[Qwen3.8-Flash-Next] Compact indexer logits workspace to improve prefill efficiency (#54915)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
This commit is contained in:
Thien Tran
2026-09-04 00:43:20 -07:00
committed by GitHub
parent 8f816a3f66
commit a5c9179e73
4 changed files with 46 additions and 16 deletions
+20 -3
View File
@@ -244,6 +244,7 @@ def test_qsa_side_metadata_marks_cudagraph_padding_inert() -> None:
num_actual_tokens=16,
num_reqs=4,
max_query_len=4,
max_seq_len=68,
query_start_loc=query_start_loc,
query_start_loc_cpu=query_start_loc.cpu(),
seq_lens=torch.tensor([68, 68, 68, 0], dtype=torch.int32, device=device),
@@ -297,6 +298,7 @@ def test_qsa_circular_buffer_metadata_keeps_only_each_requests_suffix() -> None:
num_actual_tokens=16,
num_reqs=3,
max_query_len=7,
max_seq_len=11,
query_start_loc=query_start_loc,
query_start_loc_cpu=query_start_loc.cpu(),
seq_lens=torch.tensor([9, 11, 0], dtype=torch.int32, device=device),
@@ -428,6 +430,7 @@ def test_qsa_compressed_metadata_keeps_dummy_slots_inert() -> None:
num_actual_tokens=8,
num_reqs=3,
max_query_len=5,
max_seq_len=12,
query_start_loc=query_start_loc,
query_start_loc_cpu=query_start_loc.cpu(),
seq_lens=torch.tensor([7, 0, 12], dtype=torch.int32, device=device),
@@ -673,13 +676,24 @@ def test_qsa_decode_selection_correctness(
@requires_qsa_kernels
def test_qsa_prefill_selection_correctness() -> None:
@pytest.mark.parametrize("seq_len_slack", [0, 1792])
@pytest.mark.parametrize("force_chunk", [False, True])
def test_qsa_prefill_selection_correctness(
monkeypatch: pytest.MonkeyPatch, seq_len_slack: int, force_chunk: bool
) -> None:
# page_size=24 (does not divide the 64-aligned clipped width) and an
# oversized page table, so the clipped logits width comes from
# max_seq_len, not page geometry. seq_len_slack > 0 simulates the
# spec-decode case where the bound is an over-estimate. force_chunk
# drives the logits budget to one row per chunk.
if force_chunk:
monkeypatch.setenv("VLLM_SPARSE_INDEXER_MAX_LOGITS_MB", "0")
torch.manual_seed(2)
query_lens = [3, 33]
rows, heads, head_dim = sum(query_lens), 4, 128
q = torch.randn(rows, heads, head_dim, device="cuda", dtype=torch.bfloat16)
cache = torch.randn(160, 16, 1, head_dim, device="cuda", dtype=torch.bfloat16)
page_table = torch.randperm(160, device="cuda", dtype=torch.int32).reshape(2, 80)
cache = torch.randn(128, 24, 1, head_dim, device="cuda", dtype=torch.bfloat16)
page_table = torch.randperm(128, device="cuda", dtype=torch.int32).reshape(2, 64)
token_to_req = torch.repeat_interleave(
torch.arange(2, device="cuda", dtype=torch.int32),
torch.tensor(query_lens, device="cuda"),
@@ -713,6 +727,7 @@ def test_qsa_prefill_selection_correctness() -> None:
compress_ratio,
max(query_lens),
actual,
max_seq_len=sequence_lengths.max().item() + seq_len_slack,
)
expected = _qsa_select_paged_reference(
q,
@@ -931,6 +946,7 @@ def test_qsa_split_selection_correctness(workspace_init, decode_query_len: int)
compress_ratio,
query_lens[-1],
block_indices[prefill_slice],
max_seq_len=sequence_lengths.max().item(),
)
