perf: build graph for next batch async to keep GPU busy (#11863)

* perf: build graph for next batch in parallel to keep GPU busy

This refactors the main run loop of the ollama runner to perform the main GPU
intensive tasks (Compute+Floats) in a go routine so we can prepare the next
batch in parallel to reduce the amount of time the GPU stalls waiting for the
next batch of work.

* tests: tune integration tests for ollama engine

This tunes the integration tests to focus more on models supported
by the new engine.

integration/context_test.go

@@ -22,7 +22,7 @@ func TestLongInputContext(t *testing.T) {
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
-		Model:  "llama2",
+		Model:  smol,
 		Prompt: "Oh, don’t speak to me of Austria. Perhaps I don’t understand things, but Austria never has wished, and does not wish, for war. She is betraying us! Russia alone must save Europe. Our gracious sovereign recognizes his high vocation and will be true to it. That is the one thing I have faith in! Our good and wonderful sovereign has to perform the noblest role on earth, and he is so virtuous and noble that God will not forsake him. He will fulfill his vocation and crush the hydra of revolution, which has become more terrible than ever in the person of this murderer and villain! We alone must avenge the blood of the just one.... Whom, I ask you, can we rely on?... England with her commercial spirit will not and cannot understand the Emperor Alexander’s loftiness of soul. She has refused to evacuate Malta. She wanted to find, and still seeks, some secret motive in our actions. What answer did Novosíltsev get? None. The English have not understood and cannot understand the self-abnegation of our Emperor who wants nothing for himself, but only desires the good of mankind. And what have they promised? Nothing! And what little they have promised they will not perform! Prussia has always declared that Buonaparte is invincible, and that all Europe is powerless before him.... And I don’t believe a word that Hardenburg says, or Haugwitz either. This famous Prussian neutrality is just a trap. I have faith only in God and the lofty destiny of our adored monarch. He will save Europe! What country is this referring to?",
 		Stream: &stream,
 		Options: map[string]any{
@@ -36,7 +36,7 @@ func TestLongInputContext(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia"}, 120*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia", "individuals", "coalition", "conflict"}, 120*time.Second, 10*time.Second)
 }
 
 func TestContextExhaustion(t *testing.T) {
@@ -49,7 +49,7 @@ func TestContextExhaustion(t *testing.T) {
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
-		Model:  "llama2",
+		Model:  smol,
 		Prompt: "Write me a story with a ton of emojis?",
 		Stream: &stream,
 		Options: map[string]any{
@@ -63,10 +63,10 @@ func TestContextExhaustion(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived"}, 120*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived", "sunny", "cloudy", "clear", "water"}, 120*time.Second, 10*time.Second)
 }
 
-// Send multiple requests with prior context and ensure the response is coherant and expected
+// Send multiple generate requests with prior context and ensure the response is coherant and expected
 func TestGenerateWithHistory(t *testing.T) {
 	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
 	req, resp := GenerateRequests()
@@ -111,5 +111,56 @@ func TestGenerateWithHistory(t *testing.T) {
 		}(i)
 	}
 	wg.Wait()
-
+}
+
+// Send multiple chat requests with prior context and ensure the response is coherant and expected
+func TestChatWithHistory(t *testing.T) {
+	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
+	req, resp := ChatRequests()
+	numParallel := 2
+	iterLimit := 2
+
+	softTimeout, hardTimeout := getTimeouts(t)
+	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
+	defer cancel()
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+
+	// Get the server running (if applicable) warm the model up with a single initial empty request
+	slog.Info("loading", "model", modelOverride)
+	err := client.Generate(ctx,
+		&api.GenerateRequest{Model: modelOverride, KeepAlive: &api.Duration{Duration: 10 * time.Second}},
+		func(response api.GenerateResponse) error { return nil },
+	)
+	if err != nil {
+		t.Fatalf("failed to load model %s: %s", modelOverride, err)
+	}
+
+	var wg sync.WaitGroup
+	wg.Add(numParallel)
+	for i := range numParallel {
+		go func(i int) {
+			defer wg.Done()
+			k := i % len(req)
+			req[k].Model = modelOverride
+			for j := 0; j < iterLimit; j++ {
+				if time.Now().Sub(started) > softTimeout {
+					slog.Info("exceeded soft timeout, winding down test")
+					return
+				}
+				slog.Info("Starting", "thread", i, "iter", j)
+				// On slower GPUs it can take a while to process the concurrent requests
+				// so we allow a much longer initial timeout
+				assistant := DoChat(ctx, t, client, req[k], resp[k], 120*time.Second, 20*time.Second)
+				if assistant == nil {
+					t.Fatalf("didn't get an assistant response for context")
+				}
+				req[k].Messages = append(req[k].Messages,
+					*assistant,
+					api.Message{Role: "user", Content: "tell me more!"},
+				)
+			}
+		}(i)
+	}
+	wg.Wait()
 }