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

[BugFix][V1][ROCm] Triton MLA uses V0 backend on V1 engine (#19067)

Browse Source 
Signed-off-by: Tianyuan Wu <Tianyuan.Wu@amd.com>
This commit is contained in:
TY-AMD
2025-07-01 16:12:19 +08:00
committed by GitHub
parent b1c1fe35a5
commit 96453cfa83
5 changed files with 78 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
tests/kernels/attention/test_rocm_attention_selector.py
View File

@@ -35,7 +35,8 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
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"
assert (backend.get_name() == "TRITON_MLA"
or backend.get_name() == "TRITON_MLA_VLLM_V1")
# If attention backend is None
# If use_mla is true
@@ -43,7 +44,8 @@ def test_selector(monkeypatch: pytest.MonkeyPatch):
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"
assert (backend.get_name() == "TRITON_MLA"
or backend.get_name() == "TRITON_MLA_VLLM_V1")
# change the attention backend to AITER MLA
m.setenv(STR_BACKEND_ENV_VAR, "ROCM_AITER_MLA")