Merge pull request #26 from facebookresearch/25-maintenance

Skip coverage test for graph visualization

.circleci/config.yml

@@ -56,6 +56,11 @@ install_repo: &install_repo
       command: |
         pip install .
 
+install_dev_requirements: &install_dev_requirements
+  - run:
+      name: Install development packages
+      command: |
+        pip install -r requirements-dev.txt
 
 # This is an alias to run all unit tests possible on a platform.
 run_unittests: &run_unittests
@@ -85,6 +90,7 @@ jobs:
       - checkout
       - <<: *setup_venv
       - <<: *install_repo
+      - <<: *install_dev_requirements
 
       # Cache the venv directory that contains dependencies
       - restore_cache:

README.md

@@ -36,21 +36,21 @@ HTA provides the following features:
 
 HTA runs on Linux and Mac with Python >= 3.8.
 
-### Create a Conda environment (optional)
+### Setup a Conda environment (optional)
 
-To install Miniconda see [here](https://docs.conda.io/en/latest/miniconda.html).
+See [here](https://docs.conda.io/en/latest/miniconda.html) to install Miniconda.
 
-To create the environment `env_name`
+Create the environment `env_name`
 ``` bash
 conda create -n env_name
 ```
 
-To activate the environment
+Activate the environment
 ``` bash
 conda activate env_name
 ```
 
-To deactivate the environment
+Deactivate the environment
 ``` bash
 conda deactivate
 ```

docs/source/intro/installation.rst

@@ -5,13 +5,8 @@ We recommend using a Conda environment to install HTA. To install Anaconda, see
 `here <https://docs.anaconda.com/anaconda/install/index.html>`_. Holistic Trace
 Analysis runs on Linux and Mac with Python >= 3.8.
 
-**Get the HTA source code**
 
-.. code-block::
-
-  git clone https://github.com/facebookresearch/HolisticTraceAnalysis.git
-
-**Using a Conda environment**
+**Setup a Conda environment**
 
 .. code-block::
 
@@ -26,8 +21,18 @@ Analysis runs on Linux and Mac with Python >= 3.8.
 
 **Installing Holistic Trace Analysis**
 
-Execute the command below from the root of the repo
+Install using pip
+
+.. code-block::
+
+   pip install HolisticTraceAnalysis
+
+Install from source
 
 .. code-block::
 
-   pip install -e .
+  # get the source code
+  git clone https://github.com/facebookresearch/HolisticTraceAnalysis.git
+
+  # execute the command below from the root of the repo
+  pip install -e .

hta/analyzers/breakdown_analysis.py

@@ -121,7 +121,7 @@ def get_gpu_kernel_breakdown(
             inplace=True,
         )
 
-        if visualize:
+        if visualize:  # pragma: no cover
             non_zero_kernel_df = kernel_type_df[(kernel_type_df["percentage"] > 0)]
 
             fig = px.pie(
@@ -336,7 +336,7 @@ def idle_time_per_rank(trace_df: pd.DataFrame) -> Tuple[int, int, int, int]:
         result_df["non_compute_time"] = result_df["non_compute_time(ns)"] / result_df["kernel_time(ns)"]
         result_df["non_compute_time_pctg"] = round(100 * result_df["non_compute_time"], 2)
 
-        if visualize:
+        if visualize:  # pragma: no cover
             fig = px.bar(
                 result_df,
                 x="rank",
@@ -480,7 +480,7 @@ def get_idle_time_breakdown(
         result_df.rename(mapper=idle_category_name_map, axis=0, inplace=True)
         result_df.reset_index(inplace=True)
 
-        if visualize:
+        if visualize:  # pragma: no cover
             result_df["stream"] = result_df.stream.astype(str)
             ycol = "idle_time_ratio" if visualize_pctg else "idle_time"
             fig = px.bar(

hta/analyzers/communication_analysis.py

@@ -71,7 +71,7 @@ def get_comm_comp_overlap_value(trace_df: pd.DataFrame) -> float:
         result_df = pd.DataFrame(result)
         result_df["comp_comm_overlap_pctg"] = round(100 * result_df["comp_comm_overlap_ratio"], 2)
 
-        if visualize:
+        if visualize:  # pragma: no cover
             fig = px.bar(
                 result_df,
                 x="rank",

hta/analyzers/cuda_kernel_analysis.py

@@ -172,7 +172,7 @@ def _generate_frequent_pattern_results(
             'View the generated trace file using Chrome Tracing and search for "Patterns" to highlight the frequent patterns.'
         )
 
