[Bugfix] Fix DeepseekV2Model hidden_size (#46986)
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
Jee Jee Li committed
Jun 30, 2026 at 00:44 UTC
7be582697b27277e2756a3878f563fa9dfea30aa
1 file changed
+4
-4
vllm/model_executor/models/deepseek_v2.py
+4
-4
@@ -1336,7 +1336,7 @@ class DeepseekV2Model(nn.Module):
1336
quant_config = vllm_config.quant_config
1337
self.config = config
1338
self.device = current_platform.device_type
1339
-
1339
+ self.hidden_size = config.hidden_size
1340
self.vocab_size = config.vocab_size
1341
self.is_v32 = hasattr(config, "index_topk")
1342
if self.is_v32:
@@ -1353,7 +1353,7 @@ class DeepseekV2Model(nn.Module):
1353
if get_pp_group().is_first_rank:
1354
self.embed_tokens = VocabParallelEmbedding(
1355
config.vocab_size,
1356
- config.hidden_size,
1356
+ self.hidden_size,
1357
quant_config=quant_config,
1358
prefix=f"{prefix}.embed_tokens",
1359
)
@@ -1370,11 +1370,11 @@ class DeepseekV2Model(nn.Module):
1370
)
1371
1372
if get_pp_group().is_last_rank:
1373
- self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
1373
+ self.norm = RMSNorm(self.hidden_size, eps=config.rms_norm_eps)
1374
else:
1375
self.norm = PPMissingLayer()
1376
self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
1377
- ["hidden_states", "residual"], config.hidden_size
1377
+ ["hidden_states", "residual"], self.hidden_size
1378
)
1379
1380
self.aux_hidden_state_layers = tuple[int, ...]()