Skip to content

Fix scalar-indexed slices in V3.2 / Qwen3-32B + sparse_attn rope refactor #288

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 2 commits into from
May 15, 2026
+52 −35
Merged

Fix scalar-indexed slices in V3.2 / Qwen3-32B + sparse_attn rope refactor #288

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e2e9823
Fix DeepSeek V3.2 / Qwen3-32B scalar-indexed slice compile failures
zhangqi-chen May 15, 2026
a39c216
Refactor DeepSeek V4 sparse_attn rope-interleave assemble
zhangqi-chen May 15, 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
49 changes: 25 additions & 24 deletions models/deepseek/v3_2/deepseek_v3_2_decode_front.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -199,13 +199,14 @@ def deepseek_v3_2_decode_front_scope1234(
for b in pl.parallel(BATCH):
with pl.at(level=pl.Level.CORE_GROUP, name_hint="q_rope"):
ctx_len = pl.read(seq_lens, [b])
cos_lo = rope_cos[ctx_len - 1, 0 : QK_ROPE_HEAD_DIM // 2]
cos_hi = rope_cos[ctx_len - 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
sin_lo = rope_sin[ctx_len - 1, 0 : QK_ROPE_HEAD_DIM // 2]
sin_hi = rope_sin[ctx_len - 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
pos = ctx_len - 1
cos_lo = rope_cos[pos : pos + 1, 0 : QK_ROPE_HEAD_DIM // 2]
cos_hi = rope_cos[pos : pos + 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
sin_lo = rope_sin[pos : pos + 1, 0 : QK_ROPE_HEAD_DIM // 2]
sin_hi = rope_sin[pos : pos + 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
for q_col in pl.range(QK_NOPE_HEAD_DIM, NUM_HEADS * QK_HEAD_DIM, QK_HEAD_DIM):
q_lo = pl.cast(q_proj[b, q_col : q_col + QK_ROPE_HEAD_DIM // 2], target_type=pl.FP32)
q_hi = pl.cast(q_proj[b, q_col + QK_ROPE_HEAD_DIM // 2 : q_col + QK_ROPE_HEAD_DIM], target_type=pl.FP32)
q_lo = pl.cast(q_proj[b : b + 1, q_col : q_col + QK_ROPE_HEAD_DIM // 2], target_type=pl.FP32)
q_hi = pl.cast(q_proj[b : b + 1, q_col + QK_ROPE_HEAD_DIM // 2 : q_col + QK_ROPE_HEAD_DIM], target_type=pl.FP32)
q_rot_lo = pl.sub(pl.col_expand_mul(q_lo, cos_lo), pl.col_expand_mul(q_hi, sin_lo))
q_rot_hi = pl.add(pl.col_expand_mul(q_hi, cos_hi), pl.col_expand_mul(q_lo, sin_hi))
q_proj = pl.assemble(q_proj, pl.cast(q_rot_lo, target_type=pl.BF16), [b, q_col])
Expand All @@ -230,13 +231,13 @@ def deepseek_v3_2_decode_front_scope1234(
ctx_len = pl.read(seq_lens, [b])
pos = ctx_len - 1
cache_row = b * MAX_SEQ + pos
cos_lo = rope_cos[pos, 0 : QK_ROPE_HEAD_DIM // 2]
cos_hi = rope_cos[pos, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
sin_lo = rope_sin[pos, 0 : QK_ROPE_HEAD_DIM // 2]
sin_hi = rope_sin[pos, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
kv_normed_row = kv_normed_out[b, 0 : KV_LORA_RANK]
pe_lo = pl.cast(kv_a_out[b, KV_LORA_RANK : KV_LORA_RANK + QK_ROPE_HEAD_DIM // 2], target_type=pl.FP32)
pe_hi = pl.cast(kv_a_out[b, KV_LORA_RANK + QK_ROPE_HEAD_DIM // 2 : KV_LORA_RANK + QK_ROPE_HEAD_DIM], target_type=pl.FP32)
cos_lo = rope_cos[pos : pos + 1, 0 : QK_ROPE_HEAD_DIM // 2]
cos_hi = rope_cos[pos : pos + 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
