summarize_graph_break() — pytorch Function Reference

Architecture documentation for the summarize_graph_break() function in common.py from the pytorch codebase.

Function python DynamoBenchmarks Microbenchmarking called by 1

Entity Profile

DynamoBenchmarks→ Microbenchmarking→ summarize_graph_break() — pytorch Function Reference

Dependency Diagram

graph TD
  5752cb25_7ac2_c2bf_6619_ca3cbd78adbb["summarize_graph_break()"]
  bf029f88_db70_e8df_06de_66d5b0f31c56["print_summary()"]
  bf029f88_db70_e8df_06de_66d5b0f31c56 -->|calls| 5752cb25_7ac2_c2bf_6619_ca3cbd78adbb
  style 5752cb25_7ac2_c2bf_6619_ca3cbd78adbb fill:#6366f1,stroke:#818cf8,color:#fff

Relationship Graph

Source Code

benchmarks/dynamo/common.py lines 581–602

def summarize_graph_break(filename):
    """
    Sorts and de-dupes the graphs breaks on the reason string. Note that this
    function is just a best effort to reduce the logging information. We could
    miss some graph breaks because of de-duping. We can further refine this
    function as need arises.
    """
    log_file = f"{filename.rstrip('.csv')}_graph_breaks.csv"
    if os.path.exists(log_file):
        df = pd.read_csv(log_file)
        df = df.sort_values("reason").drop_duplicates(subset="reason")

        # Specialize for multi tensor sgd as reason is not identical
        multi_tensor_sgd_row = df.loc[df["reason"].str.contains("_multi_tensor_sgd")]
        if len(multi_tensor_sgd_row):
            df = df[
                ~df["reason"].str.contains("_multi_tensor_sgd")
            ]  # Drop all sgd rows
            df = pd.concat(
                [df, pd.DataFrame([multi_tensor_sgd_row.iloc[0]])], axis=0
            )  # Add back a single row
        df.to_csv(f"{log_file.rstrip('.csv')}_deduped.csv", index=False)