before_code stringlengths 14 465k | reviewer_comment stringlengths 16 64.5k | after_code stringlengths 9 467k | diff_context stringlengths 0 97k | file_path stringlengths 5 226 | comment_line int32 0 26 | language stringclasses 37
values | quality_score float32 0.07 1 | comment_type stringclasses 9
values | comment_length int32 16 64.5k | before_lines int32 1 17.2k | after_lines int32 1 12.1k | is_negative bool 2
classes | pr_title stringlengths 1 308 | pr_number int32 1 299k | repo_name stringclasses 533
values | repo_stars int64 321 419k | repo_language stringclasses 27
values | reviewer_username stringlengths 0 39 | author_username stringlengths 2 39 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
select {
case <-ctx.Done():
return ctx.Err()
default:
}
req, err := srv.Recv()
if err == io.EOF {
return nil
}
if err != nil {
return status.Errorf(codes.Unknown, "cannot receive stream request: %v", err)
}
resp := &extProcPb.ProcessingResponse{}
switch v := req.Request.(type) {
case... | in the configuration in `config/gateway/gateway.yaml` we still use buffer for request, is this still consistent? | ctx := srv.Context()
requestID := uuid.New().String()
for {
select {
case <-ctx.Done():
return ctx.Err()
default:
}
req, err := srv.Recv()
if err == io.EOF {
return nil
}
if err != nil {
return status.Errorf(codes.Unknown, "cannot receive stream request: %v", err)
}
resp := &extProcPb... | @@ -120,13 +124,13 @@ func (s *Server) Process(srv extProcPb.ExternalProcessor_ProcessServer) error {
resp, user, rpm, routingStrategy = s.HandleRequestHeaders(ctx, requestID, req)
case *extProcPb.ProcessingRequest_RequestBody:
- resp, targetPodIP = s.HandleRequestBody(ctx, requestID, req, user, routingStrat... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.571 | question | 113 | 51 | 51 | false | Add envoy gateway streaming support | 377 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
}}},
err.Error())
}
defer func() {
go func() {
s.cache.AddRequestTrace(res.Model, res.Usage.PromptTokens, res.Usage.CompletionTokens)
}()
}()
headers := []*configPb.HeaderValueOption{}
if user.Name != "" {
tpm, err := s.ratelimiter.Incr(ctx, fmt.Sprintf("%v_TPM_CURRENT", user), int64(res.Usage.Tot... | seems it's binary logic, why not just use if..else.. |
return &extProcPb.ProcessingResponse{
Response: &extProcPb.ProcessingResponse_ResponseHeaders{
ResponseHeaders: &extProcPb.HeadersResponse{
Response: &extProcPb.CommonResponse{
HeaderMutation: &extProcPb.HeaderMutation{
SetHeaders: headers,
},
ClearRouteCache: true,
},
},
},
}
... | @@ -311,64 +337,97 @@ func (s *Server) HandleResponseHeaders(ctx context.Context, requestID string, re
}
}
-func (s *Server) HandleResponseBody(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest, user utils.User, rpm int64, targetPodIP string) *extProcPb.ProcessingResponse {
- klog.Infof("--- ... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.286 | suggestion | 52 | 51 | 51 | false | Add envoy gateway streaming support | 377 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
Key: "target-pod",
RawValue: []byte(targetPodIP),
},
},
)
}
return &extProcPb.ProcessingResponse{
Response: &extProcPb.ProcessingResponse_ResponseBody{
ResponseBody: &extProcPb.BodyResponse{
Response: &extProcPb.CommonResponse{
HeaderMutation: &extProcPb.HeaderMutation{
Se... | nit: in future, we can add some post-processor interface here. I think update rate limit would be one of them. This makes the code cleaner. | Key: "x-streaming-error", RawValue: []byte("true"),
}}},
err.Error())
}
case false:
if err := json.Unmarshal(b.ResponseBody.Body, &res); err != nil {
klog.ErrorS(err, "error to unmarshal response", "requestID", requestID, "responseBody", string(b.ResponseBody.GetBody()))
return generateErrorResp... | @@ -311,64 +337,97 @@ func (s *Server) HandleResponseHeaders(ctx context.Context, requestID string, re
}
}
-func (s *Server) HandleResponseBody(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest, user utils.User, rpm int64, targetPodIP string) *extProcPb.ProcessingResponse {
- klog.Infof("--- ... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.357 | nitpick | 139 | 51 | 51 | false | Add envoy gateway streaming support | 377 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
var ok bool
var jsonMap map[string]interface{}
body := req.Request.(*extProcPb.ProcessingRequest_RequestBody)
if err := json.Unmarshal(body.RequestBody.GetBody(), &jsonMap); err != nil {
return generateErrorResponse(envoyTypePb.StatusCode_InternalServerError,
[]*configPb.HeaderValueOption{{Header: &configPb.H... | Here, it requires user to specify a header for stream usage `x-stream-options`? can it be compatible with openai stream sdk? does openai-sdk automatically register this header? |
func (s *Server) HandleRequestBody(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest, user utils.User, routingStrategy string) (*extProcPb.ProcessingResponse, string, bool) {
klog.InfoS("-- In RequestBody processing ...", "requestID", requestID)
var model, targetPodIP string
var ok, stream boo... | @@ -203,25 +207,46 @@ func (s *Server) HandleRequestHeaders(ctx context.Context, requestID string, req
}, user, rpm, routingStrategy
}
-func (s *Server) HandleRequestBody(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest, user utils.User, routingStrategy string) (*extProcPb.ProcessingResponse... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.571 | question | 177 | 51 | 51 | false | Add envoy gateway streaming support | 377 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
*/
package common
// ScalingContext defines the generalized common that holds all necessary data for scaling calculations.
type ScalingContext interface {
GetTargetValue() float64
GetUpFluctuationTolerance() float64
GetDownFluctuationTolerance() float64
GetMaxScaleUpRate() float64
GetMaxScaleDownRate() float64
... | What does `UpdateByPaTypes` mean? it tries to build the context from the object? | */
package common
import (
"strconv"
autoscalingv1alpha1 "github.com/aibrix/aibrix/api/autoscaling/v1alpha1"
"k8s.io/klog/v2"
)
const (
AutoscalingLabelPrefix = "autoscaling.aibrix.ai/"
maxScaleUpRateLabel = AutoscalingLabelPrefix + "max-scale-up-rate"
maxScaleDownRateLabel = AutoscalingLabelPrefix + "max... | @@ -24,6 +37,7 @@ type ScalingContext interface {
GetMaxScaleUpRate() float64
GetMaxScaleDownRate() float64
GetCurrentUsePerPod() float64
+ UpdateByPaTypes(pa *autoscalingv1alpha1.PodAutoscaler) error | pkg/controller/podautoscaler/common/context.go | 26 | Go | 0.5 | question | 80 | 51 | 51 | false | Enable setting PodAutoscaler configuration via YAML labels | 409 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kr11 |
time.sleep(wait_time)
next_start += interval_sec
end_time = time.time()
logging.info(f"Benchmark completed in {end_time - start_time:.2f} seconds")
else:
try:
interval_sec = args.w_interval_sec
with open(args.workload, 'r') as file:... | minor: Let's use logging for consistency | if wait_time > 0:
time.sleep(wait_time)
next_start += interval_sec
end_time = time.time()
logging.info(f"Benchmark completed in {end_time - start_time:.2f} seconds")
if __name__ == "__main__":
parser = FlexibleArgumentParser(description='Benchmark the performance of... | @@ -235,69 +235,96 @@ def sample_sharegpt_requests(
def main(args):
+ FROM_GENERATOR = args.workload is not None
WORKLOAD_MODE = args.concurrency is not None
tokenizer = get_tokenizer(args.model, trust_remote_code=True)
- if args.dataset_path is not None:
- logging.info(f"Start to sample {ar... | benchmarks/autoscaling/benchmark.py | 26 | Python | 0.143 | suggestion | 40 | 51 | 43 | false | AIBrix Workload Generator / Scenario Simulator | 428 | vllm-project/aibrix | 2,572 | Go | Jeffwan | happyandslow |
else:
prompt_len = len(tokenizer(PROMPT).input_ids)
input_requests = [(PROMPT, prompt_len, args.output_len, None)] * args.num_prompts
logging.info(f"Samples results: {input_requests[0]}")
start_time = time.time()
if WORKLOAD_MODE:
logging.info... | the client side code is exact same as generator mode. Should we refactor this part?
either accept generated trace or leverage `generate_workload` to generate the load, then starts the benchmarking. | logging.info(f"Samples results: {input_requests[0]}")
openai_endpoints = build_openai_endpoints(args.deployment_endpoints)
openai_clients = build_openai_clients(openai_endpoints)
start_time = time.time()
if WORKLOAD_MODE:
logging.info(f"Starting benchmark for {args.num_prompts} prompts wit... | @@ -235,69 +235,96 @@ def sample_sharegpt_requests(
def main(args):
+ FROM_GENERATOR = args.workload is not None
WORKLOAD_MODE = args.concurrency is not None
tokenizer = get_tokenizer(args.model, trust_remote_code=True)
- if args.dataset_path is not None:
- logging.info(f"Start to sample {ar... | benchmarks/autoscaling/benchmark.py | 26 | Python | 0.643 | refactor | 202 | 51 | 51 | false | AIBrix Workload Generator / Scenario Simulator | 428 | vllm-project/aibrix | 2,572 | Go | Jeffwan | happyandslow |
prompt = dataset[i][0]
prompt_token_ids = tokenizer(prompt).input_ids
completion = dataset[i][1]
completion_token_ids = tokenizer(completion).input_ids
prompt_len = len(prompt_token_ids)
output_len = len(completion_token_ids) if fixed_output_len is None else fixed_output_... | lot of clients/experiments has this similar methods. can we refactor this part? If you do not like to make it now, let's have a brief chat. | prompt_token_ids = tokenizer(prompt).input_ids
completion = dataset[i][1]
completion_token_ids = tokenizer(completion).input_ids
prompt_len = len(prompt_token_ids)
output_len = len(completion_token_ids) if fixed_output_len is None else fixed_output_len
if prompt_len < 4 o... | @@ -219,36 +220,61 @@ def main(args):
if args.num_prompts < len(openai_endpoints) * args.concurrency:
num_prompts = len(openai_endpoints) * args.concurrency
- if args.dataset_path is not None:
- logging.info(f"Start to sample {num_prompts} prompts from {args.dataset_path}")
- input_requ... | benchmarks/lora/benchmark.py | 26 | Python | 0.429 | refactor | 140 | 51 | 51 | false | AIBrix Workload Generator / Scenario Simulator | 428 | vllm-project/aibrix | 2,572 | Go | Jeffwan | happyandslow |
)
const (
metricServerDefaultMetricWindow = time.Minute
)
type KPAMetricsClient struct {
fetcher MetricFetcher
// collectionsMutex protects access to both panicWindowDict and stableWindowDict,
// ensuring thread-safe read and write operations. It uses a read-write mutex to
// allow multiple concurrent reads whi... | client is not singleton now? it only response for one metric? |
const (
metricServerDefaultMetricWindow = time.Minute
paGranularity = time.Second
)
type KPAMetricsClient struct {
fetcher MetricFetcher
// collectionsMutex protects access to both panicWindowDict and stableWindowDict,
// ensuring thread-safe read and write operations. It uses a read-write mut... | @@ -54,42 +54,28 @@ type KPAMetricsClient struct {
// are collected and processed within the sliding window.
granularity time.Duration
// the difference between stable and panic metrics is the time window range
- panicWindowDict map[NamespaceNameMetric]*aggregation.TimeWindow
- stableWindowDict map[NamespaceName... | pkg/controller/podautoscaler/metrics/client.go | 26 | Go | 0.214 | question | 61 | 51 | 51 | false | Refactor Scaler: Resolve Issues with Metric Parameter Updates in Multiple KPAs | 437 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kr11 |
func (c *KPAMetricsClient) UpdateMetricIntoWindow(metricKey NamespaceNameMetric, now time.Time, metricValue float64) error {
// Add to panic and stable windows; create a new window if not present in the map
// Ensure that panicWindowDict and stableWindowDict maps are checked and updated
updateWindow := func(windowDi... | how does it know the metrics now? | func (c *KPAMetricsClient) UpdateMetricIntoWindow(now time.Time, metricValue float64) error {
c.panicWindow.Record(now, metricValue)
c.stableWindow.Record(now, metricValue)
return nil
}
func (c *KPAMetricsClient) UpdatePodListMetric(metricValues []float64, metricKey NamespaceNameMetric, now time.Time) error {
retu... | @@ -112,7 +98,7 @@ func (c *KPAMetricsClient) UpdateMetrics(now time.Time, metricKey NamespaceNameM
defer c.collectionsMutex.Unlock()
// Update metrics into the window for tracking
- err := c.UpdateMetricIntoWindow(metricKey, now, sumMetricValue)
+ err := c.UpdateMetricIntoWindow(now, sumMetricValue) | pkg/controller/podautoscaler/metrics/client.go | 26 | Go | 0.071 | question | 34 | 51 | 51 | false | Refactor Scaler: Resolve Issues with Metric Parameter Updates in Multiple KPAs | 437 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kr11 |
var (
DefaultRequeueDuration = 10 * time.Second
)
// Add creates a new PodAutoscaler Controller and adds it to the Manager with default RBAC.
// The Manager will set fields on the Controller and Start it when the Manager is Started.
func Add(mgr manager.Manager) error {
r, err := newReconciler(mgr)
if err != nil {... | em. in this case, does it mean multiple HPA, KPA or APA scalers are flatten?
We used to have 3 singleton autoscaler |
var (
DefaultRequeueDuration = 10 * time.Second
)
// Add creates a new PodAutoscaler Controller and adds it to the Manager with default RBAC.
