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 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
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 ... | No issues found. |
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 ... | @@ -44,6 +44,9 @@ func IsRayClusterAvailable(cluster *rayclusterv1.RayCluster, minReadySeconds int
headPodReadyCond := meta.FindStatusCondition(cluster.Status.Conditions, string(rayclusterv1.HeadPodReady))
provisionedCond := meta.FindStatusCondition(cluster.Status.Conditions, string(rayclusterv1.RayClusterProvisi... | pkg/utils/raycluster.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bug] fix: condition nil panic in FindStatusCondition func | 1,078 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
// and requeue the Deployment.
if d.Status.CollisionCount == nil {
d.Status.CollisionCount = new(int32)
}
preCollisionCount := *d.Status.CollisionCount
*d.Status.CollisionCount++
// Update the collisionCount for the Deployment and let it requeue by returning the original
// error.
dErr := r.Status().... | No issues found. | // and requeue the Deployment.
if d.Status.CollisionCount == nil {
d.Status.CollisionCount = new(int32)
}
preCollisionCount := *d.Status.CollisionCount
*d.Status.CollisionCount++
// Update the collisionCount for the Deployment and let it requeue by returning the original
// error.
dErr := r.Status().... | @@ -192,6 +192,13 @@ func (r *RayClusterFleetReconciler) getNewReplicaSet(ctx context.Context, d *orc
// new ReplicaSet does not exist, create one.
newRSTemplate := *d.Spec.Template.DeepCopy()
+
+ // Add rayclusterfleet name to all pod label
+ newRSTemplate.Spec.HeadGroupSpec.Template.Labels = labelsutil.CloneAnd... | pkg/controller/rayclusterfleet/sync.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bug]fix: add scale subresource to rayclusterfleet | 1,082 | vllm-project/aibrix | 2,572 | Go | zhixian82 | |
`)
t.Run("Parse simple metric with model labels", func(t *testing.T) {
value, err := ParseMetricFromBody(body, "vllm:num_requests_waiting")
assert.NoError(t, err)
assert.Equal(t, 2.0, value)
})
}
func TestExtractBucketBoundary(t *testing.T) {
line := `vllm:time_per_output_token_seconds_bucket{le="0.1",model_... | No issues found. | `)
t.Run("Parse simple metric with model labels", func(t *testing.T) {
value, err := ParseMetricFromBody(body, "vllm:num_requests_waiting")
assert.NoError(t, err)
assert.Equal(t, 2.0, value)
})
}
func TestExtractBucketBoundary(t *testing.T) {
line := `vllm:time_per_output_token_seconds_bucket{le="0.1",model_... | @@ -17,7 +17,13 @@ limitations under the License.
package metrics
import (
+ "context"
+ "io"
+ "net/http"
+ "net/http/httptest"
+ "strings"
"testing"
+ "time"
"github.com/stretchr/testify/assert"
)
@@ -89,3 +95,88 @@ func TestBuildQueryWithModelLabel(t *testing.T) {
assert.Contains(t, query, `model_name=... | pkg/metrics/utils_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | feature: use worker pool management for periodic metrics update | 1,096 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
}
...
]
}
"""
try:
data = json.loads(json_str)
temp_members: List[Member] = []
temp_slots = SortedList(key=lambda x: x[0])
for node in data["nodes"]:
slots = node.pop("slots")
... | No issues found. | }
...
]
}
"""
try:
data = json.loads(json_str)
temp_members: List[Member] = []
temp_slots = SortedList(key=lambda x: x[0])
for node in data["nodes"]:
slots = node.pop("slots")
... | @@ -17,6 +17,7 @@
from dataclasses import dataclass
from typing import Any, Dict, List, Sequence, Tuple, TypeVar
+import torch
from sortedcontainers import SortedList
from ...common.absl_logging import getLogger
@@ -27,7 +28,6 @@
Connector,
ConnectorConfig,
ConnectorFeature,
- ConnectorRegiste... | python/aibrix_kvcache/aibrix_kvcache/l2/placement/placement.py | 0 | Python | 1 | none | 16 | 51 | 51 | true | [Feature] KVCache: Suppport InfiniStore GID and enhance cluster mode | 1,106 | vllm-project/aibrix | 2,572 | Go | DwyaneShi | |
},
Status: v1.PodStatus{
PodIP: fmt.Sprintf("10.0.0.%d", i+1),
Conditions: []v1.PodCondition{
{
Type: v1.PodReady,
Status: v1.ConditionTrue,
},
},
},
}
switch readyMap[i] {
case 1:
pod.Status.PodIP = ""
case 2:
pod.Status.Conditions[0].Status = v1.ConditionFals... | No issues found. | },
Status: v1.PodStatus{
PodIP: fmt.Sprintf("10.0.0.%d", i+1),
Conditions: []v1.PodCondition{
{
Type: v1.PodReady,
Status: v1.ConditionTrue,
},
},
},
}
switch readyMap[i] {
case 1:
pod.Status.PodIP = ""
case 2:
pod.Status.Conditions[0].Status = v1.ConditionFals... | @@ -18,10 +18,12 @@ package utils
import (
"fmt"
"math/rand"
+ "testing"
"time"
. "github.com/onsi/ginkgo"
. "github.com/onsi/gomega"
+ "github.com/stretchr/testify/assert"
v1 "k8s.io/api/core/v1"
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
)
@@ -81,3 +83,59 @@ var _ = Describe("Pod", func() {
Ex... | pkg/utils/pod_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Add support for custom inference engine port | 1,140 | vllm-project/aibrix | 2,572 | Go | varungup90 | |
else:
return JSONResponse({"message": "Deployment optimization already started"})
except Exception as e:
return JSONResponse(
{"error": f"Error starting deployment optimization: {e}"}, status_code=500
)
@app.route("/monitor/{namespace}/{deployment_name}", methods=["... | No issues found. | else:
return JSONResponse({"message": "Deployment optimization already started"})
except Exception as e:
return JSONResponse(
{"error": f"Error starting deployment optimization: {e}"}, status_code=500
)
@app.route("/monitor/{namespace}/{deployment_name}", methods=["... | @@ -81,9 +81,12 @@ def new_deployment(deployment):
min_replicas = int(label)
return DeploymentStates(
- deployment.metadata.name,
- deployment.spec.replicas if deployment.spec.replicas is not None else 0,
- min_replicas,
+ name=deployment.metadata.name,
+ namespace... | python/aibrix/aibrix/gpu_optimizer/app.py | 0 | Python | 1 | none | 16 | 51 | 51 | true | [Misc] [gpu_optimizer] add namespace info for log | 1,149 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
},
{
message: "/v1/chat/completions json unmarhsal no messages",
requestPath: "/v1/chat/completions",
requestBody: []byte(`{"model": "llama2-7b"}`),
statusCode: envoyTypePb.StatusCode_BadRequest,
},
{
message: "/v1/chat/completions json unmarhsal valid messages",
requestPath: "/v1/cha... | No issues found. | },
{
message: "/v1/chat/completions json unmarhsal no messages",
requestPath: "/v1/chat/completions",
requestBody: []byte(`{"model": "llama2-7b"}`),
statusCode: envoyTypePb.StatusCode_BadRequest,
},
{
message: "/v1/chat/completions json unmarhsal valid messages",
requestPath: "/v1/cha... | @@ -67,6 +67,20 @@ func Test_ValidateRequestBody(t *testing.T) {
messages: "this is system say this is test",
statusCode: envoyTypePb.StatusCode_OK,
},
+ {
+ message: "/v1/chat/completions json unmarhsal invalid messages with complex content",
+ requestPath: "/v1/chat/completions",
+ requestBo... | pkg/plugins/gateway/util_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bug] fix: error when parse complex content in completion body | 1,160 | vllm-project/aibrix | 2,572 | Go | firebook | |
//go:build !race
/*
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 w... | No issues found. | //go:build !race
/*
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 w... | @@ -1,3 +1,5 @@
+//go:build !race
+
/*
Copyright 2024 The Aibrix Team.
