Skip to content

Megatron GLM evaluation #45

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

Draft
pawalt wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

Megatron GLM evaluation #45

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
28 changes: 28 additions & 0 deletions megatron/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,6 +74,33 @@ This command:

**Memory Note:** The default configuration (mbs=1, full recompute) is chosen for long-context examples. Increase micro batch size only if you have headroom, or increase tensor/pipeline parallelism.

### 4. Evaluation

Run Megatron native evaluation on the latest LoRA checkpoint using the
existing LongMIT-128K dataset:

```bash
modal run modal_train.py::eval_lora
```

Optional overrides:

```bash
# Use a specific checkpoint subdirectory under /checkpoints
modal run modal_train.py::eval_lora --checkpoint-subdir glm47_lora

# Point directly to a checkpoint dir or use a run id suffix
modal run modal_train.py::eval_lora --checkpoint-dir /checkpoints/glm47_lora
modal run modal_train.py::eval_lora --run-id <run-id>

# Adjust evaluation settings
modal run modal_train.py::eval_lora --context 128k --lora-rank 128 --eval-iters 2
```

The evaluator will re-use the existing preprocessed dataset at
`/data/longmit-128k`. If `validation.jsonl` is missing, it creates a
lightweight link to `training.jsonl` for the duration of the run.

## Training Configuration

### Parallelism Strategy
Expand Down Expand Up @@ -152,5 +179,6 @@ nvidia-smi
|------|-------------|
| `modal_train.py` | Modal app with download, convert, prep, and train functions |
| `train.py` | Training script executed via torchrun on each node |
| `eval.py` | Evaluation script executed via torchrun on each node |
| `memory_calc.py` | GPU memory calculator for exploring parallelism configurations |

Loading