remove megatron-lm, no longer pip installable (microsoft#3389)

* remove megatron-lm, no longer pip installable * Add skips to tests that require megatron-lm and can't be run currently. * formatting * Formatting --------- Co-authored-by: Logan Adams <[email protected]>
kazemnejad · Apr 28, 2023 · a094c97 · a094c97
1 parent f7d71ec
commit a094c97
Show file tree

Hide file tree

Showing 4 changed files with 10 additions and 1 deletion.
diff --git a/requirements/requirements-dev.txt b/requirements/requirements-dev.txt
@@ -2,7 +2,6 @@ clang-format==16.0.2
 docutils<0.18
 future
 importlib-metadata>=4
-megatron-lm==1.1.5
 pre-commit>=2.20.0
 pytest
 pytest-forked

diff --git a/tests/unit/compression/test_compression.py b/tests/unit/compression/test_compression.py
@@ -224,6 +224,7 @@ def test_linear_layer_compress(self, tmpdir):
         assert isinstance(compressed_model.layer[0].attention.self.key, LinearLayer_Compress)
         assert isinstance(compressed_model.layer[0].attention.self.value, LinearLayer_Compress)
 
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_mpu_compress(self, tmpdir):
         if not required_maximum_torch_version(major_version=1, minor_version=13):
             pytest.skip("megatron not compatible with torch >1.13")

diff --git a/tests/unit/model_parallelism/test_configurable_parallel_mp.py b/tests/unit/model_parallelism/test_configurable_parallel_mp.py
@@ -60,6 +60,7 @@ def inputs(self, bs=1, seq_len=20):
 class TestConfigurableMP(ConfigurableMP):
 
     @pytest.mark.world_size(1)
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_gpt2_basic(self, tmpdir, inputs):
         args_defaults = {
             'num_layers': 2,
@@ -87,6 +88,7 @@ def test_gpt2_basic(self, tmpdir, inputs):
                               atol=1e-07), f"Baseline output {baseline} is not equal to save-then-load output {test}"
 
     @pytest.mark.world_size(2)
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_gpt2_mp2_no_resize(self, tmpdir, inputs):
         args_defaults = {
             'num_layers': 2,
@@ -148,6 +150,7 @@ def run(self, inputs, class_tmpdir):
 class TestConfigurableResizeMP(ConfigurableMP):
     world_size = [1, 4]
 
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test(self, baseline_mp2, inputs, class_tmpdir):
         args_defaults = {
             'num_layers': 2,

diff --git a/tests/unit/model_parallelism/test_configurable_parallel_pp.py b/tests/unit/model_parallelism/test_configurable_parallel_pp.py
@@ -69,6 +69,7 @@ class TestConfigurablePP(ConfigurablePP):
     pp_size = 2
     world_size = 4  # mp_size * pp_size
 
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_pp_basic(self, inputs, tmpdir):
         # basic test case, mp_size=2, pp_size=2, verify ckpt saving/loading.
         args_defaults = {
@@ -234,30 +235,35 @@ def _test(self, inputs, class_tmpdir, checkpoint_tag, mp_size, pp_size, mp_resiz
     # These tests are divided by baseline model worldsize and test model worldsize
     @pytest.mark.world_size(1)
     @pytest.mark.parametrize("mp_size, pp_size, mp_resize, pp_resize", [(1, 2, 1, 1)])
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_world_size_2to1(self, inputs, class_tmpdir, checkpoint_tag, baseline_ws2, mp_size, pp_size, mp_resize,
                              pp_resize):
         self._test(inputs, class_tmpdir, checkpoint_tag, mp_size, pp_size, mp_resize, pp_resize)
 
     @pytest.mark.world_size(1)
     @pytest.mark.parametrize("mp_size, pp_size, mp_resize, pp_resize", [(2, 2, 1, 1)])
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_world_size_4to1(self, inputs, class_tmpdir, checkpoint_tag, baseline_ws4, mp_size, pp_size, mp_resize,
                              pp_resize):
         self._test(inputs, class_tmpdir, checkpoint_tag, mp_size, pp_size, mp_resize, pp_resize)
 
     @pytest.mark.world_size(2)
     @pytest.mark.parametrize("mp_size, pp_size, mp_resize, pp_resize", [(2, 2, 2, 1)])
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_world_size_4to2(self, inputs, class_tmpdir, checkpoint_tag, baseline_ws4, mp_size, pp_size, mp_resize,
                              pp_resize):
         self._test(inputs, class_tmpdir, checkpoint_tag, mp_size, pp_size, mp_resize, pp_resize)
 
     @pytest.mark.world_size(4)
     @pytest.mark.parametrize("mp_size, pp_size, mp_resize, pp_resize", [(1, 1, 2, 2)])
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_world_size_1to4(self, inputs, class_tmpdir, checkpoint_tag, baseline_ws1, mp_size, pp_size, mp_resize,
                              pp_resize):
         self._test(inputs, class_tmpdir, checkpoint_tag, mp_size, pp_size, mp_resize, pp_resize)
 
     @pytest.mark.world_size(4)
     @pytest.mark.parametrize("mp_size, pp_size, mp_resize, pp_resize", [(1, 2, 1, 4), (2, 1, 2, 2)])
+    @pytest.mark.skip(reason="megatron-lm is currently broken so this test cannot be run.")
     def test_world_size_2to4(self, inputs, class_tmpdir, checkpoint_tag, baseline_ws2, mp_size, pp_size, mp_resize,
                              pp_resize):
         self._test(inputs, class_tmpdir, checkpoint_tag, mp_size, pp_size, mp_resize, pp_resize)