Files
llama-swap/proxy/processgroup.go
T
Benson Wong 02e015fa49
Linux CI / run-tests (push) Failing after 14m56s
Windows CI / run-tests (push) Has been cancelled
Introduce new routing backend (#790)
This is a huge backend change that essentially started with rewriting
the concurrency handling for processes and blew up to a refactor of the
entire application. In short these are the improvements:

**Better state and life cycle management:** 

Life cycle management of processes has always been the trickiest part of
the code. Juggling mutex locks between multiple locations to reduce race
conditions was complex. Too complex for my feeble brain to build a
simple mental model around as llama-swap gained more features. All of
that has been refactored. Most of the locks are gone, replaced with a
single run() that owns all state changes. There is one place to start
from now to understand and extend routing logic.

The improved life cycle management makes it easier to implement more
complex swap optimization strategies in the future like #727.

**Collation of requests:**

llama-swap previously handled requests and swapping in the order they
came in. For example requests for models in this order ABCABC would
result in 5 swaps. Now those requests are handled in this order AABBCC.
The result is less time waiting for swap under a high churn request
queue. This fixes #588 #612.

A possible future enhancement is to support a starvation parameter so
swap can be forced when models have been waiting too long.

**Shared base implementation for groups and swap matrix:** 

During the refactor it became clear that much of the swapping logic was
shared between these two implementations. That is not surprising
considering the swap matrix was added many moons after groups. Now they
share a common base and their specific swap strategies are implemented
into the swapPlanner interface.

Requests for bespoke or specific swapping scenarios is a common theme in
the issues. Now users can implement whatever bespoke and weird swapping
strategy they want in their own fork. Just ask your agent of choice to
implement swapPlanner. I'll still remaining more conservative on what
actually lands in core llama-swap and will continue to evaluate PRs if
the changes is good for everyone or just one specific use case.

**AI / Agentic Disclosure:** 

I paid very close attention to the low level swap concurrency design and
implementation. It's important to keep that essential part reliable,
boring and no surprises. Backwards compatibility was also maintained,
even the one way non-exclusive group model loading behaviour that people
have rightly pointed out be a weird design decision.

With the underlying swap core done the web server, api and UI sitting on
top were largely ported over with Claude Code and Opus 4.7 in multiple
phases. If you're curious I kept the changes in docs/newrouter-todo.md.
I did several passes to make sure things weren't left behind.

However, even frontier LLMs at the time of this PR still make small
decisions that don't make a lot of sense. They get shit wrong all the
time, just in small subtle way.

That said, there's likely to be some new bugs introduced with this
massive refactor. I'm fairly confident that there's no major
architectural flaws that would cause goal seeking agents to make dumb,
ugly code decisions.

For a little while the legacy llama-swap will be available under
cmd/legacy/llama-swap. The plan is to eventually delete that entry point
as well as the proxy package.

On a bit of a personal note, this PR is exciting and a bit sad for me. I
hand wrote much of the original code and this PR ultimately replaces
much of it. While the old code served as a good reference for the agent
to implement the new stuff it still a bit sad to eventually delete it
all.
2026-05-28 21:47:01 -07:00

