Support fixed sampling sigmas in FlowMatchEulerDiscreteScheduler

[naykun]

What does this PR do?

Distilled checkpoints can require a fixed sampling grid. Add sample_sigmas to FlowMatchEulerDiscreteScheduler so that grid can be saved in the scheduler config and used without applying shifting or sigma conversions again. Explicit sigmas or timesteps retain their existing behavior and override the config.

Update QwenImage21Pipeline to use the configured grid and its step count when no explicit sigmas are supplied. This also takes precedence over the pipeline's num_inference_steps argument. Add regression tests for the configured schedule and explicit sigma overrides.

Validation

• Rebased onto upstream main at 578c9b2c6; both original patches are unchanged.
• Ten additional local scheduler checks passed: conversion bypass, inversion, Euler stepping, state reset, step-count mismatch, explicit overrides, and config round-trip.
• make style and make fix-copies completed; unrelated formatter changes were discarded.
• Three pipeline regression cases passed on CPU (configured schedule with default/explicit step counts and explicit sigma override).

Self-review

No blocking correctness issues or unused added code found for the supported fixed-schedule inputs. The scheduler branch is exercised through set_timesteps, and the pipeline delegates denoising to the real scheduler.

Documentation suggestion for review: the pipeline's argument docstring could explicitly state that configured sample_sigmas takes precedence over num_inference_steps. This precedence is covered by the regression tests and described above. The scheduler's terminal-sigma documentation could also mention the existing invert_sigmas=True convention (terminal 1).

[yiyixuxu]

thanks, i left a comment
also i want to ask: does this checkpoint works only with this exact grid? e.g. would it still work well if user just pass num_inference_steps = 4?

[yiyixuxu]

i think it's easier to make the sample_sigmas a pipeline config, instead of a scheduler one, because we try to keep our our scheduler to a standard set of config/arguments so that people can swap them.

also, I think we would not need to make any changes to scheduler when we pass sigmas directly if we just also disable dynamic shifting/shift-terminal stretching etc in the scheduler config for distilled checkpoint here (but for the distilled checkpoint instead) https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/scheduler/scheduler_config.json, e.g. we can just set use_dynamic_shifting=False shift=1.0 etc

[naykun]

Implemented the pipeline-config approach in #4, following the suggestion here.

sample_sigmas now lives in model_index.json, and the scheduler code is unchanged. The distilled checkpoint config disables shifting/stretching/conversions; the pipeline passes the grid through the existing sigmas argument. The new PR also adds save/load and override tests and documents the checkpoint configuration.

On the real accelerated checkpoint (2048 × 2048, 8 steps, bf16, one fixed prompt/seed), the new approach exactly matched this PR's implementation in sigmas, timesteps, every denoising latent, and final RGBA pixels (maximum difference 0).

On the four-step question: quality with a different grid has not been evaluated. In the current implementation, num_inference_steps=4 alone still uses the configured eight-point grid; an explicit sigmas list overrides it.

[naykun]

thanks, i left a comment also i want to ask: does this checkpoint works only with this exact grid? e.g. would it still work well if user just pass num_inference_steps = 4?

thank you @yiyixuxu , yes currently we find with this specific grid the quality will be the best. but I think the community can explore other settings based on our recommendation.