[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:
Jakub Byczkowski
2026-09-10 17:47:29 +08:00
committed by GitHub
co-authored by Copilot Kunshang Ji
parent fe26c705f7
commit 6fd08c45e6
+3 -1
View File
@@ -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)