rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-06-05 21:59:24 +02:00
Code Issues Projects Releases Wiki Activity

Better AutoGPTQ fallback

Browse Source
This commit is contained in:
Henk
2023-08-10 18:10:48 +02:00
parent f2d7ef3aca
commit 9c7ebe3b04
1 changed files with 20 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
modeling/inference_models/gptq_hf_torch/class.py
View File

@@ -169,6 +169,7 @@ class model_backend(HFTorchInferenceModel):
self.init_model_config()
self.lazy_load = True
self.implementation = "occam"
gpulayers = self.breakmodel_config.gpu_blocks
@@ -323,22 +324,25 @@ class model_backend(HFTorchInferenceModel):
enable=self.lazy_load,
dematerialized_modules=False,
):
try:
if model_type == "gptj":
model = load_quant_offload_device_map(gptj_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "gpt_neox":
model = load_quant_offload_device_map(gptneox_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "llama":
model = load_quant_offload_device_map(llama_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "opt":
model = load_quant_offload_device_map(opt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "mpt":
model = load_quant_offload_device_map(mpt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "gpt_bigcode":
model = load_quant_offload_device_map(bigcode_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias).half()
else:
raise RuntimeError("Model not supported by Occam's GPTQ")
except:
if self.implementation == "occam":
try:
if model_type == "gptj":
model = load_quant_offload_device_map(gptj_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "gpt_neox":
model = load_quant_offload_device_map(gptneox_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "llama":
model = load_quant_offload_device_map(llama_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "opt":
model = load_quant_offload_device_map(opt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_tseype == "mpt":
model = load_quant_offload_device_map(mpt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
elif model_type == "gpt_bigcode":
model = load_quant_offload_device_map(bigcode_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias).half()
else:
raise RuntimeError("Model not supported by Occam's GPTQ")
except:
self.implementation = "AutoGPTQ"
if self.implementation == "AutoGPTQ":
try:
import auto_gptq
from auto_gptq import AutoGPTQForCausalLM