biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Fix cpu offload testing for gptq/awq/ct (#15648)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-03-28 10:29:32 -06:00
committed by GitHub
parent 432cf22a6a
commit 47e9038d23
2 changed files with 42 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
tests/quantization/test_cpu_offload.py
View File

@@ -33,7 +33,9 @@ def test_cpu_offload_fp8():
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
reason="gptq_marlin is not supported on this GPU type.")
def test_cpu_offload_gptq():
def test_cpu_offload_gptq(monkeypatch):
# This quant method is sensitive to dummy weights, so we force real weights
monkeypatch.setenv('VLLM_TEST_FORCE_LOAD_FORMAT', 'auto')
# Test GPTQ Marlin
compare_two_settings("Qwen/Qwen2-1.5B-Instruct-GPTQ-Int4", [],
["--cpu-offload-gb", "1"],
@@ -47,7 +49,9 @@ def test_cpu_offload_gptq():
@pytest.mark.skipif(not is_quant_method_supported("awq_marlin"),
reason="awq_marlin is not supported on this GPU type.")
def test_cpu_offload_awq():
def test_cpu_offload_awq(monkeypatch):
# This quant method is sensitive to dummy weights, so we force real weights
monkeypatch.setenv('VLLM_TEST_FORCE_LOAD_FORMAT', 'auto')
# Test AWQ Marlin
compare_two_settings("Qwen/Qwen2-1.5B-Instruct-AWQ", [],
["--cpu-offload-gb", "1"],
@@ -61,7 +65,9 @@ def test_cpu_offload_awq():
@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin"),
reason="gptq_marlin is not supported on this GPU type.")
def test_cpu_offload_compressed_tensors():
def test_cpu_offload_compressed_tensors(monkeypatch):
# This quant method is sensitive to dummy weights, so we force real weights
monkeypatch.setenv('VLLM_TEST_FORCE_LOAD_FORMAT', 'auto')
# Test wNa16
compare_two_settings("nm-testing/tinyllama-oneshot-w4a16-channel-v2", [],
["--cpu-offload-gb", "1"],