// The Manager will set fields on the Controller and Start it when the Manager is Started.
func Add(mgr manager.Manager) error {
r, err := newReconciler(mgr)
if err != nil {... | @@ -74,36 +74,9 @@ func newReconciler(mgr manager.Manager) (reconcile.Reconciler, error) {
Mapper: mgr.GetRESTMapper(),
resyncInterval: 10 * time.Second, // TODO: this should be override by an environment variable
eventCh: make(chan event.GenericEvent),
+ AutoscalerMap: make(map[metrics.Namesp... | pkg/controller/podautoscaler/podautoscaler_controller.go | 26 | Go | 0.429 | suggestion | 118 | 51 | 51 | false | Refactor Scaler: Resolve Issues with Metric Parameter Updates in Multiple KPAs | 437 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kr11 |
},
{
"name": "gpu_cache_usage_perc",
"type": "gauge",
"description": "GPU KV-cache usage. 1 means 100 percent usage.",
"value": overrides.get(
"gpu_cache_usage_perc", gpu_cache_usage_perc
)
},
]
# Generate all m... | I forgot to share with you the lora metrics as well. Do you need any help here? | "description": "Average prefill throughput in tokens/s.",
"value": overrides.get("avg_prompt_throughput", avg_prompt_throughput)
},
{
"name": "avg_generation_throughput_toks_per_s",
"type": "gauge",
"description": "Average generation throughput... | @@ -386,17 +394,12 @@ def metrics():
]
# Generate all metrics
- metrics_output = "".join(
- generate_counter_gauge_metric(metric["name"], metric["type"], metric["description"], model_name,
- metric["value"])
- for metric in simple_metrics
- )
-
- l... | docs/development/app/app.py | 26 | Python | 0.214 | question | 79 | 51 | 51 | false | Support metrics multi labels for different models | 450 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
},
{
"name": "avg_generation_throughput_toks_per_s",
"type": "gauge",
"description": "Average generation throughput in tokens/s.",
"value": overrides.get("avg_generation_throughput", avg_generation_throughput)
},
{
"name": "gpu_... | I think this is due to my careless. `lora-model-1` was for testing but not real production workloads, we should correct it here later | "description": "Number of requests waiting to be processed.",
"value": overrides.get("waiting", waiting)
},
{
"name": "avg_prompt_throughput_toks_per_s",
"type": "gauge",
"description": "Average prefill throughput in tokens/s.",
"va... | @@ -386,17 +394,12 @@ def metrics():
]
# Generate all metrics
- metrics_output = "".join(
- generate_counter_gauge_metric(metric["name"], metric["type"], metric["description"], model_name,
- metric["value"])
- for metric in simple_metrics
- )
-
- l... | docs/development/app/app.py | 26 | Python | 0.571 | suggestion | 133 | 51 | 51 | false | Support metrics multi labels for different models | 450 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
"k8s.io/client-go/rest"
"k8s.io/client-go/tools/cache"
"k8s.io/klog/v2"
modelv1alpha1 "github.com/aibrix/aibrix/api/model/v1alpha1"
v1alpha1 "github.com/aibrix/aibrix/pkg/client/clientset/versioned"
v1alpha1scheme "github.com/aibrix/aibrix/pkg/client/clientset/versioned/scheme"
"github.com/aibrix/aibrix/pkg/met... | Seems `PodModelMetrics` can not be replaced by `PodMetrics` and `ModelToPodMapping` together. right? we need another dimension or hierarchy to support lora case | "k8s.io/client-go/tools/cache"
"k8s.io/klog/v2"
modelv1alpha1 "github.com/aibrix/aibrix/api/model/v1alpha1"
v1alpha1 "github.com/aibrix/aibrix/pkg/client/clientset/versioned"
v1alpha1scheme "github.com/aibrix/aibrix/pkg/client/clientset/versioned/scheme"
"github.com/aibrix/aibrix/pkg/metrics"
prometheusv1 "gith... | @@ -61,10 +61,11 @@ type Cache struct {
metrics map[string]interface{}
ModelMetrics map[string]map[string]interface{}
Pods map[string]*v1.Pod
- PodMetrics map[string]map[string]metrics.MetricValue // pod_name: map[metric_name]metric_val
- PodToModelMapping map[string]map[string]... | pkg/cache/cache.go | 26 | Go | 0.571 | suggestion | 160 | 51 | 51 | false | Support metrics multi labels for different models | 450 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
continue
}
value := metricValue.GetHistogramValue()
c.PodMetrics[pod.Name][metricName] = &metrics.HistogramMetricValue{
Sum: value.Sum,
Count: value.Count,
Buckets: value.Buckets,
}
klog.V(5).InfoS("Successfully parsed metrics", "metric", metricName, "PodIP", pod.Status.PodIP, "Por... | using `expfmt` makes a lot of sense. Great | continue
}
podName := pod.Name
if len(c.PodMetrics[podName]) == 0 {
c.PodMetrics[podName] = map[string]metrics.MetricValue{}
}
if len(c.PodModelMetrics[podName]) == 0 {
c.PodModelMetrics[podName] = make(map[string]map[string]metrics.MetricValue)
}
// We should use the primary container port. In ... | @@ -538,41 +587,83 @@ func (c *Cache) updatePodMetrics() {
}
}()
- body, err := io.ReadAll(resp.Body)
+ // TODO: the metrics should come from those router subscribers in future
+ var parser expfmt.TextParser | pkg/cache/cache.go | 26 | Go | 0.214 | suggestion | 42 | 51 | 51 | false | Support metrics multi labels for different models | 450 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
continue
}
histogramValue := &metrics.HistogramMetricValue{
Sum: metricValue.Sum,
Count: metricValue.Count,
Buckets: metricValue.Buckets,
}
err = c.updatePodRecord(podName, modelName, metricName, scope, histogramValue)
if err != nil {
klog.Errorf("Failed to update met... | Technically, this might not be populated if pod just arrive. but I think this is ok, we may miss few metrics until the pod is added into the map. | continue
}
histogramValue := &metrics.HistogramMetricValue{
Sum: metricValue.Sum,
Count: metricValue.Count,
Buckets: metricValue.Buckets,
}
err = c.updatePodRecord(podName, modelName, metricName, scope, histogramValue)
if err != nil {
klog.Errorf("Failed to update met... | @@ -672,39 +672,43 @@ func (c *Cache) updatePodMetrics() {
}
for _, metricName := range prometheusMetricNames {
- // TODO: Get another model PromQL value from the same pod
- modelName := pod.Labels["model.aibrix.ai/name"]
- queryLabels := map[string]string{
- "model_name": modelName,
- "instance": ... | pkg/cache/cache.go | 26 | Go | 0.357 | suggestion | 146 | 51 | 51 | false | Fix multi models metric result in PromQL | 458 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
initial_err_perc = 0.5,
err_step = 0.05
)
if sampled_requests: # Only add non-empty groups
grouped_requests.append((ts, sampled_requests))
ts += interval_ms
if ts > duration_ms:
break
# Move to the next tim... | I suggest to have all `--trace-type` listed here. | sampled_requests = sample_sharegpt_requests_len_range(
df = sharegpt_df,
num_requests = len(input_lens),
input_lens = input_lens,
output_lens = output_lens,
initial_err_perc = 0.5,
err_step = 0.05
)
... | @@ -180,64 +180,64 @@ def generate_from_azure_csv(file_path: str,
)
if sampled_requests: # Only add non-empty groups
- grouped_requests.append(sampled_requests)
-
+ grouped_requests.append((ts, sampled_requests))
+ ts += interval_ms
+ if t... | benchmarks/generator/workload_generator.py | 26 | Python | 0.357 | suggestion | 49 | 51 | 51 | false | Refactor generator to generate time-based traces | 478 | vllm-project/aibrix | 2,572 | Go | Jeffwan | happyandslow |
err_step = 0.05
)
if sampled_requests: # Only add non-empty groups
grouped_requests.append((ts, sampled_requests))
ts += interval_ms
if ts > duration_ms:
break
# Move to the next time range
current_time += time_ran... | Let's add more details in the `help`. what kind of trace-type needs such file etc. | df = sharegpt_df,
num_requests = len(input_lens),
input_lens = input_lens,
output_lens = output_lens,
initial_err_perc = 0.5,
err_step = 0.05
)
if sampled_requests: # Only add non-empty gro... | @@ -180,64 +180,64 @@ def generate_from_azure_csv(file_path: str,
)
if sampled_requests: # Only add non-empty groups
- grouped_requests.append(sampled_requests)
-
+ grouped_requests.append((ts, sampled_requests))
+ ts += interval_ms
+ if t... | benchmarks/generator/workload_generator.py | 26 | Python | 0.5 | suggestion | 84 | 51 | 51 | false | Refactor generator to generate time-based traces | 478 | vllm-project/aibrix | 2,572 | Go | Jeffwan | happyandslow |
)
if sampled_requests: # Only add non-empty groups
grouped_requests.append((ts, sampled_requests))
ts += interval_ms
if ts > duration_ms:
break
# Move to the next time range
current_time += time_range
# Print or process group... | let's change the default one to something else. like https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct. llama7b can not be publicly accessible | num_requests = len(input_lens),
input_lens = input_lens,
output_lens = output_lens,
initial_err_perc = 0.5,
err_step = 0.05
)
if sampled_requests: # Only add non-empty groups
grouped_requests.a... | @@ -180,64 +180,64 @@ def generate_from_azure_csv(file_path: str,
)
if sampled_requests: # Only add non-empty groups
- grouped_requests.append(sampled_requests)
-
+ grouped_requests.append((ts, sampled_requests))
+ ts += interval_ms
+ if t... | benchmarks/generator/workload_generator.py | 26 | Python | 0.429 | suggestion | 146 | 51 | 51 | false | Refactor generator to generate time-based traces | 478 | vllm-project/aibrix | 2,572 | Go | Jeffwan | happyandslow |
current_time += time_range
# Print or process grouped_requests as needed
# Save to file
grouped_requests = make_serializable(grouped_requests)
save_workload(grouped_requests, output_file)
return grouped_requests
if __name__ == '__main__':
parser = argparse.ArgumentParser(desc... | suggest to use `second` as the time unit. for duration, normally, we at least generate 120-600 etc. won't be any case less than 1s | grouped_requests.append((ts, sampled_requests))
ts += interval_ms
if ts > duration_ms:
break
# Move to the next time range
current_time += time_range
# Print or process grouped_requests as needed
# Save to file
grouped_requests = make_serializable(gro... | @@ -180,64 +180,64 @@ def generate_from_azure_csv(file_path: str,
)
if sampled_requests: # Only add non-empty groups
- grouped_requests.append(sampled_requests)
-
+ grouped_requests.append((ts, sampled_requests))
+ ts += interval_ms
+ if t... | benchmarks/generator/workload_generator.py | 26 | Python | 0.571 | suggestion | 130 | 51 | 51 | false | Refactor generator to generate time-based traces | 478 | vllm-project/aibrix | 2,572 | Go | Jeffwan | happyandslow |
*/
package metrics
const (
NumRequestsRunning = "num_requests_running"
NumRequestsWaiting = "num_requests_waiting"
NumRequestsSwapped = "num_requests_swapped"
AvgPromptThroughputToksPerS = "avg_prompt_throughput_toks_per_s"
AvgGenerationThroughputTok... | the default metric name is `gpu_cache_usage_perc`. without `_pod`. should we make it aligned? | */
package metrics
const (
NumRequestsRunning = "num_requests_running"
NumRequestsWaiting = "num_requests_waiting"
NumRequestsSwapped = "num_requests_swapped"
AvgPromptThroughputToksPerS = "avg_prompt_throughput_toks_per_s"
AvgGenerationThroughputTok... | @@ -17,20 +17,32 @@ limitations under the License.
package metrics
const (
- NumRequestsRunning = "num_requests_running"
- NumRequestsWaiting = "num_requests_waiting"
- NumRequestsSwapped = "num_requests_swapped"
- AvgPromptThroughputToksPerS = "avg_prompt_throughput_toks_... | pkg/metrics/metrics.go | 26 | Go | 0.429 | performance | 93 | 51 | 51 | false | Add metrics that scheduling needed | 486 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
return 0, "", currentTimestamp, fmt.Errorf("can not calculate metrics for scale %s", pa.Spec.ScaleTargetRef.Name)
}
// refer to knative-serving.
// In pkg/reconciler/autoscaling/kpa/kpa.go:198, kpa maintains a list of deciders into multi-scaler, each of them corresponds to a pa (PodAutoscaler).
// We create or updat... | the indent doesn't look correct? | // In pkg/reconciler/autoscaling/kpa/kpa.go:198, kpa maintains a list of deciders into multi-scaler, each of them corresponds to a pa (PodAutoscaler).
// We create or update the scaler instance according to the pa passed in
func (r *PodAutoscalerReconciler) updateScalerSpec(ctx context.Context, pa autoscalingv1alpha1.P... | @@ -616,7 +627,7 @@ func (r *PodAutoscalerReconciler) updateScalerSpec(ctx context.Context, pa autos
func (r *PodAutoscalerReconciler) updateMetricsForScale(ctx context.Context, pa autoscalingv1alpha1.PodAutoscaler, scale *unstructured.Unstructured) (err error) {
currentTimestamp := time.Now()
- metricKey := NewNa... | pkg/controller/podautoscaler/podautoscaler_controller.go | 26 | Go | 0.071 | question | 32 | 51 | 51 | false | [Bug] handle metricKey creation with MetricsSources | 488 | vllm-project/aibrix | 2,572 | Go | Jeffwan | nwangfw |
logger.V(4).Info("Obtained selector and get ReadyPodsCount", "selector", labelsSelector, "originalReadyPodsCount", originalReadyPodsCount)
var metricKey metrics.NamespaceNameMetric
if len(pa.Spec.MetricsSources) > 0 {
metricKey = metrics.NewNamespaceNameMetric(pa.Namespace, pa.Spec.ScaleTargetRef.Name, pa.Spec.Me... | for such issues, please raise the concern. here it just use the 1st one. Is it possible to have multiple keys? just like HPA does, in that case, this will fail.
| }
logger.V(4).Info("Obtained selector and get ReadyPodsCount", "selector", labelsSelector, "originalReadyPodsCount", originalReadyPodsCount)
// Calculate the desired number of pods using the autoscaler logic.
autoScaler, ok := r.AutoscalerMap[metricKey]
if !ok {
return 0, "", currentTimestamp, fmt.Errorf("unsu... | @@ -602,11 +602,18 @@ func (r *PodAutoscalerReconciler) computeReplicasForMetrics(ctx context.Context,
return 0, "", currentTimestamp, fmt.Errorf("can not calculate metrics for scale %s", pa.Spec.ScaleTargetRef.Name)
}
+func createMetricKey(pa autoscalingv1alpha1.PodAutoscaler) metrics.NamespaceNameMetric {
+ i... | pkg/controller/podautoscaler/podautoscaler_controller.go | 26 | Go | 0.429 | bug | 162 | 51 | 51 | false | [Bug] handle metricKey creation with MetricsSources | 488 | vllm-project/aibrix | 2,572 | Go | Jeffwan | nwangfw |
var metricKey metrics.NamespaceNameMetric
if len(pa.Spec.MetricsSources) > 0 {
metricKey = metrics.NewNamespaceNameMetric(pa.Namespace, pa.Spec.ScaleTargetRef.Name, pa.Spec.MetricsSources[0].Name)
} else {
metricKey = metrics.NewNamespaceNameMetric(pa.Namespace, pa.Spec.ScaleTargetRef.Name, pa.Spec.TargetMetric)... | why don't you update this method but creating a new one? | logger.V(4).Info("Obtained selector and get ReadyPodsCount", "selector", labelsSelector, "originalReadyPodsCount", originalReadyPodsCount)
// Calculate the desired number of pods using the autoscaler logic.
autoScaler, ok := r.AutoscalerMap[metricKey]
if !ok {
return 0, "", currentTimestamp, fmt.Errorf("unsuppor... | @@ -602,11 +602,18 @@ func (r *PodAutoscalerReconciler) computeReplicasForMetrics(ctx context.Context,
return 0, "", currentTimestamp, fmt.Errorf("can not calculate metrics for scale %s", pa.Spec.ScaleTargetRef.Name)
}
+func createMetricKey(pa autoscalingv1alpha1.PodAutoscaler) metrics.NamespaceNameMetric {
+ i... | pkg/controller/podautoscaler/podautoscaler_controller.go | 26 | Go | 0.214 | question | 56 | 51 | 51 | false | [Bug] handle metricKey creation with MetricsSources | 488 | vllm-project/aibrix | 2,572 | Go | Jeffwan | nwangfw |
See the License for the specific language governing permissions and
limitations under the License.
*/
package metrics
import (
"context"
"time"
"k8s.io/apimachinery/pkg/types"
v1 "k8s.io/api/core/v1"
autoscalingv1alpha1 "github.com/aibrix/aibrix/api/autoscaling/v1alpha1"
)
// NamespaceNameMetric contains the... | minor: we should leave a comment here like a TODO to remind user this is due to a limitation etc. | See the License for the specific language governing permissions and
limitations under the License.
