Hi there, found a bug on HFInference.chatCompletionStream implementation specified below:

Environment

Library: @huggingface/inference
version: "^2.8.0"
Function: chatCompletionStream in streamingRequest

Problem

In the HFInference.chatCompletionStream implementation, the streamingRequest function throws errors incorrectly when the data.error field is an object. Specifically, the following code causes the error message to display [object Object] instead of the actual error content:

for (const event of events) {
  if (event.data.length > 0) {
    if (event.data === "[DONE]") {
      return;
    }
    const data = JSON.parse(event.data);
    if (typeof data === "object" && data !== null && "error" in data) {
      throw new Error(data.error); // This causes [object Object] to be thrown
    }
    yield data as T;
  }
}

The error should contain the message field or be stringified to provide readable error information, such as:

Error: Input validation error: inputs tokens + max_new_tokens must be <= 4096. Given: 48 inputs tokens and 4096 max_new_tokens

Suggested Fix

change
throw new Error(data.error);
to explicitly stringify the error message:
throw new Error(data.error.message || JSON.stringify(data.error));

Thanks!