xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-01-23 15:34:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

[BugFix] Fix Llama4 Pipeline Parallelism Assert Error (#28577)

Browse Source 
Co-authored-by: Dezhan Tu <dztu@meta.com>
This commit is contained in:
Dezhan 2025-11-20 02:52:36 -08:00 committed by GitHub
parent fb8851f254
commit dc45efc8ef
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 7 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/models/llama4.py
View File

@ -53,6 +53,7 @@ from vllm.model_executor.models.utils import sequence_parallel_chunk
from .llama import LlamaForCausalLM, LlamaMLP, LlamaModel
from .utils import (
AutoWeightsLoader,
PPMissingLayer,
extract_layer_index,
fast_topk,
is_pp_missing_parameter,
@ -729,6 +730,9 @@ class Llama4ForCausalLM(LlamaForCausalLM, MixtureOfExperts):
self.moe_layers = []
example_moe = None
for layer in self.model.layers:
if isinstance(layer, PPMissingLayer):
continue
assert isinstance(layer, Llama4DecoderLayer)
if isinstance(layer.feed_forward, Llama4MoE):
# Pick last one layer since the first ones may be dense layers.
@ -765,6 +769,9 @@ class Llama4ForCausalLM(LlamaForCausalLM, MixtureOfExperts):
self.num_local_physical_experts = num_local_physical_experts
self.num_redundant_experts = num_physical_experts - self.num_logical_experts
for layer in self.model.layers:
if isinstance(layer, PPMissingLayer):
continue
if isinstance(layer.feed_forward, Llama4MoE):
moe = layer.feed_forward
moe.n_local_physical_experts = num_local_physical_experts