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

VLLM_USE_TRITON_FLASH_ATTN V0 variable deprecation (#27611)

Browse Source 
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
This commit is contained in:
Andreas Karatzas
2025-11-11 20:34:36 -06:00
committed by GitHub
parent 7f829be7d3
commit 9f0247cfa4
15 changed files with 12 additions and 1588 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/models/language/pooling/test_classification.py
View File

@@ -27,13 +27,7 @@ def test_models(
example_prompts,
model: str,
dtype: str,
monkeypatch,
) -> None:
if current_platform.is_rocm():
# ROCm Triton FA does not currently support sliding window attention
# switch to use ROCm CK FA backend
monkeypatch.setenv("VLLM_USE_TRITON_FLASH_ATTN", "False")
with vllm_runner(model, max_model_len=512, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.classify(example_prompts)

7
tests/models/language/pooling/test_embedding.py
View File

@@ -4,7 +4,6 @@
import pytest
from vllm.config import PoolerConfig
from vllm.platforms import current_platform
from ...utils import check_embeddings_close
@@ -51,13 +50,7 @@ def test_models(
vllm_runner,
example_prompts,
model,
monkeypatch,
) -> None:
if model == "BAAI/bge-multilingual-gemma2" and current_platform.is_rocm():
# ROCm Triton FA does not currently support sliding window attention
# switch to use ROCm CK FA backend
monkeypatch.setenv("VLLM_USE_TRITON_FLASH_ATTN", "False")
vllm_extra_kwargs = {}
if model == "ssmits/Qwen2-7B-Instruct-embed-base":
vllm_extra_kwargs["pooler_config"] = PoolerConfig(

13
tests/models/language/pooling/test_mm_classifier_conversion.py
View File

@@ -2,18 +2,11 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.config.pooler import PoolerConfig
from vllm.platforms import current_platform
def test_idefics_multimodal(
vllm_runner,
monkeypatch,
) -> None:
if current_platform.is_rocm():
# ROCm Triton FA does not currently support sliding window attention
# switch to use ROCm CK FA backend
monkeypatch.setenv("VLLM_USE_TRITON_FLASH_ATTN", "False")
prompts = [
"Hello, my name is",
"The president of the United States is",
@@ -59,13 +52,7 @@ def update_config(config):
def test_gemma_multimodal(
vllm_runner,
monkeypatch,
) -> None:
if current_platform.is_rocm():
# ROCm Triton FA does not currently support sliding window attention
# switch to use ROCm CK FA backend
monkeypatch.setenv("VLLM_USE_TRITON_FLASH_ATTN", "False")
messages = [
{
"role": "system",

6
tests/models/language/pooling/test_reward.py
View File

@@ -76,7 +76,6 @@ def test_prm_models(
math_step_prompts,
model: str,
dtype: str,
monkeypatch,
) -> None:
check_transformers_version(
"Qwen/Qwen2.5-Math-PRM-7B", max_transformers_version="4.53.2"
@@ -85,11 +84,6 @@ def test_prm_models(
if current_platform.is_cpu():
pytest.skip("CPU only supports V1")
if current_platform.is_rocm():
# ROCm Triton FA does not currently support sliding window attention
# switch to use ROCm CK FA backend
monkeypatch.setenv("VLLM_USE_TRITON_FLASH_ATTN", "False")
with vllm_runner(model, max_model_len=1024, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.reward(math_step_prompts)