| pkg/types/types_test.go | 0 | Go | 1 | none | 16 | 31 | 31 | true | Add race condition check in unit-test CI and add test-coverage cmd in Makefile | 1,169 | vllm-project/aibrix | 2,572 | Go | varungup90 | |
func (m *ModelRouter) addRouteFromRayClusterFleet(obj interface{}) {
fleet := obj.(*orchestrationv1alpha1.RayClusterFleet)
m.createHTTPRoute(fleet.Namespace, fleet.Labels)
}
func (m *ModelRouter) deleteRouteFromRayClusterFleet(obj interface{}) {
fleet, ok := obj.(*orchestrationv1alpha1.RayClusterFleet)
if !ok {
... | No issues found. |
func (m *ModelRouter) addRouteFromRayClusterFleet(obj interface{}) {
fleet := obj.(*orchestrationv1alpha1.RayClusterFleet)
m.createHTTPRoute(fleet.Namespace, fleet.Labels)
}
func (m *ModelRouter) deleteRouteFromRayClusterFleet(obj interface{}) {
fleet, ok := obj.(*orchestrationv1alpha1.RayClusterFleet)
if !ok {
... | @@ -18,6 +18,7 @@ package modelrouter
import (
"context"
+ "errors"
"fmt"
"strconv"
@@ -105,6 +106,12 @@ func Add(mgr manager.Manager, runtimeConfig config.RuntimeConfig) error {
DeleteFunc: modelRouter.deleteRouteFromRayClusterFleet,
})
+ klog.Info("Waiting for caches to sync")
+ if ok := cacher.Wait... | pkg/controller/modelrouter/modelrouter_controller.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc]: ensure cache sync before starting controller reconcile | 1,219 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
# 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... | No issues found. | # 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... | @@ -12,6 +12,20 @@
# See the License for the specific language governing permissions and
# limitations under the License.
-from .rdma import AddrFamily, DeviceRequest, GIDType, RDMATransport
-__all__ = ["AddrFamily", "DeviceRequest", "GIDType", "RDMATransport"]
+# A global flag to indicate whether current platfor... | python/aibrix_kvcache/aibrix_kvcache/transport/__init__.py | 0 | Python | 1 | none | 16 | 32 | 32 | true | [Fix] KVCache: enhance rdma auto-detection | 1,276 | vllm-project/aibrix | 2,572 | Go | DwyaneShi | |
# 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... | No issues found. | # 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... | @@ -14,7 +14,6 @@
from .async_base import AsyncBase
from .cached_pyobject import CachedPyObjectBase
-from .nvtx import nvtx_range
from .object_pool import ObjectPool
-__all__ = ["AsyncBase", "nvtx_range", "ObjectPool", "CachedPyObjectBase"]
+__all__ = ["AsyncBase", "ObjectPool", "CachedPyObjectBase"] | python/aibrix_kvcache/aibrix_kvcache/common/__init__.py | 0 | Python | 1 | none | 16 | 20 | 20 | true | [Feature] KVCache: enhance profiling | 1,278 | vllm-project/aibrix | 2,572 | Go | DwyaneShi | |
def test_allocating_heterogeneous(allocator):
"""Test allocating with heterogeneous sizes."""
if hasattr(allocator, "num_memory_regions"):
assert allocator.num_memory_regions == 1024
sizes = [random.randint(8, 1024) for _ in range(31)]
status = allocator.alloc(sizes)
allocator.assert_consi... | No issues found. |
def test_allocating_heterogeneous(allocator):
"""Test allocating with heterogeneous sizes."""
if hasattr(allocator, "num_memory_regions"):
assert allocator.num_memory_regions == 1024
sizes = [random.randint(8, 1024) for _ in range(31)]
status = allocator.alloc(sizes)
allocator.assert_consi... | @@ -17,19 +17,21 @@
import pytest
from aibrix_kvcache.memory import TensorPoolAllocator
+from aibrix_kvcache.memory.allocator import CoalescingPoolAllocator
@pytest.fixture
def allocator(compact_layout_enabled):
# use a small slab size for testing
TensorPoolAllocator.SLAB_MAX_NBYTES = 1024
Tens... | python/aibrix_kvcache/tests/test_tensor_pool_allocator.py | 0 | Python | 1 | none | 16 | 51 | 51 | true | [Feature] KVCache: optimize allocator for compact layout | 1,288 | vllm-project/aibrix | 2,572 | Go | DwyaneShi | |
/*
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, softwar... | No issues found. | /*
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, softwar... | @@ -0,0 +1,48 @@
+/*
+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 ag... | pkg/cache/utils.go | 0 | Go | 1 | none | 16 | 49 | 49 | true | [Feature] Adding raw metrics name conversion in metrics | 1,293 | vllm-project/aibrix | 2,572 | Go | happyandslow | |
require.NoError(t, orchestrationv1alpha1.AddToScheme(scheme))
tests := []struct {
name string
description string
roleSet *orchestrationv1alpha1.RoleSet
role *orchestrationv1alpha1.RoleSpec
existingPods []*v1.Pod
expectedError bool
expectedCreate int
expectedDelete int... | No issues found. | require.NoError(t, orchestrationv1alpha1.AddToScheme(scheme))
tests := []struct {
name string
description string
roleSet *orchestrationv1alpha1.RoleSet
role *orchestrationv1alpha1.RoleSpec
existingPods []*v1.Pod
expectedError bool
expectedCreate int
expectedDelete int... | @@ -0,0 +1,445 @@
+/*
+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 a... | pkg/controller/roleset/rolesyncer_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc] Add unit test code coverage of rolesyncer | 1,296 | vllm-project/aibrix | 2,572 | Go | vie-serendipity | |
