Skip to content

fix vptq cache dir and model name #160

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
wejoncy merged 1 commit into from
Feb 11, 2025
Merged

fix vptq cache dir and model name #160

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions qllm/quantization/config_builder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -147,5 +147,6 @@ def build_config(args):
with open(args.quant_config, 'r') as fp:
dict_config = json.load(fp)
config = VPTQConfig.from_dict(dict_config)
config.model_name = args.load + args.model # one of them is empty

return config
17 changes: 8 additions & 9 deletions qllm/quantization/vptq/quant_vptq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@ def __init__(self, config) -> None:
self.quant_config.output_dir = Path(self.quant_config.output_dir) / self.quant_config.model_name
for k, v in self.quant_config.layer_config.to_dict().items():
setattr(self.quant_config, k, v)
self.quant_cache_dir = Path(f"{self.quant_config.output_dir}/quant_cache")

def set_tokenizer(self, tokenizer):
self.tokenizer = tokenizer
Expand Down Expand Up @@ -89,7 +90,7 @@ def collect_hessian_pre(self, model, model_prefix, dev):
from .qllm_hessian import process_collect_hessian
sample_args = self.quant_config.hessian_config
sample_args.base_model = self.quant_config.model_name
sample_args.save_path = f"./hessian_path/{sample_args.base_model}_{sample_args.devset_size}_{sample_args.ctx_size}"
sample_args.save_path = f"{self.quant_config.output_dir}/hessian_path/{sample_args.base_model}_{sample_args.devset_size}_{sample_args.ctx_size}"

self.quant_config.hessian_path = sample_args.save_path
self.quant_config.inv_hessian_path = sample_args.save_path+"_inv"
Expand Down Expand Up @@ -123,21 +124,20 @@ def parallel_quantize(self, quantize_layer, attention_layers, num_gpus, dev):

pbar = tqdm.tqdm(total=len(attention_layers), desc=f"running VPTQ on {num_gpus} GPUs")
output_queue = theading_queue.Queue()
quant_tmp = Path("quant_tmp")
for i in range(num_gpus):
output_queue.put(i) # poison pill
def fetch_next_task(future):
comm_utils.clear_memory()
pbar.update(1)
pbar.set_postfix_str(f'gpu memory: {torch.cuda.memory_allocated(future.gpu_idx)/1024**3:.2f}GB')
output_queue.put(future.gpu_idx)
torch.save(future.result(), quant_tmp/f"layer_{future.layer_idx}.pt")
torch.save(future.result(), self.quant_cache_dir/f"layer_{future.layer_idx}.pt")

for layer_idx,layer in enumerate(attention_layers):
if (quant_tmp/f"layer_{layer_idx}.pt").exists():
if (self.quant_cache_dir/f"layer_{layer_idx}.pt").exists():
import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)
attention_layers[layer_idx] = torch.load(quant_tmp / f"layer_{layer_idx}.pt", weights_only=False)
attention_layers[layer_idx] = torch.load(self.quant_cache_dir / f"layer_{layer_idx}.pt", weights_only=False)
pbar.update(1)
continue
free_gpu_id = output_queue.get()
Expand Down Expand Up @@ -178,15 +178,14 @@ def do_quantize(self, model, dataloader, model_prefix, dev):
vptq_quantizer = InternalVPTQQuantizer()
quantize_layer = vptq_quantizer.quantize_layer
quantizers = {}
quant_tmp = Path("quant_tmp")
quant_tmp.mkdir(exist_ok=True)
self.quant_cache_dir.mkdir(exist_ok=True)

if num_gpus > 1:
self.parallel_quantize(quantize_layer, attention_layers, num_gpus, dev)
else:
for layer_idx in tqdm.trange((len(attention_layers)), desc="running VPTQ"):
if (quant_tmp/f"layer_{layer_idx}.pt").exists():
attention_layers[layer_idx] = torch.load(quant_tmp / f"layer_{layer_idx}.pt", weights_only=False)
if (self.quant_cache_dir/f"layer_{layer_idx}.pt").exists():
attention_layers[layer_idx] = torch.load(self.quant_cache_dir / f"layer_{layer_idx}.pt", weights_only=False)
continue
attention_layers[layer_idx] = quantize_layer(
(attention_layers[layer_idx], layer_idx), self.quant_config, self.quant_config,
Expand Down