Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Run the Q-GaLore tests from the normal pytest job, not a workflow of their own (#10898)

Daniel Han7ef4ab437ee0
diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index 44043db6864..98a538e0326 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -7,7 +7,7 @@
#
# Why a separate workflow:
# - studio-backend-ci.yml's "Repo tests (CPU)" job already auto-discovers
-# tests/ minus tests/qlora, tests/saving, tests/utils, tests/sh. The 21
+# tests/ minus tests/qlora, tests/saving, tests/utils, tests/sh. The 24
# Bucket-A tests below live inside those --ignore dirs (CPU-runnable but
# historically excluded with their GPU siblings); pulling them out into
# a sibling job keeps the existing 760-passed baseline stable while we
@@ -305,6 +305,7 @@ jobs:
tests/utils/test_attention_masks.py \
tests/utils/test_packing.py \
tests/utils/test_trunc_normal_patch.py \
+ tests/utils/test_q_galore.py \
tests/python/test_fast_language_model_text_only.py
# Collect from the zoo checkout, not from here: this repo ships a real
# tests/ package, so a zoo test doing "from tests.foo import bar" binds
@@ -441,6 +442,7 @@ jobs:
tests/utils/test_attention_masks.py \
tests/utils/test_packing.py \
tests/utils/test_trunc_normal_patch.py \
+ tests/utils/test_q_galore.py \
tests/python/test_fast_language_model_text_only.py \
tests/test_bad_mappings_redirect.py \
tests/test_ollama_eos_token_order.py \
@@ -2332,7 +2334,7 @@ jobs:
pip show unsloth_zoo
echo "::endgroup::"
echo "Consolidated job done. Coverage:"
- echo " - 21 unsloth Bucket-A tests under tests/saving/ + tests/utils/"
+ echo " - 24 unsloth Bucket-A tests under tests/saving/ + tests/utils/"
echo " - 168 dataset_num_proc worker-bound guards (#2693)"
echo " - unsloth_zoo @ ${UNSLOTH_ZOO_REF} pytest tests/ (5 GPU cases deselected)"
echo " - unsloth_zoo.compiler.test_apply_fused_lm_head"
diff --git a/.github/workflows/q-galore-tests.yml b/.github/workflows/q-galore-tests.yml
deleted file mode 100644
index 567b8051180..00000000000
--- a/.github/workflows/q-galore-tests.yml
+++ /dev/null
@@ -1,52 +0,0 @@
-name: Q-GaLore CPU tests
-
-on:
- push:
- branches: [main]
- paths:
- - "unsloth/optimizers/**"
- - "tests/utils/test_q_galore.py"
- - ".github/workflows/q-galore-tests.yml"
- pull_request:
- paths:
- - "unsloth/optimizers/**"
- - "tests/utils/test_q_galore.py"
- - ".github/workflows/q-galore-tests.yml"
- workflow_dispatch:
-
-concurrency:
- group: ${{ github.workflow }}-${{ github.ref }}-${{ github.event_name }}-${{ github.ref == 'refs/heads/main' && github.sha || '' }}
- cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
-
-permissions:
- contents: read
-
-jobs:
- q-galore:
- runs-on: ubuntu-latest
- timeout-minutes: 10
- strategy:
- fail-fast: false
- matrix:
- # bitsandbytes is not API-stable across the range pyproject allows, so pinning one
- # version tests one point of it. 0.45.5 is the floor, 0.49.2 the last with
- # percentile_clipping / block_wise, 0.50.2 the first without them and with CPU
- # optimizer kernels. This PR's tests are quantizer-only and touch no bitsandbytes
- # entry point; the constructor coverage that uses these pins arrives with #10874.
- bitsandbytes: ["0.45.5", "0.49.2", "0.50.2"]
- name: q-galore (bitsandbytes ${{ matrix.bitsandbytes }})
- steps:
- - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
- with:
- persist-credentials: false
- - uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
- with:
- python-version: "3.12"
- - name: Install CPU optimizer test dependencies
- run: |
- python -m pip install --index-url https://download.pytorch.org/whl/cpu 'torch==2.10.0'
- python -m pip install 'pytest==9.0.3' 'bitsandbytes==${{ matrix.bitsandbytes }}'
- - name: Run Q-GaLore regressions
- # No --confcutdir: this file loads the optimizers through importlib specs, so the
- # GPU-spoof conftest neither helps nor hinders it. Verified both ways.
- run: python -m pytest tests/utils/test_q_galore.py -q