-        if visualize:
+        if visualize:  # pragma: no cover
             vis_df = patterns_df[:top_k].copy()
             # show the pattern in multiple lines in visualization
             vis_df["pattern"] = vis_df["pattern"].str.replace("|", "<br>", regex=False)
@@ -259,7 +259,7 @@ def _compress_kernel_names(row):
     @classmethod
     def visualize_cuda_kernel_launch_stats(
         cls, rank: int, df: pd.DataFrame, runtime_cutoff: int, launch_delay_cutoff: int
-    ) -> None:
+    ) -> None:  # pragma: no cover
         short_kernels = df[(df["cpu_duration"] <= runtime_cutoff) & (df["gpu_duration"] < df["cpu_duration"])]
         runtime_outliers = df[df["cpu_duration"] > runtime_cutoff]
         launch_delay_outliers = df[df["launch_delay"] > launch_delay_cutoff]
@@ -369,7 +369,7 @@ def cuda_kernel_launch_stats(
             renamed_df = joined_df.rename(columns={"dur_x": "cpu_duration", "dur_y": "gpu_duration"})
             events_df = renamed_df[["correlation", "cpu_duration", "gpu_duration", "launch_delay"]]
 
-            if visualize:
+            if visualize:  # pragma: no cover
                 cls.visualize_cuda_kernel_launch_stats(rank, events_df, runtime_cutoff, launch_delay_cutoff)
 
             result_dict[rank] = events_df

hta/analyzers/straggler.py

@@ -110,7 +110,7 @@ def _compute_normalized_start_time_of_significant_comm_kernels(
     long_comm_kernels = df.loc[
         (df["stream"] > 0) & (df["iteration"] > 0) & (df["dur"] >= min_duration) & (df["name"].isin(comm_op_ids))
     ]
-    if visualize:
+    if visualize:  # pragma: no cover
         plot_timeline_gpu_kernels(
             f"Timeline of Communication Kernels Longer Than {min_normalized_duration * 100:.2f}% of Iteration Time\n",
             long_comm_kernels,
@@ -135,7 +135,7 @@ def _compute_normalized_start_time_of_significant_comm_kernels(
         (last_long_comm_kernels["stream"].eq(stream)) & (last_long_comm_kernels["name"].eq(name))
     ]
 
-    if visualize:
+    if visualize:  # pragma: no cover
         plot_timeline_gpu_kernels(
             f"Timeline of Candidate Kernels (Iterations={iterations})",
             candidate_metric_kernels,
@@ -202,7 +202,7 @@ def _get_top_k_stragglers_with_metric(
         metric[col] = pd.to_numeric(metric[col])
 
     color_map = {"Yes": "red", "No": "blue"}
-    if visualize:
+    if visualize:  # pragma: no cover
         fig = px.bar(
             metric,
             x="rank",

hta/trace_diff.py

@@ -485,7 +485,7 @@ def visualize_duration_diff(
             width=1000,
             height=1000,
         )
-        if show_image:
+        if show_image:  # pragma: no cover
             fig.show()
-        if export_image_path:
+        if export_image_path:  # pragma: no cover
             fig.write_image(export_image_path)

requirements-dev.txt

@@ -0,0 +1,7 @@
+# Python dependencies required for development
+jupyterlab>=3.5.1
+kaleido>=0.2.1
+numpy>=1.23.5
+pandas>=1.5.2
+plotly>=5.11.0
+coverage>=7.0.2

requirements.txt

@@ -1,7 +1,5 @@
-# Python dependencies required for development
+# Python dependencies required for users
 jupyterlab>=3.5.1
-kaleido>=0.2.1
 numpy>=1.23.5
 pandas>=1.5.2
 plotly>=5.11.0
-coverage>=7.0.2

setup.py

@@ -29,7 +29,7 @@ def fetch_requirements():
         "Programming Language :: Python :: 3.9",
         "Programming Language :: Python :: 3.10",
         "Programming Language :: Python :: 3.11",
-        "License :: OSI Approved :: BSD License",
+        "License :: OSI Approved :: MIT License",
         "Topic :: Scientific/Engineering :: Artificial Intelligence",
         "Operating System :: OS Independent",
     ],

tests/test_call_stack.py

@@ -101,5 +101,5 @@ def test_node_depth(self):
         self.assertDictEqual(depth_from_csg, depth_from_nodes)
 
 
-if __name__ == "__main__":
+if __name__ == "__main__":  # pragma: no cover
     unittest.main()

tests/test_config.py

@@ -64,5 +64,5 @@ def test_get_config_default_values(self):
         self.assertEqual(config.get_config("d", 10), 10)
 
 
-if __name__ == "__main__":
+if __name__ == "__main__":  # pragma: no cover
     unittest.main()

tests/test_symbol_table.py

@@ -69,5 +69,5 @@ def test_query_symbols_multi_processes(self):
         self.assertTrue(all(is_consistent))
 
 
-if __name__ == "__main__":
+if __name__ == "__main__":  # pragma: no cover
     unittest.main()

tests/test_trace_analysis.py

@@ -256,5 +256,5 @@ def test_get_idle_time_breakdown(self):
             )
 
 
-if __name__ == "__main__":
+if __name__ == "__main__":  # pragma: no cover
     unittest.main()

tests/test_trace_parse.py

@@ -97,5 +97,5 @@ def test_trace_iteration(self) -> None:
             self.assertDictEqual(gpu_kernels_per_iteration, correlated_cpu_ops_per_iteration)
 
 
-if __name__ == "__main__":
+if __name__ == "__main__":  # pragma: no cover
     unittest.main()