*/
package metrics
import (
"context"
"time"
"k8s.io/apimachinery/pkg/types"
v1 "k8s.io/api/core/v1"
autoscalingv1alpha1 "github.com/aibrix/aibrix/api/autoscaling/v1alpha1"
)
// NamespaceNameMetric contains the... | @@ -33,8 +33,19 @@ type NamespaceNameMetric struct {
MetricName string
}
-func NewNamespaceNameMetric(namespace string, name string, metricName string) NamespaceNameMetric {
- return NamespaceNameMetric{NamespacedName: types.NamespacedName{Namespace: namespace, Name: name}, MetricName: metricName}
+func NewNamespa... | pkg/controller/podautoscaler/metrics/interface.go | 26 | Go | 0.286 | nitpick | 98 | 51 | 51 | false | [Bug] handle metricKey creation with MetricsSources | 488 | vllm-project/aibrix | 2,572 | Go | Jeffwan | nwangfw |
# Copyright 2024 The Aibrix Team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing... | are we sure that s3 based tos downloader will always perform better tos native one? If not, can we keep this code for a while? let's say we have tosv1 and tosv1 downloader co-exist a while. Once we are sure that v2 works better, we can deprecated v1 | # Copyright 2024 The Aibrix Team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing... | @@ -1,171 +0,0 @@
-# Copyright 2024 The Aibrix Team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-# http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable ... | python/aibrix/aibrix/downloader/tos.py | 14 | Python | 0.5 | question | 249 | 39 | 39 | false | [perf] Refact tos downloader in Runtime | 510 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
aws_secret_access_key: "SK****",,
}
"""
pass
@lru_cache()
def _is_directory(self) -> bool:
"""Check if model_uri is a directory."""
if self.bucket_path.endswith("/"):
return True
objects_out = self.client.list_objects_v2(
... | I can not remember the details here. `file.endswith("/")` means folder. here, it tries to filter all directories but left all files? but the method name is `_directory_list`? | aws_secret_access_key: "SK****",,
}
"""
pass
@lru_cache()
def _is_directory(self) -> bool:
"""Check if model_uri is a directory."""
if self.bucket_path.endswith("/"):
return True
objects_out = self.client.list_objects_v2(
... | @@ -123,7 +121,9 @@ def _directory_list(self, path: str) -> List[str]:
Bucket=self.bucket_name, Delimiter="/", Prefix=path
)
contents = objects_out.get("Contents", [])
- return [content.get("Key") for content in contents]
+ files = [content.get("Key") for content in contents... | python/aibrix/aibrix/downloader/s3.py | 26 | Python | 0.5 | question | 175 | 51 | 51 | false | [perf] Refact tos downloader in Runtime | 510 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
), "S3 bucket name is not set."
assert (
self.bucket_path is not None and self.bucket_path != ""
), "S3 bucket path is not set."
try:
self.client.head_bucket(Bucket=self.bucket_name)
except Exception as e:
assert False, f"S3 bucket {self.bucket... | BTW, can we keep the tos file and move TOSDownloader to tos.py? | # Copyright 2024 The Aibrix Team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing... | @@ -182,3 +182,86 @@ def download_progress(bytes_transferred):
Callback=download_progress if self.enable_progress_bar else None,
)
save_meta_data(meta_data_file, etag)
+
+
+class S3Downloader(S3BaseDownloader):
+ def __init__(
+ self,
+ model_uri,
+ mod... | python/aibrix/aibrix/downloader/s3.py | 26 | Python | 0.286 | question | 63 | 51 | 230 | false | [perf] Refact tos downloader in Runtime | 510 | vllm-project/aibrix | 2,572 | Go | Jeffwan | brosoul |
/*
Copyright 2024 The Aibrix Team.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, softwar... | Let's still use `GetPaMetricSources`? limitation could be documented. No need to expose the implemnetation in the function method names |
// +kubebuilder:object:root=true
// +k8s:deepcopy-gen:interfaces=k8s.io/apimachinery/pkg/runtime.Object
// PodAutoscalerList contains a list of PodAutoscaler
type PodAutoscalerList struct {
metav1.TypeMeta `json:",inline"`
metav1.ListMeta `json:"metadata,omitempty"`
Items []PodAutoscaler `json:"items"`
}... | @@ -142,3 +164,11 @@ const (
// QPS is the requests per second reaching the Pod.
QPS = "qps"
)
+
+// GetFirstPaMetricSources Currently, we don't support metric resources that are more than one yet.
+func GetFirstPaMetricSources(pa PodAutoscaler) (MetricSource, error) { | api/autoscaling/v1alpha1/podautoscaler_types.go | 26 | Go | 0.571 | suggestion | 135 | 145 | 32 | false | Refactor metric source for customized protocol, port and path | 511 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kr11 |
// APA represents the AiBrix Pod Autoscaling Algorithm
APA ScalingStrategyType = "APA"
)
// MetricSource defines an endpoint and path from which metrics are collected.
type MetricSource struct {
// e.g. service1.example.com
Endpoint string `json:"endpoint"`
// e.g. /api/metrics/cpu
Path string `json:"path"`
// ... | we discussed the Endpoint here. Should we build the endpoint from Host:Port/Path or keep using it ? I kind of forgot the discussion conclusion | )
type MetricSourceType string
const (
// POD need to scan all k8s pods to fetch the data
POD MetricSourceType = "pod"
// DOMAIN only need to access specified domain
DOMAIN MetricSourceType = "domain"
)
type ProtocolType string
const (
HTTP ProtocolType = "http"
HTTPS ProtocolType = "https"
)
// MetricSourc... | @@ -86,14 +84,38 @@ const (
APA ScalingStrategyType = "APA"
)
+type MetricSourceType string
+
+const (
+ // POD need to scan all k8s pods to fetch the data
+ POD MetricSourceType = "pod"
+ // DOMAIN only need to access specified domain
+ DOMAIN MetricSourceType = "domain"
+)
+
+type ProtocolType string
+
+const (
... | api/autoscaling/v1alpha1/podautoscaler_types.go | 26 | Go | 0.429 | suggestion | 142 | 51 | 51 | false | Refactor metric source for customized protocol, port and path | 511 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kr11 |
// Implement a timeout for the reconciliation process.
ctx, cancel := context.WithTimeout(ctx, 10*time.Second)
defer cancel()
klog.V(4).InfoS("Reconciling PodAutoscaler", "obj", req.NamespacedName)
var pa autoscalingv1alpha1.PodAutoscaler
if err := r.Get(ctx, req.NamespacedName, &pa); err != nil {
if errors.I... | there's some other cases like other objects sometimes get enqueued (not PodAutoscaler object), in this case, `deleteScaler` need to handle such case. | }
}
//+kubebuilder:rbac:groups=autoscaling.aibrix.ai,resources=podautoscalers,verbs=get;list;watch;create;update;patch;delete
//+kubebuilder:rbac:groups=autoscaling.aibrix.ai,resources=podautoscalers/status,verbs=get;update;patch
//+kubebuilder:rbac:groups=autoscaling.aibrix.ai,resources=podautoscalers/finalizers,ver... | @@ -146,8 +162,9 @@ func (r *PodAutoscalerReconciler) Reconcile(ctx context.Context, req ctrl.Reques
var pa autoscalingv1alpha1.PodAutoscaler
if err := r.Get(ctx, req.NamespacedName, &pa); err != nil {
if errors.IsNotFound(err) {
- // Object might have been deleted after reconcile request, ignore and return.
-... | pkg/controller/podautoscaler/podautoscaler_controller.go | 26 | Go | 0.643 | suggestion | 149 | 51 | 51 | false | clean memory scaler object when pa crd is deleted | 520 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kr11 |
type MetricType string
const (
ResourceMetrics MetricType = "resource"
CustomMetrics MetricType = "custom"
RawMetrics MetricType = "raw"
)
// MetricFetcher defines an interface for fetching metrics. it could be Kubernetes metrics or Pod prometheus metrics.
type MetricFetcher interface {
// Obseleted: Call ... | this is acceptable at this moment, we could consider RestMetricsFetcher interface later to better support testing | type MetricType string
const (
ResourceMetrics MetricType = "resource"
CustomMetrics MetricType = "custom"
RawMetrics MetricType = "raw"
)
// MetricFetcher defines an interface for fetching metrics. it could be Kubernetes metrics or Pod prometheus metrics.
type MetricFetcher interface {
// Obseleted: Call ... | @@ -59,7 +59,10 @@ func (f *abstractMetricsFetcher) FetchMetric(ctx context.Context, pod v1.Pod, me
}
// RestMetricsFetcher implements MetricFetcher to fetch metrics from Pod's /metrics endpoint.
-type RestMetricsFetcher struct{}
+type RestMetricsFetcher struct {
+ // For unit test purpose only
+ test_url_setter fu... | pkg/controller/podautoscaler/metrics/fetcher.go | 26 | Go | 0.429 | suggestion | 113 | 51 | 51 | false | [Bug] Accumulated bug fix on controller manager, mock app configuration, and gpu optimizer. | 522 | vllm-project/aibrix | 2,572 | Go | Jeffwan | zhangjyr |
}
return podList, nil
}
func CountReadyPods(podList *v1.PodList) (int64, error) {
if podList == nil || len(podList.Items) == 0 {
return 0, nil
}
readyPodCount := 0
for _, pod := range podList.Items {
isReady := IsPodReady(&pod)
if pod.Status.Phase == v1.PodRunning && isReady {
readyPodCount++
}
klo... | is this method being used by other files? like some other controllers other than cache.go? I am a little bit concerned on the impact if we have other scenarios.
the key difference of this change is `terminated` pod?
| }
return podList, nil
}
func CountReadyPods(podList *v1.PodList) (int64, error) {
if podList == nil || len(podList.Items) == 0 {
return 0, nil
}
readyPodCount := 0
for _, pod := range podList.Items {
isReady := IsPodReady(&pod)
if pod.Status.Phase == v1.PodRunning && isReady {
readyPodCount++
}
klo... | @@ -144,9 +144,12 @@ func CountReadyPods(podList *v1.PodList) (int64, error) {
func FilterReadyPods(pods map[string]*v1.Pod) []*v1.Pod {
var readyPods []*v1.Pod
for _, pod := range pods {
- if pod.Status.PodIP != "" && IsPodReady(pod) { | pkg/utils/pod.go | 26 | Go | 0.643 | performance | 221 | 32 | 35 | false | Handle terminating pod for request routing | 549 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
klog.ErrorS(err, "error to unmarshal response", "requestID", requestID, "requestBody", string(body.RequestBody.GetBody()))
return generateErrorResponse(envoyTypePb.StatusCode_InternalServerError,
[]*configPb.HeaderValueOption{{Header: &configPb.HeaderValue{
Key: "x-request-body-processing-error", RawValue: [... | it filters the terminated pods here? | klog.ErrorS(err, "error to unmarshal response", "requestID", requestID, "requestBody", string(body.RequestBody.GetBody()))
return generateErrorResponse(envoyTypePb.StatusCode_InternalServerError,
[]*configPb.HeaderValueOption{{Header: &configPb.HeaderValue{
Key: "x-request-body-processing-error", RawValue: [... | @@ -243,6 +243,15 @@ func (s *Server) HandleRequestBody(ctx context.Context, requestID string, req *e
fmt.Sprintf("model %s does not exist", model)), targetPodIP, stream
}
+ // early reject if no pods are ready to accept request for a model
+ pods, err := s.cache.GetPodsForModel(model) | pkg/plugins/gateway/gateway.go | 26 | Go | 0.071 | question | 37 | 51 | 51 | false | Handle terminating pod for request routing | 549 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
"description": "Average generation throughput in tokens/s.",
"value": overrides.get("avg_generation_throughput", avg_generation_throughput)
},
{
"name": "gpu_cache_usage_perc",
"type": "gauge",
"description": "GPU KV-cache usage. 1 means 100 pe... | do you have dynamic lora information, should it be a rendered template with lora names? | "description": "Average generation throughput in tokens/s.",
"value": overrides.get("avg_generation_throughput", avg_generation_throughput)
},
{
"name": "gpu_cache_usage_perc",
"type": "gauge",
"description": "GPU KV-cache usage. 1 means 100 pe... | @@ -535,7 +535,11 @@ def metrics():
]
# Generate all metrics
- metrics_output = ""
+ metrics_output = """
+# HELP vllm:lora_requests_info Running stats on lora requests.
+# TYPE vllm:lora_requests_info gauge
+vllm:lora_requests_info{max_lora="1",running_lora_adapters="text2sql-lora-1",waiting_lora_ada... | development/app/app.py | 26 | Python | 0.286 | question | 88 | 51 | 51 | false | Mock app log fix for missing metrics warning | 564 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
if !ok {
continue
}
if model["id"] == instance.Name {
return true, nil
}
}
return false, nil
}
// Separate method to load the LoRA adapter
func (r *ModelAdapterReconciler) loadModelAdapter(host string, instance *modelv1alpha1.ModelAdapter) error {
artifactURL := instance.Spec.ArtifactURL
if strings.... | We can communicate later on how to also support the model on TOS. At present, the `/v1/model/download` interface provided on the runtime should come in handy. Let's create an issue to discuss about it | var response map[string]interface{}
if err := json.NewDecoder(resp.Body).Decode(&response); err != nil {
return false, err
}
data, ok := response["data"].([]interface{})
if !ok {
return false, errors.New("invalid data format")
}
for _, item := range data {
model, ok := item.(map[string]interface{})
if ... | @@ -599,10 +611,11 @@ func (r *ModelAdapterReconciler) modelAdapterExists(host string, instance *model
}
// Separate method to load the LoRA adapter
-func (r *ModelAdapterReconciler) loadModelAdapter(host string, instance *modelv1alpha1.ModelAdapter) error {
+func (r *ModelAdapterReconciler) loadModelAdapter(url st... | pkg/controller/modeladapter/modeladapter_controller.go | 26 | Go | 0.643 | suggestion | 200 | 51 | 51 | false | Support model registration flow using aibrix runtime api | 580 | vllm-project/aibrix | 2,572 | Go | brosoul | Jeffwan |
prometheus_multiproc_dir = os.environ["PROMETHEUS_MULTIPROC_DIR"]
# Note: ensure it will be automatically cleaned up upon exit.
path = Path(prometheus_multiproc_dir)
path.mkdir(parents=True, exist_ok=True)
if path.is_dir():
for item in path.iterdir():
if item.is_dir():
... | Is it better to add a conversion of `lower()` for `request_headers` here? | else:
prometheus_multiproc_dir = os.environ["PROMETHEUS_MULTIPROC_DIR"]
# Note: ensure it will be automatically cleaned up upon exit.
path = Path(prometheus_multiproc_dir)
path.mkdir(parents=True, exist_ok=True)
if path.is_dir():
for item in path.iterdir():
if item.is_di... | @@ -56,8 +57,27 @@ def initial_prometheus_multiproc_dir():
os.environ["PROMETHEUS_MULTIPROC_DIR"] = envs.PROMETHEUS_MULTIPROC_DIR
+# filter_and_set_headers only keeps necessary headers
+def filter_and_set_headers(request_headers):
+ # Create a new Headers object to ensure proper case-insensitive handling
+ ... | python/aibrix/aibrix/app.py | 26 | Python | 0.429 | question | 73 | 51 | 51 | false | Support model registration flow using aibrix runtime api | 580 | vllm-project/aibrix | 2,572 | Go | brosoul | Jeffwan |
},
{
"name": "gpu_cache_usage_perc",
"type": "gauge",
"description": "GPU KV-cache usage. 1 means 100 percent usage.",
"value": overrides.get(
"gpu_cache_usage_perc", gpu_cache_usage_perc
)
},
{
"name... | em. does user register a different name here? | },
{
"name": "gpu_cache_usage_perc",
"type": "gauge",
"description": "GPU KV-cache usage. 1 means 100 percent usage.",
"value": overrides.get(
"gpu_cache_usage_perc", gpu_cache_usage_perc
)
},
{
"name... | @@ -566,16 +566,19 @@ def metrics():
]
# Generate all metrics
- metrics_output = """
-# HELP vllm:lora_requests_info Running stats on lora requests.
