[Perf][KVConnector][Mooncake] Vectorize prepare_value on the KV load path (#48531)

Signed-off-by: girasoley <girasoley@inferact.ai>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: girasoley <girasoley@inferact.ai>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Summer Yang
2026-07-23 18:12:00 +08:00
committed by GitHub
co-authored by Claude Fable 5 girasoley OpenAI Codex
parent f83de6d44c
commit a4904ba903
4 changed files with 211 additions and 21 deletions
@@ -0,0 +1,90 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Tests for ChunkedTokenDatabase.prepare_values."""
import random
import pytest
from vllm.distributed.kv_transfer.kv_connector.v1.mooncake.store.data import (
ChunkedTokenDatabase,
KeyMetadata,
)
from vllm.utils.math_utils import cdiv
BLOCK_SIZE = 128
def _reference_prepare_value(
db: ChunkedTokenDatabase, start: int, end: int, block_ids: list[int]
) -> tuple[list[int], list[int], int]:
"""Compute a token range with the original scalar implementation."""
addr_list = []
size_list = []
block_id = block_ids[start // db.block_size]
length = len(db.block_len)
for index, base_addr in enumerate(db.kv_caches_base_addr):
addr = base_addr + block_id * db.block_len[index % length]
assert (end - start) % db.block_size == 0
size = db.block_len[index % length] * cdiv(end - start, db.block_size)
addr_list.append(addr)
size_list.append(size)
return addr_list, size_list, block_id
def _make_db(num_regions: int, num_block_lens: int) -> ChunkedTokenDatabase:
md = KeyMetadata(model_name="t", tp_rank=1, pcp_rank=0, dcp_rank=0, pp_rank=0)
db = ChunkedTokenDatabase(md, BLOCK_SIZE)
db.set_kv_caches_base_addr(
[0x7F00_0000_0000 + i * (1 << 30) for i in range(num_regions)]
)
# Exercise repeated block lengths when there are more cache regions.
db.set_block_len([30_208 + 512 * i for i in range(num_block_lens)])
return db
@pytest.mark.parametrize("num_regions,num_block_lens", [(96, 96), (96, 2), (1, 1)])
def test_prepare_values_matches_reference(num_regions: int, num_block_lens: int):
db = _make_db(num_regions, num_block_lens)
rng = random.Random(0)
n_blocks = 300
block_ids = [rng.randrange(0, 1 << 20) for _ in range(n_blocks)]
chunks = []
b = 0
while b < n_blocks - 4:
span = rng.choice([1, 1, 1, 2, 4])
chunks.append((b * BLOCK_SIZE, (b + span) * BLOCK_SIZE))
b += span + rng.choice([0, 1])
addrs, sizes, bids = db.prepare_values(chunks, block_ids)
assert len(addrs) == len(sizes) == len(bids) == len(chunks)
for (start, end), addr, size, bid in zip(chunks, addrs, sizes, bids):
ref_addr, ref_size, ref_bid = _reference_prepare_value(
db, start, end, block_ids
)
assert addr == ref_addr
assert size == ref_size
assert bid == ref_bid
# Native bindings require Python ints rather than numpy scalars.
assert all(type(a) is int for a in addr)
assert type(bid) is int
def test_prepare_value_single_matches_reference():
db = _make_db(8, 8)
block_ids = list(range(64))
got = db.prepare_value(5 * BLOCK_SIZE, 7 * BLOCK_SIZE, block_ids)
assert got == _reference_prepare_value(
db, 5 * BLOCK_SIZE, 7 * BLOCK_SIZE, block_ids
)
def test_prepare_values_empty():
db = _make_db(4, 4)
assert db.prepare_values([], [1, 2, 3]) == ([], [], [])
def test_prepare_values_rejects_unaligned_chunk():
db = _make_db(4, 4)
with pytest.raises(AssertionError):
db.prepare_values([(0, BLOCK_SIZE + 1)], [0, 1])
@@ -655,6 +655,68 @@ def test_store_sending_thread_delta_saves_only_new_masked_chunks():
assert masked_hashes == [b"a2".hex()]
def test_store_sending_thread_prepares_missing_chunks_once_per_group():
store = MagicMock()
store.batch_is_exist.return_value = [0, 1, 0, 1, 0, 0]
store.batch_put_from_multi_buffers.return_value = [256, 256, 512, 512]
coord = SimpleNamespace(
lcm_block_size=16,
store_mask=lambda token_len, start_token, num_prompt_tokens=None: (
None,
None,
),
)
db0 = ChunkedTokenDatabase(
KeyMetadata("test-model", 0, 0, 0, 0, group_id=0),
block_size=16,
)
db0.set_kv_caches_base_addr([0x1000])
db0.set_block_len([256])
db0.prepare_values = MagicMock(wraps=db0.prepare_values)
db0.prepare_value = MagicMock(side_effect=AssertionError("scalar path called"))
db1 = ChunkedTokenDatabase(
KeyMetadata("test-model", 0, 0, 0, 0, group_id=1),
block_size=16,
)
db1.set_kv_caches_base_addr([0x2000])
db1.set_block_len([512])
db1.prepare_values = MagicMock(wraps=db1.prepare_values)
db1.prepare_value = MagicMock(side_effect=AssertionError("scalar path called"))
thread = _make_store_sending_thread(
store,
coord=coord,
token_databases=[db0, db1],
)
thread.add_stored_request("req-a")
thread._handle_request(
ReqMeta(
req_id="req-a",
token_len_chunk=48,
block_ids=([0, 1, 2], [2, 1, 0]),
block_hashes=[b"a0", b"a1", b"a2"],
can_save=True,
)
)
db0.prepare_value.assert_not_called()
db1.prepare_value.assert_not_called()
db0.prepare_values.assert_called_once_with([(0, 16), (32, 48)], [0, 1, 2])
db1.prepare_values.assert_called_once_with([(16, 32), (32, 48)], [2, 1, 0])
keys, addrs, sizes, _ = store.batch_put_from_multi_buffers.call_args.args
assert [key.rsplit("@", 1)[-1] for key in keys] == [
"6130",
"6132",
"6131",
"6132",
]
assert addrs == [[0x1000], [0x1200], [0x2200], [0x2000]]
assert sizes == [[256], [256], [512], [512]]
def test_store_sending_thread_only_skips_on_no_available_handle():
store = MagicMock()
store.batch_is_exist.side_effect = lambda keys: [0] * len(keys)
@@ -9,6 +9,7 @@ from collections.abc import Iterable, Sequence
from dataclasses import dataclass
from typing import cast
import numpy as np
import torch
from vllm.distributed.kv_transfer.kv_connector.v1.base import (
@@ -196,22 +197,46 @@ class ChunkedTokenDatabase:
def prepare_value(
self, start: int, end: int, block_ids: list[int]
) -> tuple[list[int], list[int], int]:
"""Compute memory addresses and sizes for a token range.
