Sync from v0.13

2026-01-19 10:38:50 +08:00
parent b2ef04d792
commit 5aef6c175a
3714 changed files with 854317 additions and 89342 deletions
--- a/tests/quantization/test_cpu_offload.py
+++ b/tests/quantization/test_cpu_offload.py
@@ -0,0 +1,73 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+# Expanded quantized model tests for CPU offloading
+# Base tests: tests/basic_correctness/test_cpu_offload.py
+
+import pytest
+
+from tests.quantization.utils import is_quant_method_supported
+
+from ..utils import compare_two_settings
+
+
+@pytest.mark.skipif(
+    not is_quant_method_supported("fp8"),
+    reason="fp8 is not supported on this GPU type.",
+)
+def test_cpu_offload_fp8():
+    # Test loading a quantized checkpoint
+    compare_two_settings(
+        "neuralmagic/Qwen2-1.5B-Instruct-FP8",
+        ["--enforce_eager"],
+        ["--enforce_eager", "--cpu-offload-gb", "1"],
+        max_wait_seconds=480,
+    )
+
+
+@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(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",
+        ["--enforce_eager"],
+        ["--enforce_eager", "--cpu-offload-gb", "1"],
+        max_wait_seconds=480,
+    )
+
+
+@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(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",
+        ["--enforce_eager"],
+        ["--enforce_eager", "--cpu-offload-gb", "1"],
+        max_wait_seconds=480,
+    )
+
+
+@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(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",
+        ["--enforce_eager"],
+        ["--enforce_eager", "--cpu-offload-gb", "1"],
+        max_wait_seconds=480,
+    )