-# TYPE vllm:lora_requests_info gauge
-vllm:lora_requests_info{max_lora="1",running_lora_adapters="text2sql-lora-1",waiting_lora_adapters=""} 1
-"""
+ me... | development/app/app.py | 26 | Python | 0.143 | question | 45 | 51 | 51 | false | Update routers to use GetPodModelMetric api and misc cleanup in metri… | 590 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
if len(evt.Choices) == 0 {
// Do not overwrite model, res can be empty.
usage = evt.Usage
}
}
if err := streaming.Err(); err != nil {
klog.ErrorS(err, "error to unmarshal response", "requestID", requestID, "responseBody", string(b.ResponseBody.GetBody()))
complete = true
return generateErrorR... | This check is used to handle error messages from the vLLM model. I assumed "model" from request header should equal to "res.Model" returned from the response. If that is not the case, replacing with len(res.Model) == 0 is ok. | if len(evt.Choices) == 0 {
// Do not overwrite model, res can be empty.
usage = evt.Usage
}
}
if err := streaming.Err(); err != nil {
klog.ErrorS(err, "error to unmarshal response", "requestID", requestID, "responseBody", string(b.ResponseBody.GetBody()))
complete = true
return generateErrorR... | @@ -410,17 +410,18 @@ func (s *Server) HandleResponseBody(ctx context.Context, requestID string, req *
Key: "x-error-response-unmarshal", RawValue: []byte("true"),
}}},
err.Error()), complete
- } else if res.Model != model { | pkg/plugins/gateway/gateway.go | 26 | Go | 0.643 | suggestion | 226 | 51 | 51 | false | Patch release for older vllm engine lora support in gateway plugins | 599 | vllm-project/aibrix | 2,572 | Go | zhangjyr | varungup90 |
"github.com/google/uuid"
"github.com/openai/openai-go"
"github.com/openai/openai-go/packages/ssestream"
"github.com/redis/go-redis/v9"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/status"
v1 "k8s.io/api/core/v1"
"k8s.io/client-go/kubernetes"
"k8s.io/klog/v2"
"github.com/aibrix/aibrix/pkg/cache"
r... | x-error-user and x-error-streaming |
"github.com/google/uuid"
"github.com/openai/openai-go"
"github.com/openai/openai-go/packages/ssestream"
"github.com/redis/go-redis/v9"
"google.golang.org/grpc/codes"
"google.golang.org/grpc/status"
v1 "k8s.io/api/core/v1"
"k8s.io/client-go/kubernetes"
"k8s.io/klog/v2"
"github.com/aibrix/aibrix/pkg/cache"
r... | @@ -48,14 +48,72 @@ import (
healthPb "google.golang.org/grpc/health/grpc_health_v1"
)
+const (
+ HeaderIncorrectRouting = "x-incorrect-routing-strategy"
+
+ // General Error Headers
+ HeaderUserError = "x-user-error" | pkg/plugins/gateway/gateway.go | 26 | Go | 0.286 | suggestion | 34 | 51 | 51 | false | [misc] Polish gateway code with better structure | 645 | vllm-project/aibrix | 2,572 | Go | varungup90 | Jeffwan |
const (
HeaderIncorrectRouting = "x-incorrect-routing-strategy"
// General Error Headers
HeaderUserError = "x-user-error"
HeaderRequestBodyProcessingError = "x-request-body-processing-error"
HeaderErrorRouting = "x-error-routing"
HeaderStreamingError = "x-streaming-erro... | right now words are not in order, we can it this way
x-error-*
x-update-*
x-exceeded-* | const (
HeaderErrorInvalidRouting = "x-error-invalid-routing-strategy"
// General Error Headers
HeaderErrorUser = "x-error-user"
HeaderErrorRouting = "x-error-routing"
HeaderErrorRequestBodyProcessing = "x-error-request-body-processing"
HeaderErrorResponseUnmarshal = "x-error-r... | @@ -48,14 +48,72 @@ import (
healthPb "google.golang.org/grpc/health/grpc_health_v1"
)
+const (
+ HeaderIncorrectRouting = "x-incorrect-routing-strategy"
+
+ // General Error Headers
+ HeaderUserError = "x-user-error"
+ HeaderRequestBodyProcessingError = "x-request-body-processing-error"
+ HeaderE... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.357 | suggestion | 91 | 51 | 51 | false | [misc] Polish gateway code with better structure | 645 | vllm-project/aibrix | 2,572 | Go | varungup90 | Jeffwan |
if !s.cache.CheckModelExists(model) {
klog.ErrorS(nil, "model doesn't exist in cache, probably wrong model name", "requestID", requestID, "model", model)
return generateErrorResponse(envoyTypePb.StatusCode_BadRequest,
[]*configPb.HeaderValueOption{{Header: &configPb.HeaderValue{
Key: HeaderNoModelDeployment... | both stream ones are errrors, should be updated to x-error-stream-* | if !s.cache.CheckModelExists(model) {
klog.ErrorS(nil, "model doesn't exist in cache, probably wrong model name", "requestID", requestID, "model", model)
return generateErrorResponse(envoyTypePb.StatusCode_BadRequest,
[]*configPb.HeaderValueOption{{Header: &configPb.HeaderValue{
Key: HeaderErrorNoModelBacke... | @@ -267,15 +331,15 @@ func (s *Server) HandleRequestBody(ctx context.Context, requestID string, req *e
klog.ErrorS(nil, "no stream option available", "requestID", requestID, "jsonMap", jsonMap)
return generateErrorResponse(envoyTypePb.StatusCode_InternalServerError,
[]*configPb.HeaderValueOption{{Header: &... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.429 | suggestion | 67 | 51 | 51 | false | [misc] Polish gateway code with better structure | 645 | vllm-project/aibrix | 2,572 | Go | varungup90 | Jeffwan |
// Spec defines the desired behavior of the PodAutoscaler.
Spec PodAutoscalerSpec `json:"spec,omitempty"`
// Status represents the current information about the PodAutoscaler.
Status PodAutoscalerStatus `json:"status,omitempty"`
}
// PodAutoscalerSpec defines the desired state of PodAutoscaler
type PodAutoscalerS... | user has to specify one metric to help autoscaling decide the next steps. I think this should be required? | Spec PodAutoscalerSpec `json:"spec,omitempty"`
// Status represents the current information about the PodAutoscaler.
Status PodAutoscalerStatus `json:"status,omitempty"`
}
// PodAutoscalerSpec defines the desired state of PodAutoscaler
type PodAutoscalerSpec struct {
// INSERT ADDITIONAL SPEC FIELDS - desired sta... | @@ -57,16 +56,19 @@ type PodAutoscalerSpec struct {
//PodSelector *metav1.LabelSelector `json:"podSelector,omitempty"`
// MinReplicas is the minimum number of replicas to which the target can be scaled down.
+ // +optional
MinReplicas *int32 `json:"minReplicas,omitempty"`
// MaxReplicas is the maximum numbe... | api/autoscaling/v1alpha1/podautoscaler_types.go | 26 | Go | 0.429 | question | 106 | 51 | 51 | false | fix api scheme | 674 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kerthcet |
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS... | The original design here is to refer a model. but we didn't add the model CRD here (due to the reason I synced with you earlier).
I feel there're two options
1. Once we have the model api, baseModel can ref that object.
2. at this moment, we probably can leverage PodSelector or objectRef to upgrade the `PodSelector... | you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS... | @@ -21,15 +21,13 @@ import (
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
)
-// EDIT THIS FILE! THIS IS SCAFFOLDING FOR YOU TO OWN!
// NOTE: json tags are required. Any new fields you add must have json tags for the fields to be serialized.
// ModelAdapterSpec defines the desired state of ModelAdapter
type... | api/model/v1alpha1/modeladapter_types.go | 26 | Go | 0.643 | suggestion | 439 | 51 | 51 | false | fix api scheme | 674 | vllm-project/aibrix | 2,572 | Go | Jeffwan | kerthcet |
option.WithMaxRetries(0),
respOpt,
)
}
func validateInference(t *testing.T, modelName string) {
client := createOpenAIClient(gatewayURL, apiKey)
validateInferenceWithClient(t, client, modelName)
}
func validateInferenceWithClient(t *testing.T, client *openai.Client, modelName string) {
chatCompletion, err := ... | I can not find reference here, is this being used? | /*
Copyright 2024 The Aibrix Team.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, softwar... | @@ -101,3 +110,38 @@ func createOpenAIClientWithRoutingStrategy(baseURL, apiKey, routingStrategy stri
respOpt,
)
}
+
+func validateInference(t *testing.T, modelName string) {
+ client := createOpenAIClient(gatewayURL, apiKey)
+ validateInferenceWithClient(t, client, modelName)
+}
+
+func validateInferenceWithClie... | test/e2e/util.go | 26 | Go | 0.286 | question | 50 | 40 | 133 | false | Add model adapter e2e tests | 701 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
return &extProcPb.ProcessingResponse{
Response: &extProcPb.ProcessingResponse_ResponseHeaders{
ResponseHeaders: &extProcPb.HeadersResponse{
Response: &extProcPb.CommonResponse{
HeaderMutation: &extProcPb.HeaderMutation{
SetHeaders: headers,
},
ClearRouteCache: true,
},
},
},
}
}... | in this case, we do not have any meaningful information right? | Header: &configPb.HeaderValue{
Key: HeaderTargetPod,
RawValue: []byte(targetPodIP),
},
})
}
var isProcessingError bool
var processingErrorCode int
for _, headerValue := range b.ResponseHeaders.Headers.Headers {
if headerValue.Key == ":status" {
code, _ := strconv.Atoi(string(headerValue.R... | @@ -421,6 +423,39 @@ func (s *Server) HandleResponseHeaders(ctx context.Context, requestID string, re
})
}
+ var isProcessingError bool
+ var processingErrorCode int
+ for _, headerValue := range b.ResponseHeaders.Headers.Headers {
+ if headerValue.Key == ":status" {
+ code, _ := strconv.Atoi(string(headerVal... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.214 | question | 63 | 51 | 51 | false | Process response headers in gateway | 703 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
package e2e
import (
"context"
"fmt"
"testing"
"time"
"github.com/stretchr/testify/assert"
modelv1alpha1 "github.com/vllm-project/aibrix/api/model/v1alpha1"
v1alpha1 "github.com/vllm-project/aibrix/pkg/client/clientset/versioned"
apierrors "k8s.io/apimachinery/pkg/api/errors"
v1 "k8s.io/apimachinery/pkg/apis... | @gangmuk reports missing license header failed the check. However, license check job was not triggered. not sure what's the problem. https://github.com/vllm-project/aibrix/issues/723
good to see it's added now | /*
Copyright 2024 The Aibrix Team.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
d... | @@ -1,3 +1,19 @@
+/* | test/e2e/model_adapter_test.go | 1 | Go | 0.429 | bug | 212 | 26 | 26 | false | Process response headers in gateway | 703 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
Key: HeaderWentIntoReqHeaders,
RawValue: []byte("true"),
},
}}
if targetPodIP != "" {
headers = append(headers, &configPb.HeaderValueOption{
Header: &configPb.HeaderValue{
Key: HeaderTargetPod,
RawValue: []byte(targetPodIP),
},
})
}
return &extProcPb.ProcessingResponse{
Respon... | 30x redirection are successful codes. Seems we will never have such case? Is it possible to have 20x? here? better to have a helper method and let's consider those cases |
func (s *Server) HandleResponseHeaders(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest, targetPodIP string) (*extProcPb.ProcessingResponse, bool, int) {
klog.InfoS("-- In ResponseHeaders processing ...", "requestID", requestID)
b := req.Request.(*extProcPb.ProcessingRequest_ResponseHeaders)
... | @@ -421,6 +428,24 @@ func (s *Server) HandleResponseHeaders(ctx context.Context, requestID string, re
})
}
+ var isProcessingError bool
+ var processingErrorCode int
+ for _, headerValue := range b.ResponseHeaders.Headers.Headers {
+ if headerValue.Key == ":status" {
+ code, _ := strconv.Atoi(string(headerVal... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.429 | refactor | 170 | 51 | 51 | false | Process response headers in gateway | 703 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
)
text_chunks = []
prompt_tokens = 0
output_tokens = 0
total_tokens = 0
async for chunk in stream:
if chunk.choices:
if chunk.choices[0].delta.content is not None:
if not first_response_time:
... | result record seems like a little bit duplicated. this is ok at this moment. | )
text_chunks = []
prompt_tokens = 0
output_tokens = 0
total_tokens = 0
async for chunk in stream:
if chunk.choices:
if chunk.choices[0].delta.content is not None:
if not first_response_time:
... | @@ -46,25 +46,43 @@ async def send_request_streaming(client: openai.AsyncOpenAI,
output_tokens = chunk.usage.completion_tokens
total_tokens = chunk.usage.total_tokens
response = "".join(text_chunks)
- response_time = asyncio.get_event_loop().time()
- late... | benchmarks/client/client.py | 26 | Python | 0.214 | refactor | 77 | 51 | 51 | false | Recording failed requests in benchmark client | 727 | vllm-project/aibrix | 2,572 | Go | Jeffwan | gangmuk |
func (n *TreeNode) GetDepth() int {
return n.depth
}
func (n *TreeNode) GetID() int {
return n.id
}
func (n *TreeNode) GetChildren() map[int]*TreeNode {
return n.children
}
func (n *TreeNode) ResetEvictedPods() {
n.evictedPods = make(map[int]bool)
}
func (n *TreeNode) ResetCachedPods() {
