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 16, 02:00 UTC
[Misc] Fix Deepseek V2 fp8 kv-scale remapping (#11947)
Signed-off-by: Yida Wu <yidawu@alumni.cmu.edu>
Concurrensee committed
Jan 13, 2025 at 01:05 UTC
cf6bbcb49324c24fc0f6f9381400c299c9c2d7ac
1 file changed
+7
-1
vllm/model_executor/models/deepseek_v2.py
+7
-1
@@ -45,7 +45,8 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
45
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
46
from vllm.model_executor.layers.vocab_parallel_embedding import (
47
ParallelLMHead, VocabParallelEmbedding)
48
-from vllm.model_executor.model_loader.weight_utils import default_weight_loader
48
+from vllm.model_executor.model_loader.weight_utils import (
49
+ default_weight_loader, maybe_remap_kv_scale_name)
50
from vllm.model_executor.sampling_metadata import SamplingMetadata
51
from vllm.sequence import IntermediateTensors
52
@@ -635,6 +636,11 @@ class DeepseekV2ForCausalLM(nn.Module, SupportsPP):
636
if name.endswith(".bias") and name not in params_dict:
637
continue
638
639
+ # Remapping the name of FP8 kv-scale.
640
+ name = maybe_remap_kv_scale_name(name, params_dict)
641
+ if name is None:
642
+ continue
643
+
644
if is_pp_missing_parameter(name, self):
645
continue
646