[Fix]Allow prompt and prior_token_ids to be provided simultaneously in GlmImagePipeline

[JaredforReal]

What does this PR do?

Previously, GlmImagePipeline.check_inputs() enforced mutual exclusion between prompt and prior_token_ids, raising a ValueError if both were provided. This was unnecessarily restrictive — there is a valid use case where a user wants to supply pre-computed prior_token_ids (to skip the expensive AR generation step) while still passing prompt so that prompt_embeds (glyph embeddings) can be derived from it automatically.

This PR relaxes that constraint:

• Removed the check that raises when both prompt and prior_token_ids are provided.
• Relaxed the prior_token_ids requires prompt_embeds check — now either prompt or prompt_embeds satisfies the requirement, since prompt will be used to generate prompt_embeds via encode_prompt() downstream.

Supported input combinations after this change

Inputs	Behavior	Status
prompt only	AR generates prior_token_ids; glyph encoder generates prompt_embeds	✅ (unchanged)
prior_token_ids + prompt_embeds	Both used directly, AR step skipped	✅ (unchanged)
prompt + prior_token_ids	prior_token_ids used directly (AR skipped); prompt generates prompt_embeds	✅ new
Neither prompt nor prior_token_ids	❌ ValueError	✅ (unchanged)
prior_token_ids alone (no prompt/prompt_embeds)	❌ ValueError	✅ (unchanged)
prompt + prompt_embeds	❌ ValueError	✅ (unchanged)

Tests

Added two new tests to test_glm_image.py:

• test_prompt_with_prior_token_ids: End-to-end test that first generates prior_token_ids via generate_prior_tokens(), then runs the full pipeline with both prompt and prior_token_ids provided together.
• test_check_inputs_rejects_invalid_combinations: Unit test verifying that the three invalid input combinations listed above still correctly raise ValueError.

All existing tests continue to pass.

[JaredforReal]

@yiyixuxu PTAL, thanks

[Copilot]

Pull request overview

This PR updates the GLM image pipeline input validation to allow supplying prompt together with precomputed prior_token_ids, enabling users to skip the AR prior generation step while still deriving glyph prompt_embeds from the prompt.

Changes:

• Removed the mutual-exclusion validation between prompt and prior_token_ids in GlmImagePipeline.check_inputs().
• Relaxed the prior_token_ids validation to accept either prompt or prompt_embeds as the source for glyph embeddings.
• Added tests covering the newly supported prompt + prior_token_ids combination and validating invalid input combinations.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
src/diffusers/pipelines/glm_image/pipeline_glm_image.py	Adjusts check_inputs() to permit prompt and prior_token_ids together and updates validation/error paths.
tests/pipelines/glm_image/test_glm_image.py	Adds new tests for the supported combo and for rejecting invalid input combinations.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

inputs_prompt_only (and its generator) is created but never used. This makes the test misleading (it implies a full pipeline run in step 1) and adds dead code; consider removing it or actually using it to run pipe(**inputs_prompt_only) if that's the intended setup.

Suggested change

	inputs_prompt_only = {
	"prompt": "A photo of a cat",
	"generator": generator,
	"num_inference_steps": 2,
	"guidance_scale": 1.5,
	"height": height,
	"width": width,
	"max_sequence_length": 16,
	"output_type": "pt",
	}
	prior_token_ids, _, _ = pipe.generate_prior_tokens(
	prompt="A photo of a cat",
	height=height,
	width=width,
	device=torch.device(device),
	generator=torch.Generator(device=device).manual_seed(0),
	prior_token_ids, _, _ = pipe.generate_prior_tokens(
	prompt="A photo of a cat",
	height=height,
	width=width,
	device=torch.device(device),
	generator=generator,

[Copilot]

The prior_token_ids validation at the end of check_inputs is effectively unreachable because an earlier check already raises whenever both prompt and prompt_embeds are None. If the goal is a more specific error when prior_token_ids is provided without any prompt inputs, consider folding this condition into the earlier prompt/prompt_embeds validation (or removing this block to avoid redundant code).

Suggested change

	if prior_token_ids is not None and prompt_embeds is None and prompt is None:
	raise ValueError("`prompt_embeds` or `prompt` must also be provided with `prior_token_ids`.")

[Copilot]

The error raised when prompt is None and prior_token_ids is None can be confusing if a user provided prompt_embeds: earlier validation allows prompt_embeds in place of prompt, but this message implies prompt is required. Consider clarifying the message to indicate that prompt_embeds alone is not sufficient because prompt is needed to generate prior_token_ids when they are not provided.

Suggested change

	if prompt is None and prior_token_ids is None:
	if prompt is None and prior_token_ids is None:
	# At this point, `prompt_embeds` is guaranteed to be not None (the earlier check
	# `elif prompt is None and prompt_embeds is None` would already have raised),
	# so clarify that `prompt_embeds` alone is not sufficient to derive `prior_token_ids`.
	if prompt_embeds is not None:
	raise ValueError(
	"You have provided `prompt_embeds` without `prior_token_ids`, but a text `prompt` is also "
	"required so that `prior_token_ids` can be generated. Please provide either a text `prompt` "
	"so the pipeline can compute `prior_token_ids`, or pass `prior_token_ids` explicitly."
	)