THUDM / THUDM/slime

An error occurred while converting Llama3.2-1B-Instruct to Megatron ckpt.

Open
#1,430 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
8.5k
Forks
1.3k
Avg merge
5h 36m
Merged PRs (30d)
22

Description

I need to train the Llama 3.2-1B-Instruct model, but there is no configuration file for it in slime/scripts/models, which prevents me from converting it into a Megatron checkpoint. I have added a custom configuration file llama3.2-1B-Instruct.sh with the following parameters:

MODEL_ARGS=(
   --swiglu
   --num-layers 16
   --hidden-size 2048
   --ffn-hidden-size 8192
   --num-attention-heads 32
   --group-query-attention
   --num-query-groups 8
   --max-position-embeddings 131072
   --use-rotary-position-embeddings
   --disable-bias-linear
   --normalization "RMSNorm"
   --norm-epsilon 1e-5
   --rotary-base 500000
   --vocab-size 128256
   --kv-channels 64
   --use-rope-scaling
   --rotary-scaling-factor 32.0
)

However, during the training process, the sample.response is garbled. I suspect that there was an error during the conversion to the Megatron checkpoint. How can I resolve this issue?

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reviewing the custom configuration in slime/scripts/models and the conversion process that produces the Megatron checkpoint. Reproduce the training case with Llama 3.2-1B-Instruct, verify whether the checkpoint conversion is valid, and confirm that sample.response is no longer garbled.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, shell
Domain
machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.