Alanxtl commented on code in PR #878: URL: https://github.com/apache/dubbo-go-pixiu/pull/878#discussion_r2779145903
########## pkg/filter/ai/kvcache/config.go: ########## @@ -0,0 +1,150 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package kvcache + +import ( + "fmt" + "strings" + "time" +) + +type Config struct { + Enabled bool `yaml:"enabled" json:"enabled" mapstructure:"enabled"` + VLLMEndpoint string `yaml:"vllm_endpoint" json:"vllm_endpoint" mapstructure:"vllm_endpoint"` + LMCacheEndpoint string `yaml:"lmcache_endpoint" json:"lmcache_endpoint" mapstructure:"lmcache_endpoint"` + DefaultModel string `yaml:"default_model" json:"default_model" mapstructure:"default_model"` + RequestTimeout time.Duration `yaml:"request_timeout" json:"request_timeout" mapstructure:"request_timeout"` + LookupRoutingTimeout time.Duration `yaml:"lookup_routing_timeout" json:"lookup_routing_timeout" mapstructure:"lookup_routing_timeout"` + HotWindow time.Duration `yaml:"hot_window" json:"hot_window" mapstructure:"hot_window"` + HotMaxRecords int `yaml:"hot_max_records" json:"hot_max_records" mapstructure:"hot_max_records"` + MaxIdleConns int `yaml:"max_idle_conns" json:"max_idle_conns" mapstructure:"max_idle_conns"` + MaxIdleConnsPerHost int `yaml:"max_idle_conns_per_host" json:"max_idle_conns_per_host" mapstructure:"max_idle_conns_per_host"` + MaxConnsPerHost int `yaml:"max_conns_per_host" json:"max_conns_per_host" mapstructure:"max_conns_per_host"` + TokenCache TokenCacheConfig `yaml:"token_cache" json:"token_cache" mapstructure:"token_cache"` + CacheStrategy CacheStrategyConfig `yaml:"cache_strategy" json:"cache_strategy" mapstructure:"cache_strategy"` + CircuitBreaker CircuitBreakerConfig `yaml:"circuit_breaker" json:"circuit_breaker" mapstructure:"circuit_breaker"` + Retry RetryConfig `yaml:"retry" json:"retry" mapstructure:"retry"` +} + +type TokenCacheConfig struct { + MaxSize int `yaml:"max_size" json:"max_size" mapstructure:"max_size"` + TTL time.Duration `yaml:"ttl" json:"ttl" mapstructure:"ttl"` + Enabled bool `yaml:"enabled" json:"enabled" mapstructure:"enabled"` +} + +type CacheStrategyConfig struct { + EnableCompression bool `yaml:"enable_compression" json:"enable_compression" mapstructure:"enable_compression"` + EnablePinning bool `yaml:"enable_pinning" json:"enable_pinning" mapstructure:"enable_pinning"` + EnableEviction bool `yaml:"enable_eviction" json:"enable_eviction" mapstructure:"enable_eviction"` + MemoryThreshold float64 `yaml:"memory_threshold" json:"memory_threshold" mapstructure:"memory_threshold"` + HotContentThreshold int `yaml:"hot_content_threshold" json:"hot_content_threshold" mapstructure:"hot_content_threshold"` + LoadThreshold float64 `yaml:"load_threshold" json:"load_threshold" mapstructure:"load_threshold"` + PinInstanceID string `yaml:"pin_instance_id" json:"pin_instance_id" mapstructure:"pin_instance_id"` + PinLocation string `yaml:"pin_location" json:"pin_location" mapstructure:"pin_location"` + CompressInstanceID string `yaml:"compress_instance_id" json:"compress_instance_id" mapstructure:"compress_instance_id"` + CompressLocation string `yaml:"compress_location" json:"compress_location" mapstructure:"compress_location"` + CompressMethod string `yaml:"compress_method" json:"compress_method" mapstructure:"compress_method"` + EvictInstanceID string `yaml:"evict_instance_id" json:"evict_instance_id" mapstructure:"evict_instance_id"` +} + +type RetryConfig struct { + MaxAttempts int `yaml:"max_attempts" json:"max_attempts" mapstructure:"max_attempts"` + BaseBackoff time.Duration `yaml:"base_backoff" json:"base_backoff" mapstructure:"base_backoff"` + MaxBackoff time.Duration `yaml:"max_backoff" json:"max_backoff" mapstructure:"max_backoff"` +} + +func (c *Config) Validate() error { + if !c.Enabled { + return nil + } + if strings.TrimSpace(c.VLLMEndpoint) == "" { + return fmt.Errorf("kvcache: vllm_endpoint is required when enabled") + } + if strings.TrimSpace(c.LMCacheEndpoint) == "" { + return fmt.Errorf("kvcache: lmcache_endpoint is required when enabled") + } + if c.TokenCache.MaxSize < 0 { + return fmt.Errorf("kvcache: token_cache.max_size must be >= 0") + } + if c.CacheStrategy.MemoryThreshold < 0 || c.CacheStrategy.MemoryThreshold > 1 { + return fmt.Errorf("kvcache: cache_strategy.memory_threshold must be between 0 and 1") + } + if c.CacheStrategy.LoadThreshold < 0 || c.CacheStrategy.LoadThreshold > 1 { + return fmt.Errorf("kvcache: cache_strategy.load_threshold must be between 0 and 1") + } + if c.CacheStrategy.HotContentThreshold < 0 { + return fmt.Errorf("kvcache: cache_strategy.hot_content_threshold must be >= 0") + } + if c.Retry.MaxAttempts < 0 { + return fmt.Errorf("kvcache: retry.max_attempts must be >= 0") + } + if c.Retry.BaseBackoff < 0 || c.Retry.MaxBackoff < 0 { + return fmt.Errorf("kvcache: retry backoff durations must be >= 0") + } + if c.HotWindow < 0 { + return fmt.Errorf("kvcache: hot_window must be >= 0") + } + if c.HotMaxRecords < 0 { + return fmt.Errorf("kvcache: hot_max_records must be >= 0") + } + return nil +} + +func (c *Config) ApplyDefaults() { + if c.RequestTimeout <= 0 { + c.RequestTimeout = 2 * time.Second Review Comment: 把常量都抽出去 ########## pkg/filter/llm/proxy/filter.go: ########## @@ -98,6 +98,18 @@ type ( } ) +func getPreferredEndpointID(hc *contexthttp.HttpContext) string { + if hc == nil || hc.Params == nil { + return "" + } + val, ok := hc.Params[constant.LLMPreferredEndpointID] + if !ok { + return "" + } + endpointID, _ := val.(string) + return endpointID Review Comment: ```suggestion if val, ok := hc.Params[constant.LLMPreferredEndpointID]; !ok { return "" } if val, ok := val.(string); ok { return val } return endpointID ``` ########## pkg/filter/ai/kvcache/circuit_breaker.go: ########## @@ -0,0 +1,106 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package kvcache + +import ( + "errors" + "sync" + "time" +) + +var ErrCircuitBreakerOpen = errors.New("kvcache circuit breaker open") + +type CircuitState int + +const ( + CircuitClosed CircuitState = iota Review Comment: not suggest using iota ########## pkg/common/constant/key.go: ########## @@ -59,8 +59,11 @@ const ( LLMProxyFilter = "dgp.filter.llm.proxy" LLMTokenizerFilter = "dgp.filter.llm.tokenizer" + AIKVCacheFilter = "dgp.filter.ai.kvcache" MCPServerFilter = "dgp.filter.mcp.mcpserver" + + LLMPreferredEndpointID = "llm_preferred_endpoint_id" Review Comment: 这个常量放到这里不合适吧 ########## pkg/filter/ai/kvcache/config.go: ########## @@ -0,0 +1,150 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package kvcache + +import ( + "fmt" + "strings" + "time" +) + +type Config struct { + Enabled bool `yaml:"enabled" json:"enabled" mapstructure:"enabled"` + VLLMEndpoint string `yaml:"vllm_endpoint" json:"vllm_endpoint" mapstructure:"vllm_endpoint"` + LMCacheEndpoint string `yaml:"lmcache_endpoint" json:"lmcache_endpoint" mapstructure:"lmcache_endpoint"` + DefaultModel string `yaml:"default_model" json:"default_model" mapstructure:"default_model"` + RequestTimeout time.Duration `yaml:"request_timeout" json:"request_timeout" mapstructure:"request_timeout"` + LookupRoutingTimeout time.Duration `yaml:"lookup_routing_timeout" json:"lookup_routing_timeout" mapstructure:"lookup_routing_timeout"` + HotWindow time.Duration `yaml:"hot_window" json:"hot_window" mapstructure:"hot_window"` + HotMaxRecords int `yaml:"hot_max_records" json:"hot_max_records" mapstructure:"hot_max_records"` + MaxIdleConns int `yaml:"max_idle_conns" json:"max_idle_conns" mapstructure:"max_idle_conns"` + MaxIdleConnsPerHost int `yaml:"max_idle_conns_per_host" json:"max_idle_conns_per_host" mapstructure:"max_idle_conns_per_host"` + MaxConnsPerHost int `yaml:"max_conns_per_host" json:"max_conns_per_host" mapstructure:"max_conns_per_host"` + TokenCache TokenCacheConfig `yaml:"token_cache" json:"token_cache" mapstructure:"token_cache"` + CacheStrategy CacheStrategyConfig `yaml:"cache_strategy" json:"cache_strategy" mapstructure:"cache_strategy"` + CircuitBreaker CircuitBreakerConfig `yaml:"circuit_breaker" json:"circuit_breaker" mapstructure:"circuit_breaker"` + Retry RetryConfig `yaml:"retry" json:"retry" mapstructure:"retry"` +} + +type TokenCacheConfig struct { + MaxSize int `yaml:"max_size" json:"max_size" mapstructure:"max_size"` + TTL time.Duration `yaml:"ttl" json:"ttl" mapstructure:"ttl"` + Enabled bool `yaml:"enabled" json:"enabled" mapstructure:"enabled"` +} + +type CacheStrategyConfig struct { + EnableCompression bool `yaml:"enable_compression" json:"enable_compression" mapstructure:"enable_compression"` + EnablePinning bool `yaml:"enable_pinning" json:"enable_pinning" mapstructure:"enable_pinning"` + EnableEviction bool `yaml:"enable_eviction" json:"enable_eviction" mapstructure:"enable_eviction"` + MemoryThreshold float64 `yaml:"memory_threshold" json:"memory_threshold" mapstructure:"memory_threshold"` + HotContentThreshold int `yaml:"hot_content_threshold" json:"hot_content_threshold" mapstructure:"hot_content_threshold"` + LoadThreshold float64 `yaml:"load_threshold" json:"load_threshold" mapstructure:"load_threshold"` + PinInstanceID string `yaml:"pin_instance_id" json:"pin_instance_id" mapstructure:"pin_instance_id"` + PinLocation string `yaml:"pin_location" json:"pin_location" mapstructure:"pin_location"` + CompressInstanceID string `yaml:"compress_instance_id" json:"compress_instance_id" mapstructure:"compress_instance_id"` + CompressLocation string `yaml:"compress_location" json:"compress_location" mapstructure:"compress_location"` + CompressMethod string `yaml:"compress_method" json:"compress_method" mapstructure:"compress_method"` + EvictInstanceID string `yaml:"evict_instance_id" json:"evict_instance_id" mapstructure:"evict_instance_id"` +} + +type RetryConfig struct { + MaxAttempts int `yaml:"max_attempts" json:"max_attempts" mapstructure:"max_attempts"` + BaseBackoff time.Duration `yaml:"base_backoff" json:"base_backoff" mapstructure:"base_backoff"` + MaxBackoff time.Duration `yaml:"max_backoff" json:"max_backoff" mapstructure:"max_backoff"` +} Review Comment: 这个retry能不能复用pixiu已有的retry呀 ########## pkg/filter/ai/kvcache/config.go: ########## @@ -0,0 +1,150 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package kvcache + +import ( + "fmt" + "strings" + "time" +) + +type Config struct { + Enabled bool `yaml:"enabled" json:"enabled" mapstructure:"enabled"` + VLLMEndpoint string `yaml:"vllm_endpoint" json:"vllm_endpoint" mapstructure:"vllm_endpoint"` + LMCacheEndpoint string `yaml:"lmcache_endpoint" json:"lmcache_endpoint" mapstructure:"lmcache_endpoint"` + DefaultModel