[Bugfix] Restore GLM-Image unshifted timesteps after set_timesteps (#7953)

Signed-off-by: sinksilk <785976238@qq.com>

tests/diffusion/models/glm_image/test_glm_image_retrieve_timesteps.py

@@ -0,0 +1,71 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""CPU tests for GLM-Image timestep handoff through retrieve_timesteps.
+
+GLM-Image integrates with resolution-shifted sigmas but conditions the DiT on
+unshifted timesteps. diffusers >= 0.40 FlowMatchEulerDiscreteScheduler.set_timesteps
+overwrites caller-supplied timesteps with `sigmas * num_train_timesteps`;
+retrieve_timesteps must restore the provided schedule.
+"""
+
+import numpy as np
+import pytest
+import torch
+from diffusers import FlowMatchEulerDiscreteScheduler
+
+from vllm_omni.diffusion.models.glm_image.pipeline_glm_image import (
+    calculate_shift,
+    retrieve_timesteps,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.diffusion]
+
+
+def _glm_image_scheduler() -> FlowMatchEulerDiscreteScheduler:
+    # Match scheduler/scheduler_config.json shipped with zai-org/GLM-Image.
+    return FlowMatchEulerDiscreteScheduler(
+        num_train_timesteps=1000,
+        use_dynamic_shifting=True,
+        time_shift_type="linear",
+        base_shift=0.25,
+        max_shift=0.75,
+    )
+
+
+def test_retrieve_timesteps_preserves_caller_unshifted_timesteps():
+    """Caller timesteps must survive set_timesteps when both args are passed."""
+    num_inference_steps = 50
+    # 1024x1024 -> image_seq_len = ((1024/8)*(1024/8)) / (2**2) = 4096
+    image_seq_len = 4096
+    scheduler = _glm_image_scheduler()
+
+    timesteps_array = np.linspace(scheduler.config.num_train_timesteps, 1.0, num_inference_steps + 1)[:-1]
+    timesteps_array = timesteps_array.astype(np.int64).astype(np.float32)
+    sigmas = timesteps_array / scheduler.config.num_train_timesteps
+    mu = calculate_shift(
+        image_seq_len,
+        scheduler.config.get("base_image_seq_len", 256),
+        scheduler.config.get("base_shift", 0.25),
+        scheduler.config.get("max_shift", 0.75),
+    )
+    assert mu == pytest.approx(3.25)
+
+    provided = timesteps_array.tolist()
+    got, n_steps = retrieve_timesteps(
+        scheduler,
+        num_inference_steps,
+        "cpu",
+        provided,
+        sigmas.tolist(),
+        mu=mu,
+    )
+
+    expected = torch.as_tensor(provided, dtype=torch.float32)
+    assert n_steps == num_inference_steps
+    assert got.shape == expected.shape
+    torch.testing.assert_close(got.cpu(), expected, rtol=0, atol=0)
+    torch.testing.assert_close(scheduler.timesteps.cpu(), expected, rtol=0, atol=0)
+
+    # Integration still uses the resolution-shifted sigma schedule.
+    shifted = torch.as_tensor(sigmas, dtype=torch.float32)
+    assert not torch.allclose(scheduler.sigmas[:-1].cpu(), shifted, rtol=0, atol=0)

vllm_omni/diffusion/models/glm_image/pipeline_glm_image.py

@@ -193,6 +193,10 @@ def retrieve_timesteps(
                 f" timestep or sigma schedules. Please check whether you are using the correct scheduler."
             )
         scheduler.set_timesteps(timesteps=timesteps, sigmas=sigmas, device=device, **kwargs)
+        # GLM-Image conditions the DiT on the *unshifted* timesteps while integrating with the
+        # resolution-shifted sigmas. FlowMatchEulerDiscreteScheduler.set_timesteps (diffusers >= 0.40)
+        # overwrites provided timesteps with `sigmas * num_train_timesteps`, so restore them here.
+        scheduler.timesteps = torch.as_tensor(timesteps, dtype=torch.float32, device=scheduler.timesteps.device)
         timesteps = scheduler.timesteps
         num_inference_steps = len(timesteps)
     elif timesteps is not None: