Files
Daniel HiltgenandGitHub ab9295d03c mlx: support concurrent generation (#19197)
* mlx: batch requests through a shared scheduler

Use one request queue and scheduler for generation, decisions, embeddings
and speech, including parallelism one. Share compatible model operations
while keeping request caches, sampling state and response streams independent.

Interleave prompt chunks with decoding and retain speculative decoding when
generation runs alone. Cover isolation, cancellation, slow readers, cache
reuse and mixed requests with unit and integration tests. Remove the stale
llama-server architecture restriction on parallel requests.

* mlx: prevent cache snapshots from retaining batch peers

Copy each request’s initial batched K/V row so a sliding-cache snapshot cannot retain the whole batch allocation while accounting for only one row.

Exclude grammar compilation waits from prompt timing and reuse row-selected attention masks across layers.

Return errors for requests rejected before admission and strengthen numerical and compiled-speech regression checks.

* address comments
2026-11-10 13:24:12 -08:00

264 lines
8.9 KiB
Go

//go:build integration && release
package integration
import (
"context"
"encoding/json"
"errors"
"fmt"
"maps"
"os"
"slices"
"strings"
"testing"
"github.com/google/go-cmp/cmp"
"github.com/google/go-cmp/cmp/cmpopts"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/decision"
"github.com/ollama/ollama/internal/decisiontest"
"github.com/ollama/ollama/llm"
"github.com/ollama/ollama/types/model"
)
func runAPISystemOne(t *testing.T, modelName string) {
path := os.Getenv("OLLAMA_TEST_DECISION_FILE")
if path == "" {
path = "../.cache/decision-data/regression.jsonl"
}
runDecisionCorpus(t, modelName, path)
}
func runAPISystemOneVision(t *testing.T, modelName string) {
abbeyRoad, docs, _ := decodeTestImages(t)
var cases []decisiontest.Case
for _, fixture := range []struct {
name string
images []api.ImageData
choice string
laptop bool
}{
{"road", []api.ImageData{abbeyRoad}, "crossing", false},
{"desk", []api.ImageData{docs}, "working", true},
{"road_then_desk", []api.ImageData{abbeyRoad, docs}, "crossing", false},
{"desk_then_road", []api.ImageData{docs, abbeyRoad}, "working", true},
} {
var req decision.Request
if err := json.Unmarshal([]byte(`{
"state":"Inspect the first attached image.",
"questions":{
"activity":{"type":"choice","instructions":"What are the animals doing in the first image?","criteria":{"crossing":"Walking across a road","working":"Working or resting at a desk"}},
"laptop":{"type":"noul","instructions":"Is there a laptop computer in the first image?","criteria":{}}
}
}`), &req); err != nil {
t.Fatal(err)
}
req.Images = fixture.images
cases = append(cases, decisiontest.Case{ID: fixture.name, Dataset: "ollama-vision", Request: req, Expected: map[string]decisiontest.Expected{
"activity": {Choice: &fixture.choice}, "laptop": {Noul: &fixture.laptop},
}})
}
runDecisionCases(t, modelName, decisiontest.Corpus{Cases: cases})
}
func runAPISystemOneAudio(t *testing.T, modelName string) {
question := decodeTestAudio(t)
abbeyRoad, docs, _ := decodeTestImages(t)
var cases []decisiontest.Case
for _, fixture := range []struct {
name string
media []api.ImageData
activity string
}{
{"audio", []api.ImageData{question}, ""},
{"road_then_audio", []api.ImageData{abbeyRoad, question}, "crossing"},
{"audio_then_desk", []api.ImageData{question, docs}, "working"},
} {
questions := `"topic":{"type":"choice","instructions":"What does the caller ask about in the recording?","criteria":{"sky":"Why the sky is blue","billing":"A charge on their bill","delivery":"A late delivery"}},
"question":{"type":"noul","instructions":"Does the caller ask a question in the recording?","criteria":{}}`
topic, asks := "sky", true
expected := map[string]decisiontest.Expected{"topic": {Choice: &topic}, "question": {Noul: &asks}}
if fixture.activity != "" {
questions += `,"activity":{"type":"choice","instructions":"What are the animals in the picture doing?","criteria":{"crossing":"Walking across a road","working":"Working or resting at a desk"}}`
expected["activity"] = decisiontest.Expected{Choice: &fixture.activity}
}
var req decision.Request
if err := json.Unmarshal([]byte(`{"state":"A caller left a voice message with the attached recording.","questions":{`+questions+`}}`), &req); err != nil {
t.Fatal(err)
}
req.Images = fixture.media
cases = append(cases, decisiontest.Case{ID: fixture.name, Dataset: "ollama-audio", Request: req, Expected: expected})
}
runDecisionCases(t, modelName, decisiontest.Corpus{Cases: cases})
}
func runDecisionCorpus(t *testing.T, modelName, path string) {
if _, err := os.Stat(path); errors.Is(err, os.ErrNotExist) {
t.Skipf("decision corpus unavailable: %s", path)
} else if err != nil {
t.Fatal(err)
}
corpus, err := decisiontest.Load(path)
if err != nil {
t.Fatal(err)
}
