Added a redundancy for a failing AMD CI (#5)

- Added a redundancy for a failing AMD CI (2772ee98a08427949d3424cc1f3ec96554419594)


Co-authored-by: Rémi Ouazan Reboul <ror@users.noreply.huggingface.co>

data.py

@@ -98,11 +98,26 @@ def read_one_dataframe(json_path: str, device_label: str) -> tuple[pd.DataFrame,
     df[f"failed_single_no_{device_label}"] = df["failures"].apply(lambda x: len(x["single"]) if "single" in x else 0)
     return df, df_upload_date
 
+def get_first_working_df(file_list: list[str]) -> str:
+    for file in file_list:
+        job_links = file.rsplit('/', 1)[0] + "/job_links.json"
+        try:
+            links = pd.read_json(f"hf://{job_links}", typ="series")
+            has_one_working_link = any(links.values)
+        except Exception as e:
+            logger.error(f"Could not read job links from {job_links}: {e}")
+            has_one_working_link = False
+        if has_one_working_link:
+            return file
+        logger.warning(f"Skipping {file} as it has no working job links.")
+    raise RuntimeError("Could not find any working dataframe in the provided list.")
+
 def get_distant_data() -> tuple[pd.DataFrame, str]:
     # Retrieve AMD dataframe
     amd_src = "hf://datasets/optimum-amd/transformers_daily_ci/**/runs/**/ci_results_run_models_gpu/model_results.json"
     files_amd = sorted(fs.glob(amd_src, refresh=True), reverse=True)
-    df_amd, date_df_amd = read_one_dataframe(f"hf://{files_amd[0]}", "amd")
+    file_amd = get_first_working_df(files_amd)
+    df_amd, date_df_amd = read_one_dataframe(f"hf://{file_amd}", "amd")
     # Retrieve NVIDIA dataframe, which pattern should be:
     # hf://datasets/hf-internal-testing`/transformers_daily_ci/raw/main/YYYY-MM-DD/ci_results_run_models_gpu/model_results.json 
     nvidia_src = "hf://datasets/hf-internal-testing/transformers_daily_ci/*/ci_results_run_models_gpu/model_results.json"