Fix FlowMatchEulerDiscreteScheduler double-shift in set_timesteps

[Ricardo-M-L]

What does this PR do?

Fixes #13243 — FlowMatchEulerDiscreteScheduler.set_timesteps applies the static timestep shift a second time when use_dynamic_shifting=False, so calling set_timesteps(num_train_timesteps) on a scheduler produces a different sigma schedule than the one __init__ cached.

Root cause

__init__ records sigma_min/sigma_max from the post-shift sigma range:

sigmas = timesteps / num_train_timesteps
if not use_dynamic_shifting:
    sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)  # ← shift applied
...
self.sigma_min = self.sigmas[-1].item()  # ← post-shift value
self.sigma_max = self.sigmas[0].item()   # ← post-shift value

set_timesteps then rebuilds sigmas via linspace(_sigma_to_t(sigma_max), _sigma_to_t(sigma_min), N) — already in shifted space — and applies the same shift again:

sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas)

Reproduction

from diffusers import FlowMatchEulerDiscreteScheduler
import torch

s = FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, shift=3.0, use_dynamic_shifting=False)
init_sigmas = s.sigmas.clone()
s.set_timesteps(num_inference_steps=1000)

print(init_sigmas[-3:])       # tensor([0.0089, 0.0060, 0.0030])
print(s.sigmas[-4:-1])        # tensor([0.0148, 0.0119, 0.0089])  ← different
torch.allclose(init_sigmas, s.sigmas[:-1])  # False

The same args produce two different schedules.

Fix

Record the pre-shift sigmas in __init__ and derive sigma_min/sigma_max from them. set_timesteps then regenerates the linear sigma range correctly and the static shift is applied exactly once. self.sigmas is unchanged (still post-shift); only the cached endpoints change.

Test

Added tests/schedulers/test_scheduler_flow_match_euler_discrete.py:

• test_set_timesteps_matches_init_with_static_shift — fails on main, passes with the fix.
• test_dynamic_shifting_is_unaffected — pins that the use_dynamic_shifting=True path is untouched (no static shift in __init__, so this code path never had the bug).

Before submitting

• This PR fixes a bug — links [BUG] FlowMatchEulerDiscreteScheduler.__init__ computes sigma_min/sigma_max after shift, causing duplicate shift in set_timesteps #13243 with a clear repro.
• Did you read the contributor guideline?
• Did you write any new necessary tests? — Yes, regression test added.

Who can review?

@yiyixuxu @DN6