string `yaml:"default_model" json:"default_model" mapstructure:"default_model"` + RequestTimeout time.Duration `yaml:"request_timeout" json:"request_timeout" mapstructure:"request_timeout"` + LookupRoutingTimeout time.Duration `yaml:"lookup_routing_timeout" json:"lookup_routing_timeout" mapstructure:"lookup_routing_timeout"` + HotWindow time.Duration `yaml:"hot_window" json:"hot_window" mapstructure:"hot_window"` + HotMaxRecords int `yaml:"hot_max_records" json:"hot_max_records" mapstructure:"hot_max_records"` + MaxIdleConns int `yaml:"max_idle_conns" json:"max_idle_conns" mapstructure:"max_idle_conns"` + MaxIdleConnsPerHost int `yaml:"max_idle_conns_per_host" json:"max_idle_conns_per_host" mapstructure:"max_idle_conns_per_host"` + MaxConnsPerHost int `yaml:"max_conns_per_host" json:"max_conns_per_host" mapstructure:"max_conns_per_host"` + TokenCache TokenCacheConfig `yaml:"token_cache" json:"token_cache" mapstructure:"token_cache"` + CacheStrategy CacheStrategyConfig `yaml:"cache_strategy" json:"cache_strategy" mapstructure:"cache_strategy"` + CircuitBreaker CircuitBreakerConfig `yaml:"circuit_breaker" json:"circuit_breaker" mapstructure:"circuit_breaker"` + Retry RetryConfig `yaml:"retry" json:"retry" mapstructure:"retry"` +} + +type TokenCacheConfig struct { + MaxSize int `yaml:"max_size" json:"max_size" mapstructure:"max_size"` + TTL time.Duration `yaml:"ttl" json:"ttl" mapstructure:"ttl"` + Enabled bool `yaml:"enabled" json:"enabled" mapstructure:"enabled"` +} + +type CacheStrategyConfig struct { + EnableCompression bool `yaml:"enable_compression" json:"enable_compression" mapstructure:"enable_compression"` + EnablePinning bool `yaml:"enable_pinning" json:"enable_pinning" mapstructure:"enable_pinning"` + EnableEviction bool `yaml:"enable_eviction" json:"enable_eviction" mapstructure:"enable_eviction"` + MemoryThreshold float64 `yaml:"memory_threshold" json:"memory_threshold" mapstructure:"memory_threshold"` + HotContentThreshold int `yaml:"hot_content_threshold" json:"hot_content_threshold" mapstructure:"hot_content_threshold"` + LoadThreshold float64 `yaml:"load_threshold" json:"load_threshold" mapstructure:"load_threshold"` + PinInstanceID string `yaml:"pin_instance_id" json:"pin_instance_id" mapstructure:"pin_instance_id"` + PinLocation string `yaml:"pin_location" json:"pin_location" mapstructure:"pin_location"` + CompressInstanceID string `yaml:"compress_instance_id" json:"compress_instance_id" mapstructure:"compress_instance_id"` + CompressLocation string `yaml:"compress_location" json:"compress_location" mapstructure:"compress_location"` + CompressMethod string `yaml:"compress_method" json:"compress_method" mapstructure:"compress_method"` + EvictInstanceID string `yaml:"evict_instance_id" json:"evict_instance_id" mapstructure:"evict_instance_id"` +} + +type RetryConfig struct { + MaxAttempts int `yaml:"max_attempts" json:"max_attempts" mapstructure:"max_attempts"` + BaseBackoff time.Duration `yaml:"base_backoff" json:"base_backoff" mapstructure:"base_backoff"` + MaxBackoff time.Duration `yaml:"max_backoff" json:"max_backoff" mapstructure:"max_backoff"` +} + +func (c *Config) Validate() error { + if !c.Enabled { + return nil + } + if strings.TrimSpace(c.VLLMEndpoint) == "" { + return fmt.Errorf("kvcache: vllm_endpoint is required when enabled") Review Comment: 把error的格式都改成这样的格式 ``` [kvcache] vllm_endpoint is required when enabled ``` ########## pkg/filter/ai/kvcache/token_manager.go: ########## @@ -0,0 +1,305 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package kvcache + +import ( + "context" + "crypto/sha256" + "encoding/hex" + "encoding/json" + "fmt" + "strings" + "sync" + "sync/atomic" + "time" +) + +import ( + "github.com/go-resty/resty/v2" +) + +type TokenManager struct { + httpClient *resty.Client + endpoint string + cache sync.Map + config TokenCacheConfig + circuitBreaker *CircuitBreaker + + cacheSize int64 + hitCount int64 + missCount int64 + + hotWindow time.Duration + hotMax int + hotMu sync.Mutex + hotMap map[string][]time.Time +} + +type TokenizeRequest struct { + Model string `json:"model,omitempty"` + Prompt string `json:"prompt"` +} + +type TokenizeResponse struct { + Count int `json:"count"` + Tokens []int `json:"tokens"` + MaxLen int `json:"max_model_len"` +} + +type tokenCacheEntry struct { + tokens []int + expiresAt time.Time +} + +func NewTokenManager(endpoint string, httpClient *resty.Client, cfg TokenCacheConfig, cb *CircuitBreaker, hotWindow time.Duration, hotMax int) *TokenManager { + return &TokenManager{ + httpClient: httpClient, + endpoint: endpoint, + config: cfg, + circuitBreaker: cb, + hotWindow: hotWindow, + hotMax: hotMax, + hotMap: make(map[string][]time.Time), + } +} + +func (tm *TokenManager) GetTokens(ctx context.Context, model string, prompt string, rawBody []byte) ([]int, error) { + cacheKey := tm.cacheKey(model, prompt) + if tm.config.Enabled { + if tokens, ok := tm.loadCache(cacheKey); ok { + atomic.AddInt64(&tm.hitCount, 1) + return tokens, nil + } + atomic.AddInt64(&tm.missCount, 1) + } + + tokens, err := tm.tokenize(ctx, model, prompt, rawBody) + if err != nil { + return nil, err + } + + if tm.config.Enabled { + tm.storeCache(cacheKey, tokens) + } + return tokens, nil +} + +func (tm *TokenManager) GetCachedTokens(model string, prompt string) ([]int, bool) { + if !tm.config.Enabled { + return nil, false + } + cacheKey := tm.cacheKey(model, prompt) + tokens, ok := tm.loadCache(cacheKey) + if ok { + atomic.AddInt64(&tm.hitCount, 1) + } else { + atomic.AddInt64(&tm.missCount, 1) + } + return tokens, ok +} + +func (tm *TokenManager) InvalidateCache(model string, prompt string) { + cacheKey := tm.cacheKey(model, prompt) + tm.deleteCache(cacheKey) +} + +func (tm *TokenManager) GetCacheStats() CacheStats { + size := atomic.LoadInt64(&tm.cacheSize) + hit := atomic.LoadInt64(&tm.hitCount) + miss := atomic.LoadInt64(&tm.missCount) + total := hit + miss + var hitRate float64 + if total > 0 { + hitRate = float64(hit) / float64(total) + } + return CacheStats{ + Size: int(size), + HitRate: hitRate, + HitCount: hit, + MissCount: miss, + } +} + +func (tm *TokenManager) tokenize(ctx context.Context, model string, prompt string, rawBody []byte) ([]int, error) { + var tokens []int + err := tm.execute(ctx, func() error { + body, err := tm.buildTokenizeBody(model, prompt, rawBody) + if err != nil { + return err + } + resp, err := tm.doTokenizeRequest(ctx, body) + if err != nil { + return err + } + tokens = resp.Tokens + return nil + }) + if err != nil { + return nil, err + } + return tokens, nil +} + +func (tm *TokenManager) buildTokenizeBody(model string, prompt string, rawBody []byte) (any, error) { + if len(rawBody) > 0 { + return rawBody, nil + } + return TokenizeRequest{Model: model, Prompt: prompt}, nil +} + +func (tm *TokenManager) doTokenizeRequest(ctx context.Context, body any) (*TokenizeResponse, error) { Review Comment: 这个Tokenize是什么概念呀 -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