actual = torch.empty(
(rows, token_topk + compress_ratio - 1), device="cuda", dtype=torch.int32
@@ -983,6 +999,7 @@ def test_qsa_selection_handles_no_complete_compressed_blocks(workspace_init) ->
compress_ratio=4,
max_query_len=2,
block_indices=block_indices,
max_seq_len=64, # clamps to the page-table capacity
)
selected = torch.empty((2, 2051), device="cuda", dtype=torch.int32)
qsa_indexer_ops.expand_qsa_block_indices(
@@ -591,6 +591,7 @@ class QSAForwardMetadata(AttentionMetadata):
num_prefill_tokens: int
max_query_len: int
decode_query_len: int
max_seq_len: int
storage_block_size: int
compress_ratio: int
@@ -717,6 +718,7 @@ class QSAMetadataBuilder(AttentionMetadataBuilder[QSAForwardMetadata]):
num_prefill_tokens=num_prefill_tokens,
max_query_len=common_attn_metadata.max_query_len,
decode_query_len=decode_query_len,
max_seq_len=common_attn_metadata.max_seq_len,
storage_block_size=self.storage_block_size,
compress_ratio=self.compress_ratio,
)
@@ -405,6 +405,7 @@ class QSAIndexer(nn.Module):
self.compress_ratio,
compressed_metadata.max_query_len,
block_indices[prefill_slice],
compressed_metadata.max_seq_len,
)
expand_qsa_block_indices(
block_indices,
+23 -13
View File
@@ -97,7 +97,7 @@ def _qsa_mqa_paged_uniform_kernel(
scores,
(BLOCK_N, DECODE_QUERY_LEN_PADDED, NUM_HEADS_PADDED),
)
score = tl.sum(scores, axis=2) / HEAD_DIM**0.5
score = tl.sum(scores, axis=2)
tl.store(
logits_ptr + rows[None, :] * stride_logits_row + columns[:, None],
score,
@@ -123,8 +123,8 @@ def _qsa_mqa_paged_prefill_kernel(
stride_logits_row,
num_rows,
query_offset,
page_table_width,
PAGE_SIZE: tl.constexpr,
PAGE_TABLE_WIDTH: tl.constexpr,
NUM_HEADS: tl.constexpr,
HEAD_DIM: tl.constexpr,
TILE_R: tl.constexpr,
@@ -132,7 +132,7 @@ def _qsa_mqa_paged_prefill_kernel(
K_TILES: tl.constexpr,
STAGES: tl.constexpr,
) -> None:
NUM_COLUMNS: tl.constexpr = PAGE_TABLE_WIDTH * PAGE_SIZE
num_columns = page_table_width * PAGE_SIZE
NUM_HEADS_PADDED: tl.constexpr = triton.next_power_of_2(NUM_HEADS)
# tl.dot requires a reduction dimension of at least 16.
BLOCK_D: tl.constexpr = max(16, triton.next_power_of_2(HEAD_DIM))
@@ -163,7 +163,7 @@ def _qsa_mqa_paged_prefill_kernel(
if k_tile_start * BLOCK_N >= max_visible:
return
k_tile_end = tl.minimum(k_tile_start + K_TILES, tl.cdiv(max_visible, BLOCK_N))
k_tile_end = tl.minimum(k_tile_end, tl.cdiv(NUM_COLUMNS, BLOCK_N))
k_tile_end = tl.minimum(k_tile_end, tl.cdiv(num_columns, BLOCK_N))
dims = tl.arange(0, BLOCK_D)
m = tl.arange(0, TILE_R * NUM_HEADS_PADDED)
@@ -184,7 +184,7 @@ def _qsa_mqa_paged_prefill_kernel(
for tile in tl.range(k_tile_start, k_tile_end, num_stages=STAGES):
columns = tile * BLOCK_N + column_offsets
live = columns < max_visible
logical_page = tl.minimum(columns // PAGE_SIZE, PAGE_TABLE_WIDTH - 1)
logical_page = tl.minimum(columns // PAGE_SIZE, page_table_width - 1)
page_offset = columns % PAGE_SIZE
physical_page = tl.load(
page_table_ptr + request * stride_table_req + logical_page,
@@ -202,11 +202,11 @@ def _qsa_mqa_paged_prefill_kernel(
)
scores = tl.dot(keys, query, out_dtype=tl.float32)
