MTP: `test_generate_with_mtp` skips for every checkpoint but Inkling, and `MtpModel.forward` raises with the default `logits_processor=None`

[pjordanandrsn]

System info

• transformers on main (c49429ed). Both are static—neither needs a run to see.

Who can help?

@Cyrilvallez—@zucchini-nlp asked in
#47912 that
MTP-model-specific things go in their own issue tagged to you.

Two things noticed while working on #47912, the EOS-past-stop bug in the shared assisted-decoding path.
Neither is that bug; both are MTP-side.

1. test_generate_with_mtp skips for every checkpoint except Inkling

tests/generation/test_utils.py:2483 gates on the model's _keys_to_ignore_on_load_unexpected:

keys_to_ignore_unexpected = model_class._keys_to_ignore_on_load_unexpected or []
# If we don't have any mtp patterns, skip
if not hasattr(config.get_text_config(), "num_mtp_layers") or not any(
    "mtp" in x or re.search(r"layers\.\d+", x) is not None for x in keys_to_ignore_unexpected
):
    self.skipTest("No MTP keys registered")

Those entries are regexes, so their dots are escaped in the source, and re.search(r"layers\.\d+", x) wants
a literal . where the string has \.:

model	_keys_to_ignore_on_load_unexpected	gate
deepseek_v3	[r"model\.layers\.61.*"]	skips
glm4_moe	[r"model\.layers\.92.*", r"model\.layers\.46.*"]	skips
solar_open	None	skips
inkling	[r"model\.mtp\..*"]	runs, on "mtp" in x

So the model.layers.{num_hidden_layers+1} half of what the docstring says it covers is never exercised—the
only model reaching the test is the one registering an mtp. pattern. re.search(r"layers\\?\.\d+", x)
matches both spellings.

Worth saying it would not have caught #47912 either way: it asserts only that generation raises nothing.

2. MtpModel.forward raises UnboundLocalError when logits_processor is None

src/transformers/modeling_layers.py:540:

if logits_processor is not None and full_input_ids is not None:
    next_token_scores = logits_processor(full_input_ids, next_token_logits.to(torch.float32))
if do_sample:
    probs = nn.functional.softmax(next_token_scores, dim=-1, dtype=torch.float32)
    next_mtp_token = torch.multinomial(probs, num_samples=1)
else:
    next_mtp_token = torch.argmax(next_token_scores, dim=-1, keepdim=True)

next_token_scores is bound only inside the guard and read unconditionally two lines below, so the call
raises UnboundLocalError whenever either argument is None. MTPCandidateGenerator.__init__ declares
logits_processor: Optional["LogitsProcessorList"] = None, so constructing one directly and drafting hits
it.

Not reachable through generate, which always passes a LogitsProcessorList and supplies full_input_ids
at candidate_generator.py:1529—so it is a trap for direct construction only. The tests in #48002 pass an
explicit LogitsProcessorList() for this reason. Falling back to next_token_logits, or making the
argument required, would both do.

---

AI-assisted: found and written up with Claude Code; both claims re-checked against main before filing.

[pjordanandrsn]

Correction: item 2 is already fixed on main. #48007 (240b8330) removed next_token_scores entirely—the processor result now re-assigns next_token_logits, which is always bound—so MtpModel.forward no longer raises with logits_processor=None.

Item 1 still stands on current main (0650ff35): the gate at tests/generation/test_utils.py:2489 skips deepseek_v3 and glm4_moe with "No MTP keys registered" (re-checked by running them), leaving inkling the only model the test runs for. So this issue is just the regex now.