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

[FEAT][ROCm] Integrate Fused MoE Kernels from AITER (#14967)

Browse Source 
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
vllmellm
2025-03-26 16:30:30 +08:00
committed by GitHub
parent 781d056280
commit 5ebf66748b
9 changed files with 391 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
tests/quantization/test_fp8.py
View File

@@ -23,8 +23,14 @@ MODELS = [
reason="FP8 is not supported on this GPU type.")
@pytest.mark.parametrize("model_id", MODELS)
@pytest.mark.parametrize("force_marlin", [False, True])
@pytest.mark.parametrize(
"use_rocm_aiter", [True, False] if current_platform.is_rocm() else [False])
def test_model_load_and_run(vllm_runner, model_id: str, force_marlin: bool,
monkeypatch) -> None:
use_rocm_aiter: bool, monkeypatch) -> None:
if use_rocm_aiter:
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
if force_marlin:
monkeypatch.setenv("VLLM_TEST_FORCE_FP8_MARLIN", "1")
@@ -47,7 +53,13 @@ KV_CACHE_MODELS = [
@pytest.mark.skipif(not is_quant_method_supported("fp8"),
reason="FP8 is not supported on this GPU type.")
@pytest.mark.parametrize("model_id", KV_CACHE_MODELS)
def test_kv_cache_model_load_and_run(vllm_runner, model_id: str, monkeypatch):
@pytest.mark.parametrize(
"use_rocm_aiter", [True, False] if current_platform.is_rocm() else [False])
def test_kv_cache_model_load_and_run(vllm_runner, model_id: str,
use_rocm_aiter: bool, monkeypatch):
if use_rocm_aiter:
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
# vllm_runner.apply_model() relies on V0 internals.
monkeypatch.setenv("VLLM_USE_V1", "0")
with vllm_runner(model_id, kv_cache_dtype="fp8") as llm:
@@ -86,8 +98,13 @@ def test_kv_cache_model_load_and_run(vllm_runner, model_id: str, monkeypatch):
reason="FP8 is not supported on this GPU type.")
@pytest.mark.parametrize("kv_cache_dtype", ["auto", "fp8"])
@pytest.mark.parametrize("force_marlin", [False, True])
@pytest.mark.parametrize(
"use_rocm_aiter", [True, False] if current_platform.is_rocm() else [False])
def test_load_fp16_model(vllm_runner, kv_cache_dtype: str, force_marlin: bool,
monkeypatch) -> None:
use_rocm_aiter: bool, monkeypatch) -> None:
if use_rocm_aiter:
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
# vllm_runner.apply_model() relies on V0 internals.
monkeypatch.setenv("VLLM_USE_V1", "0")