Defer pe embedding on Spectrum forecast steps
This commit is contained in:
@@ -388,12 +388,6 @@ def _run_flux_forward_with_spectrum(
|
||||
# Spectrum forecasts the final image-token feature right before final_layer.
|
||||
expected_feature_shape = (img.shape[0], img.shape[1], *img.shape[2:])
|
||||
|
||||
if img_ids is not None:
|
||||
ids = torch.cat((txt_ids, img_ids), dim=1)
|
||||
pe = inner.pe_embedder(ids)
|
||||
else:
|
||||
pe = None
|
||||
|
||||
vec_orig = vec
|
||||
txt_vec = vec
|
||||
extra_kwargs = {}
|
||||
@@ -446,6 +440,12 @@ def _run_flux_forward_with_spectrum(
|
||||
pred_feature = _sanitize_forecast_feature_for_final_layer(pred_feature, img.dtype)
|
||||
return inner.final_layer(pred_feature, vec_orig, **final_kwargs)
|
||||
|
||||
if img_ids is not None:
|
||||
ids = torch.cat((txt_ids, img_ids), dim=1)
|
||||
pe = inner.pe_embedder(ids)
|
||||
else:
|
||||
pe = None
|
||||
|
||||
if inner.params.global_modulation:
|
||||
vec = (inner.double_stream_modulation_img(vec_orig), inner.double_stream_modulation_txt(txt_vec))
|
||||
|
||||
|
||||
Reference in New Issue
Block a user