multimodalart HF Staff commited on
Commit
1cae2c0
·
1 Parent(s): 9d32de2

Clone the locked soundtrack back in: the next step overwrites the rows in place

Browse files
Files changed (1) hide show
  1. h3_a2v_blocks.py +4 -1
h3_a2v_blocks.py CHANGED
@@ -280,7 +280,10 @@ class MiniMaxH3A2VLoopSchedulerStep(MiniMaxH3LoopSchedulerStep):
280
  return components, block_state
281
 
282
  if block_state.audio_condition_mode == LOCKED:
283
- block_state.audio_latents = given.to(block_state.audio_latents)
 
 
 
284
  else:
285
  # The rows the *next* forward reads are at the next entry of the audio schedule; past the last one the
286
  # schedule has reached sigma = 0, i.e. `t = 1`, which returns the soundtrack itself.
 
280
  return components, block_state
281
 
282
  if block_state.audio_condition_mode == LOCKED:
283
+ # `.clone()` is load-bearing: `Tensor.to` is a no-op when the dtype and device already match, so without
284
+ # it the rows *are* `given_audio_rows`, and the next step's in-place `audio_latents[0:] = ...` overwrites
285
+ # the soundtrack itself. The locked leg then drifts into the model's own audio from step 1 on.
286
+ block_state.audio_latents = given.to(block_state.audio_latents).clone()
287
  else:
288
  # The rows the *next* forward reads are at the next entry of the audio schedule; past the last one the
289
  # schedule has reached sigma = 0, i.e. `t = 1`, which returns the soundtrack itself.