Skip to content
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

Use string based tokenizer in prefix cache #774

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
+37 −48
Open

Use string based tokenizer in prefix cache #774

Show file tree
Hide file tree
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
8 changes: 3 additions & 5 deletions pkg/plugins/gateway/algorithms/prefix_cache.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@ import (
"fmt"
"math/rand"
"strconv"
"strings"

"github.com/vllm-project/aibrix/pkg/plugins/gateway/prefixcacheindexer"
"github.com/vllm-project/aibrix/pkg/utils"
Expand Down Expand Up @@ -72,10 +73,7 @@ func (p prefixCacheRouter) Route(ctx context.Context, pods map[string]*v1.Pod, m
}
}

tokens, err := utils.TokenizeInputText(message)
if err != nil {
return "", err
}
tokens := strings.Split(message, " ")
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we have some interface here so we can probably configure it for differenc scenarios? My concern is I think it's ok to split by " " for english,, for some other languages like chinese, this might not be helpful.


var targetPod *v1.Pod
matchedTokens, unMatchedTokens, matchedPods := p.prefixCacheIndexer.MatchPrefix(tokens, model, readyPods)
Expand All @@ -96,7 +94,7 @@ func (p prefixCacheRouter) Route(ctx context.Context, pods map[string]*v1.Pod, m
for _, p := range readyPods {
readyPodNames = append(readyPodNames, p.Status.PodIP)
}
klog.InfoS("prefix cache route",
klog.V(4).InfoS("prefix cache route",
"message", message,
"tokens", tokens,
"matched_tokens", matchedTokens,
Expand Down
30 changes: 14 additions & 16 deletions pkg/plugins/gateway/prefixcacheindexer/hash.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,8 +17,6 @@ limitations under the License.
package prefixcacheindexer

import (
"bytes"
"encoding/binary"
"math/rand"
"strconv"
"sync"
Expand Down Expand Up @@ -121,7 +119,7 @@ func NewPrefixHashTable() PrefixCacheIndexer {

// returns matchedTokens, unMatchedTokens, matchedPods
// TODO: add an interface with multiple implementations such as hash or radix tree
func (c *PrefixHashTable) MatchPrefix(tokens []int, model string, pods []*v1.Pod) ([]int, []int, []*v1.Pod) {
func (c *PrefixHashTable) MatchPrefix(tokens []string, model string, pods []*v1.Pod) ([]string, []string, []*v1.Pod) {
c.mu.RLock()
defer c.mu.RUnlock()
var block, lastMatchedBlock Block
Expand All @@ -134,8 +132,9 @@ func (c *PrefixHashTable) MatchPrefix(tokens []int, model string, pods []*v1.Pod
end = len(tokens)
}

chunk := tokens[i:end]
_, _ = c.hash.Write(IntArrayToByteArray(chunk))
for _, b := range stringArrayToByteArray(tokens[i:end]) {
_, _ = c.hash.Write(b)
}
prefixHash := c.hash.Sum64()
c.hash.ResetWithSeed(c.seed)
block, ok = c.blocks[prefixHash]
Expand Down Expand Up @@ -164,7 +163,7 @@ func (c *PrefixHashTable) MatchPrefix(tokens []int, model string, pods []*v1.Pod
return matchedTokens, unMatchedTokens, matchedPods
}

func (c *PrefixHashTable) AddPrefix(unMatchedTokens []int, model, pod string) {
func (c *PrefixHashTable) AddPrefix(unMatchedTokens []string, model, pod string) {
c.mu.Lock()
defer c.mu.Unlock()

Expand All @@ -174,8 +173,9 @@ func (c *PrefixHashTable) AddPrefix(unMatchedTokens []int, model, pod string) {
end = len(unMatchedTokens)
}

chunk := unMatchedTokens[i:end]
_, _ = c.hash.Write(IntArrayToByteArray(chunk))
for _, b := range stringArrayToByteArray(unMatchedTokens[i:end]) {
_, _ = c.hash.Write(b)
}
prefixHash := c.hash.Sum64()
c.hash.ResetWithSeed(c.seed)
block, ok := c.blocks[prefixHash]
Expand Down Expand Up @@ -209,13 +209,11 @@ func (c *PrefixHashTable) Evict(now time.Time) {
}
}

