Fix AttributeError in _get_finish_reason_for_choice when finish_reason is a string

[happywhaler]

Fix AttributeError in _get_finish_reason_for_choice when finish_reason is a string

Description

This PR fixes a bug in _get_finish_reason_for_choice() that causes an AttributeError when the model returns finish_reason as a string instead of an enum.

Some models (e.g., Mistral Large 3 on Azure AI Foundry MaaS) return finish_reason as a raw string ("stop") rather than a CompletionsFinishReason enum. The current implementation unconditionally calls .value on finish_reason, which fails with:

AttributeError: 'str' object has no attribute 'value'

Root Cause

In tracing.py, _get_finish_reason_for_choice() assumes finish_reason is always an enum:

def _get_finish_reason_for_choice(self, choice):
    finish_reason = getattr(choice, "finish_reason", None)
    if finish_reason is None:
        return "none"
    return finish_reason.value  # Fails when finish_reason is already a string

However, the similar method _get_finish_reasons() already handles both cases correctly:

if hasattr(finish_reason, "value"):
    finish_reasons.append(finish_reason.value)
elif isinstance(finish_reason, str):
    finish_reasons.append(finish_reason)

Fix

Apply the same defensive pattern to _get_finish_reason_for_choice():

def _get_finish_reason_for_choice(self, choice):
    finish_reason = getattr(choice, "finish_reason", None)
    if finish_reason is None:
        return "none"
    elif hasattr(finish_reason, "value"):
        return finish_reason.value
    elif isinstance(finish_reason, str):
        return finish_reason
    else:
        return "none"

Impact

Without this fix, tracing fails for any model that returns string finish_reason, causing the entire API call to fail. Users must disable tracing (AZURE_SDK_TRACING_IMPLEMENTATION=none or tracing_enabled=False) as a workaround.

[github-actions]

Thank you for your contribution @happywhaler! We will review the pull request and get back to you soon.

[Copilot]

Pull request overview

This PR fixes tracing robustness in azure-ai-inference by handling finish_reason values that are returned as raw strings (instead of enums), preventing AttributeError during span/event creation for certain models.

Changes:

• Update _get_finish_reason_for_choice() to return "none", finish_reason.value, or the raw finish_reason string depending on the runtime type.
• Align _get_finish_reason_for_choice() behavior with the already-defensive _get_finish_reasons() implementation.

[Copilot]

_get_finish_reason_for_choice now duplicates the same enum/string normalization logic already present in _get_finish_reasons. To avoid the two implementations drifting over time, consider extracting a small helper (e.g., _normalize_finish_reason(finish_reason)) and using it from both places (or have _get_finish_reasons call _get_finish_reason_for_choice per choice).

[Copilot]

This fix changes behavior for choices where finish_reason is already a str, but the existing tracing tests appear to only validate the serialized output (and likely exercise enum-backed finish_reason). Please add a focused test that constructs a minimal choice object with finish_reason="stop" (string) and asserts tracing emits "stop" without raising (covering the reported AttributeError regression).