Fix ValueError in _encode_prompt truncation check for prompts at the CLIP token limit

[Osamaali313]

Summary

StableDiffusionPipeline._encode_prompt crashes on any prompt that reaches the CLIP token limit.

The truncation-warning guard (python_coreml_stable_diffusion/pipeline.py) is:

if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not np.equal(
        text_input_ids, untruncated_ids
):

np.equal returns an element-wise boolean array, so not <array> raises:

• ValueError: The truth value of an array with more than one element is ambiguous when the shapes match, and
• ValueError: operands could not be broadcast together with shapes (1,77) (1,90) when they differ — i.e. the real truncation case.

text_input_ids is always padded to tokenizer.model_max_length (77) and untruncated_ids uses padding="longest", so this branch is entered whenever a prompt tokenizes to ≥ 77 tokens. Such prompts crash at prompt-encoding time, before generation — including precisely the truncation scenario the warning is meant to surface. (Short prompts short-circuit on the length check, which is why this isn't always hit.)

This is a port of the upstream diffusers check, which uses torch.equal(...) — a function returning a scalar Python bool, so not torch.equal(...) is valid. The numpy port should use the scalar-returning np.array_equal.

Fix

-if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not np.equal(
+if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not np.array_equal(
         text_input_ids, untruncated_ids
 ):

np.array_equal returns a scalar bool and returns False for differing shapes (matching torch.equal semantics), so the warning fires correctly instead of raising.

Validation

Reproduced with faithful tokenizer shapes (text_input_ids=[1,77]):

case	np.equal (current)	np.array_equal (fixed)
no truncation ([1,77])	ValueError: ambiguous truth value	no warning
truncation ([1,90])	ValueError: broadcast	warns (correct)
short prompt ([1,5])	no warning	no warning

The existing tests/ are end-to-end integration tests (full Core ML conversion + model downloads + Swift CLI), so this guard cannot be exercised as an isolated unit test; the table above is the standalone repro.