from .hpkv import HPKVConnector
return HPKVConnector.from_envs(conn_id, executor, **kwargs)
elif backend_name == "PRIS":
from .pris import PrisConnector
return PrisConnector.from_envs(conn_id, executor, **kwargs)
elif backend_name == "MOCK":
... | No issues found. | from .hpkv import HPKVConnector
return HPKVConnector.from_envs(conn_id, executor, **kwargs)
elif backend_name == "PRIS":
from .pris import PrisConnector
return PrisConnector.from_envs(conn_id, executor, **kwargs)
elif backend_name == "MOCK":
... | @@ -92,6 +92,10 @@ def create(
from .hpkv import HPKVConnector
return HPKVConnector.from_envs(conn_id, executor, **kwargs)
+ elif backend_name == "PRIS":
+ from .pris import PrisConnector
+
+ return PrisConnector.from_envs(conn_id, executor, **kwargs)
el... | python/aibrix_kvcache/aibrix_kvcache/l2/connectors/connector.py | 0 | Python | 1 | none | 16 | 51 | 51 | true | [Feature] KVCache: add Pris connector | 1,303 | vllm-project/aibrix | 2,572 | Go | DwyaneShi | |
statusCode: envoyTypePb.StatusCode_OK,
headers: []*configPb.HeaderValueOption{
{
Header: &configPb.HeaderValue{
Key: HeaderRoutingStrategy,
RawValue: []byte("invalid-router"),
},
},
{
Header: &configPb.HeaderValue{
Key: HeaderTargetPod,
RawV... | No issues found. | statusCode: envoyTypePb.StatusCode_OK,
headers: []*configPb.HeaderValueOption{
{
Header: &configPb.HeaderValue{
Key: HeaderRoutingStrategy,
RawValue: []byte("invalid-router"),
},
},
{
Header: &configPb.HeaderValue{
Key: HeaderTargetPod,
RawV... | @@ -19,12 +19,11 @@ package gateway
import (
"context"
"testing"
+ "time"
"github.com/stretchr/testify/assert"
"github.com/stretchr/testify/mock"
- "time"
-
configPb "github.com/envoyproxy/go-control-plane/envoy/config/core/v3"
extProcPb "github.com/envoyproxy/go-control-plane/envoy/service/ext_proc/v3... | pkg/plugins/gateway/gateway_req_body_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [feat] Add prefill-decode disaggregation support in aibrix router | 1,309 | vllm-project/aibrix | 2,572 | Go | varungup90 | |
desiredUpdated: 1,
},
{
currentReplicas: 2,
updatedReplicas: 2,
desiredReplicas: 5,
desiredCurrent: 3,
desiredUpdated: 2,
},
{
currentReplicas: 5,
updatedReplicas: 5,
desiredReplicas: 3,
desiredCurrent: 2,
desiredUpdated: 1,
},
{
currentReplicas: 0,
updatedReplica... | No issues found. | desiredUpdated: 1,
},
{
currentReplicas: 2,
updatedReplicas: 2,
desiredReplicas: 5,
desiredCurrent: 3,
desiredUpdated: 2,
},
{
currentReplicas: 5,
updatedReplicas: 5,
desiredReplicas: 3,
desiredCurrent: 2,
desiredUpdated: 1,
},
{
currentReplicas: 0,
updatedReplica... | @@ -16,7 +16,21 @@ limitations under the License.
package stormservice
-import "testing"
+import (
+ "context"
+ "reflect"
+ "testing"
+
+ corev1 "k8s.io/api/core/v1"
+ metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+ "k8s.io/apimachinery/pkg/runtime"
+ "k8s.io/client-go/tools/record"
+ "sigs.k8s.io/controller-runt... | pkg/controller/stormservice/sync_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc] enhance headless service sync with update logic | 1,311 | vllm-project/aibrix | 2,572 | Go | omerap12 | |
roleSet: &orchestrationv1alpha1.RoleSet{
Spec: orchestrationv1alpha1.RoleSetSpec{
Roles: []orchestrationv1alpha1.RoleSpec{
{Name: "role1", Replicas: int32Ptr(3)},
{Name: "role2", Replicas: int32Ptr(2)},
},
},
Status: orchestrationv1alpha1.RoleSetStatus{
Roles: []orchestrationv1... | No issues found. | roleSet: &orchestrationv1alpha1.RoleSet{
Spec: orchestrationv1alpha1.RoleSetSpec{
Roles: []orchestrationv1alpha1.RoleSpec{
{Name: "role1", Replicas: int32Ptr(3)},
{Name: "role2", Replicas: int32Ptr(2)},
},
},
Status: orchestrationv1alpha1.RoleSetStatus{
Roles: []orchestrationv1... | @@ -22,6 +22,7 @@ import (
"github.com/stretchr/testify/assert"
orchestrationv1alpha1 "github.com/vllm-project/aibrix/api/orchestration/v1alpha1"
"github.com/vllm-project/aibrix/pkg/controller/constants"
+ utils "github.com/vllm-project/aibrix/pkg/controller/util/orchestration"
corev1 "k8s.io/api/core/v1"
met... | pkg/controller/stormservice/utils_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Fix nil slice issue and add UT coverage for stormservice utils | 1,314 | vllm-project/aibrix | 2,572 | Go | Jeffwan | |
func (m *MockCache) GetMetricValueByPod(namespace string, podName string, metricName string) (metrics.MetricValue, error) {
args := m.Called(namespace, podName, metricName)
return args.Get(0).(metrics.MetricValue), args.Error(1)
}
func (m *MockCache) GetMetricValueByPodModel(namespace string, podName string, model ... | No issues found. |
func (m *MockCache) GetMetricValueByPod(namespace string, podName string, metricName string) (metrics.MetricValue, error) {
args := m.Called(namespace, podName, metricName)
return args.Get(0).(metrics.MetricValue), args.Error(1)
}
func (m *MockCache) GetMetricValueByPodModel(namespace string, podName string, model ... | @@ -17,11 +17,22 @@ limitations under the License.