sin_lo = rope_sin[pos : pos + 1, 0 : QK_ROPE_HEAD_DIM // 2]
sin_hi = rope_sin[pos : pos + 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
kv_normed_row = kv_normed_out[b : b + 1, 0 : KV_LORA_RANK]
pe_lo = pl.cast(kv_a_out[b : b + 1, KV_LORA_RANK : KV_LORA_RANK + QK_ROPE_HEAD_DIM // 2], target_type=pl.FP32)
pe_hi = pl.cast(kv_a_out[b : b + 1, KV_LORA_RANK + QK_ROPE_HEAD_DIM // 2 : KV_LORA_RANK + QK_ROPE_HEAD_DIM], target_type=pl.FP32)
pe_rot_lo = pl.sub(pl.col_expand_mul(pe_lo, cos_lo), pl.col_expand_mul(pe_hi, sin_lo))
pe_rot_hi = pl.add(pl.col_expand_mul(pe_hi, cos_hi), pl.col_expand_mul(pe_lo, sin_hi))
kv_cache = pl.assemble(kv_cache, kv_normed_row, [cache_row, 0])
Expand Down Expand Up @@ -303,19 +304,19 @@ def deepseek_v3_2_decode_front_scope1234(
for b in pl.parallel(BATCH):
with pl.at(level=pl.Level.CORE_GROUP, name_hint="s2_idx_rope"):
pos = pl.read(seq_lens, [b]) - 1
cos_lo = rope_cos[pos, 0 : QK_ROPE_HEAD_DIM // 2]
cos_hi = rope_cos[pos, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
sin_lo = rope_sin[pos, 0 : QK_ROPE_HEAD_DIM // 2]
sin_hi = rope_sin[pos, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
cos_lo = rope_cos[pos : pos + 1, 0 : QK_ROPE_HEAD_DIM // 2]
cos_hi = rope_cos[pos : pos + 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
sin_lo = rope_sin[pos : pos + 1, 0 : QK_ROPE_HEAD_DIM // 2]
sin_hi = rope_sin[pos : pos + 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM]
for q_col in pl.range(0, INDEX_Q_OUT, INDEX_HEAD_DIM):
s2_q_lo = pl.cast(q_idx_full[b, q_col : q_col + QK_ROPE_HEAD_DIM // 2], target_type=pl.FP32)
s2_q_hi = pl.cast(q_idx_full[b, q_col + QK_ROPE_HEAD_DIM // 2 : q_col + QK_ROPE_HEAD_DIM], target_type=pl.FP32)
s2_q_lo = pl.cast(q_idx_full[b : b + 1, q_col : q_col + QK_ROPE_HEAD_DIM // 2], target_type=pl.FP32)
s2_q_hi = pl.cast(q_idx_full[b : b + 1, q_col + QK_ROPE_HEAD_DIM // 2 : q_col + QK_ROPE_HEAD_DIM], target_type=pl.FP32)
s2_q_rot_lo = pl.sub(pl.col_expand_mul(s2_q_lo, cos_lo), pl.col_expand_mul(s2_q_hi, sin_lo))
s2_q_rot_hi = pl.add(pl.col_expand_mul(s2_q_hi, cos_hi), pl.col_expand_mul(s2_q_lo, sin_hi))
q_idx_full = pl.assemble(q_idx_full, pl.cast(s2_q_rot_lo, target_type=pl.BF16), [b, q_col])
q_idx_full = pl.assemble(q_idx_full, pl.cast(s2_q_rot_hi, target_type=pl.BF16), [b, q_col + QK_ROPE_HEAD_DIM // 2])
s2_k_lo = pl.cast(k_idx[b, 0 : QK_ROPE_HEAD_DIM // 2], target_type=pl.FP32)
s2_k_hi = pl.cast(k_idx[b, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM], target_type=pl.FP32)
s2_k_lo = pl.cast(k_idx[b : b + 1, 0 : QK_ROPE_HEAD_DIM // 2], target_type=pl.FP32)
s2_k_hi = pl.cast(k_idx[b : b + 1, QK_ROPE_HEAD_DIM // 2 : QK_ROPE_HEAD_DIM], target_type=pl.FP32)
s2_k_rot_lo = pl.sub(pl.col_expand_mul(s2_k_lo, cos_lo), pl.col_expand_mul(s2_k_hi, sin_lo))
s2_k_rot_hi = pl.add(pl.col_expand_mul(s2_k_hi, cos_hi), pl.col_expand_mul(s2_k_lo, sin_hi))
k_idx = pl.assemble(k_idx, pl.cast(s2_k_rot_lo, target_type=pl.BF16), [b, 0])
Expand Down Expand Up @@ -435,7 +436,7 @@ def deepseek_v3_2_decode_front_scope1234(

