Add Peer Model Support (#438)

This PR allows a single llama-swap to be the central proxy for models served by other inference servers. The peer servers can be another llama-swap or any API that supports the /v1/* inference endpoint.

Updates: #433, #299
Closes: #296
This commit is contained in:
Benson Wong
2025-12-27 20:18:06 -08:00
committed by GitHub
parent 9864f9f517
commit 22e098ac8b
15 changed files with 1389 additions and 141 deletions
+137 -90
View File
@@ -50,6 +50,9 @@ type ProxyManager struct {
buildDate string
commit string
version string
// peer proxy see: #296, #433
peerProxy *PeerProxy
}
func New(proxyConfig config.Config) *ProxyManager {
@@ -133,6 +136,12 @@ func New(proxyConfig config.Config) *ProxyManager {
maxMetrics = proxyConfig.MetricsMaxInMemory
}
peerProxy, err := NewPeerProxy(proxyConfig.Peers, proxyLogger)
if err != nil {
proxyLogger.Errorf("Disabling Peering. Failed to create proxy peers: %v", err)
peerProxy = nil
}
pm := &ProxyManager{
config: proxyConfig,
ginEngine: gin.New(),
@@ -151,6 +160,8 @@ func New(proxyConfig config.Config) *ProxyManager {
buildDate: "unknown",
commit: "abcd1234",
version: "0",
peerProxy: peerProxy,
}
// create the process groups
@@ -166,22 +177,29 @@ func New(proxyConfig config.Config) *ProxyManager {
// do it in the background, don't block startup -- not sure if good idea yet
go func() {
discardWriter := &DiscardWriter{}
for _, realModelName := range proxyConfig.Hooks.OnStartup.Preload {
proxyLogger.Infof("Preloading model: %s", realModelName)
processGroup, _, err := pm.swapProcessGroup(realModelName)
for _, preloadModelName := range proxyConfig.Hooks.OnStartup.Preload {
modelID, ok := proxyConfig.RealModelName(preloadModelName)
if !ok {
proxyLogger.Warnf("Preload model %s not found in config", preloadModelName)
continue
}
proxyLogger.Infof("Preloading model: %s", modelID)
processGroup, err := pm.swapProcessGroup(modelID)
if err != nil {
event.Emit(ModelPreloadedEvent{
ModelName: realModelName,
ModelName: modelID,
Success: false,
})
proxyLogger.Errorf("Failed to preload model %s: %v", realModelName, err)
proxyLogger.Errorf("Failed to preload model %s: %v", modelID, err)
continue
} else {
req, _ := http.NewRequest("GET", "/", nil)
processGroup.ProxyRequest(realModelName, discardWriter, req)
processGroup.ProxyRequest(modelID, discardWriter, req)
event.Emit(ModelPreloadedEvent{
ModelName: realModelName,
ModelName: modelID,
Success: true,
})
}
@@ -399,16 +417,10 @@ func (pm *ProxyManager) Shutdown() {
pm.shutdownCancel()
}
func (pm *ProxyManager) swapProcessGroup(requestedModel string) (*ProcessGroup, string, error) {
// de-alias the real model name and get a real one
realModelName, found := pm.config.RealModelName(requestedModel)
if !found {
return nil, realModelName, fmt.Errorf("could not find real modelID for %s", requestedModel)
}
func (pm *ProxyManager) swapProcessGroup(realModelName string) (*ProcessGroup, error) {
processGroup := pm.findGroupByModelName(realModelName)
if processGroup == nil {
return nil, realModelName, fmt.Errorf("could not find process group for model %s", requestedModel)
return nil, fmt.Errorf("could not find process group for model %s", realModelName)
}
if processGroup.exclusive {
@@ -420,54 +432,71 @@ func (pm *ProxyManager) swapProcessGroup(requestedModel string) (*ProcessGroup,
}
}
