Skip to content

[ROCm] Env variable to trigger custom PA #15557

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Mar 27, 2025
Merged

[ROCm] Env variable to trigger custom PA #15557

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion vllm/attention/backends/rocm_flash_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -908,4 +908,5 @@ def _use_rocm_custom_paged_attention(qtype: torch.dtype, head_size: int,
and (qtype == torch.half or qtype == torch.bfloat16)
and (head_size == 64 or head_size == 128)
and (block_size == 16 or block_size == 32)
and (gqa_ratio >= 1 and gqa_ratio <= 16) and max_seq_len <= 32768)
and (gqa_ratio >= 1 and gqa_ratio <= 16) and max_seq_len <= 32768
and envs.VLLM_ROCM_CUSTOM_PAGED_ATTN)
6 changes: 6 additions & 0 deletions vllm/envs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,6 +78,7 @@
VLLM_ROCM_USE_AITER_RMSNORM: bool = True
VLLM_ROCM_FP8_PADDING: bool = True
VLLM_ROCM_MOE_PADDING: bool = True
VLLM_ROCM_CUSTOM_PAGED_ATTN: bool = True
VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
VLLM_DISABLE_COMPILE_CACHE: bool = False
Expand Down Expand Up @@ -541,6 +542,11 @@ def maybe_convert_int(value: Optional[str]) -> Optional[int]:
"VLLM_ROCM_MOE_PADDING":
lambda: bool(int(os.getenv("VLLM_ROCM_MOE_PADDING", "1"))),

# custom paged attention kernel for MI3* cards
"VLLM_ROCM_CUSTOM_PAGED_ATTN":
lambda: (os.getenv("VLLM_ROCM_CUSTOM_PAGED_ATTN", "True").lower() in
("true", "1")),

# Divisor for dynamic query scale factor calculation for FP8 KV Cache
"Q_SCALE_CONSTANT":
lambda: int(os.getenv("Q_SCALE_CONSTANT", "200")),
Expand Down