[wip] Kylel/debug olmo core eval

[kyleclo]

there were some bugs in launching a command of this pattern (custom oe eval branch, using gantry)

olmo-cookbook-eval evaluate "$CHECKPOINT_PATH" \
    --tasks "$TASK" \
    --cluster "$CLUSTER" \
    --num-gpus 2 \
    --model-backend vllm \
    --partition-size 8 \
    --use-hf-token \
    --dashboard "$DASHBOARD" \
    --workspace "$WORKSPACE" \
    --vllm-use-v1-spec \
    --vllm-memory-utilization=0.8 \
    --model-args max_length=4096,trust_remote_code=true \
    --use-gantry \
    --oe-eval-branch "$OE_EVAL_BRANCH" \
    $DRY_RUN_FLAG

this change handles tracking of that git branch (and handles a sub-case if the custom branch specified is 'main')

[cursor]

Bug: Missing error check on subprocess git command

The subprocess.run call for git rev-parse --abbrev-ref HEAD is missing check=True, unlike all other subprocess calls in this function. If this git command fails, the code continues silently with result.stdout potentially empty or containing unexpected output. The subsequent comparison current_branch == commit_branch could then incorrectly fall through to the else branch, masking the actual error.