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

[FEAT][ROCm]: Support AITER MLA (#15893)

Browse Source 
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: qli88 <qiang.li2@amd.com>
This commit is contained in:
vllmellm
2025-04-23 00:31:13 +08:00
committed by GitHub
parent f34410715f
commit 30bc3e0f66
9 changed files with 668 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
tests/kernels/test_rocm_attention_selector.py
View File

@@ -28,7 +28,34 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
assert (backend.get_name() == "ROCM_FLASH"
or backend.get_name() == "TRITON_ATTN_VLLM_V1")
# mla test for deepseek related
# MLA test for deepseek related
# change the attention backend to triton MLA
m.setenv(STR_BACKEND_ENV_VAR, "TRITON_MLA")
backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False,
False, True)
assert backend.get_name() == "TRITON_MLA"
# If attention backend is None
# If use_mla is true
# The selected backend is triton MLA
m.setenv(STR_BACKEND_ENV_VAR, None)
backend = get_attn_backend(576, torch.bfloat16, "auto", 16, False,
False, True)
assert backend.get_name() == "TRITON_MLA"
# change the attention backend to AITER MLA
m.setenv(STR_BACKEND_ENV_VAR, "ROCM_AITER_MLA")
backend = get_attn_backend(576, torch.bfloat16, "auto", 1, False,
False, True)
assert backend.get_name() == "ROCM_AITER_MLA"
# If attention backend is None
# If use_mla is true
# If VLLM_ROCM_USE_AITER is enabled
# The selected backend is ROCM_AITER_MLA
m.setenv(STR_BACKEND_ENV_VAR, None)
m.setenv("VLLM_ROCM_USE_AITER", "1")
backend = get_attn_backend(576, torch.bfloat16, "auto", 1, False,
False, True)
assert backend.get_name() == "ROCM_AITER_MLA"