Skip to content

Perf: fuse HCA merge, inverse RoPE, and packing #1102

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
zhangqi-chen merged 1 commit into from
Sep 2, 2026
Merged

Perf: fuse HCA merge, inverse RoPE, and packing #1102

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
40 changes: 5 additions & 35 deletions models/deepseek_v4_flash_dspark/decode_hca.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,3 @@
# Copyright (c) PyPTO Contributors.
# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
# CANN Open Software License Agreement Version 2.0 (the "License").
Expand Down Expand Up @@ -105,7 +105,6 @@
H = M.num_attention_heads
HEAD_DIM = M.head_dim
ROPE_HEAD_DIM = M.qk_rope_head_dim
NOPE_HEAD_DIM = M.nope_head_dim
Q_LORA = M.q_lora_rank
WIN = M.sliding_window
SOFTMAX_SCALE = M.softmax_scale
Expand Down Expand Up @@ -291,24 +290,21 @@
attention_local_flat = pl.create_tensor([ATTENTION_WINDOW_ROWS, O_GROUP_IN], dtype=pl.BF16)
attn_out = pl.create_tensor([t_dim, D], dtype=pl.BF16)
with pl.scope():
(
stream_heads,
rope_cos_il, rope_sin_signed, rope_swap_idx,
stream_heads_tid, rope_swap_tid, rope_cs_tid,
) = sparse_attn_hca(
attention_grouped = pl.create_tensor([O_GROUPS * LOCAL_T_PAD, O_GROUP_IN], dtype=pl.BF16)
attention_grouped, heads_tid = sparse_attn_hca(
q, kv_cache, window_swa_indices, window_swa_lens,
cmp_kv, cmp_block_table,
position_ids_local, kv_seq_lens,
attn_sink, freqs_cos_local, freqs_sin_local,
attention_grouped,
cache_ready_dep,
)

attention_grouped = pl.create_tensor([O_GROUPS * LOCAL_T_PAD, O_GROUP_IN], dtype=pl.BF16)
pack_work_count = (t_dim // ATTENTION_PUBLISH_T_TILE) * (H // H_TILE)
with pl.spmd(
ATTENTION_PUBLISH_WORKERS,
name_hint="hca_stream_pack_publish",
deps=[stream_heads_tid, rope_swap_tid, rope_cs_tid],
name_hint="hca_stream_publish",
deps=[heads_tid],
) as publish_tid:
worker = pl.tile.get_block_idx()
for pack_work in pl.range(worker, pack_work_count, ATTENTION_PUBLISH_WORKERS):
Expand All @@ -320,32 +316,6 @@
destination_rank = global_group0 // LOCAL_O_GROUPS
local_group0 = global_group0 - destination_rank * LOCAL_O_GROUPS

for stream_dt in pl.unroll(ATTENTION_PUBLISH_T_TILE):
stream_t = stream_t0 + stream_dt
stream_state_row = stream_t * H + stream_h0
stream_output = stream_heads[stream_state_row : stream_state_row + H_TILE, 0:HEAD_DIM]
stream_bf16 = pl.cast(stream_output, target_type=pl.BF16, mode="rint")
stream_rope = stream_output[0:H_TILE, NOPE_HEAD_DIM:HEAD_DIM]
stream_cos_il = rope_cos_il[stream_t : stream_t + 1, 0:ROPE_HEAD_DIM]
stream_sin_signed = rope_sin_signed[stream_t : stream_t + 1, 0:ROPE_HEAD_DIM]
stream_swap_zero = pl.full([H_TILE, ROPE_HEAD_DIM], dtype=pl.INT32, value=0)
stream_swap_idx = pl.col_expand_add(stream_swap_zero, rope_swap_idx[0:1, 0:ROPE_HEAD_DIM])
stream_swapped = pl.gather(stream_rope, dim=-1, index=stream_swap_idx)
stream_rot = pl.add(
pl.col_expand_mul(stream_rope, stream_cos_il),
pl.col_expand_mul(stream_swapped, stream_sin_signed),
)
n_rope_bf16 = pl.cast(stream_rot, target_type=pl.BF16, mode="rint")
n_full_bf16 = pl.concat(stream_bf16[0:H_TILE, 0:NOPE_HEAD_DIM], n_rope_bf16)
for n_hi in pl.unroll(H_TILE):
n_head = stream_h0 + n_hi
source_row = (n_head // HEADS_PER_GROUP) * T_PAD + stream_t
source_col = (n_head % HEADS_PER_GROUP) * HEAD_DIM
attention_grouped[
source_row : source_row + 1,
source_col : source_col + HEAD_DIM,
] = n_full_bf16[n_hi : n_hi + 1, 0:HEAD_DIM]

