fix: convert boundary_ratio from timestep space to index space (#32)

* fix: convert boundary_ratio from timestep space to index space via inverse sigma_shift

* fix: use table lookup for boundary index instead of inverse sigma_shift formula

Replace the analytical inverse sigma_shift formula with a direct table
lookup on scheduler.timesteps. This avoids floating-point precision
issues (e.g. 0.24999 vs 0.25) and the sigma_min != 0 approximation
error, giving an exact boundary at the intended timestep threshold.

Author: Auraithm

mova/diffusion/pipelines/mova_train.py

@@ -1386,10 +1386,13 @@ def training_step(
             mode_scale=1.0,
             independent_timesteps=False,
         )
+        # Look up the index boundary directly from the scheduler's timestep table.
+        # boundary_ratio=0.9 means timestep 900; count how many timesteps >= 900 to get the index fraction.
+        boundary = (self.scheduler.timesteps >= self.boundary_ratio * self.scheduler.num_train_timesteps).sum().item() / self.scheduler.num_train_timesteps
         if global_step % 2 == 0:
-            timestep_config.max_timestep_boundary = self.boundary_ratio
+            timestep_config.max_timestep_boundary = boundary
         else:
-            timestep_config.min_timestep_boundary = self.boundary_ratio
+            timestep_config.min_timestep_boundary = boundary
 
         timestep, audio_timestep = self.sample_timestep_pair(timestep_config)
         timestep = timestep.to(device=device)