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 7, 02:45 UTC
[Bugfix] Fix compressed-tensors quantization failure for DeepSeek-R1 on MI300x (#36247)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
vllmellm committed
Mar 7, 2026 at 17:26 UTC
ee8a29511fc69e3f0f6291fa6ff1cf6e47f7750d
1 file changed
+2
-2
vllm/model_executor/models/deepseek_v2.py
+2
-2
@@ -756,7 +756,7 @@ direct_register_custom_op(
756
)
757
758
759
-class DeepSeekV2FusedQkvAProj(MergedColumnParallelLinear):
759
+class DeepSeekV2FusedQkvAProjLinear(MergedColumnParallelLinear):
760
def __init__(
761
self,
762
input_size: int,
@@ -848,7 +848,7 @@ class DeepseekV2MLAAttention(nn.Module):
848
self.max_position_embeddings = max_position_embeddings
849
850
if self.q_lora_rank is not None:
851
- self.fused_qkv_a_proj = DeepSeekV2FusedQkvAProj(
851
+ self.fused_qkv_a_proj = DeepSeekV2FusedQkvAProjLinear(
852
self.hidden_size,
853
[self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
854
quant_config=quant_config,