for group_slot in pl.unroll(PUBLISH_GROUPS):
source_row = (global_group0 + group_slot) * T_PAD + stream_t0
target_row = ((local_group0 + group_slot) * GROUP_T_PAD + tp_rank * local_t + stream_t0)
Expand Down
218 changes: 107 additions & 111 deletions models/deepseek_v4_flash_dspark/decode_sparse_attn_hca.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,3 @@
# Copyright (c) PyPTO Contributors.
# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
# CANN Open Software License Agreement Version 2.0 (the "License").
Expand Down Expand Up @@ -72,6 +72,7 @@
ROPE_INTERLEAVE_TILE = 2 * ROPE_TILE
ROPE_CS_T_TILE = 8
T_PAD = ((T + 16 - 1) // 16) * 16
MERGE_WORKERS = 48
ATTENTION_PUBLISH_WORKERS = 48
ATTENTION_PUBLISH_T_TILE = 8
LOCAL_O_GROUPS = O_GROUPS // TP
Expand All @@ -95,6 +96,8 @@
raise ValueError("each request must contain complete inverse-RoPE token tiles")
if H_TILE % HEADS_PER_GROUP != 0:
raise ValueError(f"HCA head tile {H_TILE} must contain complete output groups")
if PUBLISH_GROUPS != 2:
raise ValueError("HCA merge pack expects two output groups per head tile")
if O_GROUPS % TP != 0:
raise ValueError(f"output groups {O_GROUPS} must be divisible by TP size {TP}")
if LOCAL_O_GROUPS % PUBLISH_GROUPS != 0:
Expand All @@ -116,9 +119,10 @@
attn_sink: pl.Tensor[[H], pl.FP32],
freqs_cos: pl.Tensor[[T_DYN, ROPE_DIM], pl.BF16],
freqs_sin: pl.Tensor[[T_DYN, ROPE_DIM], pl.BF16],
o_packed_heads: pl.Tensor[[O_GROUPS * T_PAD, O_GROUP_IN], pl.BF16],
cache_ready_dep: pl.Scalar[pl.TASK_ID],
):
"""Run the raw and compressed branches, merge them, and build inverse-RoPE metadata."""
"""Merge raw and compressed attention into grouped inverse-RoPE heads."""
t_dim = pl.tensor.dim(q, 0)
rope_cs_blocks = t_dim // ROPE_CS_T_TILE
ori_block_num = pl.tensor.dim(ori_kv, 0)
Expand All @@ -144,10 +148,8 @@
cmp_partial_o = pl.create_tensor([cmp_partial_rows, HEAD_DIM], dtype=pl.FP32)
rope_cos_il = pl.create_tensor([T_PAD, ROPE_DIM], dtype=pl.FP32)
rope_sin_signed = pl.create_tensor([T_PAD, ROPE_DIM], dtype=pl.FP32)
rope_swap_idx = pl.create_tensor([1, ROPE_DIM], dtype=pl.INT32)
raw_branch_tids = pl.array.create(1, pl.TASK_ID)
cmp_branch_tids = pl.array.create(1, pl.TASK_ID)
rope_swap_tids = pl.array.create(1, pl.TASK_ID)
rope_cs_tids = pl.array.create(1, pl.TASK_ID)

with pl.scope():
Expand Down Expand Up @@ -253,21 +255,6 @@
)
pl.store(valid_mask, [valid_t0, 0], raw_valid)

with pl.spmd(1, name_hint="rope_swap") as rope_swap_tid:
sw_block = pl.tile.get_block_idx()
sw_one = pl.full([1, ROPE_DIM], dtype=pl.FP32, value=1.0)
sw_index_i32 = pl.arange(0, [1, ROPE_DIM], dtype=pl.INT32)
sw_index = pl.cast(sw_index_i32, target_type=pl.FP32)
sw_col = pl.col_expand_mul(sw_one, sw_index)
sw_dup = pl.mul(sw_col, 0.5)
sw_dup_i32 = pl.cast(sw_dup, target_type=pl.INT32, mode="trunc")
sw_dup_f = pl.cast(sw_dup_i32, target_type=pl.FP32)
sw_lane = pl.sub(sw_col, pl.mul(sw_dup_f, 2.0))
sw_next = pl.add(sw_col, 1.0)
sw_lane_offset = pl.mul(sw_lane, 2.0)
sw_swap = pl.sub(sw_next, sw_lane_offset)
rope_swap_idx[sw_block : sw_block + 1, 0:ROPE_DIM] = pl.cast(sw_swap, target_type=pl.INT32)

with pl.spmd(HALF_ROPE // ROPE_TILE, name_hint="rope_cs") as rope_cs_tid:
cp = pl.tile.get_block_idx()
cp_r0 = cp * ROPE_TILE
Expand Down Expand Up @@ -388,7 +375,6 @@
pl.store(stream_o_new, [stream_state_row, 0], stream_heads)

raw_branch_tids[0] = raw_heads_tid
rope_swap_tids[0] = rope_swap_tid
rope_cs_tids[0] = rope_cs_tid

with pl.scope():
Expand Down Expand Up @@ -509,62 +495,110 @@
cmp_branch_tids[0] = cmp_qk_tid

with pl.spmd(
stream_block_count,
name_hint="hca_stream_merge",
deps=[raw_branch_tids[0], cmp_branch_tids[0]],
) as stream_heads_tid:
stream_idx = pl.tile.get_block_idx()
stream_t = stream_idx // (H // H_TILE)
stream_h_tile = stream_idx - stream_t * (H // H_TILE)
stream_h0 = stream_h_tile * H_TILE
stream_state_row = stream_t * H + stream_h0
stream_m = pl.load(stream_state_m, [stream_state_row, 0], [H_TILE, 1], target_memory=pl.MemorySpace.Vec)
stream_l = pl.load(stream_state_l, [stream_state_row, 0], [H_TILE, 1], target_memory=pl.MemorySpace.Vec)
stream_o = pl.load(stream_heads, [stream_state_row, 0], [H_TILE, HEAD_DIM], target_memory=pl.MemorySpace.Vec)
stream_token_base = stream_t * (H // H_TILE) * cmp_work_count * H_TILE
for stream_work in pl.range(cmp_work_count):
stream_partial_row = (stream_token_base + stream_h_tile * cmp_work_count * H_TILE + stream_work * H_TILE)
stream_cmp_m_aligned = pl.load(
cmp_partial_m,
[stream_partial_row, 0],
[H_TILE, 8],
MERGE_WORKERS,
name_hint="hca_stream_merge_pack",
deps=[raw_branch_tids[0], cmp_branch_tids[0], rope_cs_tids[0]],
) as heads_tid:
worker = pl.tile.get_block_idx()
for stream_idx in pl.range(worker, stream_block_count, MERGE_WORKERS):
merge_t = stream_idx // (H // H_TILE)
merge_h_tile = stream_idx - merge_t * (H // H_TILE)
merge_h0 = merge_h_tile * H_TILE
merge_state_row = merge_t * H + merge_h0
stream_m = pl.load(stream_state_m, [merge_state_row, 0], [H_TILE, 1], target_memory=pl.MemorySpace.Vec)
stream_l = pl.load(stream_state_l, [merge_state_row, 0], [H_TILE, 1], target_memory=pl.MemorySpace.Vec)
stream_o = pl.load(stream_heads, [merge_state_row, 0], [H_TILE, HEAD_DIM], target_memory=pl.MemorySpace.Vec)
stream_token_base = merge_t * (H // H_TILE) * cmp_work_count * H_TILE
for stream_work in pl.range(cmp_work_count):
stream_partial_row = (
stream_token_base
+ merge_h_tile * cmp_work_count * H_TILE
+ stream_work * H_TILE
)
stream_cmp_m_aligned = pl.load(
cmp_partial_m,
[stream_partial_row, 0],
[H_TILE, 8],
target_memory=pl.MemorySpace.Vec,
)
stream_cmp_l_aligned = pl.load(
cmp_partial_l,
[stream_partial_row, 0],
[H_TILE, 8],
target_memory=pl.MemorySpace.Vec,
)
stream_cmp_m = stream_cmp_m_aligned[0:H_TILE, 0:1]
stream_cmp_l = stream_cmp_l_aligned[0:H_TILE, 0:1]
stream_cmp_o = pl.load(
cmp_partial_o,
[stream_partial_row, 0],
[H_TILE, HEAD_DIM],