n.cachedPods = make(m... | you can update the comments later. |
func (n *TreeNode) GetDepth() int {
return n.depth
}
func (n *TreeNode) GetID() int {
return n.id
}
func (n *TreeNode) GetChildren() map[int]*TreeNode {
return n.children
}
func (n *TreeNode) ResetEvictedPods() {
n.evictedPods = make(map[int]bool)
}
func (n *TreeNode) ResetCachedPods() {
n.cachedPods = make(m... | @@ -72,6 +112,114 @@ func (n *TreeNode) GetDepth() int {
return n.depth
}
+func (n *TreeNode) GetID() int {
+ return n.id
+}
+
+func (n *TreeNode) GetChildren() map[int]*TreeNode {
+ return n.children
+}
+
+func (n *TreeNode) ResetEvictedPods() {
+ n.evictedPods = make(map[int]bool)
+}
+
+func (n *TreeNode) ResetC... | pkg/plugins/gateway/prefixcacheindexer/tree.go | 26 | Go | 0.071 | suggestion | 34 | 51 | 51 | false | Improve thread safety for TreeNode data structure and refactor related codes | 730 | vllm-project/aibrix | 2,572 | Go | varungup90 | gangmuk |
client = openai.AsyncOpenAI(
api_key=args.api_key,
base_url=args.endpoint + "/v1",
)
if args.routing_strategy is not None:
client.default_headers["routing-strategy"] = args.routing_strategy
if not args.streaming:
logging.info("Using batch c... | em. that's something we didn't review clearly. good fix | cur_time = time.time()
target_time = base_time + ts / 1000.0
logging.warning(f"Prepare to launch {len(requests)} batched tasks after {target_time - cur_time}")
if target_time > cur_time:
await asyncio.sleep(target_time - cur_time)
formatted_prompts = [wrap_prompt_as_c... | @@ -199,7 +219,9 @@ def main(args):
base_url=args.endpoint + "/v1",
)
if args.routing_strategy is not None:
- client.default_headers["routing-strategy"] = args.routing_strategy
+ client = client.with_options( | benchmarks/client/client.py | 26 | Python | 0.214 | suggestion | 55 | 45 | 51 | false | [fix] fixing bug in using AsyncOpenAI client (header setting, token counting, etc) | 738 | vllm-project/aibrix | 2,572 | Go | Jeffwan | gangmuk |
current_value = round(trend + noise)
if pattern_config['only_rise']:
current_value = max(prev_value, current_value)
prev_value = current_value
return current_value, prev_value
assert duration_ms is not None and interval_ms is not None, \
"duration_ms and inte... | Concurrency could be the misleading term here. As far as I know, concurrency or concurrent requests usually means (the number of) requests that are concurrently running in the system. For example, concurrency control is controlling that. By that definition, workload generator can't control concurrency. So what it meant... | current_value = round(trend + noise)
if pattern_config['only_rise']:
current_value = max(prev_value, current_value)
prev_value = current_value
return current_value, prev_value
assert duration_ms is not None and interval_ms is not None, \
"duration_ms and inte... | @@ -242,32 +246,33 @@ def math_function(t, pattern_config, length, prev_value):
previous_input_len = -1
previous_output_len = -1
ts = 0
- base_req_id = 0
- sharegpt_df = load_requests(dataset_path=prompt_file_path, tokenizer=tokenizer)
+ rps_dist = []
+ input_token_len_dist = []
+ ou... | benchmarks/generator/workload_generator.py | 26 | Python | 0.5 | suggestion | 395 | 51 | 51 | false | Reconfigure workload generator for predefined synthetic patterns | 771 | vllm-project/aibrix | 2,572 | Go | gangmuk | happyandslow |
klog.Infof("using default prefix cache match threshold percent: %d", defaultPrefixCacheMatchThresholdPercent)
return defaultPrefixCacheMatchThresholdPercent
}
type prefixCacheRouter struct {
prefixCacheIndexer prefixcacheindexer.PrefixCacheIndexer
}
func NewPrefixCacheRouter() (Router, error) {
return prefixCache... | Can we have some interface here so we can probably configure it for differenc scenarios? My concern is I think it's ok to split by `" "` for english,, for some other languages like chinese, this might not be helpful. | if value != "" {
intValue, err := strconv.Atoi(value)
if err != nil || intValue <= 0 || intValue > 100 {
klog.Infof("invalid AIBRIX_PREFIX_CACHE_MATCH_THRESHOLD_PERCENT: %s, valid value between 0 and 100, failing back to default", value)
} else {
klog.Infof("using AIBRIX_PREFIX_CACHE_MATCH_THRESHOLD_PERCEN... | @@ -72,10 +73,7 @@ func (p prefixCacheRouter) Route(ctx context.Context, pods map[string]*v1.Pod, m
}
}
- tokens, err := utils.TokenizeInputText(message)
- if err != nil {
- return "", err
- }
+ tokens := strings.Split(message, " ") | pkg/plugins/gateway/algorithms/prefix_cache.go | 26 | Go | 0.571 | question | 218 | 51 | 51 | false | Use string based tokenizer in prefix cache | 774 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
type Cache struct {
mu sync.RWMutex
redisClient *redis.Client
prometheusApi prometheusv1.API
initialized bool
subscribers []metrics.MetricSubscriber
metrics map[string]interface{}
ModelMetrics map[string]map[string]interface{}
Pods map[string]*v1.... | I am thinking if we want to add a aibrix label like `orchestration.aibrix.ai/server-pod`. autoscaling need that feature as well. We can start with ray specific label and replace with our own later as a separate issue
| type Cache struct {
mu sync.RWMutex
redisClient *redis.Client
prometheusApi prometheusv1.API
initialized bool
subscribers []metrics.MetricSubscriber
metrics map[string]interface{}
ModelMetrics map[string]map[string]interface{}
Pods map[string]*v1.... | @@ -75,6 +75,7 @@ type Block struct {
const (
modelIdentifier = "model.aibrix.ai/name"
+ nodeType = "ray.io/node-type" | pkg/cache/cache.go | 26 | Go | 0.643 | suggestion | 218 | 51 | 51 | false | Ignore worker pods for gateway routing | 776 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
"net/http"
"github.com/gorilla/mux"
"github.com/redis/go-redis/v9"
"github.com/vllm-project/aibrix/pkg/cache"
"github.com/vllm-project/aibrix/pkg/utils"
"k8s.io/klog/v2"
)
type httpServer struct {
redisClient *redis.Client
cache *cache.Cache
}
func NewHTTPServer(addr string, redis *redis.Client) *http.... | These are user related handlers. | "net/http"
"github.com/gorilla/mux"
"github.com/redis/go-redis/v9"
"github.com/vllm-project/aibrix/pkg/cache"
"github.com/vllm-project/aibrix/pkg/utils"
"k8s.io/klog/v2"
)
type httpServer struct {
redisClient *redis.Client
cache *cache.Cache
}
func NewHTTPServer(addr string, redis *redis.Client) *http.... | @@ -45,10 +45,12 @@ func NewHTTPServer(addr string, redis *redis.Client) *http.Server {
cache: c,
}
r := mux.NewRouter()
+ // Rate Limit related handlers | pkg/metadata/handlers.go | 26 | Go | 0.286 | suggestion | 32 | 51 | 51 | false | Support OpenAI api style /v1/models response | 829 | vllm-project/aibrix | 2,572 | Go | varungup90 | Jeffwan |
// +kubebuilder:validation:Optional
// +kubebuilder:default:="IfNotPresent"
ImagePullPolicy string `json:"imagePullPolicy,omitempty"`
// shared memory size for kvcach
// +kubebuilder:validation:Optional
// +kubebuilder:default:=""
SharedMemorySize string `json:"sharedMemorySize,omitempty"`
// kvcache environm... | we can use annotation to control it for short term, we can have an `v1alpha2` version for spec change later. seems your changes below doesn't leverage this yet. | // +kubebuilder:validation:Optional
// +kubebuilder:default:="IfNotPresent"
ImagePullPolicy string `json:"imagePullPolicy,omitempty"`
// shared memory size for kvcach
// +kubebuilder:validation:Optional
// +kubebuilder:default:=""
SharedMemorySize string `json:"sharedMemorySize,omitempty"`
// kvcache environm... | @@ -104,6 +104,26 @@ type CacheSpec struct {
CPU string `json:"cpu,omitempty"`
}
+type AffinityConfig struct { | api/orchestration/v1alpha1/kvcache_types.go | 26 | Go | 0.571 | suggestion | 160 | 51 | 51 | false | Added antiaffinity in kvcache crd | 865 | vllm-project/aibrix | 2,572 | Go | Jeffwan | gangmuk |
{
LabelSelector: &metav1.LabelSelector{
MatchExpressions: []metav1.LabelSelectorRequirement{
{
Key: "model.aibrix.ai/name",
Operator: metav1.LabelSelectorOpIn,
Values: []string{value},
},
},
},
TopologyKey: "kubernetes.io/hostname",
},
}... | what's the usage here? do you want to apply anti-affinity to other engine pods or just kv cache pod?
the cache pod doesn't have `model.aibrix.ai/name` labels, right? | {
LabelSelector: &metav1.LabelSelector{
MatchExpressions: []metav1.LabelSelectorRequirement{
{
Key: "model.aibrix.ai/name",
Operator: metav1.LabelSelectorOpIn,
Values: []string{value},
},
},
},
TopologyKey: "kubernetes.io/hostname",
},
}... | @@ -463,6 +465,52 @@ func (r *KVCacheReconciler) reconcileDeployment(ctx context.Context, kvCache *or
affinity.PodAffinity = podAffinity
}
+ // Add pod anti-affinity handling
+ if value, ok := kvCache.Annotations[KVCacheAnnotationPodAntiAffinityKey]; ok {
+ podAntiAffinity := &corev1.PodAntiAffinity{
+ Requir... | pkg/controller/kvcache/kvcache_controller.go | 26 | Go | 0.5 | performance | 170 | 51 | 51 | false | Added antiaffinity in kvcache crd | 865 | vllm-project/aibrix | 2,572 | Go | Jeffwan | gangmuk |
func TestLRUStore_TTL(t *testing.T) {
store := NewLRUStore[string, string](2, 2*time.Second, 1*time.Second)
// Test TTL expiration
store.Put("key1", "value1")
time.Sleep(3 * time.Second) // Wait for TTL to expire
if _, ok := store.Get("key1"); ok {
t.Errorf("expected key1 to be expired")
}
}
func TestLRUSto... | Can you also add a test to validate least recently used key is evicted? |
func TestLRUStore_TTL(t *testing.T) {
store := NewLRUStore[string, string](2, 2*time.Second, 1*time.Second, DefaultGetCurrentTime)
// Test TTL expiration
store.Put("key1", "value1")
time.Sleep(3 * time.Second) // Wait for TTL to expire
if _, ok := store.Get("key1"); ok {
t.Errorf("expected key1 to be expired"... | @@ -0,0 +1,105 @@
+/*
+Copyright 2025 The Aibrix Team.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agre... | pkg/plugins/gateway/cache/lru_store_test.go | 26 | Go | 0.286 | question | 71 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
... | increase the default block size to 1 million. | http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
... | @@ -23,24 +23,42 @@ import (
"time"
"github.com/cespare/xxhash/v2"
+ "github.com/vllm-project/aibrix/pkg/plugins/gateway/cache"
"github.com/vllm-project/aibrix/pkg/utils"
v1 "k8s.io/api/core/v1"
"k8s.io/klog/v2"
)
const (
+ defaultPrefixCacheBlockNumber = 100 | pkg/plugins/gateway/prefixcacheindexer/hash.go | 26 | Go | 0.071 | suggestion | 45 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package pr... | you can rename the variable to defaultPrefixCacheEvictionInternalInSec | Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package pr... | @@ -23,24 +23,42 @@ import (
"time"
"github.com/cespare/xxhash/v2"
+ "github.com/vllm-project/aibrix/pkg/plugins/gateway/cache"
"github.com/vllm-project/aibrix/pkg/utils"
v1 "k8s.io/api/core/v1"
"k8s.io/klog/v2"
)
const (
+ defaultPrefixCacheBlockNumber = 100
defaultPrefixCacheBlockSize ... | pkg/plugins/gateway/prefixcacheindexer/hash.go | 26 | Go | 0.286 | performance | 70 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and... | Right now the structure is as follows
hash1 -> model1 -> pod1 + pod1 last access time
-> model1 -> pod2 + pod2 last access time
-> model2 -> pod3 + pod3 last access time
-> blocks last access time
Right now eviction is based of last blocks access time. I kept the eviction l... |
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and... | @@ -0,0 +1,159 @@
+/*
+Copyright 2025 The Aibrix Team.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agre... | pkg/plugins/gateway/cache/lru_store.go | 26 | Go | 0.5 | suggestion | 664 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
}
func (e *LRUStore[K, V]) evict(now time.Time) {
var keysToEvict []K
e.Lock()
for key, entry := range e.freeTable {
if now.Sub(entry.lastAccessTime) > e.ttl {
keysToEvict = append(keysToEvict, key)
}
}
e.Unlock()
for _, key := range keysToEvict {
e.Lock()
if entry, exists := e.freeTable[key]; exist... | here. I am curious whether it need key level lock? could it be for loop level? | e.lruList.moveToHead(entry)
return entry.Value, true
}
var zero V
return zero, false
}
func (e *LRUStore[K, V]) Len() int {
e.RLock()
defer e.RUnlock()
return len(e.freeTable)
}
func (e *LRUStore[K, V]) evict(now time.Time) {
var keysToEvict []K
e.RLock()
for key, entry := range e.freeTable {
if now.S... | @@ -0,0 +1,159 @@
+/*
+Copyright 2025 The Aibrix Team.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agre... | pkg/plugins/gateway/cache/lru_store.go | 26 | Go | 0.286 | question | 79 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | Jeffwan | vie-serendipity |
/*
Copyright 2025 The Aibrix Team.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
d... | if this is empty, tail.prev==head. does it still remove elements?