# Stage 3.3: Run sort32 + mrgsort.
with pl.at(level=pl.Level.CORE_GROUP, name_hint="s3_sort"):
s3_score_row = scores[b, :]
s3_score_row = scores[b : b + 1, :]
idx_init = pl.tensor.arange(0, [1, SORT_LEN], dtype=pl.UINT32)
s3_sorted_t = pl.tensor.sort32(s3_score_row, idx_init)
s3_sorted_t = pl.tensor.mrgsort(s3_sorted_t, block_len=64)
Expand Down Expand Up @@ -472,7 +473,7 @@ def deepseek_v3_2_decode_front_scope1234(

# Stage 4.1: Load q_pe and project q_nope into latent space.
with pl.at(level=pl.Level.CORE_GROUP, name_hint="s4_q_pe_load"):
q_pe = pl.cast(q_proj[b, q_col + QK_NOPE_HEAD_DIM : q_col + QK_HEAD_DIM], target_type=pl.FP32)
q_pe = pl.cast(q_proj[b : b + 1, q_col + QK_NOPE_HEAD_DIM : q_col + QK_HEAD_DIM], target_type=pl.FP32)
q_pe_batch = pl.col_expand(
pl.full([MATMUL_ROW_PAD, QK_ROPE_HEAD_DIM], dtype=pl.FP32, value=0.0),
q_pe,
Expand All @@ -481,7 +482,7 @@ def deepseek_v3_2_decode_front_scope1234(
pl.full([MATMUL_ROW_PAD, QK_NOPE_HEAD_DIM], dtype=pl.FP32, value=0.0),
target_type=pl.BF16,
)
q_nope_padded = pl.col_expand(q_nope_padded, q_proj[b, q_col : q_col + QK_NOPE_HEAD_DIM])
q_nope_padded = pl.col_expand(q_nope_padded, q_proj[b : b + 1, q_col : q_col + QK_NOPE_HEAD_DIM])
q_nope_latent_batch = pl.full([MATMUL_ROW_PAD, KV_LORA_RANK], dtype=pl.FP32, value=0.0)

# Stage 4.2: Project q_nope to latent space chunk-by-chunk.
Expand Down
30 changes: 23 additions & 7 deletions models/deepseek/v4/sparse_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,6 +129,8 @@ def sparse_attn(
o_proj_odd = pl.create_tensor([T * H, HALF_ROPE], dtype=pl.FP32)
rope_even = pl.create_tensor([T * H, HALF_ROPE], dtype=pl.BF16)
rope_odd = pl.create_tensor([T * H, HALF_ROPE], dtype=pl.BF16)
rope_even_interleave_buf = pl.create_tensor([T * H, ROPE_DIM], dtype=pl.FP32)
rope_odd_interleave_buf = pl.create_tensor([T * H, ROPE_DIM], dtype=pl.FP32)
o_rope_interleave = pl.create_tensor([T * H, ROPE_DIM], dtype=pl.BF16)
o_packed = pl.create_tensor([O_GROUPS * T, O_GROUP_IN], dtype=pl.BF16)

