Fix KeyError in apply_chat_template when message has no content (#45290)

[agentspan]

Summary

Fixes #45290.

ProcessorMixin.apply_chat_template and several related code paths assumed
every message in a conversation has a content key. Assistant messages with
tool_calls and no textual content (a valid shape per the OpenAI
chat-completion spec — the assistant is requesting tool execution and has
no textual reply yet) crashed with KeyError: 'content'.

This PR introduces a single get_message_content helper in
transformers.utils.chat_template_utils and migrates every call site that
previously assumed message["content"] exists to use it. The helper defaults
missing content to either an empty list (for code that iterates over
multimodal content blocks) or an empty string (for text-only paths), so the
loop body still runs over zero items instead of either skipping the message
entirely or raising.

Why a centralized helper

Without a helper, every processor / tokenizer / serving handler that needs
to access message["content"] would have to remember the convention
independently, and any future contributor adding a new entry point could
re-introduce the same bug. The helper:

• Centralizes the OpenAI-spec convention in one place with a clear docstring
• Has a default parameter that is keyword-only so call sites are
  self-documenting (get_message_content(msg, default=[]) vs
  get_message_content(msg, default=""))
• Is the only place in the patched surface that calls message.get("content", ...)

Sites patched

All five via the new helper:

• src/transformers/processing_utils.py — main apply_chat_template loop
  that extracts images / videos / audio from message content
• src/transformers/models/smolvlm/processing_smolvlm.py — video-detection
  list comprehension across all messages
• src/transformers/cli/serving/utils.py — VLM and LLM modality message
  parser used by the OpenAI-compatible serving CLI
• src/transformers/utils/chat_template_utils.py — Jinja template renderer;
  normalises every message before passing to the template engine
• src/transformers/tokenization_mistral_common.py — _maybe_adapt_message
  helper that normalizes content format

is_valid_message consistency fix

is_valid_message (used by transformers.pipelines.base to detect
chat-format inputs) previously rejected any message that did not have a
content key. Under the new convention this would reject valid
tool-call-only assistant messages, so it now accepts messages with either a
content key OR a tool_calls key (still requiring role). The Chat
class docstring and its constructor's error message are updated to match.

Tests

New unit tests in tests/utils/test_chat_template_utils.py

• GetMessageContentTest — 5 tests covering returns-content, default-list,
  explicit-string default, explicit-list default, and keyword-only enforcement
  (positional default raises TypeError)
• IsValidMessageTest — 5 tests covering accepts role+content, accepts
  role+tool_calls without content, rejects missing role, rejects messages
  missing both content and tool_calls, rejects non-dict inputs

New regression tests across the patched surfaces

• tests/test_processing_common.py::ProcessorTesterMixin::test_apply_chat_template_with_tool_calls_no_content
• tests/cli/test_serve.py::TestProcessorInputsFromMessages::test_vlm_tool_calls_without_content
• tests/cli/test_serve.py::TestProcessorInputsFromMessages::test_llm_tool_calls_without_content
• tests/test_tokenization_mistral_common.py::TestMistralCommonBackend::test_apply_chat_template_with_tool_calls_no_content

All construct a conversation containing
{"role": "assistant", "tool_calls": [...]} with no content key and
assert no KeyError is raised.

Verification

The exact failing repro from the issue body now exits 0:

from transformers import AutoProcessor

processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct")
processor.apply_chat_template(
    [[
        {"role": "user", "content": [{"type": "text", "text": "dummy"}]},
        {"role": "assistant", "tool_calls": [{"type": "function", "function": {"name": "foo", "arguments": {}}}]},
    ]],
    tokenize=True,
)

Locally:

• 14 relevant tests pass (5 helper, 5 is_valid_message, 4 regression)
• ruff format --check and ruff check are both clean on every changed file
• A grep for raw message.get("content" / msg.get("content" across the
  patched surface returns zero hits except inside the helper definition

Stats

5 source files patched, 1 source file gains the helper, 1 test file gains
helper unit tests, 3 test files gain regression tests. +282 / -23 lines.

Closes #45290

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: smolvlm

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=45309&sha=90130a