[AMD][ROCm] Enable DeepSeek model on ROCm (#12662)

Signed-off-by: Hongxia Yang <hongxia.yang@amd.com>
Co-authored-by: Matthew Wong <Matthew.Wong2@amd.com>
This commit is contained in:
Hongxia Yang
2025-02-04 08:24:11 +00:00
committed by GitHub
co-authored by Matthew Wong
parent 4896d0c2dd
commit c36ac98d01
5 changed files with 58 additions and 1 deletions
@@ -0,0 +1,31 @@
# SPDX-License-Identifier: Apache-2.0
from unittest.mock import patch
import pytest
import torch
from tests.kernels.utils import override_backend_env_variable
from vllm.attention.selector import _cached_get_attn_backend, get_attn_backend
from vllm.platforms.rocm import RocmPlatform
@pytest.fixture(autouse=True)
def clear_cache():
"""Clear lru cache to ensure each test case runs without caching.
"""
_cached_get_attn_backend.cache_clear()
def test_selector(monkeypatch):
"""Test that the attention selector for ROCm.
"""
override_backend_env_variable(monkeypatch, "ROCM_FLASH")
with patch("vllm.attention.selector.current_platform", RocmPlatform()):
backend = get_attn_backend(16, torch.float16, torch.float16, 16, False)
assert backend.get_name() == "ROCM_FLASH"
# mla test for deepseek related
backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False,
False, True)
assert backend.get_name() == "TRITON_MLA"
+9
View File
@@ -24,6 +24,15 @@ def _create_model_runner(model: str, *args, **kwargs) -> ModelRunner:
return model_runner
def test_deepseek_mla_attn_backend_module():
model_runner = _create_model_runner(
"deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct",
trust_remote_code=True,
enable_chunked_prefill=False,
)
assert model_runner.attn_backend.__name__ == "TritonMLABackend"
@pytest.mark.parametrize("batch_size", list(range(1, 257)))
def test_prepare_prompt(batch_size):
model_runner = _create_model_runner(