[ROCm] Attention selector reordering (#36702)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com> Signed-off-by: Micah Williamson <micah.williamson@amd.com> Co-authored-by: Micah Williamson <micah.williamson@amd.com>
This commit is contained in:
committed by
GitHub
parent
09c3dc9186
commit
189ddefbfd
@@ -54,7 +54,7 @@ def mock_on_mi3xx():
|
||||
(
|
||||
{},
|
||||
None,
|
||||
AttentionBackendEnum.TRITON_ATTN.get_path(),
|
||||
AttentionBackendEnum.ROCM_ATTN.get_path(),
|
||||
),
|
||||
# Test Case 2: Explicit TRITON_ATTN backend
|
||||
(
|
||||
@@ -81,41 +81,24 @@ def mock_on_mi3xx():
|
||||
AttentionBackendEnum.ROCM_AITER_UNIFIED_ATTN.get_path(),
|
||||
),
|
||||
# Test Case 6: VLLM_ROCM_USE_AITER=1
|
||||
# (defaults to AITER FA when MHA not explicitly disabled)
|
||||
(
|
||||
{"VLLM_ROCM_USE_AITER": "1"},
|
||||
None,
|
||||
AttentionBackendEnum.ROCM_AITER_FA.get_path(),
|
||||
AttentionBackendEnum.ROCM_ATTN.get_path(),
|
||||
),
|
||||
# Test Case 7: VLLM_ROCM_USE_AITER=1 + VLLM_ROCM_USE_AITER_MHA=1
|
||||
(
|
||||
{"VLLM_ROCM_USE_AITER": "1", "VLLM_ROCM_USE_AITER_MHA": "1"},
|
||||
None,
|
||||
AttentionBackendEnum.ROCM_AITER_FA.get_path(),
|
||||
),
|
||||
# Test Case 8: VLLM_ROCM_USE_AITER=1 + VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1
|
||||
(
|
||||
{
|
||||
"VLLM_ROCM_USE_AITER": "1",
|
||||
"VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION": "1",
|
||||
},
|
||||
None,
|
||||
AttentionBackendEnum.ROCM_AITER_UNIFIED_ATTN.get_path(),
|
||||
),
|
||||
# Test Case 9: VLLM_ROCM_USE_AITER=1 + explicit TRITON_ATTN
|
||||
# Test Case 7: VLLM_ROCM_USE_AITER=1 + explicit TRITON_ATTN
|
||||
(
|
||||
{"VLLM_ROCM_USE_AITER": "1"},
|
||||
"TRITON_ATTN",
|
||||
AttentionBackendEnum.TRITON_ATTN.get_path(),
|
||||
),
|
||||
# Test Case 10: VLLM_ROCM_USE_AITER=1 + VLLM_ROCM_USE_AITER_MHA=0
|
||||
# (explicitly disabled)
|
||||
# Test Case 8: VLLM_ROCM_USE_AITER=1 + VLLM_ROCM_USE_AITER_MHA=0
|
||||
(
|
||||
{"VLLM_ROCM_USE_AITER": "1", "VLLM_ROCM_USE_AITER_MHA": "0"},
|
||||
None,
|
||||
AttentionBackendEnum.TRITON_ATTN.get_path(),
|
||||
AttentionBackendEnum.ROCM_ATTN.get_path(),
|
||||
),
|
||||
# Test Case 11: VLLM_ROCM_USE_AITER=1 + explicit ROCM_ATTN
|
||||
# Test Case 9: VLLM_ROCM_USE_AITER=1 + explicit ROCM_ATTN
|
||||
(
|
||||
{"VLLM_ROCM_USE_AITER": "1"},
|
||||
"ROCM_ATTN",
|
||||
|
||||
Reference in New Issue
Block a user