[Hardware][XPU] Register batch-invariant kernels for XPU (#41934)

Signed-off-by: tzielinski-habana <tomasz.zielinski@intel.com>
Signed-off-by: Tomasz Zielinski <85164140+tzielinski-habana@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Chendi.Xue <chendi.xue@intel.com>
This commit is contained in:
Tomasz Zielinski
2026-07-15 11:19:44 -04:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Chendi.Xue
parent 05eed72aec
commit 61141ed265
6 changed files with 127 additions and 55 deletions
+47 -21
View File
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import os
import random
from typing import NamedTuple
import pytest
import torch
@@ -11,36 +12,61 @@ from vllm.transformers_utils.config import get_config
from vllm.transformers_utils.model_arch_config_convertor import (
ModelArchConfigConvertorBase,
)
from vllm.triton_utils import HAS_TRITON
from vllm.v1.attention.backends.fa_utils import flash_attn_supports_mla
skip_unsupported = pytest.mark.skipif(
not (current_platform.is_cuda() and current_platform.has_device_capability(80)),
# Supports testing on Ampere and Ada Lovelace devices.
# Note: For devices with SM < 90, batch invariance does not support CUDA Graphs.
reason="Requires CUDA and >= Ampere (SM80)",
)
class DeviceConfig(NamedTuple):
available: bool
backends: list[str]
# Maps each device to its availability and supported backends.
DEVICE_BACKENDS: dict[str, DeviceConfig] = {
"cuda": DeviceConfig(
available=current_platform.is_cuda()
and current_platform.has_device_capability(80),
# FlashInfer backend temporarily disabled due to invariant CTA sizes.
# See FlashInfer issue #2424
backends=["FLASH_ATTN", "TRITON_ATTN", "FLEX_ATTENTION"],
),
"xpu": DeviceConfig(
available=current_platform.is_xpu() and HAS_TRITON,
backends=["TRITON_ATTN"],
),
}
DEFAULT_MODEL = "Qwen/Qwen3-1.7B"
TEST_MODEL = os.getenv("VLLM_TEST_MODEL", DEFAULT_MODEL)
BACKENDS: list[str] = [
"FLASH_ATTN",
"TRITON_ATTN",
"FLEX_ATTENTION",
]
# FlashInfer temporarily disabled due to invariant CTA sizes.
# See FlashInfer issue #2424
# if has_flashinfer():
# BACKENDS.append("FLASHINFER")
# only run MLA backends when the requested test model is itself an MLA model.
# Override backends for MLA models (MLA only supported on CUDA).
if os.getenv("VLLM_TEST_MODEL"):
config = get_config(TEST_MODEL, trust_remote_code=False)
if ModelArchConfigConvertorBase(config, config.get_text_config()).is_deepseek_mla():
BACKENDS = ["TRITON_MLA"]
if flash_attn_supports_mla():
BACKENDS.append("FLASH_ATTN_MLA")
DEVICE_BACKENDS["cuda"] = DeviceConfig(
available=DEVICE_BACKENDS["cuda"].available,
backends=["TRITON_MLA"]
+ (["FLASH_ATTN_MLA"] if flash_attn_supports_mla() else []),
)
DEVICE_BACKENDS["xpu"] = DeviceConfig(
available=DEVICE_BACKENDS["xpu"].available,
backends=[],
)
# Only include backends for devices that are actually available.
BACKENDS: list[str] = sorted(
{b for cfg in DEVICE_BACKENDS.values() if cfg.available for b in cfg.backends}
)
skip_unsupported = pytest.mark.skipif(
not any(cfg.available for cfg in DEVICE_BACKENDS.values()),
reason="Requires CUDA >= Ampere (SM80) or Intel XPU with Triton",
)
skip_if_not_cuda = pytest.mark.skipif(
not DEVICE_BACKENDS["cuda"].available,
reason="Requires CUDA >= Ampere (SM80)",
)
def _random_prompt(min_words: int = 1024, max_words: int = 1024 * 2) -> str: