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 13, 09:30 UTC
[Model] fix glm4_moe_mtp load weights with GLM-4.6 checkpoint. (#27597)
Signed-off-by: wuao.scotty <wuao.scotty@bytedance.com> Co-authored-by: wuao.scotty <wuao.scotty@bytedance.com>
wuyaoxuehun committed
Nov 12, 2025 at 17:14 UTC
d3ade61e429fcae40829587244ced9c553120703
1 file changed
+11
-4
vllm/model_executor/models/glm4_moe_mtp.py
+11
-4
@@ -256,11 +256,18 @@ class Glm4MoeMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
256
257
params_dict = dict(self.named_parameters())
258
loaded_params: set[str] = set()
259
+ spec_layer = self.model.mtp_start_layer_idx
260
for name, loaded_weight in weights:
260
- spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
261
- if spec_layer is None:
262
- continue
263
- name = self._rewrite_spec_layer_name(spec_layer, name)
261
+ if name == "lm_head.weight":
262
+ name = f"model.layers.{spec_layer}.shard_head.head.weight"
263
+ elif name == "model.embed_tokens.weight":
264
+ # This name is same with local model, rewriting is not needed.
265
+ pass
266
+ else:
267
+ spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
268
+ if spec_layer is None:
269
+ continue
270
+ name = self._rewrite_spec_layer_name(spec_layer, name)
271
for param_name, weight_name, shard_id in stacked_params_mapping:
272
# Skip non-stacked layers and experts (experts handled below).
273
if weight_name not in name: