xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-24 21:05:58 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Update starcoder2 to remap k/v scale names for kv_cache quantization (#11148)

Browse Source
This commit is contained in:
Dipika Sikka 2024-12-13 00:07:20 -05:00 committed by GitHub
parent 1efce68605
commit 3989a79824
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 6 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
vllm/model_executor/models/starcoder2.py
View File

@ -37,7 +37,8 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding)
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.model_loader.weight_utils import (
default_weight_loader, maybe_remap_kv_scale_name)
from vllm.model_executor.sampling_metadata import SamplingMetadata
from vllm.sequence import IntermediateTensors
@ -345,6 +346,10 @@ class Starcoder2ForCausalLM(nn.Module, SupportsPP):
weight_loader(param, loaded_weight, shard_id)
break
else:
name = maybe_remap_kv_scale_name(name, params_dict)
if name is None:
continue
if self.config.tie_word_embeddings and "lm_head.weight" in name:
continue
if is_pp_missing_parameter(name, self):