func IntArrayToByteArray(intArray []int) []byte {
buf := new(bytes.Buffer)
for _, val := range intArray {
err := binary.Write(buf, binary.LittleEndian, int32(val))
if err != nil {
panic(err)
}
// stringArrayToByteArray converts a string array to a byte array.
func stringArrayToByteArray(stringArray []string) [][]byte {
byteArray := make([][]byte, len(stringArray))
for i, str := range stringArray {
byteArray[i] = []byte(str)
}
return buf.Bytes()
return byteArray
}
43 changes: 18 additions & 25 deletions pkg/plugins/gateway/prefixcacheindexer/hash_test.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,14 +17,13 @@ limitations under the License.
package prefixcacheindexer

import (
"fmt"
"math/rand"
"strings"
"testing"
"time"

"github.com/cespare/xxhash/v2"
"github.com/stretchr/testify/assert"
"github.com/vllm-project/aibrix/pkg/utils"
v1 "k8s.io/api/core/v1"
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
)
Expand All @@ -42,29 +41,28 @@ func Test_PrefixHashTableE2E(t *testing.T) {
{ObjectMeta: metav1.ObjectMeta{Name: "p2"}},
}

inputText := "Hello World! What a Good Day! Good Morning! 你好世界！多么美好的一天啊！早上好！"
tokens, err := utils.TokenizeInputText(inputText)
assert.Equal(t, nil, err)
inputText := "Hello World! What a Good Day! Good day to code and learn new things in LLM!! 你好世界！ 多么美好的一天啊！"
tokens := strings.Split(inputText, " ")

matchedTokens, unMatchedTokens, matchPods := cache.MatchPrefix(tokens, "m1", pods)
assert.Equal(t, 0, len(matchedTokens))
assert.Equal(t,
[]int{9906, 4435, 0, 3639, 264, 7839, 6187, 0, 7839, 29084, 0, 220, 57668, 53901, 3574, 244, 98220, 6447, 43240, 82696, 58666, 53901, 9554, 15120, 36827, 28308, 232, 6447, 6079, 102, 17905, 53901, 6447},
[]string{"Hello", "World!", "What", "a", "Good", "Day!", "Good", "day", "to", "code", "and", "learn", "new", "things", "in", "LLM!!", "你好世界！", "多么美好的一天啊！"},
unMatchedTokens)
assert.Equal(t, 0, len(matchPods))

cache.AddPrefix(unMatchedTokens, "m1", "p1")
matchedTokens, unMatchedTokens, matchPods = cache.MatchPrefix(tokens, "m1", pods)
assert.Equal(t,
[]int{9906, 4435, 0, 3639, 264, 7839, 6187, 0, 7839, 29084, 0, 220, 57668, 53901, 3574, 244, 98220, 6447, 43240, 82696, 58666, 53901, 9554, 15120, 36827, 28308, 232, 6447, 6079, 102, 17905, 53901, 6447},
[]string{"Hello", "World!", "What", "a", "Good", "Day!", "Good", "day", "to", "code", "and", "learn", "new", "things", "in", "LLM!!", "你好世界！", "多么美好的一天啊！"},
matchedTokens)
assert.Equal(t, 0, len(unMatchedTokens))
assert.Equal(t, "p1", matchPods[0].Name)

