Could I get safe tensor without lazy loading?

[voidxb]

System Info

I see safe_open and deserialize, it seems that both two are lazy loading.
So if I don't want to load safetensor without lazy loading
how could I do, thanks

Information

• The official example scripts
• My own modified scripts

Reproduction

I use sglang, and in sglang model_loader/weight_utils.py
it load safetensors like this
if not is_all_weights_sharded: with safe_open(st_file, framework="pt") as f: for name in f.keys(): # noqa: SIM118 param = f.get_tensor(name) yield name, param else: result = load_file(st_file, device="cpu") for name, param in result.items(): yield name, param
I found it loads safe tensor too slow(about 20min+), whether is_all_weights_sharded is True
and if I prefetch safetensors before load_model(like cat * > /dev/null), it could only cost 5min
I try to use threadExecutor to parallel this code, and although get_tensor could be quick, but loading weight still cost 20min +, so I doubt that lazy loading.thanks

Expected behavior

without lazy loading

[Narsil]

load(open(filename, "r").read()).

This has been extensively discussed in many places so I won't repeat everything.
You're using some kind of network disk access using a mount point. The OS uses memory mapping assuming the underlying data is an actual disk therefore doing suboptimal decisions. Just force load everything in one go.

Having lazy-loading using shared memory is what makes safetensors fast in the first place when using real disks.