p100 GPUs.
     def test_pca_fit_transform(self):
         import unittest
         import numpy as np

tests/test_fastai.py

@@ -3,8 +3,6 @@
 import fastai
 from fastai.tabular.all import *
 
-from common import p100_exempt
-
 
 class TestFastAI(unittest.TestCase):
   # Basic import
@@ -24,7 +22,6 @@ def test_torch_tensor(self):
 
     self.assertTrue(torch.all(a == b))
 
-  @p100_exempt
   def test_tabular(self):
     dls = TabularDataLoaders.from_csv(
         "/input/tests/data/train.csv",

tests/test_flax.py

@@ -8,8 +8,6 @@
 from flax import linen as nn
 from flax.training import train_state
 
-from common import p100_exempt
-
 
 class TestFlax(unittest.TestCase):
 
@@ -19,12 +17,6 @@ def test_pooling(self):
         y = nn.pooling.pool(x, 1., mul_reduce, (2, 2), (1, 1), 'VALID')
         np.testing.assert_allclose(y, np.full((1, 2, 2, 1), 2. ** 4))
 
-    # cuDNN 9.19 (pulled in by torch 2.11) dropped the Pascal (sm_60) kernels from
-    # libcudnn_ops/cnn/adv and ships PTX for sm_121 only, so nothing can be JIT'd
-    # down to sm_60 either. Every cuDNN convolution engine fails on P100 with
-    # CUDNN_STATUS_EXECUTION_FAILED. Not fixable here: cuDNN 9.19 is a hard
-    # requirement of torch 2.11, which comes from the Colab base image.
-    @p100_exempt
     def test_cnn(self):
         class CNN(nn.Module):
             @nn.compact

tests/test_keras.py

@@ -7,15 +7,7 @@
 
 import keras
 
-from common import p100_exempt
-
 class TestKeras(unittest.TestCase):
-    # cuDNN 9.19 (pulled in by torch 2.11) dropped the Pascal (sm_60) kernels from
-    # libcudnn_ops/cnn/adv and ships PTX for sm_121 only, so nothing can be JIT'd
-    # down to sm_60 either. Every cuDNN convolution engine fails on P100 with
-    # CUDNN_STATUS_EXECUTION_FAILED. Not fixable here: cuDNN 9.19 is a hard
-    # requirement of torch 2.11, which comes from the Colab base image.
-    @p100_exempt
     def test_train(self):
         path = '/input/tests/data/mnist.npz'
         with np.load(path) as f:

tests/test_keras_cv.py

@@ -4,16 +4,9 @@
 import keras
 import numpy as np
 
-from common import p100_exempt
 from utils.kagglehub import create_test_kagglehub_server
 
 class TestKerasCV(unittest.TestCase):
-    # cuDNN 9.19 (pulled in by torch 2.11) dropped the Pascal (sm_60) kernels from
-    # libcudnn_ops/cnn/adv and ships PTX for sm_121 only, so nothing can be JIT'd
-    # down to sm_60 either. Every cuDNN convolution engine fails on P100 with
-    # CUDNN_STATUS_EXECUTION_FAILED. Not fixable here: cuDNN 9.19 is a hard
-    # requirement of torch 2.11, which comes from the Colab base image.
-    @p100_exempt
     def test_inference(self):
         with create_test_kagglehub_server():
             classifier = keras_cv.models.ImageClassifier.from_preset(

tests/test_pytorch.py

@@ -4,7 +4,7 @@
 import torch.nn as tnn
 import torch.autograd as autograd
 
-from common import gpu_test, p100_exempt
+from common import gpu_test
 
 
 class TestPyTorch(unittest.TestCase):
@@ -16,7 +16,6 @@ def test_nn(self):
         linear_torch(data_torch)
 
     @gpu_test
-    @p100_exempt
     def test_linalg(self):
         A = torch.randn(3, 3).t().to('cuda')
         B = torch.randn(3).t().to('cuda')
@@ -25,7 +24,6 @@ def test_linalg(self):
         self.assertEqual(3, result.shape[0])
 
     @gpu_test
-    @p100_exempt
     def test_gpu_computation(self):
         cuda = torch.device('cuda')  
         a = torch.tensor([1., 2.], device=cuda)
@@ -35,7 +33,6 @@ def test_gpu_computation(self):
         self.assertEqual(torch.tensor([3.], device=cuda), result)
 
     @gpu_test
-    @p100_exempt
     def test_cuda_nn(self):
         # These throw if cuda is misconfigured
         tnn.GRUCell(10,10).cuda()

tests/test_pytorch_lightning.py

@@ -5,8 +5,6 @@
 import torch.nn.functional as F
 from torch.utils.data import DataLoader, TensorDataset
 
-from common import p100_exempt
-
 
 class LitDataModule(pl.LightningDataModule):
 
@@ -61,7 +59,6 @@ class TestPytorchLightning(unittest.TestCase):
     def test_version(self):
         self.assertIsNotNone(pl.__version__)
 
-    @p100_exempt
     def test_mnist(self):
         dm = LitDataModule()
         model = LitClassifier()