Expand Down Expand Up @@ -254,7 +256,7 @@ def sparse_attn(
rope_even = pl.assemble(rope_even, pl.cast(rope_even_acc, target_type=pl.BF16, mode="rint"), [rope_head_row, 0])
rope_odd = pl.assemble(rope_odd, pl.cast(rope_odd_acc, target_type=pl.BF16, mode="rint"), [rope_head_row, 0])

with pl.at(level=pl.Level.CORE_GROUP, name_hint="cfa_proj_rope_assemble"):
with pl.at(level=pl.Level.CORE_GROUP, name_hint="cfa_proj_rope_assemble_matmul"):
for r0 in pl.range(0, HALF_ROPE, ROPE_CHUNK):
rope_even_chunk = rope_even[rope_head_row : rope_head_row + H, r0 : r0 + ROPE_CHUNK]
rope_odd_chunk = rope_odd[rope_head_row : rope_head_row + H, r0 : r0 + ROPE_CHUNK]
Expand All @@ -270,16 +272,30 @@ def sparse_attn(
b_trans=True,
out_dtype=pl.FP32,
)
rope_chunk = pl.cast(
pl.add(rope_even_interleave, rope_odd_interleave),
target_type=pl.BF16,
rope_even_interleave_buf = pl.assemble(
rope_even_interleave_buf,
rope_even_interleave,
[rope_head_row, 2 * r0],
)
o_rope_interleave = pl.assemble(
o_rope_interleave,
rope_chunk,
rope_odd_interleave_buf = pl.assemble(
rope_odd_interleave_buf,
rope_odd_interleave,
[rope_head_row, 2 * r0],
)

with pl.at(level=pl.Level.CORE_GROUP, name_hint="cfa_proj_rope_assemble_combine"):
rope_even_tile = rope_even_interleave_buf[rope_head_row : rope_head_row + H, 0 : ROPE_DIM]
rope_odd_tile = rope_odd_interleave_buf[rope_head_row : rope_head_row + H, 0 : ROPE_DIM]
rope_full = pl.cast(
pl.add(rope_even_tile, rope_odd_tile),
target_type=pl.BF16,
)
o_rope_interleave = pl.assemble(
o_rope_interleave,
rope_full,
[rope_head_row, 0],
)

for b in pl.range(B):
for h in pl.parallel(0, H, 1):
pack_head_row = b * H + h
Expand Down
8 changes: 4 additions & 4 deletions models/qwen3/32b/qwen3_32b_decode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -168,10 +168,10 @@ def qwen3_decode(
ctx_len = pl.read(seq_lens, [b])
pos = ctx_len - 1
ctx_blocks = (ctx_len + SEQ_TILE - 1) // SEQ_TILE
cos_lo = rope_cos[pos, 0 : HALF_DIM]
cos_hi = rope_cos[pos, HALF_DIM : HEAD_DIM]
sin_lo = rope_sin[pos, 0 : HALF_DIM]
sin_hi = rope_sin[pos, HALF_DIM : HEAD_DIM]
cos_lo = rope_cos[pos : pos + 1, 0 : HALF_DIM]
cos_hi = rope_cos[pos : pos + 1, HALF_DIM : HEAD_DIM]
sin_lo = rope_sin[pos : pos + 1, 0 : HALF_DIM]
sin_hi = rope_sin[pos : pos + 1, HALF_DIM : HEAD_DIM]

# Stage 1: K RoPE + cache update + V cache + Q RoPE + pad.
with pl.at(level=pl.Level.CORE_GROUP, name_hint="rope_kv_cache"):
Expand Down
Loading