[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:
co-authored by
Claude Fable 5
girasoley
OpenAI Codex
parent
f83de6d44c
commit
a4904ba903
@@ -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)
|
||||
|
||||
Reference in New Issue
Block a user