cache.Evict(time.Now().Add(60 * time.Minute))
_, unMatchedTokens, matchPods = cache.MatchPrefix(tokens, "m1", pods)
assert.Equal(t,
[]int{9906, 4435, 0, 3639, 264, 7839, 6187, 0, 7839, 29084, 0, 220, 57668, 53901, 3574, 244, 98220, 6447, 43240, 82696, 58666, 53901, 9554, 15120, 36827, 28308, 232, 6447, 6079, 102, 17905, 53901, 6447},
[]string{"Hello", "World!", "What", "a", "Good", "Day!", "Good", "day", "to", "code", "and", "learn", "new", "things", "in", "LLM!!", "你好世界！", "多么美好的一天啊！"},
unMatchedTokens)
assert.Equal(t, 0, len(matchPods))
}
Expand All @@ -78,8 +76,8 @@ func Test_MatchPrefix(t *testing.T) {
cache PrefixHashTable
model string
pods []*v1.Pod
matchTokens []int
unMatchTokens []int
matchTokens []string
unMatchTokens []string
matchPods []*v1.Pod
}{
{
Expand All @@ -95,16 +93,16 @@ func Test_MatchPrefix(t *testing.T) {
{ObjectMeta: metav1.ObjectMeta{Name: "p1"}},
{ObjectMeta: metav1.ObjectMeta{Name: "p2"}},
},
matchTokens: []int{},
unMatchTokens: []int{9906, 4435, 0, 3639, 264, 7839, 6187, 0, 220, 57668, 53901, 3574, 244, 98220, 6447, 43240, 82696, 58666, 53901, 9554, 15120, 36827, 28308, 232, 6447},
matchTokens: []string{},
unMatchTokens: []string{"Hello", "World!", "What", "a", "Good", "Day!", "你好世界！多么美好的一天啊！"},
matchPods: nil,
},
{
name: "token length more than prefix block size, one prefix block exist in the cache",
inputText: "Hello World! What a Good Day! 你好世界！多么美好的一天啊！",
inputText: "Hello World! What a Good Day! Good day to code and learn new things in LLM!! 你好世界！多么美好的一天啊！",
cache: PrefixHashTable{
blocks: map[uint64]Block{
8954089069687757318: {
8439316938363978324: {
modelToPods: map[string]map[string]time.Time{
"m1": {
"p1": time.Now(),
Expand All @@ -121,24 +119,19 @@ func Test_MatchPrefix(t *testing.T) {
{ObjectMeta: metav1.ObjectMeta{Name: "p1"}},
{ObjectMeta: metav1.ObjectMeta{Name: "p2"}},
},
matchTokens: []int{9906, 4435, 0, 3639, 264, 7839, 6187, 0, 220, 57668, 53901, 3574, 244, 98220, 6447, 43240},
unMatchTokens: []int{82696, 58666, 53901, 9554, 15120, 36827, 28308, 232, 6447},
matchTokens: []string{"Hello", "World!", "What", "a", "Good", "Day!", "Good", "day", "to", "code", "and", "learn", "new", "things", "in", "LLM!!"},
unMatchTokens: []string{"你好世界！多么美好的一天啊！"},
matchPods: []*v1.Pod{
{ObjectMeta: metav1.ObjectMeta{Name: "p1"}},
},
},
}

for _, tt := range tests {
tokens, err := utils.TokenizeInputText(tt.inputText)
assert.Equal(t, nil, err)
fmt.Println(len(tokens))
fmt.Println(tokens)

matchTokens, unMatchTokens, matchPods := tt.cache.MatchPrefix(tokens, tt.model, tt.pods)
matchTokens, unMatchTokens, matchPods := tt.cache.MatchPrefix(strings.Split(tt.inputText, " "), tt.model, tt.pods)

assert.Equal(t, tt.matchTokens, matchTokens)
assert.Equal(t, tt.unMatchTokens, unMatchTokens)
assert.Equal(t, tt.matchPods, matchPods)
assert.Equal(t, tt.matchTokens, matchTokens, tt.name)
assert.Equal(t, tt.unMatchTokens, unMatchTokens, tt.name)
assert.Equal(t, tt.matchPods, matchPods, tt.name)
}
}
4 changes: 2 additions & 2 deletions pkg/plugins/gateway/prefixcacheindexer/indexer.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,10 +25,10 @@ import (
type PrefixCacheIndexer interface {
// MatchPrefix matches the longest prefix sequence for input request (passed as input tokens)
// and returns matched prefix (as tokens), remaining unmatched input request (as tokens) and pods matching the prefix
MatchPrefix(inputTokens []int, model string, pods []*v1.Pod) (matchedTokens []int, unMatchedTokens []int, matchedPods []*v1.Pod)
MatchPrefix(inputTokens []string, model string, pods []*v1.Pod) (matchedTokens []string, unMatchedTokens []string, matchedPods []*v1.Pod)

// AddPrefix adds tokens in internal prefix cache indexer to be used by future requests
AddPrefix(tokens []int, model, pod string)
AddPrefix(tokens []string, model, pod string)

// Evict is invoked at fixed internal to clean up expired tokens from prefix cache.
// TODO: Add max blocks to cache, add LRU policy along with TTL and add performance benchmark tests.
Expand Down
Loading