[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>
Jakub Byczkowski committed
Sep 10, 2026 at 11:47 UTC
6fd08c45e696999b403b64835ae44d322efaab35
1 file changed
+3
-1
vllm/model_executor/models/falcon_h1.py
+3
-1
@@ -449,8 +449,10 @@ class FalconH1Model(nn.Module):
449
self.start_layer, self.end_layer, self.layers = make_layers(
450
config.num_hidden_layers, get_layer, prefix=f"{prefix}.layers"
451
)
452
+ # FalconH1ParallelHybrid recomputes residual internally each layer,
453
+ # it is never threaded across the PP boundary.
454
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
453
- ["hidden_states", "residual"], config.hidden_size
455
+ ["hidden_states"], config.hidden_size
456
)
457
if get_pp_group().is_last_rank:
458
self.final_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)