Skip to content

[ROCm][Perf] Fold scalar KV scales in MiniMax-M3 sparse attention #52060

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

Open
akii96 wants to merge 1 commit into
base: main
Choose a base branch
from
+27 −18
Open

[ROCm][Perf] Fold scalar KV scales in MiniMax-M3 sparse attention #52060

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
15 changes: 9 additions & 6 deletions vllm/models/minimax_m3/amd/ops/sparse_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,6 +117,11 @@ def _gqa_sparse_fwd_kernel(
SUB_K: tl.constexpr, # CDNA only: KV sub-tile width (see _IS_MI3XX)
):
sm_scale_log2e = sm_scale * 1.4426950409
# Scalar KV scales are loop invariant. Fold them into the QK scale and
# normalized output instead of scaling every K/V element.
if KV_SCALE_MODE == 1:
sm_scale_log2e *= tl.load(k_scale_ptr)
value_scale = tl.load(v_scale_ptr)
pid_q = tl.program_id(0)
pid_kh = tl.program_id(1)
pid_b = tl.program_id(2)
Expand Down Expand Up @@ -176,9 +181,7 @@ def _gqa_sparse_fwd_kernel(
)
if USE_FP8:
k_sub = k_sub.to(q.dtype)
if KV_SCALE_MODE == 1:
k_sub = (k_sub * tl.load(k_scale_ptr)).to(q.dtype)
elif KV_SCALE_MODE == 2:
if KV_SCALE_MODE == 2:
k_scale = tl.load(
k_scale_ptr
+ pid_kh * stride_ks_h
Expand Down Expand Up @@ -212,9 +215,7 @@ def _gqa_sparse_fwd_kernel(
)
if USE_FP8:
v_sub = v_sub.to(q.dtype)
if KV_SCALE_MODE == 1:
v_sub = (v_sub * tl.load(v_scale_ptr)).to(q.dtype)
elif KV_SCALE_MODE == 2:
if KV_SCALE_MODE == 2:
v_scale = tl.load(
v_scale_ptr
+ pid_kh * stride_vs_h
Expand All @@ -227,6 +228,8 @@ def _gqa_sparse_fwd_kernel(
m_i = m_ij
lse_i = m_ij + tl.log2(tl.exp2(lse_i - m_ij) + l_ij)
acc_o = acc_o * tl.exp2(m_i - lse_i)[:, None]
if KV_SCALE_MODE == 1:
acc_o *= value_scale
acc_o = tl.reshape(acc_o, BLOCK_SIZE_Q, BLOCK_SIZE_H, BLOCK_SIZE_D)
o_ptrs = tl.make_block_ptr(
base=o_ptr + q_start * stride_on + pid_h * stride_oh,
Expand Down
30 changes: 18 additions & 12 deletions vllm/models/minimax_m3/common/ops/sparse_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,6 +94,11 @@ def _gqa_sparse_fwd_kernel(
KV_SCALE_MODE: tl.constexpr, # 0: none, 1: scalar, 2: [kv_head, token]
):
sm_scale_log2e = sm_scale * 1.4426950409
# Scalar KV scales are loop invariant. Fold them into the QK scale and
# normalized output instead of scaling every K/V element.
if KV_SCALE_MODE == 1:
sm_scale_log2e *= tl.load(k_scale_ptr)
value_scale = tl.load(v_scale_ptr)
pid_q = tl.program_id(0)
pid_kh = tl.program_id(1)
pid_b = tl.program_id(2)
Expand Down Expand Up @@ -155,9 +160,7 @@ def _gqa_sparse_fwd_kernel(
)
if USE_FP8:
k = k.to(q.dtype)
if KV_SCALE_MODE == 1:
k = (k * tl.load(k_scale_ptr)).to(q.dtype)
elif KV_SCALE_MODE == 2:
if KV_SCALE_MODE == 2:
k_scale = tl.load(
k_scale_ptr
+ pid_kh * stride_ks_h
Expand Down Expand Up @@ -187,9 +190,7 @@ def _gqa_sparse_fwd_kernel(
)
if USE_FP8:
v = v.to(q.dtype)
if KV_SCALE_MODE == 1:
v = (v * tl.load(v_scale_ptr)).to(q.dtype)
elif KV_SCALE_MODE == 2:
if KV_SCALE_MODE == 2:
v_scale = tl.load(
v_scale_ptr
+ pid_kh * stride_vs_h
Expand All @@ -202,6 +203,8 @@ def _gqa_sparse_fwd_kernel(
m_i = m_ij
lse_i = m_ij + tl.log2(tl.exp2(lse_i - m_ij) + l_ij)
acc_o = acc_o * tl.exp2(m_i - lse_i)[:, None]
if KV_SCALE_MODE == 1:
acc_o *= value_scale
acc_o = tl.reshape(acc_o, BLOCK_SIZE_Q, BLOCK_SIZE_H, BLOCK_SIZE_D)
o_ptrs = tl.make_block_ptr(
base=o_ptr + q_start * stride_on + pid_h * stride_oh,
Expand Down Expand Up @@ -278,6 +281,11 @@ def _gqa_sparse_decode_kernel(
USE_PDL: tl.constexpr,
):
sm_scale_log2e = sm_scale * 1.4426950409
# Scalar KV scales are loop invariant. Fold them into the QK scale and
# normalized output instead of scaling every K/V element.
if KV_SCALE_MODE == 1:
sm_scale_log2e *= tl.load(k_scale_ptr)
value_scale = tl.load(v_scale_ptr)
# split-K over the topk dimension: pid(0) folds (query-token, chunk).
pid_bc, pid_kh = tl.program_id(0), tl.program_id(1)
pid_b = pid_bc % total_q
Expand Down Expand Up @@ -341,9 +349,7 @@ def _gqa_sparse_decode_kernel(
)
if USE_FP8:
k = k.to(q.dtype)
if KV_SCALE_MODE == 1:
k = (k * tl.load(k_scale_ptr)).to(q.dtype)
elif KV_SCALE_MODE == 2:
if KV_SCALE_MODE == 2:
k_scale = tl.load(
k_scale_ptr
+ pid_kh * stride_ks_h
Expand All @@ -370,9 +376,7 @@ def _gqa_sparse_decode_kernel(
)
if USE_FP8:
v = v.to(q.dtype)
if KV_SCALE_MODE == 1:
v = (v * tl.load(v_scale_ptr)).to(q.dtype)
elif KV_SCALE_MODE == 2:
if KV_SCALE_MODE == 2:
v_scale = tl.load(
v_scale_ptr
+ pid_kh * stride_vs_h
Expand All @@ -392,6 +396,8 @@ def _gqa_sparse_decode_kernel(
# can hit 0 * NaN. All-empty padded rows may still produce NaNs in merge.
scale = tl.where(lse_i > float("-inf"), tl.exp2(m_i - lse_i), tl.zeros_like(lse_i))
acc_o = acc_o * scale[:, None]
if KV_SCALE_MODE == 1:
acc_o *= value_scale
o_ptrs = tl.make_block_ptr(
base=o_ptr + pid_c * stride_o_c + pid_b * stride_o_b + pid_h * stride_o_h,
shape=(gqa_group_size, head_dim),
Expand Down
Loading