Skip to content

add pagged attention nhd for aiter_backend #269

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
apinge wants to merge 3 commits into
base: Qwen3.5_v0.5.9
Choose a base branch
from
Draft

add pagged attention nhd for aiter_backend #269

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5e0fcf1
add pagged attention nhd for aiter_backend
apinge Apr 28, 2026
3a38ce9
revert unused modification
apinge Apr 28, 2026
f742fe9
remove comment
apinge Apr 28, 2026
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
59 changes: 46 additions & 13 deletions python/sglang/srt/layers/attention/aiter_backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,7 @@
mla_prefill_ps_asm_fwd,
mla_reduce_v1,
paged_attention_ragged,
paged_attention_ragged_nhd,
)
from aiter.mla import mla_decode_fwd, mla_prefill_fwd
except ImportError:
Expand Down Expand Up @@ -96,8 +97,7 @@ class ForwardMetadata:

global_workspace_buffer = None

_AITER_PARTITION_SIZE_ROCM = 256

_AITER_PARTITION_SIZE_ROCM = 256 # 256 512 1024 for paged_attention_ragged_nhd

class AiterAttnBackend(AttentionBackend):
def __init__(
Expand Down Expand Up @@ -187,17 +187,28 @@ def __init__(
self.max_num_partitions = (
self.max_context_len + _AITER_PARTITION_SIZE_ROCM - 1
) // _AITER_PARTITION_SIZE_ROCM

nbyes_per_qo_elem = torch.finfo(torch.float32).bits // 8

if not self.use_mla:
self.workspace_buffer = torch.empty(
(max_bs * self.num_head * self.max_num_partitions * self.head_dim)
* nbyes_per_qo_elem
+ 2 * (max_bs * self.num_head * self.max_num_partitions) * 4,
dtype=torch.uint8,
device=self.device,
)
# self.workspace_buffer = torch.empty(
# (max_bs * self.num_head * self.max_num_partitions * self.head_dim)
# * nbyes_per_qo_elem
# + 2 * (max_bs * self.num_head * self.max_num_partitions) * 4,
# dtype=torch.uint8,
# device=self.device,
# ).contiguous()
# aiter pa_ragged_nhd: three tensors (no single uint8 blob + host pointer math).
_nhp = max_bs * self.num_head * self.max_num_partitions
self.pa_nhd_exp_sums = torch.empty(
(_nhp,), dtype=torch.float32, device=self.device
).contiguous()
self.pa_nhd_max_logits = torch.empty(
(_nhp,), dtype=torch.float32, device=self.device
).contiguous()
self.pa_nhd_tmp_out = torch.empty(
(_nhp * self.head_dim,), dtype=torch.bfloat16, device=self.device
).contiguous()


self.scale = float(1.0 / (self.head_dim**0.5))
self.k_scale = self.v_scale = torch.tensor([1.0], dtype=torch.float32).to(
Expand Down Expand Up @@ -1770,10 +1781,11 @@ def forward_decode(

k_cache = k_cache.to(dtype)
v_cache = v_cache.to(dtype)

paged_attention_ragged(
paged_attention_ragged_nhd(
o.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
self.workspace_buffer,
self.pa_nhd_exp_sums,
self.pa_nhd_max_logits,
self.pa_nhd_tmp_out,
q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
k_cache.view(-1, 1, layer.tp_k_head_num, layer.qk_head_dim),
v_cache.view(-1, 1, layer.tp_v_head_num, layer.v_head_dim),
Expand All @@ -1792,6 +1804,27 @@ def forward_decode(
None,
_AITER_PARTITION_SIZE_ROCM,
)
# paged_attention_ragged(
# o.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
# self.workspace_buffer,
# q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
# k_cache.view(-1, 1, layer.tp_k_head_num, layer.qk_head_dim),
# v_cache.view(-1, 1, layer.tp_v_head_num, layer.v_head_dim),
# self.scale,
# self.forward_metadata.kv_indptr,
# self.forward_metadata.kv_indices,
# self.kv_last_page_len,
# 1,
# self.max_num_partitions,
# None,
# "auto",
# "NHD",
# self.logits_soft_cap,
# self.k_scale,
# self.v_scale,
# None,
# _AITER_PARTITION_SIZE_ROCM,
# )

return o

Expand Down
Loading