package gateway
import (
+ "context"
+
"github.com/stretchr/testify/mock"
"github.com/vllm-project/aibrix/pkg/cache"
"github.com/vllm-project/aibrix/pkg/metrics"
"github.com/vllm-project/aibrix/pkg/types"
v1 "k8s.io/api/core/v1"
+ metav1 "k8s.io/apimachin... | pkg/plugins/gateway/gateway_test_helpers.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Add check for missing httproute before request start | 1,344 | vllm-project/aibrix | 2,572 | Go | varungup90 | |
"github.com/vllm-project/aibrix/pkg/types"
"github.com/vllm-project/aibrix/pkg/utils"
)
func (s *Server) HandleRequestBody(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest, user utils.User) (*extProcPb.ProcessingResponse, string, *types.RoutingContext, bool, int64) {
var term int64 // Identif... | No issues found. | "github.com/vllm-project/aibrix/pkg/types"
"github.com/vllm-project/aibrix/pkg/utils"
)
func (s *Server) HandleRequestBody(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest, user utils.User) (*extProcPb.ProcessingResponse, string, *types.RoutingContext, bool, int64) {
var term int64 // Identif... | @@ -19,6 +19,7 @@ package gateway
import (
"context"
"fmt"
+ "strconv"
"k8s.io/klog/v2"
@@ -81,7 +82,8 @@ func (s *Server) HandleRequestBody(ctx context.Context, requestID string, req *e
}
headers = buildEnvoyProxyHeaders(headers,
HeaderRoutingStrategy, string(routingAlgorithm),
- HeaderTargetPod... | pkg/plugins/gateway/gateway_req_body.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Bug fixes for PD disaggregation routing | 1,354 | vllm-project/aibrix | 2,572 | Go | varungup90 | |
{ObjectMeta: metav1.ObjectMeta{Labels: map[string]string{"role-name": "other"}}},
{ObjectMeta: metav1.ObjectMeta{Labels: map[string]string{"role-name": "prefill", PodGroupIndex: "1"}}},
},
expectPrefill: 1,
expectDecode: 1,
},
}
r := pdRouter{}
for _, tt := range tests {
t.Run(tt.name, func(t ... | No issues found. | {ObjectMeta: metav1.ObjectMeta{Labels: map[string]string{"role-name": "other"}}},
{ObjectMeta: metav1.ObjectMeta{Labels: map[string]string{"role-name": "prefill", PodGroupIndex: "1"}}},
},
expectPrefill: 1,
expectDecode: 1,
},
}
r := pdRouter{}
for _, tt := range tests {
t.Run(tt.name, func(t ... | @@ -49,9 +49,9 @@ func TestPDRouter_Route(t *testing.T) {
{
name: "successful routing with both prefill and decode pods",
readyPods: []*v1.Pod{
- {ObjectMeta: metav1.ObjectMeta{Labels: map[string]string{"role-name": "prefill"}, Name: "prefill-1"}, Status: v1.PodStatus{PodIP: "127.0.0.1",
+ {ObjectMeta:... | pkg/plugins/gateway/algorithms/pd_disaggregation_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Select PD workers in same roleset | 1,409 | vllm-project/aibrix | 2,572 | Go | varungup90 | |
constants.StormServiceNameLabelKey: "test-service",
},
},
Spec: orchestrationv1alpha1.RoleSetSpec{
SchedulingStrategy: orchestrationv1alpha1.SchedulingStrategy{
PodGroup: &schedv1alpha1.PodGroupSpec{
MinMember: 3,
},
},
Roles: []orchestrationv1alpha1.RoleSpec{
{
Name: "test-rol... | No issues found. | constants.StormServiceNameLabelKey: "test-service",
},
},
Spec: orchestrationv1alpha1.RoleSetSpec{
SchedulingStrategy: orchestrationv1alpha1.SchedulingStrategy{
PodGroup: &schedv1alpha1.PodGroupSpec{
MinMember: 3,
},
},
Roles: []orchestrationv1alpha1.RoleSpec{
{
Name: "test-rol... | @@ -17,6 +17,7 @@ limitations under the License.
package roleset
import (
+ "reflect"
"testing"
"github.com/stretchr/testify/assert"
@@ -292,3 +293,107 @@ func makeNotReadyPod(name string) *corev1.Pod {
pod.Status.Conditions[0].Status = corev1.ConditionFalse
return pod
}
+
+func TestSortRolesByUpgradeOrd... | pkg/controller/roleset/utils_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Feat: Support role upgrade sequences in stormservice | 1,432 | vllm-project/aibrix | 2,572 | Go | omerap12 | |
testEnv *envtest.Environment // Test environment that spins up etcd and kube-apiserver
ctx context.Context // Context to control the lifecycle of the controller manager
cancel context.CancelFunc // Function to cancel the context and stop the manager
stopCh chan struct{} // Channel to sig... | No issues found. | testEnv *envtest.Environment // Test environment that spins up etcd and kube-apiserver
ctx context.Context // Context to control the lifecycle of the controller manager
cancel context.CancelFunc // Function to cancel the context and stop the manager
stopCh chan struct{} // Channel to sig... | @@ -0,0 +1,185 @@
+/*
+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 a... | test/integration/controller/suit_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Test]: add controller integration test framework | 1,448 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
},
ginkgo.Entry("normal RoleSet create and update replicas with sequential update strategy",
&testValidatingCase{
makeRoleSet: func() *orchestrationapi.RoleSet {
podTemplate := corev1.PodTemplateSpec{
ObjectMeta: metav1.ObjectMeta{
Labels: map[string]string{
"app": "nginx",
... | No issues found. | },
ginkgo.Entry("normal RoleSet create and update replicas with sequential update strategy",
&testValidatingCase{
makeRoleSet: func() *orchestrationapi.RoleSet {
podTemplate := corev1.PodTemplateSpec{
ObjectMeta: metav1.ObjectMeta{
Labels: map[string]string{
"app": "nginx",
... | @@ -0,0 +1,226 @@
+/*
+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 a... | test/integration/controller/roleset_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | add rolset integration test | 1,491 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
}),
},
podMetrics: map[string]map[string]metrics.MetricValue{
"p1": {metrics.GPUBusyTimeRatio: &metrics.SimpleMetricValue{Value: 0.5}},
},
expectErr: false,
expectMsgs: []string{"1.1.1.1:8000"},
},
{
name: "all pods have no metrics - should use random fallback",
readyPods: []*v1.Pod{
... | No issues found. | }),
},
podMetrics: map[string]map[string]metrics.MetricValue{
"p1": {metrics.GPUBusyTimeRatio: &metrics.SimpleMetricValue{Value: 0.5}},
},
expectErr: false,
expectMsgs: []string{"1.1.1.1:8000"},
},
{
name: "all pods have no metrics - should use random fallback",
readyPods: []*v1.Pod{
... | @@ -0,0 +1,241 @@
+/*
+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 a... | pkg/plugins/gateway/algorithms/least_busy_time_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc] Add unit tests for least_busy_time | 1,495 | vllm-project/aibrix | 2,572 | Go | omerap12 | |
Value: roleSet.Annotations[constants.RoleSetIndexAnnotationKey],
}
envMap[constants.RoleNameEnvKey] = v1.EnvVar{
Name: constants.RoleNameEnvKey,
Value: role.Name,
}
envMap[constants.RoleTemplateHashEnvKey] = v1.EnvVar{
Name: constants.RoleTemplateHashEnvKey,
Value: templateHash,
}
if roleIndex != n... | No issues found. | Value: roleSet.Annotations[constants.RoleSetIndexAnnotationKey],
}
envMap[constants.RoleNameEnvKey] = v1.EnvVar{
