ISA

Author: AlexanderDokuchaev

.github/workflows/conformance_weight_compression.yml

@@ -16,7 +16,7 @@ on:
 
 jobs:
   examples-cpu:
-    name: Test exmaples CPU [${{ matrix.group }}/4]
+    name: Weight compression [${{ matrix.group }}/2]
     runs-on: ubuntu-22.04-16-cores
     strategy:
       fail-fast: false
@@ -58,6 +58,7 @@ jobs:
             ${{ github.event.inputs.pytest_args || '' }}
         env:
           TQDM_DISABLE: 1
+          HOME_HF: "/home/runner/hf_home"
       - name: Upload artifact
         uses: actions/upload-artifact@v4
         if: ${{ !cancelled() }}

tests/post_training/.test_durations

@@ -1,2 +1,13 @@
 {
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_awq_backup_mode_none_backend_OV]": 269.111,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_data_aware_awq_scale_estimation_backend_OV]": 421.082,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_data_aware_awq_scale_estimation_stateful_backend_OV]": 374.507,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_data_aware_awq_stateful_backend_OV]": 243.878,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_data_aware_backend_OV]": 190.344,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_data_aware_gptq_backend_OV]": 1261.154,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_data_aware_lora_stateful_backend_OV]": 483.905,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_data_free_backend_OV]": 196.905,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_int8_data_free_backend_TORCH]": 154.724,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_NF4_scale_estimation_stateful_per_channel_backend_OV]": 256.375,
+    "tests/post_training/test_quantize_conformance.py::test_weight_compression[tinyllama_scale_estimation_per_channel_backend_OV]": 258.622,
 }

tests/post_training/test_quantize_conformance.py

@@ -9,6 +9,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import os
 import re
 import time
 import traceback
@@ -30,6 +31,8 @@
 from tests.post_training.pipelines.base import BaseTestPipeline
 from tests.post_training.pipelines.base import RunInfo
 
+os.environ["ONEDNN_MAX_CPU_ISA"] = "AVX2"
+
 DATA_ROOT = Path(__file__).parent / "data"