pytest skips for tests requiring certain ops (#411)

* add pytest skips around tests that require certain ops to be installed

tests/unit/test_adam_acuracy.py

@@ -6,8 +6,12 @@ import numpy as np
 import pytest
 import copy
 
+import deepspeed
 from deepspeed.ops.adam import DeepSpeedCPUAdam
 
+if not deepspeed.ops.__installed_ops__['cpu-adam']:
+    pytest.skip("cpu-adam is not installed", allow_module_level=True)
+
 
 def check_equal(first, second, atol=1e-2, verbose=False):
     x = first.detach().numpy()

tests/unit/test_checkpointing.py

 import torch
+
 import torch.distributed as dist
 
 import deepspeed
@@ -151,6 +152,8 @@ def checkpoint_correctness_verification(args,
         compare_lr_scheduler_states(trained_model, loaded_model)
 
 
+@pytest.mark.skipif(not deepspeed.ops.__installed_ops__['lamb'],
+                    reason="lamb is not installed")
 def test_checkpoint_unfused_optimizer(tmpdir):
     config_dict = {
         "train_batch_size": 2,
@@ -264,6 +267,9 @@ def test_checkpoint_fused_optimizer(tmpdir):
                               'deepspeed_adam'),
                          ])
 def test_checkpoint_zero_optimizer(tmpdir, zero_stage, use_cpu_offload, adam_optimizer):
+    if use_cpu_offload and not deepspeed.ops.__installed_ops__['cpu-adam']:
+        pytest.skip("cpu-adam is not installed")
+
     config_dict = {
         "train_batch_size": 2,
         "steps_per_print": 1,
@@ -320,6 +326,9 @@ def test_checkpoint_zero_no_optimizer(tmpdir,
                                       zero_stage,
                                       use_cpu_offload,
                                       adam_optimizer):
+    if use_cpu_offload and not deepspeed.ops.__installed_ops__['cpu-adam']:
+        pytest.skip("cpu-adam is not installed")
+
     config_dict = {
         "train_batch_size": 2,
         "steps_per_print": 1,
@@ -379,6 +388,9 @@ def test_checkpoint_zero_no_optimizer(tmpdir,
                               'deepspeed_adam'),
                          ])
 def test_checkpoint_lr_scheduler(tmpdir, zero_stage, use_cpu_offload, adam_optimizer):
+    if use_cpu_offload and not deepspeed.ops.__installed_ops__['cpu-adam']:
+        pytest.skip("cpu-adam is not installed")
+
     config_dict = {
         "train_batch_size": 2,
         "steps_per_print": 1,
@@ -450,6 +462,9 @@ def test_checkpoint_lr_scheduler(tmpdir, zero_stage, use_cpu_offload, adam_optim
                               'deepspeed_adam'),
                          ])
 def test_checkpoint_no_lr_scheduler(tmpdir, zero_stage, use_cpu_offload, adam_optimizer):
+    if use_cpu_offload and not deepspeed.ops.__installed_ops__['cpu-adam']:
+        pytest.skip("cpu-adam is not installed")
+
     config_dict = {
         "train_batch_size": 2,
         "steps_per_print": 1,

tests/unit/test_cuda_backward.py

@@ -12,9 +12,13 @@ from modelingpreln import BertEncoder as BertEncoderPreln
 from modeling import BertEncoder as BertEncoderPostln
 from modeling import BertConfig, BertLayerNorm
 from deepspeed import DeepSpeedTransformerLayer, DeepSpeedTransformerConfig
+import deepspeed
 
 import sys
 
+if not deepspeed.ops.__installed_ops__['transformer']:
+    pytest.skip("transformer kernels are not installed", allow_module_level=True)
+
 
 def check_equal(first, second, atol=1e-2, verbose=False):
     diction_x = {}

tests/unit/test_cuda_forward.py

@@ -12,9 +12,13 @@ from modelingpreln import BertEncoder as BertEncoderPreln
 from modeling import BertEncoder as BertEncoderPostln
 from modeling import BertLayerNorm, BertConfig
 from deepspeed import DeepSpeedTransformerLayer, DeepSpeedTransformerConfig
+import deepspeed
 
 import sys
 
+if not deepspeed.ops.__installed_ops__['transformer']:
+    pytest.skip("transformer kernels are not installed", allow_module_level=True)
+
 
 def check_equal(first, second, atol=1e-2, verbose=False):
     if verbose:

tests/unit/test_dynamic_loss_scale.py

@@ -8,6 +8,9 @@ import numpy as np
 from common import distributed_test
 from simple_model import SimpleModel, args_from_dict
 