Name: constants.RoleNameEnvKey,
Value: role.Name,
}
envMap[constants.RoleTemplateHashEnvKey] = v1.EnvVar{
Name: constants.RoleTemplateHashEnvKey,
Value: templateHash,
}
if roleIndex != n... | @@ -26,6 +26,7 @@ import (
"k8s.io/apimachinery/pkg/labels"
"k8s.io/apimachinery/pkg/selection"
"k8s.io/apimachinery/pkg/util/sets"
+ "k8s.io/klog/v2"
"sigs.k8s.io/controller-runtime/pkg/client"
orchestrationv1alpha1 "github.com/vllm-project/aibrix/api/orchestration/v1alpha1"
@@ -357,8 +358,13 @@ func creat... | pkg/controller/roleset/utils.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc]: ignore AlreadyExists and NotFound in controller operations | 1,507 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
},
},
},
},
},
}
},
}),
ginkgo.Entry("apply StormService with sidecar injection annotation "+
"and sidecar runtime image annotation", &testDefaultingCase{
stormservice: func() *orchestrationapi.StormService {
return makeStormServiceWithNoSidecarInjection("stormser... | No issues found. | },
},
},
},
},
}
},
}),
ginkgo.Entry("apply StormService with sidecar injection annotation "+
"and sidecar runtime image annotation", &testDefaultingCase{
stormservice: func() *orchestrationapi.StormService {
return makeStormServiceWithNoSidecarInjection("stormser... | @@ -102,6 +102,9 @@ var _ = ginkgo.Describe("stormservice default webhook", func() {
"app": "my-ai-app",
},
},
+ UpdateStrategy: orchestrationapi.StormServiceUpdateStrategy{
+ Type: orchestrationapi.RollingUpdateStormServiceStrategyType,
+ },
Stateful: true,
Templat... | test/integration/webhook/stormservice_webhook_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc]: add sync-crds makefile cmd | 1,513 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
minExpectedLatency := math.MaxFloat64
sumPromptTokens := 0.0
sumGenerationTokens := 0.0
cntPromt := 0
cntGeneration := 0
for _, pod := range readyPodList.All() {
avgPromptTokens, err := r.cache.GetMetricValueByPodModel(pod.Name, pod.Namespace, ctx.Model, metrics.AvgPromptToksPerReq)
if err != nil {
klog.E... | No issues found. | minExpectedLatency := math.MaxFloat64
sumPromptTokens := 0.0
sumGenerationTokens := 0.0
cntPromt := 0
cntGeneration := 0
for _, pod := range readyPodList.All() {
avgPromptTokens, err := r.cache.GetMetricValueByPodModel(pod.Name, pod.Namespace, ctx.Model, metrics.AvgPromptToksPerReq)
if err != nil {
klog.E... | @@ -94,30 +94,40 @@ func (r leastExpectedLatencyRouter) Route(ctx *types.RoutingContext, readyPodLis
}
// expected prefill latency
- avgPromptTokens, err := r.cache.GetMetricValueByPodModel(pod.Name, pod.Namespace, ctx.Model, metrics.AvgPromptToksPerReq)
+ avgPromptTokensMetric, err := r.cache.GetMetricValueB... | pkg/plugins/gateway/algorithms/least_latency.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bug] use fallback value to prevent divide zero error | 1,517 | vllm-project/aibrix | 2,572 | Go | zyfy29 | |
createBudget--
}
}
if _, err = createPodSetsInBatch(ctx, p.cli, toCreate); err != nil {
return err
}
if _, err = deletePodSetsInBatch(ctx, p.cli, toDelete); err != nil {
return err
}
p.printLog(roleSet, role, toCreate, toDelete)
return nil
}
func (p *PodSetRoleSyncer) AllReady(ctx context.Context, role... | No issues found. | createBudget--
}
}
if _, err = createPodSetsInBatch(ctx, p.cli, toCreate); err != nil {
return err
}
if _, err = deletePodSetsInBatch(ctx, p.cli, toDelete); err != nil {
return err
}
p.printLog(roleSet, role, toCreate, toDelete)
return nil
}
func (p *PodSetRoleSyncer) AllReady(ctx context.Context, role... | @@ -442,7 +442,7 @@ func getRolePodSets(ctx context.Context, cli client.Client, namespace, roleSetNa
func filterActivePodSets(allPodSets []*orchestrationv1alpha1.PodSet) ([]*orchestrationv1alpha1.PodSet, []*orchestrationv1alpha1.PodSet) {
var active, inactive []*orchestrationv1alpha1.PodSet
for _, podSet := range ... | pkg/controller/roleset/podset_rollsyncer.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bug] fix count ready podset condition | 1,518 | vllm-project/aibrix | 2,572 | Go | Epsilon314 | |
// Remove from all maps
for node := range nodeMap {
delete(h.histogram, node)
delete(h.nodeToCount, node)
delete(h.hitTokens, node)
delete(h.promptTokens, node)
delete(h.decodingSize, node)
delete(h.podAllocations, node)
}
}
func (h *SlidingWindowHistogram) removeOldEntries(currentTime time.Time) {
h.m... | No issues found. | // Remove from all maps
for node := range nodeMap {
delete(h.histogram, node)
delete(h.nodeToCount, node)
delete(h.hitTokens, node)
delete(h.promptTokens, node)
delete(h.decodingSize, node)
delete(h.podAllocations, node)
}
}
func (h *SlidingWindowHistogram) removeOldEntries(currentTime time.Time) {
h.m... | @@ -542,7 +542,9 @@ func (p *prefixCacheAndLoadRouter) Route(ctx *types.RoutingContext, readyPodList
targetPod = pod
}
}
- klog.InfoS("Lowest cost pod: %s", "podName", targetPod.Name)
+ if targetPod != nil {
+ klog.InfoS("Lowest cost pod: %s", "podName", targetPod.Name)
+ }
}
if targetPod == nil ... | pkg/plugins/gateway/algorithms/prefix_cache_preble.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc] (test): add unit test for prefix_cache_preble | 1,519 | vllm-project/aibrix | 2,572 | Go | zhengkezhou1 | |
blocks. Each tuple in it represents (block id, address offset).
on_mr_release: Callback function when memory region is released.
Returns:
KVCacheHandle.
"""
block_ntokens = block_spec.block_ntokens
block_nbytes = block_spec.block_nbytes
block_... | No issues found. | blocks. Each tuple in it represents (block id, address offset).
on_mr_release: Callback function when memory region is released.
Returns:
KVCacheHandle.
"""
block_ntokens = block_spec.block_ntokens
block_nbytes = block_spec.block_nbytes
block_... | @@ -100,7 +100,7 @@ def __len__(self) -> int:
def create(
blocks: Sequence[torch.Tensor],
block_spec: KVCacheBlockSpec,
- slot_mapping: Sequence[int] | torch.Tensor,
+ select: Sequence[Tuple[int, int]],
on_mr_release: Callable[[torch.Tensor, int, int], None] | None = None,
... | python/aibrix_kvcache/aibrix_kvcache/cache_handle.py | 0 | Python | 1 | none | 16 | 51 | 51 | true | [Refactor] KVCache: simplify external cache handle create API | 1,542 | vllm-project/aibrix | 2,572 | Go | DwyaneShi | |
template := orchestrationapi.RoleSetTemplateSpec{
ObjectMeta: metav1.ObjectMeta{
Labels: map[string]string{
"app": "my-ai-app",
},
},
Spec: &orchestrationapi.RoleSetSpec{
UpdateStrategy: orchestrationapi.InterleaveRoleSetStrategyType,
Roles: []orchestrationapi.RoleSpec{},
},
}
w.stor... | No issues found. | template := orchestrationapi.RoleSetTemplateSpec{
ObjectMeta: metav1.ObjectMeta{
Labels: map[string]string{
"app": "my-ai-app",
},
},
Spec: &orchestrationapi.RoleSetSpec{
UpdateStrategy: orchestrationapi.InterleaveRoleSetStrategyType,
Roles: []orchestrationapi.RoleSpec{},
},
}
w.stor... | @@ -17,9 +17,8 @@ limitations under the License.