scores = tl.reshape(scores, (BLOCK_N, TILE_R, NUM_HEADS_PADDED))
score = tl.sum(tl.maximum(scores, 0.0), axis=2) / HEAD_DIM**0.5
score = tl.sum(tl.maximum(scores, 0.0), axis=2)
store_mask = (
valid_rows[None, :]
& (columns[:, None] < visible[None, :])
& (columns[:, None] < NUM_COLUMNS)
& (columns[:, None] < num_columns)
)
tl.store(
logits_ptr + rows[None, :] * stride_logits_row + columns[:, None],
@@ -372,22 +372,25 @@ def _prefill_logits(
query_start_loc: torch.Tensor,
visible_blocks: torch.Tensor,
max_query_len: int,
logits_width: int,
query_offset: int,
num_queries: int,
) -> torch.Tensor:
assert query_start_loc.shape == (page_table.shape[0] + 1,)
assert visible_blocks.shape == (q.shape[0],)
assert 0 <= query_offset <= query_offset + num_queries <= q.shape[0]
assert 0 < logits_width <= page_table.shape[1] * k_cache.shape[1]
columns = page_table.shape[1] * k_cache.shape[1]
logits = torch.empty((num_queries, columns), dtype=torch.float32, device=q.device)
logits = torch.empty(
(num_queries, logits_width), dtype=torch.float32, device=q.device
)
TILE_R = 64
BLOCK_N = 64
K_TILES = 16
grid = (
page_table.shape[0],
triton.cdiv(min(num_queries, max_query_len), TILE_R),
triton.cdiv(columns, BLOCK_N * K_TILES),
triton.cdiv(logits_width, BLOCK_N * K_TILES),
)
_qsa_mqa_paged_prefill_kernel[grid](
q,
@@ -402,8 +405,8 @@ def _prefill_logits(
*logits.stride()[:-1],
num_queries,
query_offset,
page_table.shape[1],
PAGE_SIZE=k_cache.shape[1],
PAGE_TABLE_WIDTH=page_table.shape[1],
NUM_HEADS=q.shape[1],
HEAD_DIM=q.shape[2],
TILE_R=TILE_R,
@@ -557,6 +560,7 @@ def qsa_select_paged_prefill(
compress_ratio: int,
max_query_len: int,
block_indices: torch.Tensor,
max_seq_len: int,
) -> None:
"""Score and select compressed prefill blocks in bounded chunks.
@@ -572,16 +576,21 @@ def qsa_select_paged_prefill(
compress_ratio: Number of logical tokens represented by a cache row.
max_query_len: Maximum number of query tokens in one request.
block_indices: Compressed-index output buffer.
max_seq_len: Longest context length in the batch this step.
"""
assert token_topk % compress_ratio == 0
assert block_indices.shape == (q.shape[0], token_topk // compress_ratio)
rows = q.shape[0]
columns = page_table.shape[1] * k_cache.shape[1]
# No row scores beyond cdiv(max_seq_len, compress_ratio) compressed
# columns. Round up to 64 to keep the logits row stride
# cooperative_topk-compatible.
logits_width = triton.cdiv(triton.cdiv(max_seq_len, compress_ratio), 64) * 64
logits_width = min(max(64, logits_width), page_table.shape[1] * k_cache.shape[1])
# chunk the inputs to keep temp logits below VLLM_SPARSE_INDEXER_MAX_LOGITS_MB
max_logits_bytes = envs.VLLM_SPARSE_INDEXER_MAX_LOGITS_MB * 1024 * 1024
rows_per_chunk = max(1, max_logits_bytes // (columns * 4))
rows_per_chunk = max(1, max_logits_bytes // (logits_width * 4))
topk_workspace = torch.empty(
(_TOPK_WORKSPACE_BYTES,), dtype=torch.uint8, device=q.device
)
@@ -596,6 +605,7 @@ def qsa_select_paged_prefill(
query_start_loc,
visible_blocks,
max_query_len,
logits_width,
query_offset=query_start,
num_queries=query_end - query_start,
)