Remove deprecated VQModel/VQEncoderOutput shim (past target version 0.31)

[Dev-X25874]

src/diffusers/models/vq_model.py still contains the backward-compat shim for VQModel/VQEncoderOutput with deprecate(..., 0.31). Current version is 0.40.0.dev0, so this is well past its removal target. Happy to send a PR removing this file and its reference in the relevant init.py if maintainers are OK with it. Let me know if this is still wanted or already being handled elsewhere.

[4ktLuffy]

Confirmed on main (0.40.0.dev0): src/diffusers/models/vq_model.py still has the shim, both call sites targeting 0.31.

Worth flagging that this is more than dead weight, and that it isn't isolated.

deprecate() raises rather than warns once the target version passes — utils/deprecation_utils.py:

if version.parse(version.parse(__version__).base_version) >= version.parse(version_name):
    raise ValueError(f"The deprecation tuple {(attribute, version_name, message)} should be removed ...")

So any of these code paths that is still reachable raises ValueError when executed:

>>> from diffusers.models.vq_model import VQModel; VQModel()
ValueError: The deprecation tuple ('VQModel', '0.31', ...) should be removed since diffusers' version 0.40.0.dev0 is >= 0.31

An AST scan for deprecate() calls whose target version is below 0.40 finds 20 call sites across 8 target versions, the oldest being 0.27.0:

target	sites
0.27.0	loaders/__init__.py (2)
0.28.0	pipelines/deprecated/versatile_diffusion/modeling_text_unet.py (1)
0.31	models/vq_model.py (2) — this issue
0.33.0	models/embeddings.py (4)
0.34.0	models/embeddings.py, transformers/transformer_chroma.py (2)
0.34	pipelines/controlnet/multicontrolnet.py, lumina, lumina2 (3)
0.35.0	transformers/transformer_hidream_image.py, pipelines/hidream_image (4)
0.39.0	models/controlnets/controlnet_qwenimage.py (2)

The embeddings.py group is the one with user-facing impact — four public utilities raise on a documented argument value:

import numpy as np
from diffusers.models import embeddings as E
E.get_2d_sincos_pos_embed(embed_dim=8, grid_size=2, output_type="np")
E.get_2d_sincos_pos_embed_from_grid(8, np.zeros((2, 4)), output_type="np")
E.get_1d_sincos_pos_embed_from_grid(8, np.arange(4), output_type="np")
E.get_2d_rotary_pos_embed(8, (2, 2), (0, 0, 2, 2), output_type="np")

All four raise ValueError. Happy for @Dev-X25874 to take vq_model.py as offered — flagging the rest in case you would rather scope this as one sweep, or split the embeddings.py ones out as a separate fix since those are reachable public API.

Not verified: reachability of every one of the 20. I executed the vq_model and embeddings cases; the loaders/__init__.py pair (oldest, 0.27.0) turned out not to be exported — from diffusers.loaders import text_encoder_lora_state_dict fails with ImportError — so that pair looks like dead code rather than a live path. The remaining sites I inventoried statically without executing.

[Dev-X25874]

Thanks for digging deeper on this @4ktLuffy — good catch that it's raise, not warn, past the target version. Happy to keep my PR scoped to vq_model.py as discussed, and agree embeddings.py sounds like it deserves its own issue since it's a live bug on public API, not just cleanup.

[Dev-X25874]

cc @yiyixuxu @hlky — since sayakpaul/DN6 haven't had a chance to look yet, wondering if either of you could weigh in on whether this cleanup is still wanted.