[CI] Fit small KV-offload evals within shared memory

[taneem-ibrahim]

Purpose

Reduce cpu_offload_gib from 4 GiB to 1 GiB for the two small KV-offload GSM8K evaluations. These jobs share a constrained /dev/shm on H200 CI workers and were failing before evaluation began, despite needing substantially less than 1 GiB during execution.

Root Cause

#50094 moved CPUOffloadingSpec into /dev/shm.
#50358 correctly added capacity validation.
Some shared H200 CI workers have only 1.4–2.1 GiB available, but these small evals requested 4 GiB.

This PR fix assigns 1 GiB only to the Nemotron-H and Gemma 4 OffloadingConnector cases. Larger and simple-connector cases remain unchanged.

Test Plan and Results (on a single H200)

sudo unshare --mount --propagation private bash -c '
mount -t tmpfs -o size=2g,mode=1777 tmpfs /dev/shm'

PATH="$PWD/.venv/bin:$PATH" CUDA_VISIBLE_DEVICES=0 \
.venv/bin/python -m pytest \
tests/evals/gsm8k/test_gsm8k_offloading.py \
-k "nemotron-h-8b or gemma-4-e4b-it" -s -v

On main:

RuntimeError: Insufficient space in /dev/shm:
4095/4096 MiB required, 1405 MiB free

On the PR branch:

4 passed, 5 deselected, 14 warnings in 663.99s

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[taneem-ibrahim]

/ci run

[github-actions]

❌ @taneem-ibrahim, A reviewer with write access must run /ci run, approve the PR, or add the ready label first.

[yewentao256]

LGTM, thanks for the work!

[github-actions]

✅ @taneem-ibrahim, CI is now available for this PR.

• /ci run starts a CI build.
• /ci retry retries failed jobs in the CI build for the current PR head. If the current head has no CI build, it starts a new CI build for the current head containing only jobs that failed in the latest earlier CI build for this PR.
• /ci cancel cancels scheduled or running CI builds for this PR branch.

[yewentao256]

/ci run

[github-actions]

✅ Triggered Buildkite CI #84107 for commit 1ce18d77f8f4.