From 85fc4498cc5eed967d5ed812a8e0d95f1b0fa330 Mon Sep 17 00:00:00 2001
From: yucheng-berri <yucheng@berri.ai>
Date: Thu, 8 Oct 2026 16:22:04 -0700
Subject: [PATCH] fix(batches): an unreadable error file no longer drops the
 output line events

Costing counts an error file it could not fetch as no failed lines; line logging
refetched it and the raise aborted the whole fan-out, so the successful lines were
never emitted. The error file now reads as empty with a warning.
---
 litellm/batches/batch_line_item_logging.py    | 18 +++++++++++-
 .../batches/test_batch_line_item_logging.py   | 29 +++++++++++++++++++
 2 files changed, 46 insertions(+), 1 deletion(-)

diff --git a/litellm/batches/batch_line_item_logging.py b/litellm/batches/batch_line_item_logging.py
index 854f8c2ef34..bde63efb30f 100644
--- a/litellm/batches/batch_line_item_logging.py
+++ b/litellm/batches/batch_line_item_logging.py
@@ -419,6 +419,22 @@ async def _fetch_managed_file_or_empty(
     )
 
 
+async def _fetch_error_file_or_empty(
+    batch: LiteLLMBatch,
+    custom_llm_provider: _BatchLineProvider,
+    fetch_params: dict[str, object] | None,  # mutable-ok: batch_utils file fetch takes the shared litellm_params dict
+) -> bytes:
+    try:
+        return await _fetch_managed_file_or_empty(batch.error_file_id, custom_llm_provider, fetch_params)
+    except Exception as e:  # noqa: BLE001  # costing counts an unreadable error file as no failed lines, so the output lines still go out
+        verbose_logger.warning(
+            "batch line-item callbacks could not fetch the error file, emitting the output lines only. batch_id=%s error=%s",
+            batch.id,
+            e,
+        )
+        return b""
+
+
 async def log_batch_line_items(
     batch: LiteLLMBatch,
     custom_llm_provider: str,
@@ -488,7 +504,7 @@ async def log_batch_line_items(
         error_content: Final = (
             result_files.error
             if result_files is not None and result_files.error is not None
-            else await _fetch_managed_file_or_empty(batch.error_file_id, line_provider, fetch_params)
+            else await _fetch_error_file_or_empty(batch, line_provider, fetch_params)
         )
         for content in (output_content, error_content):
             for entry in _output_entries(content):
diff --git a/tests/unit/batches/test_batch_line_item_logging.py b/tests/unit/batches/test_batch_line_item_logging.py
index d383546b7bf..aea0ae63ebe 100644
--- a/tests/unit/batches/test_batch_line_item_logging.py
+++ b/tests/unit/batches/test_batch_line_item_logging.py
@@ -1241,6 +1241,35 @@ async def test_supplied_result_files_skip_the_output_and_error_fetch(recorder):
     assert file_mock.await_args.kwargs["file_id"] == "input-file-1"
 
 
+def _error_file_unreadable(file_id: str, **_kwargs):
+    if file_id == "error-file-1":
+        raise ValueError("error file gone")
+    return SimpleNamespace(content=_FILE_BYTES[file_id])
+
+
+@pytest.mark.asyncio
+async def test_an_unreadable_error_file_still_emits_the_output_lines(recorder):
+    file_mock: Final = AsyncMock(side_effect=_error_file_unreadable)
+    with patch(
+        "litellm.files.main.afile_content", file_mock
+    ):  # test-quality-ok: afile_content is the provider boundary; no injection seam for managed file fetch
+        emitted: Final = await log_batch_line_items(
+            batch=_batch(),
+            custom_llm_provider="openai",
+            parent=_parent_logging(),
+            model_name="gpt-4o",
+            litellm_params=None,
+            model_info=None,
+            result_files=BatchResultFiles(output=OUTPUT_JSONL, error=None),
+            claim_cache=DualCache(),
+        )
+
+    assert emitted == 1
+    assert len(recorder.success_events) == 1
+    assert len(recorder.failure_events) == 0
+    assert sorted(call.kwargs["file_id"] for call in file_mock.await_args_list) == ["error-file-1", "input-file-1"]
+
+
 @pytest.mark.asyncio
 async def test_empty_result_files_fall_back_to_fetching_everything(recorder):
     file_mock: Final = AsyncMock(side_effect=_file_content)