195 lines
4.9 KiB
Go

package proxy
import (
"fmt"
"net/http"
"slices"
"sync"
"github.com/mostlygeek/llama-swap/internal/config"
"github.com/mostlygeek/llama-swap/internal/logmon"
)
type ProcessGroup struct {
sync.Mutex
config config.Config
id string
swap bool
exclusive bool
persistent bool
proxyLogger *logmon.Monitor
upstreamLogger *logmon.Monitor
// map of current processes
processes map[string]*Process
lastUsedProcess string
// inflight tracks fast-path requests (requests for the already-selected
// model in a swap group). Fast-path requests Add(1) while holding pg.Lock
// and Done() on completion; a concurrent swap request calls inflight.Wait()
// under pg.Lock before stopping the current process. Without this tracking,
// a fast-path request that has released pg.Lock but has not yet called
// Process.inFlightRequests.Add(1) races with Stop()'s Wait() and can be
// killed mid-request.
inflight sync.WaitGroup
// testDelayFastPath is a test-only hook that, when non-nil, is invoked in
// the fast path after pg.Lock is released but before the request is
// dispatched to Process.ProxyRequest. Tests use it to park a fast-path
// request at the exact race window to deterministically reproduce the
// fast-path vs swap race.
testDelayFastPath func()
}
func NewProcessGroup(id string, config config.Config, proxyLogger *logmon.Monitor, upstreamLogger *logmon.Monitor) *ProcessGroup {
groupConfig, ok := config.Groups[id]
if !ok {
panic("Unable to find configuration for group id: " + id)
}
pg := &ProcessGroup{
id: id,
config: config,
swap: groupConfig.Swap,
exclusive: groupConfig.Exclusive,
persistent: groupConfig.Persistent,
proxyLogger: proxyLogger,
upstreamLogger: upstreamLogger,
processes: make(map[string]*Process),
}
// Create a Process for each member in the group
for _, modelID := range groupConfig.Members {
modelConfig, modelID, _ := pg.config.FindConfig(modelID)
processLogger := logmon.NewWriter(upstreamLogger)
process := NewProcess(modelID, pg.config.HealthCheckTimeout, modelConfig, processLogger, pg.proxyLogger)
pg.processes[modelID] = process
}
return pg
}
// ProxyRequest proxies a request to the specified model
func (pg *ProcessGroup) ProxyRequest(modelID string, writer http.ResponseWriter, request *http.Request) error {
if !pg.HasMember(modelID) {
return fmt.Errorf("model %s not part of group %s", modelID, pg.id)
}
if pg.swap {
pg.Lock()
if pg.lastUsedProcess != modelID {
// Wait for in-flight fast-path requests to drain before stopping
// the previous process. Without this, a fast-path request that has
// released pg.Lock but has not yet incremented
// Process.inFlightRequests races with Stop() and can be killed
// mid-request.
pg.inflight.Wait()
// is there something already running?
if pg.lastUsedProcess != "" {
pg.processes[pg.lastUsedProcess].Stop()
}
// wait for the request to the new model to be fully handled
// and prevent race conditions see issue #277
pg.processes[modelID].ProxyRequest(writer, request)
pg.lastUsedProcess = modelID
// short circuit and exit
pg.Unlock()
return nil
}
// Fast path: register this request in inflight before releasing
// pg.Lock so a concurrent swap will wait for it to complete.
pg.inflight.Add(1)
defer pg.inflight.Done()
pg.Unlock()
if pg.testDelayFastPath != nil {
pg.testDelayFastPath()
}
}
pg.processes[modelID].ProxyRequest(writer, request)
return nil
}
func (pg *ProcessGroup) HasMember(modelName string) bool {
return slices.Contains(pg.config.Groups[pg.id].Members, modelName)
}
func (pg *ProcessGroup) GetMember(modelName string) (*Process, bool) {
if pg.HasMember(modelName) {
return pg.processes[modelName], true
}
return nil, false
}
func (pg *ProcessGroup) StopProcess(modelID string, strategy StopStrategy) error {
pg.Lock()
process, exists := pg.processes[modelID]
if !exists {
pg.Unlock()
return fmt.Errorf("process not found for %s", modelID)
}
if pg.lastUsedProcess == modelID {
pg.lastUsedProcess = ""
}
pg.Unlock()
switch strategy {
case StopImmediately:
process.StopImmediately()
default:
process.Stop()
}
return nil
}
func (pg *ProcessGroup) StopProcesses(strategy StopStrategy) {
pg.Lock()
defer pg.Unlock()
if strategy != StopImmediately {
pg.inflight.Wait()
}
if len(pg.processes) == 0 {
return
}
// stop Processes in parallel
var wg sync.WaitGroup
for _, process := range pg.processes {
wg.Add(1)
go func(process *Process) {
defer wg.Done()
switch strategy {
case StopImmediately:
process.StopImmediately()
default:
process.Stop()
}
}(process)
}
wg.Wait()
}
func (pg *ProcessGroup) Shutdown() {
var wg sync.WaitGroup
for _, process := range pg.processes {
wg.Add(1)
go func(process *Process) {
defer wg.Done()
process.Shutdown()
}(process)
}
wg.Wait()
}