Skip to content
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

Tune paged attention parameters for AMD GPU. #3255

Merged
merged 5 commits into from
Feb 2, 2025
+13 −2
Merged

Tune paged attention parameters for AMD GPU. #3255

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
c145acb
Tune paged attention parameters for AMD GPU.
whchung Feb 1, 2025
9c5980e
Additional tuning for grouped page attention kernel.
whchung Feb 1, 2025
eb34954
Additional tuning for grouped paged attention kernel
whchung Feb 1, 2025
5c095e5
Segregate AMD-specific tuning with is_hip checks
whchung Feb 1, 2025
bf36b5b
Merge branch 'main' into whchung/exp-amd-pa
whchung Feb 1, 2025
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
11 changes: 9 additions & 2 deletions python/sglang/srt/layers/attention/triton_ops/decode_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,6 +181,9 @@ def _decode_att_m_fwd(
logit_cap,
):
BLOCK = 64
# [TODO] work around SGPR limit on MI3xx
if is_hip_:
BLOCK = 8
NUM_KV_SPLITS = num_kv_splits
Lk = k_buffer.shape[-1]
Lv = v_buffer.shape[-1]
Expand All @@ -194,6 +197,8 @@ def _decode_att_m_fwd(
num_warps = 4
else:
num_warps = 2
if is_hip_:
num_warps = 1

BLOCK_DMODEL = triton.next_power_of_2(Lk)
BLOCK_DV = triton.next_power_of_2(Lv)
Expand Down Expand Up @@ -433,10 +438,12 @@ def _decode_grouped_att_m_fwd(
)

extra_kargs = {}
num_stages = 2
if is_hip_:
# https://rocm.docs.amd.com/en/docs-6.2.0/how-to/llm-fine-tuning-optimization/optimizing-triton-kernel.html
# https://github.com/triton-lang/triton/blob/main/third_party/amd/backend/compiler.py
extra_kargs = {"waves_per_eu": 4, "matrix_instr_nonkdim": 16, "kpack": 2}
extra_kargs = {"waves_per_eu": 1, "matrix_instr_nonkdim": 16, "kpack": 2}
num_stages = 1

_fwd_grouped_kernel_stage1[grid](
q,
Expand Down Expand Up @@ -467,7 +474,7 @@ def _decode_grouped_att_m_fwd(
NUM_KV_SPLITS=NUM_KV_SPLITS,
logit_cap=logit_cap,
num_warps=4,
num_stages=2,
num_stages=num_stages,
Lk=Lk,
Lv=Lv,
**extra_kargs,
Expand Down
4 changes: 4 additions & 0 deletions python/sglang/srt/server_args.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -273,6 +273,10 @@ def __post_init__(self):
) and check_gguf_file(self.model_path):
self.quantization = self.load_format = "gguf"

# AMD-specific Triton attention KV splits default number
if is_hip():
self.triton_attention_num_kv_splits = 16

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@whchung we noticed that default 8 works better to most cases? Is 16 slightly better as you see, or better toward long sequences as you interested?

Copy link
Collaborator

@HaiShaw HaiShaw Feb 2, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@this PR boosts grok, etc. as observed.

@staticmethod
def add_cli_args(parser: argparse.ArgumentParser):
# Model and port args
Expand Down
Loading