return processGroup, realModelName, nil
return processGroup, nil
}
func (pm *ProxyManager) listModelsHandler(c *gin.Context) {
data := make([]gin.H, 0, len(pm.config.Models))
createdTime := time.Now().Unix()
newRecord := func(modelId string, modelConfig config.ModelConfig) gin.H {
record := gin.H{
"id": modelId,
"object": "model",
"created": createdTime,
"owned_by": "llama-swap",
}
if name := strings.TrimSpace(modelConfig.Name); name != "" {
record["name"] = name
}
if desc := strings.TrimSpace(modelConfig.Description); desc != "" {
record["description"] = desc
}
// Add metadata if present
if len(modelConfig.Metadata) > 0 {
record["meta"] = gin.H{
"llamaswap": modelConfig.Metadata,
}
}
return record
}
for id, modelConfig := range pm.config.Models {
if modelConfig.Unlisted {
continue
}
newRecord := func(modelId string) gin.H {
record := gin.H{
"id": modelId,
"object": "model",
"created": createdTime,
"owned_by": "llama-swap",
}
if name := strings.TrimSpace(modelConfig.Name); name != "" {
record["name"] = name
}
if desc := strings.TrimSpace(modelConfig.Description); desc != "" {
record["description"] = desc
}
// Add metadata if present
if len(modelConfig.Metadata) > 0 {
record["meta"] = gin.H{
"llamaswap": modelConfig.Metadata,
}
}
return record
}
data = append(data, newRecord(id))
data = append(data, newRecord(id, modelConfig))
// Include aliases
if pm.config.IncludeAliasesInList {
for _, alias := range modelConfig.Aliases {
if alias := strings.TrimSpace(alias); alias != "" {
data = append(data, newRecord(alias))
data = append(data, newRecord(alias, modelConfig))
}
}
}
}
if pm.peerProxy != nil {
for peerID, peer := range pm.peerProxy.ListPeers() {
// add peer models
for _, modelID := range peer.Models {
// Skip unlisted models if not showing them
record := newRecord(modelID, config.ModelConfig{
Name: fmt.Sprintf("%s: %s", peerID, modelID),
Metadata: map[string]any{
"peerID": peerID,
},
})
data = append(data, record)
}
}
}
// Sort by the "id" key
sort.Slice(data, func(i, j int) bool {
si, _ := data[i]["id"].(string)
@@ -506,8 +535,8 @@ func (pm *ProxyManager) findModelInPath(path string) (searchName string, realNam
searchModelName = searchModelName + "/" + part
}
if real, ok := pm.config.RealModelName(searchModelName); ok {
return searchModelName, real, "/" + strings.Join(parts[i+1:], "/"), true
if modelID, ok := pm.config.RealModelName(searchModelName); ok {
return searchModelName, modelID, "/" + strings.Join(parts[i+1:], "/"), true
}
}
@@ -517,23 +546,22 @@ func (pm *ProxyManager) findModelInPath(path string) (searchName string, realNam
func (pm *ProxyManager) proxyToUpstream(c *gin.Context) {
upstreamPath := c.Param("upstreamPath")
searchModelName, modelName, remainingPath, modelFound := pm.findModelInPath(upstreamPath)
searchModelName, modelID, remainingPath, modelFound := pm.findModelInPath(upstreamPath)
if !modelFound {
pm.sendErrorResponse(c, http.StatusBadRequest, "model id required in path")
return
}
// Check if this is exactly a model name with no additional path
// and doesn't end with a trailing slash
// Redirect /upstream/modelname to /upstream/modelname/ for URL consistency.
// This ensures relative URLs in upstream responses resolve correctly and
// provides canonical URL form. Uses 308 for POST/PUT/etc to preserve the
// HTTP method (301 would downgrade to GET).
