biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc][Quark] Upstream Quark format to VLLM (#10765)

Browse Source 
Signed-off-by: kewang-xlnx <kewang@xilinx.com>
Signed-off-by: kewang2 <kewang2@amd.com>
Co-authored-by: kewang2 <kewang2@amd.com>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
This commit is contained in:
kewang-xlnx
2025-01-16 00:05:15 +08:00
committed by GitHub
parent 5ecf3e0aaf
commit de0526f668
32 changed files with 1264 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/models/llama.py
View File

@@ -38,8 +38,6 @@ from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.quantization.compressed_tensors.utils import (
get_compressed_tensors_cache_scale)
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 (
@@ -306,6 +304,7 @@ class LlamaModel(nn.Module):
lora_config = vllm_config.lora_config
self.config = config
self.quant_config = quant_config
self.padding_idx = config.pad_token_id
lora_vocab = (lora_config.lora_extra_vocab_size *
(lora_config.max_loras or 1)) if lora_config else 0
@@ -396,12 +395,15 @@ class LlamaModel(nn.Module):
# Models trained using ColossalAI may include these tensors in
# the checkpoint. Skip them.
continue
if scale_name := get_compressed_tensors_cache_scale(name):
# Loading kv cache scales for compressed-tensors quantization
if (self.quant_config is not None and
(scale_name := self.quant_config.get_cache_scale(name))):
# Loading kv cache scales for quark and
# compressed-tensors quantization
param = params_dict[scale_name]
weight_loader = getattr(param, "weight_loader",
default_weight_loader)
loaded_weight = loaded_weight[0]
loaded_weight = (loaded_weight if loaded_weight.dim() == 0 else
loaded_weight[0])
weight_loader(param, loaded_weight)
loaded_params.add(scale_name)
continue