System Info

Summary

Multiple models across the library use an XOR (^) condition to validate input_ids and inputs_embeds inputs. This pattern has two distinct issues depending on the model:

1. Severe (11 models): XOR is paired with the error message "You cannot specify both ...", which is factually wrong when neither input is provided — the opposite situation.
2. Minor (most other models, e.g. BART): XOR is paired with the generic message "You must specify exactly one of ...", which is functionally correct but conflates two opposite error cases into one vague message.

Background and History

Before PR #43590

BartEncoder used a correct explicit 4-branch check:

if input_ids is not None and inputs_embeds is not None:
    raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
    ...  # use input_ids
elif inputs_embeds is not None:
    ...  # use inputs_embeds
else:
    raise ValueError("You have to specify either input_ids or inputs_embeds")

BartDecoder already had the severe bug (XOR + wrong message):

# Pre-#43590 BartDecoder — already broken
if (input_ids is None) ^ (inputs_embeds is not None):
    raise ValueError(
        "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
    )

After PR #43590

The refactor unified both encoder and decoder to:

if (input_ids is None) ^ (inputs_embeds is not None):
    raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

This fixed the wrong message in BartDecoder, but the XOR logic itself still conflates two opposite error cases. Meanwhile, 11 other older models were not updated and still carry the original severe bug.

Why XOR Is Problematic Here

The truth table for (input_ids is None) ^ (inputs_embeds is not None):

XOR fires for both the "too many" and "too few" cases, but these require different, opposite error messages. Using a single message for both makes it impossible for users to know whether they passed too many or too few inputs.

Issue 1 (Severe): XOR + Wrong Message — 11 Models Not Updated by #43590

These models still use XOR paired with "You cannot specify both ...", which is factually incorrect when neither input is provided:

Suggested Fix

Replace the XOR with two explicit conditions, which is unambiguous and consistent with the original BartEncoder logic before #43590:

if input_ids is not None and inputs_embeds is not None:
    raise ValueError(
        "You cannot specify both input_ids and inputs_embeds at the same time"
    )
if input_ids is None and inputs_embeds is None:
    raise ValueError(
        "You have to specify either input_ids or inputs_embeds"
    )
# No elif needed — the embed_tokens call below handles the remaining cases cleanly
if inputs_embeds is None:
    inputs_embeds = self.embed_tokens(input_ids)

This preserves the structural simplification introduced in #43590 while restoring precise, actionable error messages for each failure case.

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Reproduction (example with Blenderbot):

from transformers import BlenderbotModel, BlenderbotConfig

config = BlenderbotConfig()
model = BlenderbotModel(config)

model.model.decoder(input_ids=None, inputs_embeds=None)

Expected behavior

Pass neither input — should say "must specify one", but says "cannot specify both"

ValueError: You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time

Issue 2 (Minor): XOR + Vague Message — BART and Most Other Models

Models updated by #43590 (including BART) now use:

if (input_ids is None) ^ (inputs_embeds is not None):
    raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

This message is technically acceptable for the "neither provided" case, but is misleading for the "both provided" case, where the user would expect to be told they passed too many inputs, not that they need to pick one.