Should we check
```
if l.tail.prev == l.head {
return nil
}
``` | e.next = l.head.next
l.head.next.prev = e
l.head.next = e
}
func (l *list[K, V]) moveToHead(e *entry[K, V]) {
l.remove(e)
l.addToHead(e)
}
func (l *list[K, V]) remove(e *entry[K, V]) {
if e.prev != nil {
e.prev.next = e.next
}
if e.next != nil {
e.next.prev = e.prev
}
e.prev = nil
e.next = nil
}
func ... | @@ -0,0 +1,159 @@
+/*
+Copyright 2025 The Aibrix Team.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agre... | pkg/plugins/gateway/cache/lru_store.go | 26 | Go | 0.786 | suggestion | 142 | 160 | 30 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | Jeffwan | vie-serendipity |
/*
Copyright 2025 The Aibrix Team.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
d... | RWMutex for `Get()` or `Len()` Operations? | /*
Copyright 2025 The Aibrix Team.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
d... | @@ -0,0 +1,159 @@
+/*
+Copyright 2025 The Aibrix Team.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+ http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agre... | pkg/plugins/gateway/cache/lru_store.go | 25 | Go | 0.214 | question | 42 | 50 | 50 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | Jeffwan | vie-serendipity |
}
}
klog.Infof("using default prefix cache eviction interval: %d ms", defaultPrefixCacheEvictionInternalInMS)
return defaultPrefixCacheEvictionInternalInMS * time.Millisecond
}
func getPrefixCacheEvictionDuration() time.Duration {
value := utils.LoadEnv("AIBRIX_PREFIX_CACHE_EVICTION_DURATION_MINS", "")
if value... | Using cache interface here to replace blocks looks better. this is great | }
}
klog.Infof("using default prefix cache eviction interval: %d ms", defaultPrefixCacheEvictionInternalInSec)
return defaultPrefixCacheEvictionInternalInSec * time.Second
}
func getPrefixCacheEvictionDuration() time.Duration {
value := utils.LoadEnv("AIBRIX_PREFIX_CACHE_EVICTION_DURATION_MINS", "")
if value !=... | @@ -87,33 +105,27 @@ func getPrefixCacheEvictionDuration() time.Duration {
}
type PrefixHashTable struct {
- mu sync.RWMutex
- blocks map[uint64]Block
- hash *xxhash.Digest
- seed uint64
+ mu sync.RWMutex
+ hash *xxhash.Digest
+ seed uint64
+ store cache.Store[uint64, Block] | pkg/plugins/gateway/prefixcacheindexer/hash.go | 26 | Go | 0.286 | performance | 72 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | Jeffwan | vie-serendipity |
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
... | this value determines the size of the cache, I feel 100 is too small. We need to change default one to large values. Later, the prefix will be changed to use character instead of word or tokens, results in more blocks. 20000 token request end up with 1250 blocks. we can roughly calculate the memory size (like 1GiB) a... | http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
... | @@ -23,24 +23,42 @@ import (
"time"
"github.com/cespare/xxhash/v2"
+ "github.com/vllm-project/aibrix/pkg/plugins/gateway/cache"
"github.com/vllm-project/aibrix/pkg/utils"
v1 "k8s.io/api/core/v1"
"k8s.io/klog/v2"
)
const (
+ defaultPrefixCacheBlockNumber = 100 | pkg/plugins/gateway/prefixcacheindexer/hash.go | 26 | Go | 1 | performance | 701 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | Jeffwan | vie-serendipity |
} else {
klog.Infof("using AIBRIX_PREFIX_CACHE_BLOCK_NUMBER env value for prefix cache block number: %d", intValue)
return intValue
}
}
klog.Infof("using default prefix cache block number: %d", defaultPrefixCacheBlockNumber)
return defaultPrefixCacheBlockNumber
}
func getPrefixCacheBlockSize() int {
valu... | nit: Secs -> SECONDS or SECS (prefer first one) | } else {
klog.Infof("using AIBRIX_PREFIX_CACHE_BLOCK_NUMBER env value for prefix cache block number: %d", intValue)
return intValue
}
}
klog.Infof("using default prefix cache block number: %d", defaultPrefixCacheBlockNumber)
return defaultPrefixCacheBlockNumber
}
func getPrefixCacheBlockSize() int {
valu... | @@ -57,18 +75,18 @@ func getPrefixCacheBlockSize() int {
}
func getPrefixCacheEvictionInterval() time.Duration {
- value := utils.LoadEnv("AIBRIX_PREFIX_CACHE_EVICTION_INTERVAL_MS", "")
+ value := utils.LoadEnv("AIBRIX_PREFIX_CACHE_EVICTION_INTERVAL_Secs", "") | pkg/plugins/gateway/prefixcacheindexer/hash.go | 26 | Go | 0.143 | suggestion | 47 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
klog.Infof("using AIBRIX_PREFIX_CACHE_EVICTION_DURATION_MINS env value for prefix cache eviction duration: %d ms", intValue)
return time.Duration(intValue) * time.Minute
}
}
klog.Infof("using default prefix cache eviction duration: %d mins", defaultPrefixCacheEvictionDurationInMins)
return defaultPrefixCache... | nit: since variables are already assigned, use them directly. | klog.Infof("using AIBRIX_PREFIX_CACHE_EVICTION_DURATION_MINS env value for prefix cache eviction duration: %d ms", intValue)
return time.Duration(intValue) * time.Minute
}
}
klog.Infof("using default prefix cache eviction duration: %d mins", defaultPrefixCacheEvictionDurationInMins)
return defaultPrefixCache... | @@ -87,33 +105,27 @@ func getPrefixCacheEvictionDuration() time.Duration {
}
type PrefixHashTable struct {
- mu sync.RWMutex
- blocks map[uint64]Block
- hash *xxhash.Digest
- seed uint64
+ mu sync.RWMutex
+ hash *xxhash.Digest
+ seed uint64
+ store cache.Store[uint64, Block]
}
type Block struct {
-... | pkg/plugins/gateway/prefixcacheindexer/hash.go | 26 | Go | 0.286 | nitpick | 61 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
getPrefixCacheEvictionDuration(),
getPrefixCacheEvictionInterval()),
}
return instance
}
// returns matchedTokens, unMatchedTokens, matchedPods
func (c *PrefixHashTable) MatchPrefix(tokens []byte, model string, pods []*v1.Pod) ([]byte, []byte, []*v1.Pod) {
c.mu.Lock()
defer c.mu.Unlock()
var block, lastMat... | store is thread-safe, so we can remove lock from line 134 now. Right? | prefixCacheEvictionDuration,
prefixCacheEvictionInterval,
func() time.Time { return time.Now() }),
}
return instance
}
// returns matchedTokens, unMatchedTokens, matchedPods
func (c *PrefixHashTable) MatchPrefix(tokens []byte, model string, pods []*v1.Pod) ([]byte, []byte, []*v1.Pod) {
var block, lastMatc... | @@ -134,17 +146,15 @@ func (c *PrefixHashTable) MatchPrefix(tokens []byte, model string, pods []*v1.Po
_, _ = c.hash.Write(tokens[i:end])
prefixHash := c.hash.Sum64()
c.hash.ResetWithSeed(c.seed)
-
- block, ok = c.blocks[prefixHash]
+ block, ok = c.store.Get(prefixHash) | pkg/plugins/gateway/prefixcacheindexer/hash.go | 26 | Go | 0.429 | question | 69 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
{ObjectMeta: metav1.ObjectMeta{Name: "p1"}},
{ObjectMeta: metav1.ObjectMeta{Name: "p2"}},
}
inputText := "Hello World! What a Good Day! Good day to code and learn new things in LLM!! 你好世界！ 多么美好的一天啊！"
tokens, err := tokenizer.NewStringTokenizer().TokenizeInputText(inputText)
assert.NoError(t, err)
matchedToke... | In the test, we can not sleep for 25 seconds. Can you add a mock for time or update the env variable in test for lower value. | }
pods := []*v1.Pod{
{ObjectMeta: metav1.ObjectMeta{Name: "p1"}},
{ObjectMeta: metav1.ObjectMeta{Name: "p2"}},
}
inputText := "Hello World! What a Good Day! Good day to code and learn new things in LLM!! 你好世界！ 多么美好的一天啊！"
tokens, err := tokenizer.NewStringTokenizer().TokenizeInputText(inputText)
assert.NoErro... | @@ -63,7 +64,7 @@ func Test_PrefixHashTableE2E(t *testing.T) {
assert.Equal(t, 0, len(unMatchedTokens))
assert.Equal(t, "p1", matchPods[0].Name)
- cache.Evict(time.Now().Add(60 * time.Minute))
+ time.Sleep(25 * time.Second) | pkg/plugins/gateway/prefixcacheindexer/hash_test.go | 26 | Go | 0.357 | suggestion | 125 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
modelToPods map[string]map[string]time.Time // model_name: map[pod_name]pod_last_access_time
lastAccessTime time.Time // block_last_access_time
}
func NewPrefixHashTable() PrefixCacheIndexer {
r := rand.New(rand.NewSource(time.Now().Unix()))
seed := r.Uint64()
instance := &PrefixHashTable... | @vie-serendipity I assumed LRU store is thread-safe and removing this lock was fine. While working on PR #933 I came across the bug for concurrent read/write for blocks.
I have added the lock back, if you get a chance, can you make LRU store thread-safe? Thank you. | klog.Infof("invalid AIBRIX_PREFIX_CACHE_EVICTION_DURATION_MINS: %s, falling back to default", value)
} else {
klog.Infof("using AIBRIX_PREFIX_CACHE_EVICTION_DURATION_MINS env value for prefix cache eviction duration: %d ms", intValue)
return time.Duration(intValue) * time.Minute
}
}
klog.Infof("using def... | @@ -87,40 +105,33 @@ func getPrefixCacheEvictionDuration() time.Duration {
}
type PrefixHashTable struct {
- mu sync.RWMutex
- blocks map[uint64]Block
- hash *xxhash.Digest
- seed uint64
+ mu sync.RWMutex
+ hash *xxhash.Digest
+ seed uint64
+ store cache.Store[uint64, Block]
}
type Block struct {
-... | pkg/plugins/gateway/prefixcacheindexer/hash.go | 26 | Go | 0.571 | bug | 268 | 51 | 51 | false | [API] [Misc]: Support LRU cache with TTL for prefix cache indexer | 905 | vllm-project/aibrix | 2,572 | Go | varungup90 | vie-serendipity |
cluster := clusters[i]
wg.Add(1)
go func(cluster rayclusterv1.RayCluster) {
defer wg.Done()
if err := r.Delete(ctx, &cluster); err != nil {
if !apierrors.IsNotFound(err) {
errCh <- fmt.Errorf("failed to delete pod: %w", err)
}
}
r.Expectations.DeletionObserved(types.NamespacedName{Namespa... | using aggregated errors looks good to me. | const maxConcurrency = 10
semaphore := make(chan struct{}, maxConcurrency)
for i := 0; i < diff; i++ {
cluster := clusters[i]
semaphore <- struct{}{}
wg.Add(1)
go func(cluster rayclusterv1.RayCluster) {
defer wg.Done()
defer func() { <-semaphore }() // release the semaphore
if err := r.Delete(ctx, ... | @@ -192,13 +198,18 @@ func (r *RayClusterReplicaSetReconciler) scaleDown(ctx context.Context, replicas
}
wg.Wait()
+ close(errCh)
- select {
- case err := <-errCh:
- return err
- default:
- return nil
+ // aggregate errors if any.
+ var aggregatedErrors []error
+ for err := range errCh {
+ aggregatedErrors =... | pkg/controller/rayclusterreplicaset/rayclusterreplicaset_controller.go | 26 | Go | 0.214 | suggestion | 41 | 51 | 51 | false | [Misc] chore(raycluster): add concurrency limit and error aggregation to scaleDown | 914 | vllm-project/aibrix | 2,572 | Go | Jeffwan | googs1025 |
if err := r.updateReplicaSetStatus(ctx, replicaset, newStatus, rsKey); err != nil {
return reconcile.Result{}, err
}
return ctrl.Result{}, nil
}
// scaleUp handles RayCluster creation logic when scaling up
func (r *RayClusterReplicaSetReconciler) scaleUp(ctx context.Context, replicaset *orchestrationv1alpha1.Ray... | it's ok to have a fixed value at this moment | if err := r.updateReplicaSetStatus(ctx, replicaset, newStatus, rsKey); err != nil {
return reconcile.Result{}, err
}
return ctrl.Result{}, nil
}
// scaleUp handles RayCluster creation logic when scaling up
func (r *RayClusterReplicaSetReconciler) scaleUp(ctx context.Context, replicaset *orchestrationv1alpha1.Ray... | @@ -176,12 +176,18 @@ func (r *RayClusterReplicaSetReconciler) scaleUp(ctx context.Context, replicaset
func (r *RayClusterReplicaSetReconciler) scaleDown(ctx context.Context, replicaset *orchestrationv1alpha1.RayClusterReplicaSet, clusters []rayclusterv1.RayCluster, diff int) error {
var wg sync.WaitGroup
errCh :=... | pkg/controller/rayclusterreplicaset/rayclusterreplicaset_controller.go | 26 | Go | 0.071 | suggestion | 44 | 51 | 51 | false | [Misc] chore(raycluster): add concurrency limit and error aggregation to scaleDown | 914 | vllm-project/aibrix | 2,572 | Go | Jeffwan | googs1025 |
"time"
"github.com/vllm-project/aibrix/pkg/metrics"
"github.com/vllm-project/aibrix/pkg/types"
"github.com/vllm-project/aibrix/pkg/utils"
"k8s.io/klog/v2"
)
const (
expireWriteRequestTraceIntervalInMins = 10
)
func (c *Store) getRequestTrace(modelName string) *RequestTrace {
trace := NewRequestTrace(time.Now(... | For the log messages, you can add the requestID (add to routing context if not present) and pod name. Without these two fields, log message is not useful.
Comment applies to all log messages in this function. | "time"
"github.com/vllm-project/aibrix/pkg/metrics"
"github.com/vllm-project/aibrix/pkg/types"
"github.com/vllm-project/aibrix/pkg/utils"
"k8s.io/klog/v2"
)
const (
expireWriteRequestTraceIntervalInMins = 10
)
func (c *Store) getRequestTrace(modelName string) *RequestTrace {
trace := NewRequestTrace(time.Now(... | @@ -41,6 +43,43 @@ func (c *Store) getRequestTrace(modelName string) *RequestTrace {
return newer
}
+func (c *Store) addPodStats(ctx *types.RoutingContext) {
+ if !ctx.HasRouted() {
+ klog.Warning("request has not been routed, please route request first") | pkg/cache/cache_trace.go | 26 | Go | 0.5 | suggestion | 211 | 51 | 51 | false | [Misc] Add pods stats example: running requests. | 918 | vllm-project/aibrix | 2,572 | Go | varungup90 | zhangjyr |
"fmt"
"sync/atomic"
"time"
"github.com/vllm-project/aibrix/pkg/metrics"
"github.com/vllm-project/aibrix/pkg/types"
"github.com/vllm-project/aibrix/pkg/utils"
"k8s.io/klog/v2"
)
const (
expireWriteRequestTraceIntervalInMins = 10
)
func (c *Store) getRequestTrace(modelName string) *RequestTrace {
trace := New... | Is there a unit test to validate add pod stats, specially with concurrent requests. | "fmt"
"sync/atomic"
"time"
"github.com/vllm-project/aibrix/pkg/metrics"
"github.com/vllm-project/aibrix/pkg/types"
"github.com/vllm-project/aibrix/pkg/utils"
"k8s.io/klog/v2"
)
const (
expireWriteRequestTraceIntervalInMins = 10
)
func (c *Store) getRequestTrace(modelName string) *RequestTrace {
trace := New... | @@ -41,6 +43,43 @@ func (c *Store) getRequestTrace(modelName string) *RequestTrace {
return newer
}
+func (c *Store) addPodStats(ctx *types.RoutingContext) { | pkg/cache/cache_trace.go | 26 | Go | 0.214 | nitpick | 83 | 51 | 51 | false | [Misc] Add pods stats example: running requests. | 918 | vllm-project/aibrix | 2,572 | Go | varungup90 | zhangjyr |
func (c *Store) DoneRequestCount(ctx *types.RoutingContext, requestID string, modelName string, traceTerm int64) {
if ctx != nil {
c.donePodStats(ctx)
}
meta, ok := c.metaModels.Load(modelName)
if ok {
atomic.AddInt32(&meta.pendingRequests, -1)
}
// DoneRequest only works for current term, no need to retry.... | Right now, done request trace is called in gateway_resp_body. I identified two scenario where a request can complete before executing resp_body. You can use this PR to address this gap.
1) If error code is returned in resp_headers.
2) If the client cancels the request and error is return from here `if err := srv.Se... | func (c *Store) DoneRequestCount(ctx *types.RoutingContext, requestID string, modelName string, traceTerm int64) {
if ctx != nil {
c.donePodStats(ctx, requestID)
}
meta, ok := c.metaModels.Load(modelName)
if ok {
atomic.AddInt32(&meta.pendingRequests, -1)
}
// DoneRequest only works for current term, no nee... | @@ -214,6 +222,10 @@ func (c *Store) DoneRequestCount(ctx *types.RoutingContext, requestID string, mo
// outputTokens: Output tokens count
// traceTerm: Trace term identifier
func (c *Store) DoneRequestTrace(ctx *types.RoutingContext, requestID string, modelName string, inputTokens, outputTokens, traceTerm int64) { | pkg/cache/cache_impl.go | 26 | Go | 0.786 | suggestion | 343 | 51 | 51 | false | [Misc] Add pods stats example: running requests. | 918 | vllm-project/aibrix | 2,572 | Go | varungup90 | zhangjyr |
import argparse
import logging
import time
import asyncio
import openai
import json
import io
import traceback
from typing import List
from utils import (load_workload, wrap_prompt_as_chat_message)
logging.basicConfig(level=logging.INFO)
async def send_request_streaming(client: openai.AsyncOpenAI,
... | is multi-threading necessary for this client? | import argparse
import logging
import time
import asyncio
import openai
import json
import io
import traceback
import threading
from queue import Queue
from concurrent.futures import ThreadPoolExecutor
from typing import List
from utils import (load_workload, wrap_prompt_as_chat_message)
logging.basicConfig(level=lo... | @@ -6,25 +6,50 @@
import json
import io
import traceback
+import threading
+from queue import Queue
+from concurrent.futures import ThreadPoolExecutor
from typing import List
from utils import (load_workload, wrap_prompt_as_chat_message)
logging.basicConfig(level=logging.INFO)
+task_queue = Queue()
+thread_... | benchmarks/client/client.py | 21 | Python | 0.214 | question | 47 | 46 | 46 | false | Enable parallel client using thread pool in benchmark client | 919 | vllm-project/aibrix | 2,572 | Go | Jeffwan | happyandslow |
var controllers string
var modeladapterSchedulerPolicy string
var enableRuntimeSidecar bool
var debugMode bool
flag.StringVar(&metricsAddr, "metrics-bind-address", ":8080", "The address the metric endpoint binds to.")
flag.StringVar(&probeAddr, "health-probe-bind-address", ":8081", "The address the probe endpoint... | I am thinking should we use `model-adapter` or just `modeladapter`. I did see our controller name use hyphen. that's probably more readable since this is a user facing feature. what do you think? | var controllers string
var modeladapterSchedulerPolicy string
var enableRuntimeSidecar bool
var debugMode bool
flag.StringVar(&metricsAddr, "metrics-bind-address", ":8080", "The address the metric endpoint binds to.")
flag.StringVar(&probeAddr, "health-probe-bind-address", ":8081", "The address the probe endpoint... | @@ -131,6 +133,7 @@ func main() {
flag.StringVar(&leaderElectionId, "leader-election-id", "aibrix-controller-manager",
"leader-election-id determines the name of the resource that leader election will use for holding the leader lock, Default is aibrix-controller-manager.")
flag.StringVar(&controllers, "controlle... | cmd/controllers/main.go | 26 | Go | 0.571 | question | 195 | 51 | 51 | false | [CLI] feature(modeladapter): make modeladapter controller scheduler policy be configured | 921 | vllm-project/aibrix | 2,572 | Go | Jeffwan | googs1025 |
}
}
// ignore worker pods
oldNodeType := oldPod.Labels[nodeType]
newNodeType := newPod.Labels[nodeType]
if oldNodeType == nodeWorker || newNodeType == nodeWorker {
klog.InfoS("ignored ray worker pod", "old pod", oldPod.Name, "new pod", newPod.Name)
return
}
// Add new mappings if present
if newOk {
me... | that's a good catch. I am considering whether it's better to use switch case for `cache.DeletedFinalStateUnknown` and `v1.Pod` which makes the code cleaner?
something like this. otherwise the v1.pod-> cache.DeletedFinalStateUnknown -> v1.pod chain is kind of messy.
