[ROCm][CI][Bugfix] Multi-Modal Model Support Fixes and Attention Backend Improvements (#30270)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Andreas Karatzas
2025-12-19 02:17:27 +00:00
committed by GitHub
parent 6a09612b2e
commit 7b43db210c
7 changed files with 109 additions and 47 deletions
@@ -1,6 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Pytest configuration for vLLM tests."""
"""Pytest configuration for vLLM multimodal tests."""
import warnings
@@ -9,16 +9,13 @@ import torch
from vllm.platforms import current_platform
def pytest_configure(config):
"""Disable Flash/MemEfficient SDP on ROCm to avoid HF
Transformers accuracy issues.
"""
def pytest_collection_modifyitems(config, items):
"""Configure ROCm-specific settings based on collected tests."""
if not current_platform.is_rocm():
return
skip_patterns = ["test_granite_speech.py"]
if any(pattern in str(arg) for arg in config.args for pattern in skip_patterns):
# Skip disabling SDP for Granite Speech tests on ROCm
return
# Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
@@ -173,6 +173,13 @@ VLM_TEST_SETTINGS = {
auto_cls=AutoModelForImageTextToText,
vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
patch_hf_runner=model_utils.qwen3_vl_patch_hf_runner,
vllm_runner_kwargs={
"attention_config": {
"backend": "ROCM_AITER_FA",
},
}
if current_platform.is_rocm()
else None,
image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
marks=[
pytest.mark.core_model,
@@ -253,8 +260,19 @@ VLM_TEST_SETTINGS = {
image_size_factors=[(0.25, 0.2, 0.15)],
vllm_runner_kwargs={
"model_impl": "transformers",
# TODO: [ROCm] Revert this once issue #30167 is resolved
**(
{
"mm_processor_kwargs": {
"min_pixels": 256 * 28 * 28,
"max_pixels": 1280 * 28 * 28,
},
}
if current_platform.is_rocm()
else {}
),
},
marks=[large_gpu_mark(min_gb=32)],
marks=[large_gpu_mark(min_gb=80 if current_platform.is_rocm() else 32)],
),
#### Extended model tests
"aria": VLMTestInfo(
@@ -645,7 +663,17 @@ VLM_TEST_SETTINGS = {
hf_output_post_proc=model_utils.minimax_vl_01_hf_output,
patch_hf_runner=model_utils.minimax_vl_01_patch_hf_runner,
auto_cls=AutoModelForImageTextToText,
marks=[large_gpu_mark(min_gb=80)],
marks=[
large_gpu_mark(min_gb=80),
# TODO: [ROCm] Fix pickle issue with ROCm spawn and tp>1
pytest.mark.skipif(
current_platform.is_rocm(),
reason=(
"ROCm: Model too large for single GPU; "
"multi-GPU blocked by HF _LazyConfigMapping pickle issue with spawn"
),
),
],
),
"molmo": VLMTestInfo(
models=["allenai/Molmo-7B-D-0924"],
@@ -39,7 +39,7 @@ models = [MODEL_NAME]
def granite_speech_attention_config():
"""Return attention config for Granite Speech tests on ROCm."""
if current_platform.is_rocm():
return {"backend": "TRITON_ATTN"}
return {"backend": "ROCM_AITER_FA"}
return None
@@ -1,18 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Pytest configuration for vLLM pooling tests."""
import pytest
from vllm.platforms import current_platform
@pytest.fixture
def siglip_attention_config():
"""Return attention config for SigLIP tests on ROCm.
On ROCm, SigLIP tests require FLEX_ATTENTION backend.
"""
if current_platform.is_rocm():
return {"backend": "FLEX_ATTENTION"}
return None