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, 13:15 UTC
[BugFix] 'DeepseekV2Config' object has no attribute 'use_mla'` (#28387)
Signed-off-by: Lin, Fanli <fanli.lin@intel.com>
Fanli Lin committed
Nov 11, 2025 at 16:45 UTC
c7991269dd8fe86096a3eee5040e855801ae9665
1 file changed
+5
-1
vllm/model_executor/models/kimi_vl.py
+5
-1
@@ -456,7 +456,11 @@ class KimiVLForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):
456
(".gate_up_proj", ".gate_proj", 0),
457
(".gate_up_proj", ".up_proj", 1),
458
]
459
- if not config.use_mla:
459
+ use_mha = (
460
+ config.model_type == "deepseek"
461
+ or config.qk_nope_head_dim + config.qk_rope_head_dim == 0
462
+ )
463
+ if use_mha:
464
stacked_params_mapping += [
465
(".qkv_proj", ".q_proj", "q"),
466
(".qkv_proj", ".k_proj", "k"),