```
switch obj := obj.(type) {
case *v1.Pod:
... | }
}
// ignore worker pods
oldNodeType := oldPod.Labels[nodeType]
newNodeType := newPod.Labels[nodeType]
if oldNodeType == nodeWorker || newNodeType == nodeWorker {
klog.InfoS("ignored ray worker pod", "old pod", oldPod.Name, "new pod", newPod.Name)
return
}
// Add new mappings if present
if newOk {
me... | @@ -158,25 +158,49 @@ func (c *Store) updatePod(oldObj interface{}, newObj interface{}) {
}
func (c *Store) deletePod(obj interface{}) {
- pod := obj.(*v1.Pod)
- _, ok := pod.Labels[modelIdentifier]
- _, existed := c.metaPods.Load(pod.Name) // Make sure nothing left.
- if !ok && !existed {
+ var namespace, name str... | pkg/cache/informers.go | 26 | Go | 1 | performance | 593 | 51 | 51 | false | [BUG] cache: handle DeletedFinalStateUnknown by the delete func | 926 | vllm-project/aibrix | 2,572 | Go | Jeffwan | Iceber |
return hasLabelAndModelIdentifier(e.ObjectNew.GetLabels(), labelKey)
},
DeleteFunc: func(e event.DeleteEvent) bool {
return hasLabelAndModelIdentifier(e.Object.GetLabels(), labelKey)
},
GenericFunc: func(e event.GenericEvent) bool {
return hasLabelAndModelIdentifier(e.Object.GetLabels(), labelKey)
},... | makes sense. I did see some log populated in these scenarios which is confusing. | return hasLabelAndModelIdentifier(e.ObjectNew.GetLabels(), labelKey)
},
DeleteFunc: func(e event.DeleteEvent) bool {
return hasLabelAndModelIdentifier(e.Object.GetLabels(), labelKey)
},
GenericFunc: func(e event.GenericEvent) bool {
return hasLabelAndModelIdentifier(e.Object.GetLabels(), labelKey)
},... | @@ -122,9 +122,12 @@ func add(mgr manager.Manager, r reconcile.Reconciler) error {
Owns(&appsv1.Deployment{}).
Watches(&corev1.Pod{}, &handler.EnqueueRequestForObject{}, builder.WithPredicates(podWithLabelFilter(KVCacheLabelKeyIdentifier))).
Complete(r)
+ if err != nil { | pkg/controller/kvcache/kvcache_controller.go | 26 | Go | 0.214 | suggestion | 80 | 51 | 51 | false | [BUG] return directly when error occurs while adding the controller | 937 | vllm-project/aibrix | 2,572 | Go | Jeffwan | Iceber |
})
if err != nil {
return err
}
_, err = fleetInformer.AddEventHandler(cache.ResourceEventHandlerFuncs{
AddFunc: modelRouter.addRouteFromRayClusterFleet,
DeleteFunc: modelRouter.deleteRouteFromRayClusterFleet,
})
return err
}
type ModelRouter struct {
client.Client
Scheme *runtime.Scheme
Run... | can we also change to the switch style to make the code consistent like we discussed in the earlier PR? | })
if err != nil {
return err
}
_, err = fleetInformer.AddEventHandler(cache.ResourceEventHandlerFuncs{
AddFunc: modelRouter.addRouteFromRayClusterFleet,
DeleteFunc: modelRouter.deleteRouteFromRayClusterFleet,
})
return err
}
type ModelRouter struct {
client.Client
Scheme *runtime.Scheme
Run... | @@ -119,7 +119,17 @@ func (m *ModelRouter) addRouteFromDeployment(obj interface{}) {
}
func (m *ModelRouter) deleteRouteFromDeployment(obj interface{}) {
- deployment := obj.(*appsv1.Deployment) | pkg/controller/modelrouter/modelrouter_controller.go | 26 | Go | 0.429 | question | 103 | 51 | 51 | false | [BUG] controller: handle DeletedFinalStateUnknown by the delete func | 938 | vllm-project/aibrix | 2,572 | Go | Jeffwan | Iceber |
// newReconciler returns a new reconcile.Reconciler
func newReconciler(mgr manager.Manager, runtimeConfig config.RuntimeConfig) (reconcile.Reconciler, error) {
reconciler := &RayClusterFleetReconciler{
Client: mgr.GetClient(),
Scheme: mgr.GetScheme(),
Recorder: mgr.GetEventRecorderFor(control... | I think this is kind of typo, so changed | func newReconciler(mgr manager.Manager, runtimeConfig config.RuntimeConfig) (reconcile.Reconciler, error) {
reconciler := &RayClusterFleetReconciler{
Client: mgr.GetClient(),
Scheme: mgr.GetScheme(),
Recorder: mgr.GetEventRecorderFor(controllerName),
Expectation: expectation.NewControllerE... | @@ -84,7 +82,7 @@ func add(mgr manager.Manager, r reconcile.Reconciler) error {
return err
}
- klog.V(4).InfoS("Finished to add model-adapter-controller") | pkg/controller/rayclusterfleet/rayclusterfleet_controller.go | 26 | Go | 0.143 | nitpick | 40 | 51 | 51 | false | [Misc]: move crd check in Initialize part | 949 | vllm-project/aibrix | 2,572 | Go | googs1025 | googs1025 |
)
// validateStreamOptions validates whether stream options to include usage is set for user request
func validateStreamOptions(requestID string, user utils.User, jsonMap map[string]interface{}) *extProcPb.ProcessingResponse {
if user.Tpm != 0 {
streamOptions, ok := jsonMap["stream_options"].(map[string]interface{}... | `CheckEnvExists` doesn't seem like the best name. Maybe rename it to `LookupEnv` or just use os.LookupEnv directly?
https://github.com/vllm-project/aibrix/blob/6c325828b22098ef17240c7fbd50d032ce93449c/pkg/utils/redis.go#L33-L38
@Jeffwan What do you think? | )
// validateStreamOptions validates whether stream options to include usage is set for user request
func validateStreamOptions(requestID string, user utils.User, jsonMap map[string]interface{}) *extProcPb.ProcessingResponse {
if user.Tpm != 0 {
streamOptions, ok := jsonMap["stream_options"].(map[string]interface{}... | @@ -50,30 +50,20 @@ func validateStreamOptions(requestID string, user utils.User, jsonMap map[string
return nil
}
+var defaultRoutingStrategy, defaultRoutingStrategyEnabled = utils.CheckEnvExists(EnvRoutingAlgorithm) | pkg/plugins/gateway/util.go | 26 | Go | 0.643 | question | 259 | 51 | 51 | false | [Bug] avoid frequent lookup of the routing strategy env | 956 | vllm-project/aibrix | 2,572 | Go | Iceber | Iceber |
config, err = rest.InClusterConfig()
} else {
klog.Infof("using configuration from '%s'", kubeConfig)
config, err = clientcmd.BuildConfigFromFlags("", kubeConfig)
}
if err != nil {
panic(err)
}
cache.InitForGateway(config, stopCh, redisClient)
// Connect to K8s cluster
k8sClient, err := kubernetes.New... | we used `klog.Fatalf` earlier, should we keep it same? | } else {
klog.Infof("using configuration from '%s'", kubeConfig)
config, err = clientcmd.BuildConfigFromFlags("", kubeConfig)
}
if err != nil {
panic(err)
}
cache.InitForGateway(config, stopCh, redisClient)
// Connect to K8s cluster
k8sClient, err := kubernetes.NewForConfig(config)
if err != nil {
kl... | @@ -85,10 +86,14 @@ func main() {
if err != nil {
klog.Fatalf("failed to listen: %v", err)
}
+ gatewayK8sClient, err := versioned.NewForConfig(config)
+ if err != nil {
+ panic(err.Error()) | cmd/plugins/main.go | 26 | Go | 0.429 | question | 54 | 51 | 51 | false | Add httproute status check for response header errors | 957 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
func (s *Server) selectTargetPod(ctx *types.RoutingContext, pods types.PodList) (string, error) {
router, err := routing.Select(ctx.Algorithm)(ctx)
if err != nil {
return "", err
}
if pods.Len() == 0 {
return "", fmt.Errorf("no pods to forward request")
}
readyPods := utils.FilterRoutablePods(pods.All())
i... | why does this method return a string? does error look better? the caller can use error.Error() to wrap the message, otherwise, for validated scenarios, it need to check string == ""?
I feel change to error type as return result is better | }
func (s *Server) selectTargetPod(ctx *types.RoutingContext, pods types.PodList) (string, error) {
router, err := routing.Select(ctx.Algorithm)(ctx)
if err != nil {
return "", err
}
if pods.Len() == 0 {
return "", fmt.Errorf("no pods to forward request")
}
readyPods := utils.FilterRoutablePods(pods.All())
... | @@ -156,6 +167,48 @@ func (s *Server) selectTargetPod(ctx *types.RoutingContext, pods types.PodList)
return router.Route(ctx, &utils.PodArray{Pods: readyPods})
}
+// validateHTTPRouteStatus checks if httproute object exists and validates its conditions are true
+func (s *Server) validateHTTPRouteStatus(ctx context... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.643 | question | 242 | 51 | 51 | false | Add httproute status check for response header errors | 957 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
"k8s.io/client-go/kubernetes"
"k8s.io/klog/v2"
extProcPb "github.com/envoyproxy/go-control-plane/envoy/service/ext_proc/v3"
envoyTypePb "github.com/envoyproxy/go-control-plane/envoy/type/v3"
"github.com/vllm-project/aibrix/pkg/cache"
routing "github.com/vllm-project/aibrix/pkg/plugins/gateway/algorithms"
"githu... | Does this gateway client have caching enabled? From what I see, it should be safe now since httproute fetch is only triggered on errors. I just want to confirm that it won’t cause issues for the API server in misconfigured scenarios—e.g., if a large number of requests are sent directly to the API server. | "google.golang.org/grpc/codes"
"google.golang.org/grpc/status"
"k8s.io/client-go/kubernetes"
"k8s.io/klog/v2"
extProcPb "github.com/envoyproxy/go-control-plane/envoy/service/ext_proc/v3"
envoyTypePb "github.com/envoyproxy/go-control-plane/envoy/type/v3"
"github.com/vllm-project/aibrix/pkg/cache"
routing "githu... | @@ -38,17 +37,25 @@ import (
"github.com/vllm-project/aibrix/pkg/types"
"github.com/vllm-project/aibrix/pkg/utils"
healthPb "google.golang.org/grpc/health/grpc_health_v1"
+ metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+ gatewayv1 "sigs.k8s.io/gateway-api/apis/v1"
+ gatewayapi "sigs.k8s.io/gateway-api/pkg/client/... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.643 | question | 305 | 51 | 51 | false | Add httproute status check for response header errors | 957 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
return "", fmt.Errorf("no ready pods available for fallback")
}
if len(readyPods) == 1 {
for _, pod := range readyPods {
ctx.SetTargetPod(pod)
return ctx.TargetAddress(), nil
}
}
return router.Route(ctx, &utils.PodArray{Pods: readyPods})
}
// validateHTTPRouteStatus checks if httproute object exists a... | In our current setting, there's only one parent right? there's no need to check all parents so break here is safe | if len(readyPods) == 0 {
return "", fmt.Errorf("no ready pods available for fallback")
}
if len(readyPods) == 1 {
for _, pod := range readyPods {
ctx.SetTargetPod(pod)
return ctx.TargetAddress(), nil
}
}
return router.Route(ctx, &utils.PodArray{Pods: readyPods})
}
// validateHTTPRouteStatus checks if... | @@ -156,6 +167,48 @@ func (s *Server) selectTargetPod(ctx *types.RoutingContext, pods types.PodList)
return router.Route(ctx, &utils.PodArray{Pods: readyPods})
}
+// validateHTTPRouteStatus checks if httproute object exists and validates its conditions are true
+func (s *Server) validateHTTPRouteStatus(ctx context... | pkg/plugins/gateway/gateway.go | 26 | Go | 0.357 | suggestion | 113 | 51 | 51 | false | Add httproute status check for response header errors | 957 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
} else {
klog.Infof("using configuration from '%s'", kubeConfig)
config, err = clientcmd.BuildConfigFromFlags("", kubeConfig)
}
if err != nil {
panic(err)
}
cache.InitForGateway(config, stopCh, redisClient)
// Connect to K8s cluster
k8sClient, err := kubernetes.NewForConfig(config)
if err != nil {
kl... | this server automatically implements `Check` and `Watch` so we do not need maintain healthcheck server ourselves?
| klog.Infof("using configuration from '%s'", kubeConfig)
config, err = clientcmd.BuildConfigFromFlags("", kubeConfig)
}
if err != nil {
panic(err)
}
cache.InitForGateway(config, stopCh, redisClient)
// Connect to K8s cluster
k8sClient, err := kubernetes.NewForConfig(config)
if err != nil {
klog.Fatalf(... | @@ -87,9 +86,11 @@ func main() {
}
s := grpc.NewServer()
-
extProcPb.RegisterExternalProcessorServer(s, gateway.NewServer(redisClient, k8sClient))
- healthPb.RegisterHealthServer(s, gateway.NewHealthCheckServer())
+
+ healthCheck := health.NewServer() | cmd/plugins/main.go | 26 | Go | 0.5 | question | 118 | 51 | 51 | false | Add graceful shutdown for gateway and add liveness/readiness probes | 962 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
return errors.New("certificates are not ready")
}
}); err != nil {
setupLog.Error(err, "unable to set up ready check")
os.Exit(1)
}
setupLog.Info("starting manager")
if err := mgr.Start(ctrl.SetupSignalHandler()); err != nil {
setupLog.Error(err, "problem running manager")
os.Exit(1)
}
}
// TODO: if... | Looking at the code, I don't think the order will have any particular effect.