"""Compute memory addresses and sizes for a single token range.
Returns:
(addr_list, size_list, block_id)
"""
addr_list = []
size_list = []
block_id = block_ids[start // self.block_size]
addr_lists, size_lists, chunk_block_ids = self.prepare_values(
((start, end),), block_ids
)
return addr_lists[0], size_lists[0], chunk_block_ids[0]
def prepare_values(
self,
chunks: Sequence[tuple[int, int]],
block_ids: list[int],
) -> tuple[list[list[int]], list[list[int]], list[int]]:
"""Compute memory addresses and sizes for multiple token ranges.
Returns:
(addr_lists, size_lists, chunk_block_ids), one entry per chunk.
"""
if not chunks:
return [], [], []
base = np.asarray(self.kv_caches_base_addr, dtype=np.int64)
length = len(self.block_len)
for index, base_addr in enumerate(self.kv_caches_base_addr):
addr = base_addr + block_id * self.block_len[index % length]
assert (end - start) % self.block_size == 0
size = self.block_len[index % length] * cdiv(end - start, self.block_size)
addr_list.append(addr)
size_list.append(size)
return addr_list, size_list, block_id
blen = np.asarray(
[self.block_len[i % length] for i in range(base.shape[0])],
dtype=np.int64,
)
n = len(chunks)
starts = np.fromiter((c[0] for c in chunks), dtype=np.int64, count=n)
spans = np.fromiter((c[1] for c in chunks), dtype=np.int64, count=n) - starts
assert not (spans % self.block_size).any()
bids = np.fromiter(
(block_ids[i] for i in (starts // self.block_size).tolist()),
dtype=np.int64,
count=n,
)
addrs = base[None, :] + bids[:, None] * blen[None, :]
sizes = blen[None, :] * (spans // self.block_size)[:, None]
return addrs.tolist(), sizes.tolist(), bids.tolist()
def process_tokens(
self,
@@ -634,6 +634,21 @@ class KVCacheStoreSendingThread(KVTransferThread):
addrs: list[list[int]] = []
sizes: list[list[int]] = []
stored_events: list[BlockStored] = []
chunks_per_group: list[list[tuple[int, int]]] = [
[] for _ in self.token_databases
]
for start, end, g_idx in zip(starts, ends, group_indices, strict=True):
chunks_per_group[g_idx].append((start, end))
for g_idx, chunks in enumerate(chunks_per_group):
if not chunks:
continue
db = self.token_databases[g_idx]
group_addrs, group_sizes, _ = db.prepare_values(
chunks, block_ids_per_group[g_idx]
)
addrs.extend(group_addrs)
sizes.extend(group_sizes)
# parent_block_hash chains live within a group, not across.
if self.enable_kv_event:
prev_key_per_group: dict[int, Any] = {}
@@ -645,10 +660,6 @@ class KVCacheStoreSendingThread(KVTransferThread):
zip(starts, ends, group_indices, strict=True)
):
db = self.token_databases[g_idx]
addr, size, _ = db.prepare_value(s, e, block_ids_per_group[g_idx])
addrs.append(addr)
sizes.append(size)
if self.enable_kv_event:
token_ids = (
req_meta.token_ids[s:e]
@@ -805,19 +816,21 @@ class KVCacheStoreRecvingThread(KVTransferThread):
block_id_list: list[int] = []
for g_idx, db in enumerate(self.token_databases):
mask = load_mask_per_group[g_idx]
chunks: list[tuple[int, int]] = []
for start, end, block_hash in db.process_tokens(
token_len, req_meta.block_hashes, mask_num
):
chunk_idx = start // db.block_size
if chunk_idx >= len(mask) or not mask[chunk_idx]:
continue
addr, size, block_id = db.prepare_value(
start, end, req_meta.block_ids[g_idx]
)
key_list.append(db.key_for(block_hash))
addr_list.append(addr)
size_list.append(size)
block_id_list.append(block_id)
chunks.append((start, end))
g_addrs, g_sizes, g_block_ids = db.prepare_values(
chunks, req_meta.block_ids[g_idx]
)
addr_list.extend(g_addrs)
size_list.extend(g_sizes)
block_id_list.extend(g_block_ids)
# Rotate aligned lists by tp_rank for load balancing.
rotation = self.tp_rank % len(key_list)