package wrapper
import (
- "strconv"
-
"fmt"
+ "strconv"
corev1 "k8s.io/api/core/v1"
v1 "k8s.io/api/core/v1"
@@ -29,7 +28,7 @@ import (
"k8s.io/utils/ptr"
orchestrationapi "github.com/vllm-project/aibrix/api/orchestration/v1alpha1"
- webhook "github.com... | test/utils/wrapper/stormservice.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc]: add stormservice integration test | 1,544 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
package v1alpha1
// RoleStatusApplyConfiguration represents a declarative configuration of the RoleStatus type for use
// with apply.
type RoleStatusApplyConfiguration struct {
Name *string `json:"name,omitempty"`
Replicas *int32 `json:"replicas,omitempty"`
ReadyReplicas *int32 ... | No issues found. | package v1alpha1
// RoleStatusApplyConfiguration represents a declarative configuration of the RoleStatus type for use
// with apply.
type RoleStatusApplyConfiguration struct {
Name *string `json:"name,omitempty"`
Replicas *int32 `json:"replicas,omitempty"`
ReadyReplicas *int32 ... | @@ -0,0 +1,83 @@
+/*
+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 ag... | pkg/client/applyconfiguration/orchestration/v1alpha1/rolestatus.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [API]: add roleStatuses field in stormservice api | 1,599 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
for _, pod := range decodePods {
if podRoleSet, exists := pod.Labels[PDRoleSetIdentifier]; exists && podRoleSet == prefillRoleSet {
filteredDecodePods = append(filteredDecodePods, pod)
}
}
if len(filteredDecodePods) == 0 {
return nil
}
// prefer decode pod with least running requests
decodePod := select... | No issues found. | for _, pod := range decodePods {
if podRoleSet, exists := pod.Labels[PDRoleSetIdentifier]; exists && podRoleSet == prefillRoleSet {
filteredDecodePods = append(filteredDecodePods, pod)
}
}
if len(filteredDecodePods) == 0 {
return nil
}
// prefer decode pod with least running requests
decodePod := select... | @@ -145,11 +145,22 @@ func (r *pdRouter) evaluatePrefixCache(ctx *types.RoutingContext, prefillPods []
matchedPods, prefixHashes := r.prefixCacheIndexer.MatchPrefix(tokens, ctx.Model, readyPodsMap)
var prefillPod *v1.Pod
- if len(matchedPods) > 0 {
+ // check for load imbalance first
+ targetPod, isImbalanced := ... | pkg/plugins/gateway/algorithms/pd_disaggregation.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Improve pdRouter with load-aware routing | 1,601 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
return processingRes, complete
}
}
}
var requestEnd string
if totalTokens != 0 {
complete = true
// Count token per user.
if user.Name != "" {
tpm, err := s.ratelimiter.Incr(ctx, fmt.Sprintf("%v_TPM_CURRENT", user.Name), totalTokens)
if err != nil {
return generateErrorResponse(
envoy... | No issues found. | return processingRes, complete
}
}
}
var requestEnd string
if totalTokens != 0 {
complete = true
// Count token per user.
if user.Name != "" {
tpm, err := s.ratelimiter.Incr(ctx, fmt.Sprintf("%v_TPM_CURRENT", user.Name), totalTokens)
if err != nil {
return generateErrorResponse(
envoy... | @@ -23,6 +23,7 @@ import (
"fmt"
"io"
"net/http"
+ "strings"
"github.com/openai/openai-go"
"github.com/openai/openai-go/packages/ssestream"
@@ -37,7 +38,7 @@ import (
type OpenAIResponse struct {
Model string `json:"model"`
- Usage struct {
+ Usage *struct {
PromptTokens int64 `json:"prompt_toke... | pkg/plugins/gateway/gateway_rsp_body.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Add gateway-plugin support to generate image and video | 1,603 | vllm-project/aibrix | 2,572 | Go | varungup90 | |
setCondition(&pa, string(condition.Type), metav1.ConditionStatus(condition.Status), condition.Reason, condition.Message)
}
if err = r.Status().Update(ctx, &pa); err != nil {
klog.ErrorS(err, "Failed to update PodAutoscaler status")
}
// Return with no error and no requeue needed.
return ctrl.Result{}, nil
}
/... | No issues found. | setCondition(&pa, string(condition.Type), metav1.ConditionStatus(condition.Status), condition.Reason, condition.Message)
}
if err = r.Status().Update(ctx, &pa); err != nil {
klog.ErrorS(err, "Failed to update PodAutoscaler status")
}
// Return with no error and no requeue needed.
return ctrl.Result{}, nil
}
/... | @@ -52,7 +52,6 @@ import (
"k8s.io/apimachinery/pkg/apis/meta/v1/unstructured"
"k8s.io/apimachinery/pkg/runtime"
"k8s.io/apimachinery/pkg/runtime/schema"
- utilruntime "k8s.io/apimachinery/pkg/util/runtime"
"k8s.io/client-go/discovery"
memcache "k8s.io/client-go/discovery/cached/memory"
"k8s.io/client-go/re... | pkg/controller/podautoscaler/podautoscaler_controller.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | feat: add scaling history decision in the status spec | 1,618 | vllm-project/aibrix | 2,572 | Go | omerap12 | |
# 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... | No issues found. | # 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... | @@ -15,5 +15,6 @@
from .async_base import AsyncBase
from .cached_pyobject import CachedPyObjectBase
from .object_pool import ObjectPool
+from .threading import ConditionalLock
-__all__ = ["AsyncBase", "ObjectPool", "CachedPyObjectBase"]
+__all__ = ["AsyncBase", "ObjectPool", "CachedPyObjectBase", "ConditionalLock"... | python/aibrix_kvcache/aibrix_kvcache/common/__init__.py | 0 | Python | 1 | none | 16 | 21 | 21 | true | [Feature] KVCache: support multi-threading mode | 1,628 | vllm-project/aibrix | 2,572 | Go | DwyaneShi | |
}
// ScalingStrategyType defines the type for scaling strategies.
type ScalingStrategyType string
const (
// HPA represents the Kubernetes native Horizontal Pod Autoscaler.
HPA ScalingStrategyType = "HPA"
// KPA represents the KNative Pod Autoscaling Algorithm
KPA ScalingStrategyType = "KPA"
// APA represents ... | No issues found. | }
// ScalingStrategyType defines the type for scaling strategies.
type ScalingStrategyType string
const (
// HPA represents the Kubernetes native Horizontal Pod Autoscaler.
HPA ScalingStrategyType = "HPA"
// KPA represents the KNative Pod Autoscaling Algorithm
KPA ScalingStrategyType = "KPA"
// APA represents ... | @@ -132,14 +132,18 @@ type MetricSource struct {
// Specifies how to fetch metrics: from individual pods, Kubernetes APIs, or external services
// +kubebuilder:validation:Enum={pod,resource,custom,external,domain}
MetricSourceType MetricSourceType `json:"metricSourceType"`
- // Protocol for metric collection
+ //... | api/autoscaling/v1alpha1/podautoscaler_types.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bug]: Make optional fields in MetricSource | 1,648 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
for i := 0; i < count && len(remainingPods) > 0; i++ {
pod, err := r.scheduler.SelectPod(ctx, instance.Name, remainingPods)
if err != nil {
return nil, err
}
selectedPods = append(selectedPods, *pod)
// Remove selected pod from remaining pods to avoid selecting it again
for j, p := range remainingPod... | No issues found. |
for i := 0; i < count && len(remainingPods) > 0; i++ {
pod, err := r.scheduler.SelectPod(ctx, instance.Name, remainingPods)
if err != nil {
return nil, err
}
selectedPods = append(selectedPods, *pod)
// Remove selected pod from remaining pods to avoid selecting it again
for j, p := range remainingPod... | @@ -360,13 +360,18 @@ func (r *ModelAdapterReconciler) DoReconcile(ctx context.Context, req ctrl.Reque
oldInstance := instance.DeepCopy()
+ // Save old instances list before reconcileReplicas potentially modifies it
+ // This is needed to detect pod removal in reconcileLoading
+ oldInstances := make([]string, len... | pkg/controller/modeladapter/modeladapter_controller.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Refactor the model adapter replicas feature | 1,670 | vllm-project/aibrix | 2,572 | Go | Jeffwan | |
var anyValid bool
for _, metricSource := range metricsSources {
// Extract per-request configuration
metricKey := types.MetricKey{
Namespace: pa.Namespace,
Name: pa.Spec.ScaleTargetRef.Name,
MetricName: metricSource.TargetMetric,
PaNamespace: pa.Namespace,
PaName: pa.Name,
}
re... | No issues found. | var anyValid bool
for _, metricSource := range metricsSources {
// Extract per-request configuration
metricKey := types.MetricKey{
Namespace: pa.Namespace,
Name: pa.Spec.ScaleTargetRef.Name,
MetricName: metricSource.TargetMetric,
PaNamespace: pa.Namespace,
PaName: pa.Name,
}
re... | @@ -123,27 +123,114 @@ func (a *DefaultAutoScaler) getOrCreateAlgorithm(strategy autoscalingv1alpha1.Sc
return algo
}
+// ComputeDesiredReplicas computes desired replicas based on all metrics in MetricsSources.
+// It returns the maximum recommended replicas across all valid metrics.
func (a *DefaultAutoScaler) C... | pkg/controller/podautoscaler/autoscaler.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Feat]: support mulit metrics for podautoscaler | 1,688 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
constants.PodGroupIndexLabelKey: "0",
})
cl := fake.NewClientBuilder().WithScheme(sch).
WithObjects(p0, p1, pWrongApp).
Build()
pa := autoscalingv1alpha1.PodAutoscaler{}
pa.Namespace = ns
// Scale target is a Deployment (not StormService).
scaleObj := buildScaleObject("apps/v1", "Deployment", ns, "foo-de... | No issues found. | constants.PodGroupIndexLabelKey: "0",
})
cl := fake.NewClientBuilder().WithScheme(sch).
WithObjects(p0, p1, pWrongApp).
Build()
pa := autoscalingv1alpha1.PodAutoscaler{}
pa.Namespace = ns
// Scale target is a Deployment (not StormService).
scaleObj := buildScaleObject("apps/v1", "Deployment", ns, "foo-de... | @@ -0,0 +1,328 @@
+/*
+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 a... | pkg/controller/podautoscaler/podautoscaler_controller_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | fix: autoscaler for the podset only collect metrics from rank0 | 1,704 | vllm-project/aibrix | 2,572 | Go | scarlet25151 | |
name string
routingCtx *types.RoutingContext
responseStatus string // value of :status header
respHeaders map[string]string
expected testResponse
}
tests := []testCase{
{
name: "successful response (200)",
routingCtx: createRoutingCtx(true, map[string]string{"X-... | No issues found. | name string
routingCtx *types.RoutingContext
responseStatus string // value of :status header
respHeaders map[string]string
expected testResponse
}
tests := []testCase{
{
name: "successful response (200)",
routingCtx: createRoutingCtx(true, map[string]string{"X-... | @@ -0,0 +1,185 @@
+/*
+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 a... | pkg/plugins/gateway/gateway_rsp_headers_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | feature: add simple session affinity plugins in gateway plugin | 1,751 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
Subsystem: constants.AibrixSubsystemName,
Name: "kvcache_zmq_errors_total",
Help: "Total number of errors encountered",
},
[]string{LabelPodKey, LabelErrorType},
),
// State metrics
zmqConnectionStatus: prometheus.NewGaugeVec(
prometheus.GaugeOpts{
Subsystem: constants.AibrixSu... | No issues found. | Subsystem: constants.AibrixSubsystemName,
Name: "kvcache_zmq_errors_total",
Help: "Total number of errors encountered",
},
[]string{LabelPodKey, LabelErrorType},
),
// State metrics
zmqConnectionStatus: prometheus.NewGaugeVec(
prometheus.GaugeOpts{
Subsystem: constants.AibrixSu... | @@ -21,6 +21,8 @@ import (
"github.com/prometheus/client_golang/prometheus"
"k8s.io/klog/v2"
+
+ "github.com/vllm-project/aibrix/pkg/constants"
)
// Metric labels
@@ -97,101 +99,114 @@ func createKVCacheMetrics() *KVCacheMetrics {
// Connection metrics
zmqConnectionTotal: prometheus.NewCounterVec(
p... | pkg/cache/kvcache/metrics.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | refactor(metrics): use Subsystem for aibrix metrics instead of embedding in Name | 1,756 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
func buildRuntimeSidecarContainer(sidecarImage, engineType string) corev1.Container {
return corev1.Container{
Name: SidecarName,
Image: sidecarImage,
Command: []string{
SidecarCommand,
"--port", fmt.Sprintf("%d", SidecarPort),
},
Env: []corev1.EnvVar{
{
Name: "INFERENCE_ENGINE",
Value: en... | No issues found. | func buildRuntimeSidecarContainer(sidecarImage, engineType string) corev1.Container {
return corev1.Container{
Name: SidecarName,
Image: sidecarImage,
Command: []string{
SidecarCommand,
"--port", fmt.Sprintf("%d", SidecarPort),
},
Env: []corev1.EnvVar{
{
Name: "INFERENCE_ENGINE",
Value: en... | @@ -32,7 +32,7 @@ const (
// SidecarInjectionRuntimeImageAnnotation Annotation used to specify a custom image for the sidecar container
SidecarInjectionRuntimeImageAnnotation = "model.aibrix.ai/sidecar-runtime-image"
SidecarName = "aibrix-runtime"
- SidecarImage ... | pkg/webhook/sidecar_injection.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Misc] upgrade sidecar webhook image from v0.4.0 to v0.5.0 | 1,762 | vllm-project/aibrix | 2,572 | Go | nurali-techie | |
Controller: boolPtr(true),
BlockOwnerDeletion: boolPtr(true),
}
assert.NoError(t, ValidateControllerRef(ref))
table := []struct {
name string
ref *metav1.OwnerReference
err string
}{
{"nil ref", nil, "controllerRef is nil"},
{"empty version", &metav1.OwnerReference{Kind: "A", Controller: bo... | No issues found. | Controller: boolPtr(true),
BlockOwnerDeletion: boolPtr(true),
}
assert.NoError(t, ValidateControllerRef(ref))
table := []struct {
name string
ref *metav1.OwnerReference
err string
}{
{"nil ref", nil, "controllerRef is nil"},
{"empty version", &metav1.OwnerReference{Kind: "A", Controller: bo... | @@ -132,4 +132,12 @@ func TestMinIntHelpers(t *testing.T) {
assert.Equal(t, -5, MinInt(-5, 0))
}
+func TestShorten(t *testing.T) {
+ assert.Equal(t, "abcdefghijklmnopqrstuvwxyz", Shorten("abcdefghijklmnopqrstuvwxyz", false, false))
+ assert.Equal(t, "aijklmnopqrstuvwxyz-1234567890-1234567890-1234567890-1234567890"... | pkg/controller/util/orchestration/util_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | fix(controller): shorten resource name | 1,777 | vllm-project/aibrix | 2,572 | Go | zhixian82 | |
Obj()
},
wantStormService: func() *orchestrationapi.StormService {
return wrapper.MakeStormService("st-with-no-inject-sidecar").
