Skip to content

fix(attention): bound multi-row triton decode tiles by the LDS size on ROCm #906

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

Draft
jason-fxz wants to merge 1 commit into
base: main
Choose a base branch
from
+8 −1
Draft

fix(attention): bound multi-row triton decode tiles by the LDS size on ROCm #906

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
9 changes: 8 additions & 1 deletion python/freetoken/kernel/triton/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,6 +92,11 @@ def _optin_smem_bytes(device_index: int) -> int:
return int(getattr(props, "shared_memory_per_block_optin", 0))


@functools.lru_cache(maxsize=None)
def _rocm_lds_bytes(device_index: int) -> int:
return int(torch.cuda.get_device_properties(device_index).shared_memory_per_block)


@functools.lru_cache(maxsize=None)
def _num_sms(device_index: int) -> int:
return int(torch.cuda.get_device_properties(device_index).multi_processor_count)
Expand Down Expand Up @@ -612,8 +617,10 @@ def decode_paged_attention(
if kv_format.quantized:
block_n, num_stages, num_warps = _quantized_decode_launch(block_d, block_n, num_stages)
if rows > 1:
# NOTE: torch reads 0 opt-in shared memory on ROCm, which the launch rule takes as no limit
smem_budget = _rocm_lds_bytes(q.device.index) if is_rocm() else smem
block_n, num_stages, num_warps = _multi_row_decode_launch(
rows * block_h, block_d, kv_format.quantized, elem, k_cache.element_size(), smem
rows * block_h, block_d, kv_format.quantized, elem, k_cache.element_size(), smem_budget
)

k_scale_arg, v_scale_arg, stride_kss, stride_vss = kv_scale_args(kv_format, k_cache, v_cache, k_scale, v_scale, q)
Expand Down
Loading