if remainingPath == "/" && !strings.HasSuffix(upstreamPath, "/") {
// Build new URL with query parameters preserved
newPath := "/upstream/" + searchModelName + "/"
if c.Request.URL.RawQuery != "" {
newPath += "?" + c.Request.URL.RawQuery
}
// Use 308 for non-GET/HEAD requests to preserve method
if c.Request.Method == http.MethodGet || c.Request.Method == http.MethodHead {
c.Redirect(http.StatusMovedPermanently, newPath)
} else {
@@ -542,7 +570,7 @@ func (pm *ProxyManager) proxyToUpstream(c *gin.Context) {
return
}
processGroup, realModelName, err := pm.swapProcessGroup(modelName)
processGroup, err := pm.swapProcessGroup(modelID)
if err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
return
@@ -554,15 +582,15 @@ func (pm *ProxyManager) proxyToUpstream(c *gin.Context) {
// attempt to record metrics if it is a POST request
if pm.metricsMonitor != nil && c.Request.Method == "POST" {
if err := pm.metricsMonitor.wrapHandler(realModelName, c.Writer, c.Request, processGroup.ProxyRequest); err != nil {
if err := pm.metricsMonitor.wrapHandler(modelID, c.Writer, c.Request, processGroup.ProxyRequest); err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying metrics wrapped request: %s", err.Error()))
pm.proxyLogger.Errorf("Error proxying wrapped upstream request for model %s, path=%s", realModelName, originalPath)
pm.proxyLogger.Errorf("Error proxying wrapped upstream request for model %s, path=%s", modelID, originalPath)
return
}
} else {
if err := processGroup.ProxyRequest(realModelName, c.Writer, c.Request); err != nil {
if err := processGroup.ProxyRequest(modelID, c.Writer, c.Request); err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying request: %s", err.Error()))
pm.proxyLogger.Errorf("Error proxying upstream request for model %s, path=%s", realModelName, originalPath)
pm.proxyLogger.Errorf("Error proxying upstream request for model %s, path=%s", modelID, originalPath)
return
}
}
@@ -581,41 +609,54 @@ func (pm *ProxyManager) proxyInferenceHandler(c *gin.Context) {
return
}
realModelName, found := pm.config.RealModelName(requestedModel)
if !found {
pm.sendErrorResponse(c, http.StatusBadRequest, fmt.Sprintf("could not find real modelID for %s", requestedModel))
return
}
// Look for a matching local model first
var nextHandler func(modelID string, w http.ResponseWriter, r *http.Request) error
processGroup, _, err := pm.swapProcessGroup(realModelName)
if err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
return
}
// issue #69 allow custom model names to be sent to upstream
useModelName := pm.config.Models[realModelName].UseModelName
if useModelName != "" {
bodyBytes, err = sjson.SetBytes(bodyBytes, "model", useModelName)
modelID, found := pm.config.RealModelName(requestedModel)
if found {
processGroup, err := pm.swapProcessGroup(modelID)
if err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error rewriting model name in JSON: %s", err.Error()))
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
return
}
}
// issue #174 strip parameters from the JSON body
stripParams, err := pm.config.Models[realModelName].Filters.SanitizedStripParams()
if err != nil { // just log it and continue
pm.proxyLogger.Errorf("Error sanitizing strip params string: %s, %s", pm.config.Models[realModelName].Filters.StripParams, err.Error())
} else {
for _, param := range stripParams {
pm.proxyLogger.Debugf("<%s> stripping param: %s", realModelName, param)
bodyBytes, err = sjson.DeleteBytes(bodyBytes, param)
// issue #69 allow custom model names to be sent to upstream
useModelName := pm.config.Models[modelID].UseModelName
if useModelName != "" {
bodyBytes, err = sjson.SetBytes(bodyBytes, "model", useModelName)
if err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error deleting parameter %s from request", param))
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error rewriting model name in JSON: %s", err.Error()))
return
}
}
// issue #174 strip parameters from the JSON body
stripParams, err := pm.config.Models[modelID].Filters.SanitizedStripParams()