target_memory=pl.MemorySpace.Vec,
)
stream_m_new = pl.maximum(stream_m, stream_cmp_m)
stream_alpha = pl.exp(pl.sub(stream_m, stream_m_new))
stream_beta = pl.exp(pl.sub(stream_cmp_m, stream_m_new))
stream_l = pl.add(
pl.mul(stream_alpha, stream_l),
pl.mul(stream_beta, stream_cmp_l),
)
stream_o = pl.add(
pl.row_expand_mul(stream_o, stream_alpha),
pl.row_expand_mul(stream_cmp_o, stream_beta),
)
stream_m = stream_m_new
stream_sink = pl.load(
attn_sink_col,
[merge_h0, 0],
[H_TILE, 1],
target_memory=pl.MemorySpace.Vec,
)
stream_cmp_l_aligned = pl.load(
cmp_partial_l,
[stream_partial_row, 0],
[H_TILE, 8],
target_memory=pl.MemorySpace.Vec,
stream_sink_tile = pl.add(pl.sub(stream_m, stream_m), stream_sink)
stream_denom = pl.add(stream_l, pl.exp(pl.sub(stream_sink_tile, stream_m)))
stream_output = pl.row_expand_div(stream_o, stream_denom)
pl.store(stream_output, [merge_state_row, 0], stream_heads)
packed_stream_output = stream_heads[merge_state_row : merge_state_row + H_TILE, 0:HEAD_DIM]
stream_bf16 = pl.cast(packed_stream_output, target_type=pl.BF16, mode="rint")
stream_rope = packed_stream_output[0:H_TILE, NOPE_DIM:HEAD_DIM]
stream_cos_il = rope_cos_il[merge_t : merge_t + 1, 0:ROPE_DIM]
stream_sin_signed = rope_sin_signed[merge_t : merge_t + 1, 0:ROPE_DIM]
stream_swap_one = pl.full([1, ROPE_DIM], dtype=pl.FP32, value=1.0)
stream_swap_index = pl.cast(pl.arange(0, [1, ROPE_DIM], dtype=pl.INT32), target_type=pl.FP32)
stream_swap_col = pl.col_expand_mul(stream_swap_one, stream_swap_index)
stream_swap_dup = pl.cast(
pl.mul(stream_swap_col, 0.5),
target_type=pl.INT32,
mode="trunc",
)
stream_cmp_m = stream_cmp_m_aligned[0:H_TILE, 0:1]
stream_cmp_l = stream_cmp_l_aligned[0:H_TILE, 0:1]
stream_cmp_o = pl.load(
cmp_partial_o,
[stream_partial_row, 0],
[H_TILE, HEAD_DIM],
target_memory=pl.MemorySpace.Vec,
stream_swap_dup_f = pl.cast(stream_swap_dup, target_type=pl.FP32)
stream_swap_lane = pl.sub(stream_swap_col, pl.mul(stream_swap_dup_f, 2.0))
stream_swap = pl.sub(
pl.add(stream_swap_col, 1.0),
pl.mul(stream_swap_lane, 2.0),
)
stream_m_new = pl.maximum(stream_m, stream_cmp_m)
stream_alpha = pl.exp(pl.sub(stream_m, stream_m_new))
stream_beta = pl.exp(pl.sub(stream_cmp_m, stream_m_new))
stream_l = pl.add(pl.mul(stream_alpha, stream_l), pl.mul(stream_beta, stream_cmp_l))
stream_o = pl.add(pl.row_expand_mul(stream_o, stream_alpha), pl.row_expand_mul(stream_cmp_o, stream_beta))
stream_m = stream_m_new
stream_sink = pl.load(attn_sink_col, [stream_h0, 0], [H_TILE, 1], target_memory=pl.MemorySpace.Vec)
stream_sink_tile = pl.add(pl.sub(stream_m, stream_m), stream_sink)
stream_denom = pl.add(stream_l, pl.exp(pl.sub(stream_sink_tile, stream_m)))
stream_output = pl.row_expand_div(stream_o, stream_denom)
pl.store(stream_output, [stream_state_row, 0], stream_heads)

