proxy: fix matrix race and process stop bug (#677)

- matrix.go change logic to consider any proxy.Process not in
StateStopped or StateShutdown
- process.StopImmediately, and Stop() which called it had a subtle bug
where it only handled state transitions from StateReady to
StateStopping. StateStarting -> StateStopping was ignored completely.

fix: #670

.github/workflows/unified-docker.yml

@@ -36,6 +36,11 @@ on:
         type: boolean
         required: false
         default: true
+      push_to_ghcr:
+        description: "Push images to ghcr.io"
+        type: boolean
+        required: false
+        default: true
 
 permissions:
   contents: read
@@ -116,7 +121,7 @@ jobs:
           docker/unified/build-image.sh --${{ matrix.backend }}
 
       - name: Push to GitHub Container Registry
-        if: ${{ !env.ACT }}
+        if: ${{ !env.ACT && inputs.push_to_ghcr == true }}
         run: |
           BASE_TAG="ghcr.io/mostlygeek/llama-swap:unified-${{ matrix.backend }}"
           DATE_TAG=$(date -u +%Y-%m-%d)

docker/unified/Dockerfile

@@ -42,6 +42,7 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     build-essential cmake git python3 python3-pip libssl-dev \
     curl ca-certificates ccache make wget software-properties-common \
     libvulkan-dev glslang-tools spirv-tools vulkan-validationlayers glslc \
+    spirv-headers \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /build

proxy/matrix.go

@@ -297,7 +297,7 @@ func (m *Matrix) Shutdown() {
 	wg.Wait()
 }
 
-// RunningModels returns model names currently in StateReady.
+// RunningModels returns model names currently in an active (non-stopped) state.
 func (m *Matrix) RunningModels() []string {
 	m.Lock()
 	defer m.Unlock()
@@ -308,7 +308,7 @@ func (m *Matrix) RunningModels() []string {
 func (m *Matrix) runningModels() []string {
 	var running []string
 	for id, process := range m.processes {
-		if process.CurrentState() == StateReady {
+		if process.CurrentState() != StateStopped && process.CurrentState() != StateShutdown {
 			running = append(running, id)
 		}
 	}

proxy/process.go

@@ -432,7 +432,10 @@ func (p *Process) start() error {
 
 // Stop will wait for inflight requests to complete before stopping the process.
 func (p *Process) Stop() {
+
+	// guard to prevent multiple goroutines from stopping
 	if !isValidTransition(p.CurrentState(), StateStopping) {
+		p.proxyLogger.Debugf("<%s> Stop() suppressing invalid transition from %s to StateStopping", p.ID, p.CurrentState())
 		return
 	}
 
@@ -445,13 +448,17 @@ func (p *Process) Stop() {
 // StopImmediately will transition the process to the stopping state and stop the process with a SIGTERM.
 // If the process does not stop within the specified timeout, it will be forcefully stopped with a SIGKILL.
 func (p *Process) StopImmediately() {
-	if !isValidTransition(p.CurrentState(), StateStopping) {
+
+	// guard to prevent multiple goroutines from stopping the process
+	enterState := p.CurrentState()
+	if !isValidTransition(enterState, StateStopping) {
+		p.proxyLogger.Debugf("<%s> StopImmediate() suppressing invalid transition from %s to StateStopping", p.ID, p.CurrentState())
 		return
 	}
 
-	p.proxyLogger.Debugf("<%s> Stopping process, current state: %s", p.ID, p.CurrentState())
+	p.proxyLogger.Debugf("<%s> Stopping process, enter state: %s", p.ID, enterState)
-	if curState, err := p.swapState(StateReady, StateStopping); err != nil {
+	if curState, err := p.swapState(enterState, StateStopping); err != nil {
-		p.proxyLogger.Infof("<%s> Stop() Ready -> StateStopping err: %v, current state: %v", p.ID, err, curState)
+		p.proxyLogger.Infof("<%s> Stop() %s -> StateStopping err: %v, current state: %v", p.ID, enterState, err, curState)
 		return
 	}