From 6fd08c45e696999b403b64835ae44d322efaab35 Mon Sep 17 00:00:00 2001 From: Jakub Byczkowski Date: Thu, 10 Sep 2026 11:47:29 +0200 Subject: [PATCH] [XPU][Bugfix] Fix FalconH1 pipeline-parallel execution (#55913) Signed-off-by: Jakub Byczkowski Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> Co-authored-by: Kunshang Ji --- vllm/model_executor/models/falcon_h1.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/falcon_h1.py b/vllm/model_executor/models/falcon_h1.py index 0dca542ff3..5b9853ece1 100644 --- a/vllm/model_executor/models/falcon_h1.py +++ b/vllm/model_executor/models/falcon_h1.py @@ -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)