Skip to content

perf(dsv4 prefill): avoid indexer pool token scan #618

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
zhaozhaozz wants to merge 1 commit into
base: main
Choose a base branch
from
Open

perf(dsv4 prefill): avoid indexer pool token scan #618

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
86 changes: 22 additions & 64 deletions models/deepseek/v4/prefill_indexer_compressor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -120,6 +120,28 @@ def prefill_indexer_compressor(
pl.write(write_dst_map, [0, map_seen], pl.cast(map_slot_raw, pl.INT32))
map_seen = map_seen + 1

# Pool reads current-window tokens from state, matching DeepSeek V4 decode.
# This requires every valid prefill token to have an inner state row.
for update_idx in pl.spmd(T * PACKED_PROJ_BLOCKS, name_hint="prefill_idx_c4_state_scatter_pre"):
update_ob = update_idx % PACKED_PROJ_BLOCKS
update_t = update_idx // PACKED_PROJ_BLOCKS
update_o0 = update_ob * OUT_TILE
if update_t < num_tokens:
state_row_raw = pl.read(inner_state_slot_mapping, [update_t])
if state_row_raw >= 0:
state_row = pl.cast(state_row_raw, pl.INDEX)
update_pos = pl.read(position_ids, [update_t])
ape_slot = pl.cast(update_pos % COMPRESS_RATIO, pl.INDEX)
ape_row = ape[ape_slot : ape_slot + 1, update_o0 : update_o0 + OUT_TILE]
kv_state_flat[state_row : state_row + 1, update_o0 : update_o0 + OUT_TILE] = kv_proj_scratch[
update_t : update_t + 1,
update_o0 : update_o0 + OUT_TILE,
]
score_state_flat[state_row : state_row + 1, update_o0 : update_o0 + OUT_TILE] = pl.add(
score_proj_scratch[update_t : update_t + 1, update_o0 : update_o0 + OUT_TILE],
ape_row,
)

for pool_idx in pl.spmd(PACKED_POOL_BLOCKS, name_hint="prefill_idx_c4_softmax_pool"):
write_i = pool_idx // HEAD_BLOCKS
hb = pool_idx - write_i * HEAD_BLOCKS
Expand Down Expand Up @@ -187,43 +209,6 @@ def prefill_indexer_compressor(
HEAD_DIM + h0 : HEAD_DIM + h0 + HEAD_CHUNK,
]

for pool_t in pl.range(T):
if pool_t < num_tokens:
pool_pos = pl.read(position_ids, [pool_t])
if pool_pos <= write_pos:
if pool_pos >= prev_start:
pool_ape_slot = pl.cast(pool_pos % COMPRESS_RATIO, pl.INDEX)
if pool_pos < cur_start:
pool_slot = pl.cast(pool_pos - prev_start, pl.INDEX)
pool_ape = ape[pool_ape_slot : pool_ape_slot + 1, h0 : h0 + HEAD_CHUNK]
pool_score = pl.add(
score_proj_scratch[pool_t : pool_t + 1, h0 : h0 + HEAD_CHUNK],
pool_ape,
)
pool_kv_tile[pool_slot : pool_slot + 1, 0:HEAD_CHUNK] = kv_proj_scratch[
pool_t : pool_t + 1,
h0 : h0 + HEAD_CHUNK,
]
pool_score_tile[pool_slot : pool_slot + 1, 0:HEAD_CHUNK] = pool_score
else:
pool_slot = pl.cast(COMPRESS_RATIO + pool_pos - cur_start, pl.INDEX)
pool_ape = ape[
pool_ape_slot : pool_ape_slot + 1,
HEAD_DIM + h0 : HEAD_DIM + h0 + HEAD_CHUNK,
]
pool_score = pl.add(
score_proj_scratch[
pool_t : pool_t + 1,
HEAD_DIM + h0 : HEAD_DIM + h0 + HEAD_CHUNK,
],
pool_ape,
)
pool_kv_tile[pool_slot : pool_slot + 1, 0:HEAD_CHUNK] = kv_proj_scratch[
pool_t : pool_t + 1,
HEAD_DIM + h0 : HEAD_DIM + h0 + HEAD_CHUNK,
]
pool_score_tile[pool_slot : pool_slot + 1, 0:HEAD_CHUNK] = pool_score

init_slot = STATE_LEN - 1
mi_buf = pl.create_tensor([1, HEAD_CHUNK], dtype=pl.FP32)
li_buf = pl.create_tensor([1, HEAD_CHUNK], dtype=pl.FP32)
Expand Down Expand Up @@ -340,33 +325,6 @@ def prefill_indexer_compressor(
0:HEAD_DIM,
]

for update_idx in pl.spmd(T * PACKED_PROJ_BLOCKS, name_hint="prefill_idx_c4_state_update"):
update_ob = update_idx % PACKED_PROJ_BLOCKS
update_t = update_idx // PACKED_PROJ_BLOCKS
update_o0 = update_ob * OUT_TILE
if update_t < num_tokens:
state_row_raw = pl.read(inner_state_slot_mapping, [update_t])
if state_row_raw >= 0:
state_row = pl.cast(state_row_raw, pl.INDEX)
update_pos = pl.read(position_ids, [update_t])
ape_slot = pl.cast(update_pos % COMPRESS_RATIO, pl.INDEX)
ape_row = ape[ape_slot : ape_slot + 1, update_o0 : update_o0 + OUT_TILE]
pool_dep = pl.mul(pooled_kv[0:1, 0:OUT_TILE], 0.0)
kv_state_flat[state_row : state_row + 1, update_o0 : update_o0 + OUT_TILE] = pl.add(
kv_proj_scratch[
update_t : update_t + 1,
update_o0 : update_o0 + OUT_TILE,
],
pool_dep,
)
score_state_flat[state_row : state_row + 1, update_o0 : update_o0 + OUT_TILE] = pl.add(
pl.add(
score_proj_scratch[update_t : update_t + 1, update_o0 : update_o0 + OUT_TILE],
ape_row,
),
pool_dep,
)

idx_kv_cache = pl.reshape(idx_kv_cache_flat, [PREFILL_IDX_BLOCK_NUM, BLOCK_SIZE, 1, HEAD_DIM])
kv_state = pl.reshape(kv_state_flat, [INNER_STATE_BLOCK_NUM, INNER_STATE_BLOCK_SIZE, OUT_DIM])
score_state = pl.reshape(score_state_flat, [INNER_STATE_BLOCK_NUM, INNER_STATE_BLOCK_SIZE, OUT_DIM])
Expand Down
Loading