t.Logf("decision corpus=%s sha256=%s", path, corpus.SHA256)
runDecisionCases(t, modelName, corpus)
}
func runDecisionCases(t *testing.T, modelName string, corpus decisiontest.Corpus) {
ctx := t.Context()
client, endpoint, cleanup := InitServerConnection(ctx, t)
defer cleanup()
loadCtx, cancel := context.WithTimeout(ctx, apiTestTimeout)
defer cancel()
requireCapability(loadCtx, t, client, modelName, model.CapabilityDecision)
skipIfModelTooLargeForVRAM(loadCtx, t, client, modelName)
media := map[model.Capability]bool{}
for _, c := range corpus.Cases {
for _, data := range c.Request.Images {
if llm.DetectMediaKind(data) == llm.MediaKindAudio {
media[model.CapabilityAudio] = true
} else {
media[model.CapabilityVision] = true
}
}
}
for capability := range media {
requireCapability(loadCtx, t, client, modelName, capability)
}
endpoint = "http://" + endpoint + "/v1/systemone"
score := func(t *testing.T, req decision.Request) decisiontest.Response {
t.Helper()
req.Model = modelName
requestCtx, cancel := context.WithTimeout(ctx, apiTestTimeout)
defer cancel()
result, _, err := decisiontest.Do(requestCtx, endpoint, req)
if err != nil {
t.Fatal(err)
}
return result
}
var tally decisiontest.Tally
datasets := make(map[string]decisiontest.Tally)
distinct := make([]decisiontest.Response, min(4, len(corpus.Cases)))
t.Logf("decision regression cases=%d", len(corpus.Cases))
t.Run("corpus", func(t *testing.T) {
for i, c := range corpus.Cases {
t.Run(c.ID, func(t *testing.T) {
req := c.Request
req.Model = modelName
requestCtx, cancel := context.WithTimeout(t.Context(), apiTestTimeout)
defer cancel()
result, _, err := decisiontest.Do(requestCtx, endpoint, req)
var wrong []string
if err == nil {
err = c.CheckCache(result)
if err == nil {
wrong = c.Mismatches(result)
}
}
tally.Add(c, wrong, err)
d := datasets[c.Dataset]
d.Add(c, wrong, err)
datasets[c.Dataset] = d
if err != nil {
t.Fatal(err)
}
if i < len(distinct) {
distinct[i] = result
}
if len(wrong) == 0 {
return
}
t.Error(strings.Join(wrong, "; "))
})
}
})
t.Logf("decision score: %s, expected cases=%d", tally, len(corpus.Cases))
for _, name := range slices.Sorted(maps.Keys(datasets)) {
t.Logf("decision dataset=%s: %s", name, datasets[name])
}
if t.Failed() {
return
}
// Distinct inputs catch row mix-ups that repeating one warm input cannot.
t.Run("concurrent-distinct", func(t *testing.T) {
for i, c := range corpus.Cases[:len(distinct)] {
t.Run(c.ID, func(t *testing.T) {
t.Parallel()
got := score(t, c.Request)
if diff := cmp.Diff(distinct[i].Answers, got.Answers, cmpopts.EquateApprox(0, .01)); diff != "" {
t.Errorf("answers changed (-independent +concurrent): %s", diff)
}
if got.Usage.InputTokens != distinct[i].Usage.InputTokens {
t.Errorf("input tokens changed: %d -> %d", distinct[i].Usage.InputTokens, got.Usage.InputTokens)
}
})
}
})
input := corpus.Cases[0].Request
input.Questions = &decision.Questions{}
for _, c := range corpus.Cases {
for _, q := range c.Request.Questions.All() {
if _, ok := input.Questions.Get(q.Type); !ok {
input.Questions.Set(q.Type, q)
}
}
}
baseline := score(t, input)
unchanged := func(t *testing.T) {
t.Helper()
got := score(t, input)
// Cache/prefill shapes can shift BF16 probabilities slightly; choices
// must stay identical while numbers retain the existing 0.01 tolerance.
if diff := cmp.Diff(baseline.Answers, got.Answers, cmpopts.EquateApprox(0, .01)); diff != "" {
t.Errorf("answers changed (-baseline +got): %s", diff)
}
if got.Usage.InputTokens != baseline.Usage.InputTokens {
t.Errorf("input tokens changed: %d -> %d", baseline.Usage.InputTokens, got.Usage.InputTokens)
}
}
t.Run("repeat", unchanged)
t.Run("concurrent", func(t *testing.T) {
for i := range 4 {
t.Run(fmt.Sprint(i), func(t *testing.T) { t.Parallel(); unchanged(t) })
}
})
t.Run("after_generation", func(t *testing.T) {
info, err := client.Show(ctx, &api.ShowRequest{Model: modelName})
if err != nil {
t.Fatal(err)
}
noStream := false
err = client.Generate(ctx, &api.GenerateRequest{
Model: modelName, Prompt: "Reply with one word: hello", Stream: &noStream,
Think: &api.ThinkValue{Value: false},
Options: map[string]any{"temperature": 0, "num_predict": 8},
}, func(api.GenerateResponse) error { return nil })
if !slices.Contains(info.Capabilities, model.CapabilityCompletion) {
var status api.StatusError
if !errors.As(err, &status) || status.StatusCode != 400 {
t.Fatalf("decision-only generation: got %v, want HTTP 400", err)
}
} else if err != nil {
t.Fatal(err)
}
unchanged(t)
})
t.Run("invalid_request", func(t *testing.T) {
invalid := input
invalid.Model = modelName
invalid.Questions = nil
requestCtx, cancel := context.WithTimeout(ctx, apiTestTimeout)
defer cancel()
_, _, err := decisiontest.Do(requestCtx, endpoint, invalid)
var status api.StatusError
if !errors.As(err, &status) || status.StatusCode != 400 {
t.Fatalf("missing questions: got %v, want HTTP 400", err)
}
unchanged(t)
})
}