[diffusion] fix: fix single-step flow-match timesteps (#24708)

python/sglang/multimodal_gen/runtime/models/schedulers/flow_match_pair.py

@@ -82,6 +82,7 @@ class FlowMatchScheduler(BaseScheduler):
         if self.shift_terminal is not None:
             one_minus_z = 1 - self.sigmas
             scale_factor = one_minus_z[-1] / (1 - self.shift_terminal)
+            if scale_factor != 0:
                 self.sigmas = 1 - (one_minus_z / scale_factor)
         if self.reverse_sigmas:
             self.sigmas = 1 - self.sigmas
@@ -421,6 +422,7 @@ class FlowMatchPairScheduler(FlowMatchScheduler):
                     if self.shift_terminal is not None:
                         one_minus_z = 1 - base
                         scale_factor = one_minus_z[-1] / (1 - self.shift_terminal)
+                        if scale_factor != 0:
                             base = 1 - (one_minus_z / scale_factor)
 
                     if self.reverse_sigmas:

python/sglang/multimodal_gen/runtime/models/schedulers/scheduling_flow_match_euler_discrete.py

@@ -266,6 +266,8 @@ class FlowMatchEulerDiscreteScheduler(
         """
         one_minus_z = 1 - t
         scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal)
+        if scale_factor == 0:
+            return t
         stretched_t = 1 - (one_minus_z / scale_factor)
         return stretched_t
 

python/sglang/multimodal_gen/runtime/pipelines_core/stages/timestep_preparation.py

@@ -191,8 +191,7 @@ class TimestepPreparationStage(PipelineStage):
             and isinstance(batch.timesteps, torch.Tensor)
             and torch.isnan(batch.timesteps).any()
         ):
-            # when num-inference-steps == 1, the last sigma being 1, the 1 / last_sigma could be nan
+            # diffusers flow-match scheduler can emit NaN for one-step warmup
-            # this a workaround for warmup req only
             batch.timesteps = torch.ones(
                 (1,), dtype=torch.float32, device=get_local_torch_device()
             )