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 14, 05:30 UTC
Fix weights loading for Apertus (#24100)
Signed-off-by: Nathan Ranchin <nranchin@student.ethz.ch>
nathan committed
Sep 2, 2025 at 20:34 UTC
598bd74cf80f91d0de422d6ac994034d86b9e31c
1 file changed
+6
vllm/model_executor/models/apertus.py
+6
@@ -415,6 +415,12 @@ class ApertusModel(nn.Module):
415
(".qkv_proj", ".v_proj", "v"),
416
]
417
params_dict = dict(self.named_parameters())
418
+
419
+ # we need to load the buffers for beta and eps (XIELU)
420
+ for name, buffer in self.named_buffers():
421
+ if name.endswith(".beta") or name.endswith(".eps"):
422
+ params_dict[name] = buffer
423
+
424
loaded_params: set[str] = set()
425
for name, loaded_weight in weights:
426
if "rotary_emb.inv_freq" in name: