[XPU][Bugfix] Fix FalconH1 pipeline-parallel execution (#55913)
Signed-off-by: Jakub Byczkowski <jakub.byczkowski@intel.com> Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
co-authored by
Copilot
Kunshang Ji
parent
fe26c705f7
commit
6fd08c45e6
@@ -449,8 +449,10 @@ class FalconH1Model(nn.Module):
|
||||
self.start_layer, self.end_layer, self.layers = make_layers(
|
||||
config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers"
|
||||
)
|
||||
# FalconH1ParallelHybrid recomputes residual internally each layer,
|
||||
# it is never threaded across the PP boundary.
|
||||
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
|
||||
["hidden_states", "residual"], config.hidden_size
|
||||
["hidden_states"], config.hidden_size
|
||||
)
|
||||
if get_pp_group().is_last_rank:
|
||||
self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
||||
|
||||
Reference in New Issue
Block a user