if err != nil { // just log it and continue
pm.proxyLogger.Errorf("Error sanitizing strip params string: %s, %s", pm.config.Models[modelID].Filters.StripParams, err.Error())
} else {
for _, param := range stripParams {
pm.proxyLogger.Debugf("<%s> stripping param: %s", modelID, param)
bodyBytes, err = sjson.DeleteBytes(bodyBytes, param)
if err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error deleting parameter %s from request", param))
return
}
}
}
pm.proxyLogger.Debugf("ProxyManager using local Process for model: %s", requestedModel)
nextHandler = processGroup.ProxyRequest
} else if pm.peerProxy != nil && pm.peerProxy.HasPeerModel(requestedModel) {
pm.proxyLogger.Debugf("ProxyManager using ProxyPeer for model: %s", requestedModel)
modelID = requestedModel
nextHandler = pm.peerProxy.ProxyRequest
}
if nextHandler == nil {
pm.sendErrorResponse(c, http.StatusBadRequest, fmt.Sprintf("could not find suitable inference handler for %s", requestedModel))
return
}
c.Request.Body = io.NopCloser(bytes.NewBuffer(bodyBytes))
@@ -628,19 +669,19 @@ func (pm *ProxyManager) proxyInferenceHandler(c *gin.Context) {
// issue #366 extract values that downstream handlers may need
isStreaming := gjson.GetBytes(bodyBytes, "stream").Bool()
ctx := context.WithValue(c.Request.Context(), proxyCtxKey("streaming"), isStreaming)
ctx = context.WithValue(ctx, proxyCtxKey("model"), realModelName)
ctx = context.WithValue(ctx, proxyCtxKey("model"), modelID)
c.Request = c.Request.WithContext(ctx)
if pm.metricsMonitor != nil && c.Request.Method == "POST" {
if err := pm.metricsMonitor.wrapHandler(realModelName, c.Writer, c.Request, processGroup.ProxyRequest); err != nil {
if err := pm.metricsMonitor.wrapHandler(modelID, c.Writer, c.Request, nextHandler); err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying metrics wrapped request: %s", err.Error()))
pm.proxyLogger.Errorf("Error Proxying Metrics Wrapped Request for processGroup %s and model %s", processGroup.id, realModelName)
pm.proxyLogger.Errorf("Error Proxying Metrics Wrapped Request model %s", modelID)
return
}
} else {
if err := processGroup.ProxyRequest(realModelName, c.Writer, c.Request); err != nil {
if err := nextHandler(modelID, c.Writer, c.Request); err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying request: %s", err.Error()))
pm.proxyLogger.Errorf("Error Proxying Request for processGroup %s and model %s", processGroup.id, realModelName)
pm.proxyLogger.Errorf("Error Proxying Request for model %s", modelID)
return
}
}
@@ -660,7 +701,13 @@ func (pm *ProxyManager) proxyOAIPostFormHandler(c *gin.Context) {
return
}
processGroup, realModelName, err := pm.swapProcessGroup(requestedModel)
modelID, found := pm.config.RealModelName(requestedModel)
if !found {
pm.sendErrorResponse(c, http.StatusBadRequest, fmt.Sprintf("could not find real modelID for %s", requestedModel))
return
}
processGroup, err := pm.swapProcessGroup(modelID)
if err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
return
@@ -678,7 +725,7 @@ func (pm *ProxyManager) proxyOAIPostFormHandler(c *gin.Context) {
// If this is the model field and we have a profile, use just the model name
if key == "model" {
// # issue #69 allow custom model names to be sent to upstream
useModelName := pm.config.Models[realModelName].UseModelName
useModelName := pm.config.Models[modelID].UseModelName
if useModelName != "" {
fieldValue = useModelName
@@ -749,9 +796,9 @@ func (pm *ProxyManager) proxyOAIPostFormHandler(c *gin.Context) {
modifiedReq.ContentLength = int64(requestBuffer.Len())
// Use the modified request for proxying
if err := processGroup.ProxyRequest(realModelName, c.Writer, modifiedReq); err != nil {
if err := processGroup.ProxyRequest(modelID, c.Writer, modifiedReq); err != nil {
pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying request: %s", err.Error()))
pm.proxyLogger.Errorf("Error Proxying Request for processGroup %s and model %s", processGroup.id, realModelName)
pm.proxyLogger.Errorf("Error Proxying Request for processGroup %s and model %s", processGroup.id, modelID)
return
}
}