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

AutoGPTQ fallback

Browse Source
This commit is contained in:
Henk
2023-08-10 17:18:53 +02:00
parent 1b253ce95f
commit 54addfc234
1 changed files with 16 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
modeling/inference_models/gptq_hf_torch/class.py
View File

@@ -323,19 +323,22 @@ class model_backend(HFTorchInferenceModel):
enable=self.lazy_load, enable=self.lazy_load,
dematerialized_modules=False, dematerialized_modules=False,
): ):
if model_type == "gptj": try:
model = load_quant_offload_device_map(gptj_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias) if model_type == "gptj":
elif model_type == "gpt_neox": model = load_quant_offload_device_map(gptj_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
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 == "gpt_neox":
elif model_type == "llama": model = load_quant_offload_device_map(gptneox_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
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 == "llama":
elif model_type == "opt": model = load_quant_offload_device_map(llama_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
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 == "opt":
elif model_type == "mpt": model = load_quant_offload_device_map(opt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
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 == "mpt":
elif model_type == "gpt_bigcode": model = load_quant_offload_device_map(mpt_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias)
model = load_quant_offload_device_map(bigcode_load_quant, location, gptq_file, gptq_bits, gptq_groupsize, device_map, force_bias=v2_bias).half() elif model_type == "gpt_bigcode":
else: 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:
try: try:
import auto_gptq import auto_gptq
from auto_gptq import AutoGPTQForCausalLM from auto_gptq import AutoGPTQForCausalLM