VLLM_USE_TRITON_FLASH_ATTN V0 variable deprecation (#27611)

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 18:34:36 -08:00
committed by GitHub
parent 7f829be7d3
commit 9f0247cfa4
15 changed files with 12 additions and 1588 deletions
@@ -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)
@@ -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(
@@ -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",
@@ -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)
@@ -5,7 +5,6 @@ image, embedding, and video support for different VLMs in vLLM.
"""
import math
import os
from collections import defaultdict
from pathlib import PosixPath
@@ -38,13 +37,6 @@ from .vlm_utils.types import (
VLMTestType,
)
# This hack is needed for phi3v & paligemma models
# ROCm Triton FA can run into shared memory issues with these models,
# use other backends in the meantime
# FIXME (mattwong, gshtrasb, hongxiayan)
if current_platform.is_rocm():
os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
COMMON_BROADCAST_SETTINGS = {
"test_type": VLMTestType.IMAGE,
"dtype": "half",
@@ -11,7 +11,6 @@ from huggingface_hub import snapshot_download
from vllm.assets.image import ImageAsset
from vllm.lora.request import LoRARequest
from vllm.multimodal.image import rescale_image_size
from vllm.platforms import current_platform
from ....conftest import (
IMAGE_ASSETS,
@@ -46,12 +45,6 @@ models = [model_path]
target_dtype = "half"
# ROCm Triton FA can run into shared memory issues with these models,
# use other backends in the meantime
# FIXME (mattwong, gshtrasb, hongxiayan)
if current_platform.is_rocm():
os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
def run_test(
hf_runner: type[HfRunner],
@@ -14,7 +14,6 @@ from vllm.assets.image import ImageAsset
from vllm.logprobs import SampleLogprobs
from vllm.lora.request import LoRARequest
from vllm.multimodal.image import convert_image_mode, rescale_image_size
from vllm.platforms import current_platform
from ....conftest import (
IMAGE_ASSETS,
@@ -68,12 +67,6 @@ def vllm_to_hf_output(
target_dtype = "half"
# ROCm Triton FA can run into shared memory issues with these models,
# use other backends in the meantime
# FIXME (mattwong, gshtrasb, hongxiayan)
if current_platform.is_rocm():
os.environ["VLLM_USE_TRITON_FLASH_ATTN"] = "0"
def run_test(
hf_runner: type[HfRunner],