Let me know if I'm overlooking something. | return mgr.GetWebhookServer().StartedChecker()(req)
default:
return errors.New("webhook certificates are not ready")
}
}); err != nil {
setupLog.Error(err, "unable to set up ready check")
os.Exit(1)
}
setupLog.Info("starting manager")
if err := mgr.Start(ctrl.SetupSignalHandler()); err != nil {
set... | @@ -290,13 +301,20 @@ func main() {
}
// TODO: if the argument list will grow, we should create a ControllerSetupOptions struct instead.
-func setupControllers(mgr ctrl.Manager, runtimeConfig config.RuntimeConfig, certsReady chan struct{}, disableWebhook bool) {
+func setupControllers(mgr ctrl.Manager, runtimeConfi... | cmd/controllers/main.go | 26 | Go | 0.357 | suggestion | 122 | 39 | 40 | false | cmd/controllers: add readyz check for the webhook | 969 | vllm-project/aibrix | 2,572 | Go | Iceber | Iceber |
os.Exit(1)
}
setupLog.Info("starting cache")
stopCh := make(chan struct{})
defer close(stopCh)
var config *rest.Config
// ref: https://github.com/kubernetes-sigs/controller-runtime/issues/878#issuecomment-1002204308
kubeConfig := flag.Lookup("kubeconfig").Value.String()
if kubeConfig == "" {
setupLog.Info... | Let's revert this, we'll make sure the certsReady will always be closed. | os.Exit(1)
}
setupLog.Info("starting cache")
stopCh := make(chan struct{})
defer close(stopCh)
var config *rest.Config
// ref: https://github.com/kubernetes-sigs/controller-runtime/issues/878#issuecomment-1002204308
kubeConfig := flag.Lookup("kubeconfig").Value.String()
if kubeConfig == "" {
setupLog.Info... | @@ -248,12 +250,11 @@ func main() {
cache.Init(config, stopCh)
}
- certsReady := make(chan struct{}) | cmd/controllers/main.go | 26 | Go | 0.214 | suggestion | 72 | 51 | 51 | false | cmd/controllers: add readyz check for the webhook | 969 | vllm-project/aibrix | 2,572 | Go | kerthcet | Iceber |
return nil
}
select {
case <-certsReady:
return mgr.GetWebhookServer().StartedChecker()(req)
default:
return errors.New("certificates are not ready")
}
}); err != nil {
setupLog.Error(err, "unable to set up ready check")
os.Exit(1)
}
setupLog.Info("starting manager")
if err := mgr.Start(ctr... | Basically, certsReady doesn't mean disableWebhook, which will lead to confusion for further readers, let's not do like this. But we can remove the disableWebhook judgement here because once `disabledWebhook=false`, the certsReady will always be closed. | if err := mgr.AddReadyzCheck("readyz", func(req *http.Request) error {
if disableWebhook {
return nil
}
select {
case <-certsReady:
return mgr.GetWebhookServer().StartedChecker()(req)
default:
return errors.New("webhook certificates are not ready")
}
}); err != nil {
setupLog.Error(err, "unabl... | @@ -290,13 +301,20 @@ func main() {
}
// TODO: if the argument list will grow, we should create a ControllerSetupOptions struct instead.
-func setupControllers(mgr ctrl.Manager, runtimeConfig config.RuntimeConfig, certsReady chan struct{}, disableWebhook bool) {
+func setupControllers(mgr ctrl.Manager, runtimeConfi... | cmd/controllers/main.go | 26 | Go | 0.643 | suggestion | 252 | 46 | 47 | false | cmd/controllers: add readyz check for the webhook | 969 | vllm-project/aibrix | 2,572 | Go | kerthcet | Iceber |
)
}
func validateInference(t *testing.T, modelName string) {
client := createOpenAIClient(gatewayURL, apiKey)
validateInferenceWithClient(t, client, modelName)
}
func validateInferenceWithClient(t *testing.T, client *openai.Client, modelName string) {
chatCompletion, err := client.Chat.Completions.New(context.TOD... | We set `expectedPodCount` to be the actual expected number, and using it for traversal seems to be OK, but I'm not sure if it will be confusing? 🤔 | r.URL.Path = "/v1" + r.URL.Path
return mn(r)
}),
option.WithHeader("routing-strategy", routingStrategy),
option.WithMaxRetries(0),
respOpt,
)
}
func validateInference(t *testing.T, modelName string) {
client := createOpenAIClient(gatewayURL, apiKey)
validateInferenceWithClient(t, client, modelName)
}
... | @@ -130,3 +133,18 @@ func validateInferenceWithClient(t *testing.T, client *openai.Client, modelName
assert.NotEmpty(t, chatCompletion.Choices, "chat completion has no choices returned")
assert.NotNil(t, chatCompletion.Choices[0].Message.Content, "chat completion has no message returned")
}
+
+func validateAllPods... | test/e2e/util.go | 26 | Go | 0.5 | suggestion | 148 | 38 | 43 | false | E2E CI fix to ensure all pods are ready | 972 | vllm-project/aibrix | 2,572 | Go | googs1025 | varungup90 |
/*
Copyright 2024 The Aibrix Team.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, softwar... | I believe https://github.com/onsi/ginkgo is more solid here with like `Eventually` semantic. |
func validateInferenceWithClient(t *testing.T, client *openai.Client, modelName string) {
chatCompletion, err := client.Chat.Completions.New(context.TODO(), openai.ChatCompletionNewParams{
Messages: openai.F([]openai.ChatCompletionMessageParamUnion{
openai.UserMessage("Say this is a test"),
}),
Model: openai... | @@ -129,3 +132,17 @@ func validateInferenceWithClient(t *testing.T, client *openai.Client, modelName
assert.NotEmpty(t, chatCompletion.Choices, "chat completion has no choices returned")
assert.NotNil(t, chatCompletion.Choices[0].Message.Content, "chat completion has no message returned")
}
+
+func validateAllPods... | test/e2e/util.go | 26 | Go | 0.357 | suggestion | 92 | 149 | 30 | false | E2E CI fix to ensure all pods are ready | 972 | vllm-project/aibrix | 2,572 | Go | kerthcet | varungup90 |
/*
Copyright 2024 The Aibrix Team.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, softwar... | i am curious we already check pod status here https://github.com/vllm-project/aibrix/blob/bc6cef162416496dd84ed6972c5eecff1fcc48a8/.github/workflows/installation-tests.yml#L142 and then run `make test-e2e`, why does it work as expected and we have to check one more time here? | option.WithMaxRetries(0),
respOpt,
)
}
func validateInference(t *testing.T, modelName string) {
client := createOpenAIClient(gatewayURL, apiKey)
validateInferenceWithClient(t, client, modelName)
}
func validateInferenceWithClient(t *testing.T, client *openai.Client, modelName string) {
chatCompletion, err := ... | @@ -136,3 +139,17 @@ func validateInferenceWithClient(t *testing.T, client *openai.Client, modelName
assert.NotEmpty(t, chatCompletion.Choices, "chat completion has no choices returned")
assert.NotNil(t, chatCompletion.Choices[0].Message.Content, "chat completion has no message returned")
}
+
+func validateAllPods... | test/e2e/util.go | 26 | Go | 0.571 | question | 278 | 139 | 39 | false | E2E CI fix to ensure all pods are ready | 972 | vllm-project/aibrix | 2,572 | Go | Jeffwan | varungup90 |
nodeType = "ray.io/node-type"
nodeWorker = "worker"
)
func initCacheInformers(instance *Store, config *rest.Config, stopCh <-chan struct{}) error {
if err := v1alpha1scheme.AddToScheme(scheme.Scheme); err != nil {
return err
}
k8sClientSet, err := kubernetes.NewForConfig(config)
if err != nil {
... | HandleCrash is usually used in goroutinue or in functions that contain functions that are not sure whether a panic will occur or not.
it doesn't seem necessary in here.
| nodeWorker = "worker"
)
func initCacheInformers(instance *Store, config *rest.Config, stopCh <-chan struct{}) error {
if err := v1alpha1scheme.AddToScheme(scheme.Scheme); err != nil {
return err
}
k8sClientSet, err := kubernetes.NewForConfig(config)
if err != nil {
return err
}
crdClientSet, err := v... | @@ -59,16 +58,6 @@ func initCacheInformers(instance *Store, config *rest.Config, stopCh <-chan stru
crdFactory := crdinformers.NewSharedInformerFactoryWithOptions(crdClientSet, 0)
podInformer := factory.Core().V1().Pods().Informer()
- modelInformer := crdFactory.Model().V1alpha1().ModelAdapters().Informer()
-
- d... | pkg/cache/informers.go | 26 | Go | 0.571 | bug | 171 | 51 | 51 | false | [BUG] cache: start informer after adding the resource handler | 981 | vllm-project/aibrix | 2,572 | Go | Iceber | Iceber |
import logging
import os
import threading
from typing import Dict, Optional, Tuple
import redis
import uvicorn
from kubernetes import client, config, watch
from starlette.applications import Starlette
from starlette.responses import JSONResponse, PlainTextResponse
from aibrix.gpu_optimizer.load_monitor.load_reader im... | @weapons97 could you help format this file. seems the linter report the format issue on this file | import logging
import os
import threading
from typing import Dict, Optional, Tuple
import redis
import uvicorn
from kubernetes import client, config, watch
from starlette.applications import Starlette
from starlette.responses import JSONResponse, PlainTextResponse
from aibrix.gpu_optimizer.load_monitor.load_reader im... | @@ -35,6 +35,7 @@
MIN_REPLICAS_LABEL = "model.aibrix.ai/min_replicas"
REDIS_HOST = os.getenv("REDIS_HOST", "localhost")
REDIS_PORT = int(os.getenv("REDIS_PORT", 6379))
+REDIS_PASSWORD = os.getenv("REDIS_PASSWORD", None) | python/aibrix/aibrix/gpu_optimizer/app.py | 26 | Python | 0.286 | style | 97 | 51 | 51 | false | Update add redis pass for client | 990 | vllm-project/aibrix | 2,572 | Go | Jeffwan | weapons97 |
// mean calculates the mean of a slice of float64 numbers.
func mean(numbers []float64) float64 {
sum := 0.0
for _, number := range numbers {
sum += number
}
return sum / float64(len(numbers))
}
// standardDeviation calculates the standard deviation of a slice of float64 numbers.
func standardDeviation(numbers ... | We may have fallback in different routing algorithms. We can extract a method and indicate the use of this fallback method. | )
// mean calculates the mean of a slice of float64 numbers.
func mean(numbers []float64) float64 {
sum := 0.0
for _, number := range numbers {
sum += number
}
return sum / float64(len(numbers))
}
// standardDeviation calculates the standard deviation of a slice of float64 numbers.
func standardDeviation(number... | @@ -39,3 +45,16 @@ func standardDeviation(numbers []float64) float64 {
variance := sumOfSquares / float64(len(numbers)-1)
return math.Sqrt(variance)
}
+
+// SelectRandomPodAsFallback selects a pod randomly as a fallback.
+// This method should only be used when all other selection mechanisms have failed.
+// For e... | pkg/plugins/gateway/algorithms/util.go | 26 | Go | 0.429 | refactor | 123 | 34 | 36 | false | [Misc] fix: Optimizing Route method of the gateway algorithms | 1,001 | vllm-project/aibrix | 2,572 | Go | googs1025 | googs1025 |
)
func init() {
RegisterDelayedConstructor(RouterLeastBusyTime, NewLeastBusyTimeRouter)
}
type leastBusyTimeRouter struct {
cache cache.Cache
}
func NewLeastBusyTimeRouter() (types.Router, error) {
c, err := cache.Get()
if err != nil {
return nil, err
}
return leastBusyTimeRouter{
cache: c,
}, nil
}
fun... | ip check at line 57 can be removed. FilterRoutablePods already does it. Please remove this at other places if present. | )
func init() {
RegisterDelayedConstructor(RouterLeastBusyTime, NewLeastBusyTimeRouter)
}
type leastBusyTimeRouter struct {
cache cache.Cache
}
func NewLeastBusyTimeRouter() (types.Router, error) {
c, err := cache.Get()
if err != nil {
return nil, err
}
return leastBusyTimeRouter{
cache: c,
}, nil
}
fun... | @@ -51,15 +49,11 @@ func NewLeastBusyTimeRouter() (types.Router, error) {
}, nil
}
-func (r leastBusyTimeRouter) Route(ctx *types.RoutingContext, pods types.PodList) (string, error) {
+func (r leastBusyTimeRouter) Route(ctx *types.RoutingContext, readyPodList types.PodList) (string, error) {
var targetPod *v1.Po... | pkg/plugins/gateway/algorithms/least_busy_time.go | 26 | Go | 0.429 | refactor | 119 | 51 | 51 | false | [Misc] fix: Optimizing Route method of the gateway algorithms | 1,001 | vllm-project/aibrix | 2,572 | Go | varungup90 | googs1025 |
// mean calculates the mean of a slice of float64 numbers.
func mean(numbers []float64) float64 {
sum := 0.0
for _, number := range numbers {
sum += number
}
return sum / float64(len(numbers))
}
// standardDeviation calculates the standard deviation of a slice of float64 numbers.
func standardDeviation(numbers ... | Pass the routing context as argument and use ctx.requestID in the log for ln 54 and 57. | )
// mean calculates the mean of a slice of float64 numbers.
func mean(numbers []float64) float64 {
sum := 0.0
for _, number := range numbers {
sum += number
}
return sum / float64(len(numbers))
}
// standardDeviation calculates the standard deviation of a slice of float64 numbers.
func standardDeviation(number... | @@ -39,3 +45,16 @@ func standardDeviation(numbers []float64) float64 {
variance := sumOfSquares / float64(len(numbers)-1)
return math.Sqrt(variance)
}
+
+// SelectRandomPodAsFallback selects a pod randomly as a fallback.
+// This method should only be used when all other selection mechanisms have failed.
+// For e... | pkg/plugins/gateway/algorithms/util.go | 26 | Go | 0.286 | suggestion | 87 | 34 | 36 | false | [Misc] fix: Optimizing Route method of the gateway algorithms | 1,001 | vllm-project/aibrix | 2,572 | Go | varungup90 | googs1025 |
// mean calculates the mean of a slice of float64 numbers.
func mean(numbers []float64) float64 {
sum := 0.0
for _, number := range numbers {
sum += number
}
return sum / float64(len(numbers))
}
// standardDeviation calculates the standard deviation of a slice of float64 numbers.
func standardDeviation(numbers [... | You can do a warn log message here. And remove the warn message individually from caller just after if targetPod == nil check, this warn message does not have requestID so it is not useful eitherwise. | // mean calculates the mean of a slice of float64 numbers.
func mean(numbers []float64) float64 {
sum := 0.0
for _, number := range numbers {
sum += number
}
return sum / float64(len(numbers))
}
// standardDeviation calculates the standard deviation of a slice of float64 numbers.
func standardDeviation(numbers [... | @@ -39,3 +46,17 @@ func standardDeviation(numbers []float64) float64 {
variance := sumOfSquares / float64(len(numbers)-1)
return math.Sqrt(variance)
}
+
+// SelectRandomPodAsFallback selects a pod randomly as a fallback.
+// This method should only be used when all other selection mechanisms have failed.
+// For e... | pkg/plugins/gateway/algorithms/util.go | 26 | Go | 0.5 | suggestion | 201 | 34 | 34 | false | [Misc] fix: Optimizing Route method of the gateway algorithms | 1,001 | vllm-project/aibrix | 2,572 | Go | varungup90 | googs1025 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.