A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 21:00 UTC
[BugFix] Fix Llama4 Pipeline Parallelism Assert Error (#28577)
Co-authored-by: Dezhan Tu <dztu@meta.com>
Dezhan committed
Nov 20, 2025 at 02:52 UTC
dc45efc8ef7fc1e2571331eaf4671e1652e2a865
1 file changed
+7
vllm/model_executor/models/llama4.py
+7
@@ -53,6 +53,7 @@ from vllm.model_executor.models.utils import sequence_parallel_chunk
53
from .llama import LlamaForCausalLM, LlamaMLP, LlamaModel
54
from .utils import (
55
AutoWeightsLoader,
56
+ PPMissingLayer,
57
extract_layer_index,
58
fast_topk,
59
is_pp_missing_parameter,
@@ -729,6 +730,9 @@ class Llama4ForCausalLM(LlamaForCausalLM, MixtureOfExperts):
730
self.moe_layers = []
731
example_moe = None
732
for layer in self.model.layers:
733
+ if isinstance(layer, PPMissingLayer):
734
+ continue
735
+
736
assert isinstance(layer, Llama4DecoderLayer)
737
if isinstance(layer.feed_forward, Llama4MoE):
738
# Pick last one layer since the first ones may be dense layers.
@@ -765,6 +769,9 @@ class Llama4ForCausalLM(LlamaForCausalLM, MixtureOfExperts):
769
self.num_local_physical_experts = num_local_physical_experts
770
self.num_redundant_experts = num_physical_experts - self.num_logical_experts
771
for layer in self.model.layers:
772
+ if isinstance(layer, PPMissingLayer):
773
+ continue
774
+
775
if isinstance(layer.feed_forward, Llama4MoE):
776
moe = layer.feed_forward
777
moe.n_local_physical_experts = num_local_physical_experts