return (
stream_heads,
rope_cos_il,
rope_sin_signed,
rope_swap_idx,
stream_heads_tid,
rope_swap_tids[0],
rope_cs_tids[0],
)
stream_swap_row = pl.cast(stream_swap, target_type=pl.INT32)
stream_swap_zero = pl.full([H_TILE, ROPE_DIM], dtype=pl.INT32, value=0)
stream_swap_idx = pl.col_expand_add(stream_swap_zero, stream_swap_row)
stream_swapped = pl.gather(stream_rope, dim=-1, index=stream_swap_idx)
stream_rot = pl.add(
pl.col_expand_mul(stream_rope, stream_cos_il),
pl.col_expand_mul(stream_swapped, stream_sin_signed),
)
stream_rope_bf16 = pl.cast(stream_rot, target_type=pl.BF16, mode="rint")
stream_full_bf16 = pl.concat(
stream_bf16[0:H_TILE, 0:NOPE_DIM],
stream_rope_bf16,
)
for stream_hi in pl.unroll(H_TILE):
stream_head = merge_h0 + stream_hi
stream_pack_row = (stream_head // HEADS_PER_GROUP) * T_PAD + merge_t
stream_pack_col = (stream_head % HEADS_PER_GROUP) * HEAD_DIM
o_packed_heads[
stream_pack_row : stream_pack_row + 1,
stream_pack_col : stream_pack_col + HEAD_DIM,
] = stream_full_bf16[stream_hi : stream_hi + 1, 0:HEAD_DIM]

return o_packed_heads, heads_tid


@pl.jit.inline(auto_scope=False)
Expand All @@ -584,11 +618,7 @@
cache_ready_dep: pl.Scalar[pl.TASK_ID],
) -> tuple[pl.Tensor, pl.Scalar[pl.TASK_ID]]:
"""Write HCA heads as grouped ``[T_PAD, O_GROUP_IN]`` slabs."""
(
stream_heads,
rope_cos_il, rope_sin_signed, rope_swap_idx,
stream_heads_tid, rope_swap_tid, rope_cs_tid,
) = sparse_attn_hca(
o_packed_heads, heads_tid = sparse_attn_hca(
q,
ori_kv,
window_swa_indices,
Expand All @@ -600,43 +630,9 @@
attn_sink,
freqs_cos,
freqs_sin,
o_packed_heads,
cache_ready_dep,
)
t_dim = pl.tensor.dim(q, 0)
stream_block_count = t_dim * (H // H_TILE)

with pl.spmd(
stream_block_count,
name_hint="hca_stream_pack",
deps=[stream_heads_tid, rope_swap_tid, rope_cs_tid],
) as heads_tid:
stream_idx = pl.tile.get_block_idx()
stream_t = stream_idx // (H // H_TILE)
stream_h_tile = stream_idx - stream_t * (H // H_TILE)
stream_h0 = stream_h_tile * H_TILE
stream_state_row = stream_t * H + stream_h0
stream_output = stream_heads[stream_state_row : stream_state_row + H_TILE, 0:HEAD_DIM]
stream_bf16 = pl.cast(stream_output, target_type=pl.BF16, mode="rint")
stream_rope = stream_output[0:H_TILE, NOPE_DIM:HEAD_DIM]
stream_cos_il = rope_cos_il[stream_t : stream_t + 1, 0:ROPE_DIM]
stream_sin_signed = rope_sin_signed[stream_t : stream_t + 1, 0:ROPE_DIM]
stream_swap_zero = pl.full([H_TILE, ROPE_DIM], dtype=pl.INT32, value=0)
stream_swap_idx = pl.col_expand_add(stream_swap_zero, rope_swap_idx[0:1, 0:ROPE_DIM])
stream_swapped = pl.gather(stream_rope, dim=-1, index=stream_swap_idx)
stream_rot = pl.add(
pl.col_expand_mul(stream_rope, stream_cos_il),
pl.col_expand_mul(stream_swapped, stream_sin_signed),
)
n_rope_bf16 = pl.cast(stream_rot, target_type=pl.BF16, mode="rint")
n_full_bf16 = pl.concat(stream_bf16[0:H_TILE, 0:NOPE_DIM], n_rope_bf16)
for n_hi in pl.unroll(H_TILE):
n_head = stream_h0 + n_hi
n_pack_row = (n_head // HEADS_PER_GROUP) * T_PAD + stream_t
n_col = (n_head % HEADS_PER_GROUP) * HEAD_DIM
o_packed_heads[
n_pack_row : n_pack_row + 1,
n_col : n_col + HEAD_DIM,
] = n_full_bf16[n_hi : n_hi + 1, 0:HEAD_DIM]

return o_packed_heads, heads_tid

Expand Down
Loading