search

meta-llama / llama

Inference code for Llama models
Other
56.52k stars 9.58k forks source link

Getting 400 error on https://llama3-1.llamameta.net/Meta-Llama-3.1-405B-MP8/consolidated.00.pth #1146

Open jwatte opened 4 months ago

jwatte commented 4 months ago

Describe the bug

Using download.sh from an instance in GCP with plenty of network and storage, download of models in the llama-3.1 family works until it gets to Meta-Llama-3.1-405B-MP8, at which point it gets a 400 error. Re-trying the download still gets this error on that file.

Minimal reproducible example

./download.sh

Output

2024-07-25 00:25:49 (46.0 MB/s) - ‘./Meta-Llama-3.1-405B-MP8/tokenizer.model’ saved [2183982/2183982]

Downloading consolidated.00.pth
--2024-07-25 00:25:49--  https://llama3-1.llamameta.net/Meta-Llama-3.1-405B-MP8/consolidated.00.pth?Policy=xxxxxxxxx&Key-Pair-Id=K15QRJLYKIFSLZ&Download-Request-ID=776205421258895
Resolving llama3-1.llamameta.net (llama3-1.llamameta.net)... 3.163.165.48, 3.163.165.22, 3.163.165.9, ...
Connecting to llama3-1.llamameta.net (llama3-1.llamameta.net)|3.163.165.48|:443... connected.
HTTP request sent, awaiting response... 400 Bad Request
2024-07-25 00:25:49 ERROR 400: Bad Request.

Environment

df -h
10.125.224.2:/trainingfs   30T   14T   17T  46% /trainingfs
Keisial commented 3 months ago

Indeed. The models meta-llama-3.1-405b-instruct-mp16 and meta-llama-3.1-405b-mp16 download fine but the other models of 405B (meta-llama-3.1-405b-instruct-mp8, meta-llama-3.1-405b-instruct-fp8, meta-llama-3.1-405b-mp8 and meta-llama-3.1-405b-fp8) fail to download with a HTTP error 400 Bad request.

The other sizes of llama-3.1 have no such issues.