+lamb_available = pytest.mark.skipif(not deepspeed.ops.__installed_ops__['lamb'],
+                                    reason="lamb is not installed")
+
 
 def run_model_step(model, gradient_list):
     for value in gradient_list:
@@ -165,6 +168,7 @@ def test_fused_some_overflow(tmpdir):
     _test_fused_some_overflow(args)
 
 
+@lamb_available
 def test_unfused_no_overflow(tmpdir):
     config_dict = {
         "train_batch_size": 1,
@@ -208,6 +212,7 @@ def test_unfused_no_overflow(tmpdir):
     _test_unfused_no_overflow(args)
 
 
+@lamb_available
 def test_unfused_all_overflow(tmpdir):
     config_dict = {
         "train_batch_size": 1,
@@ -253,6 +258,7 @@ def test_unfused_all_overflow(tmpdir):
     _test_unfused_all_overflow(args)
 
 
+@lamb_available
 def test_unfused_some_overflow(tmpdir):
     config_dict = {
         "train_batch_size": 1,

tests/unit/test_fp16.py

@@ -8,7 +8,11 @@ import os
 from common import distributed_test
 from simple_model import SimpleModel, SimpleOptimizer, random_dataloader, args_from_dict
 
+lamb_available = pytest.mark.skipif(not deepspeed.ops.__installed_ops__['lamb'],
+                                    reason="lamb is not installed")
 
+
+@lamb_available
 def test_lamb_fp32_grad_clip(tmpdir):
     config_dict = {
         "train_batch_size": 2,
@@ -44,6 +48,7 @@ def test_lamb_fp32_grad_clip(tmpdir):
     _test_lamb_fp32_grad_clip(args=args, model=model, hidden_dim=hidden_dim)
 
 
+@lamb_available
 def test_lamb_fp16_basic(tmpdir):
     config_dict = {
         "train_batch_size": 2,
@@ -81,6 +86,7 @@ def test_lamb_fp16_basic(tmpdir):
     _test_lamb_fp16_basic(args=args, model=model, hidden_dim=hidden_dim)
 
 
+@lamb_available
 def test_lamb_fp16_empty_grad(tmpdir):
     config_dict = {
         "train_batch_size": 2,
@@ -228,6 +234,8 @@ def test_adamw_fp16_empty_grad(tmpdir):
                               True),
                          ])
 def test_adam_fp16_zero_onecycle_compatibility(tmpdir, zero_stage, use_cpu_offload):
+    if use_cpu_offload and not deepspeed.ops.__installed_ops__['cpu-adam']:
+        pytest.skip("cpu-adam is not installed")
     config_dict = {
         "train_batch_size": 1,
         "steps_per_print": 1,
@@ -294,6 +302,8 @@ def test_adam_fp16_zero_onecycle_compatibility(tmpdir, zero_stage, use_cpu_offlo
                               True),
                          ])
 def test_zero_static_scale(tmpdir, zero_stage, use_cpu_offload):
+    if use_cpu_offload and not deepspeed.ops.__installed_ops__['cpu-adam']:
+        pytest.skip("cpu-adam is not installed")
     config_dict = {
         "train_batch_size": 4,
         "steps_per_print": 1,
@@ -392,6 +402,8 @@ def test_zero_static_scale_deprecated_format(tmpdir):
                               True),
                          ])
 def test_zero_allow_untested_optimizer(tmpdir, zero_stage, use_cpu_offload):
+    if use_cpu_offload and not deepspeed.ops.__installed_ops__['cpu-adam']:
+        pytest.skip("cpu-adam is not installed")
     config_dict = {
         "train_batch_size": 4,
         "steps_per_print": 1,
@@ -430,6 +442,8 @@ def test_zero_allow_untested_optimizer(tmpdir, zero_stage, use_cpu_offload):
                               True),
                          ])
 def test_zero_empty_partition(tmpdir, zero_stage, use_cpu_offload):
+    if use_cpu_offload and not deepspeed.ops.__installed_ops__['cpu-adam']:
+        pytest.skip("cpu-adam is not installed")
     config_dict = {
         "train_micro_batch_size_per_gpu": 1,
         "gradient_accumulation_steps": 1,
@@ -500,6 +514,7 @@ def test_adam_amp_basic(tmpdir):
     _test_adam_amp_basic(args=args, model=model, hidden_dim=hidden_dim)
 
 
+@lamb_available
 def test_lamb_amp_basic(tmpdir):
     config_dict = {
         "train_batch_size": 2,

tests/unit/test_sparse_attention.py

@@ -5,6 +5,10 @@
 
 import pytest
 import torch
+import deepspeed
+
+if not deepspeed.ops.__installed_ops__['sparse-attn']:
+    pytest.skip("cpu-adam is not installed", allow_module_level=True)
 
 
 def test_sparse_attention_module_availability():