Namespace(ns.Name).
WithDefaultConfiguration().
Obj()
},
}),
ginkgo.Entry("apply StormService with sidecar injection annotation", &testDefaultingCase{
... | No issues found. | Obj()
},
wantStormService: func() *orchestrationapi.StormService {
return wrapper.MakeStormService("st-with-no-inject-sidecar").
Namespace(ns.Name).
WithDefaultConfiguration().
Obj()
},
}),
ginkgo.Entry("apply StormService with sidecar injection annotation", &testDefaultingCase{
... | @@ -17,6 +17,8 @@ limitations under the License.
package webhook
import (
+ "strings"
+
"github.com/google/go-cmp/cmp/cmpopts"
"github.com/onsi/ginkgo/v2"
"github.com/onsi/gomega"
@@ -74,28 +76,28 @@ var _ = ginkgo.Describe("stormservice default webhook", func() {
},
ginkgo.Entry("apply StormService wit... | test/integration/webhook/stormservice_webhook_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bug]: add validation in stormservice webhook | 1,778 | vllm-project/aibrix | 2,572 | Go | googs1025 | |
},
originalBody: `{"messages":[{"role":"user","content":"test"}]}`,
expectError: false,
expectKV: true,
},
{
name: "no kv params in response",
responseData: map[string]any{"other": "data"},
originalBody: `{"messages":[{"role":"user","content":"test"}]}`,
expectError: false,
... | No issues found. | },
originalBody: `{"messages":[{"role":"user","content":"test"}]}`,
expectError: false,
expectKV: true,
},
{
name: "no kv params in response",
responseData: map[string]any{"other": "data"},
originalBody: `{"messages":[{"role":"user","content":"test"}]}`,
expectError: false,
... | @@ -1284,3 +1284,120 @@ func TestLoadImbalanceSelectDecodePod(t *testing.T) {
})
}
}
+
+func TestIsPodSuitableForPromptLength(t *testing.T) {
+ tests := []struct {
+ name string
+ podLabels map[string]string
+ promptLength int
+ expected bool
+ }{
+ {
+ name: "no prompt length range configur... | pkg/plugins/gateway/algorithms/pd_disaggregation_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | feat:router based on multi deployment configs | 1,835 | vllm-project/aibrix | 2,572 | Go | erictanjn | |
{
name: "array of strings exceeding total tokens",
input: openai.EmbeddingNewParams{
Input: openai.EmbeddingNewParamsInputUnion{
OfArrayOfStrings: func() []string {
// Create an array that would exceed MaxTotalTokens
largeString := strings.Repeat("word ", MaxTotalTokens/2)
return []st... | No issues found. | {
name: "array of strings exceeding total tokens",
input: openai.EmbeddingNewParams{
Input: openai.EmbeddingNewParamsInputUnion{
OfArrayOfStrings: func() []string {
// Create an array that would exceed MaxTotalTokens
largeString := strings.Repeat("word ", MaxTotalTokens/2)
return []st... | @@ -334,6 +334,79 @@ func Test_ValidateRequestBody_Embeddings(t *testing.T) {
}
}
+func Test_ValidateRequestBody_Rerank(t *testing.T) {
+ testCases := []struct {
+ message string
+ requestPath string
+ requestBody []byte
+ model string
+ messages string
+ stream bool
+ user utils.Us... | pkg/plugins/gateway/util_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [feat]: support rerank models | 1,837 | vllm-project/aibrix | 2,572 | Go | sanmuny | |
assert.Equal(t, tt.routingAlgo, routingCtx.Algorithm)
// Verify both routing headers are set
foundRoutingStrategy := false
foundTargetPod := false
for _, header := range resp.GetRequestBody().GetResponse().GetHeaderMutation().GetSetHeaders() {
if header.Header.Key == HeaderRoutingStrategy {
... | No issues found. | assert.Equal(t, tt.routingAlgo, routingCtx.Algorithm)
// Verify both routing headers are set
foundRoutingStrategy := false
foundTargetPod := false
for _, header := range resp.GetRequestBody().GetResponse().GetHeaderMutation().GetSetHeaders() {
if header.Header.Key == HeaderRoutingStrategy {
... | @@ -28,12 +28,13 @@ import (
configPb "github.com/envoyproxy/go-control-plane/envoy/config/core/v3"
extProcPb "github.com/envoyproxy/go-control-plane/envoy/service/ext_proc/v3"
envoyTypePb "github.com/envoyproxy/go-control-plane/envoy/type/v3"
+ v1 "k8s.io/api/core/v1"
+ metav1 "k8s.io/apimachinery/pkg/apis/meta/... | pkg/plugins/gateway/gateway_req_body_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [bugfix] ReqBody receive stream=true when calling /v1/completion | 1,850 | vllm-project/aibrix | 2,572 | Go | DengHom | |
func (r *leastRequestRouter) SubscribedMetrics() []string {
return []string{
metrics.RealtimeNumRequestsRunning,
}
}
func selectTargetPodWithLeastRequestCount(cache cache.Cache, readyPods []*v1.Pod) *v1.Pod {
var targetPod *v1.Pod
targetPods := []string{}
minCount := math.MaxInt32
podRequestCount := getReque... | No issues found. |
func (r *leastRequestRouter) SubscribedMetrics() []string {
return []string{
metrics.RealtimeNumRequestsRunning,
}
}
func selectTargetPodWithLeastRequestCount(cache cache.Cache, readyPods []*v1.Pod) *v1.Pod {
var targetPod *v1.Pod
targetPods := []string{}
minCount := math.MaxInt32
podRequestCount := getReque... | @@ -17,8 +17,11 @@ limitations under the License.
package routingalgorithms
import (
+ "fmt"
"math"
"math/rand"
+ "strconv"
+ "strings"
"github.com/vllm-project/aibrix/pkg/cache"
"github.com/vllm-project/aibrix/pkg/metrics"
@@ -45,14 +48,19 @@ func NewLeastRequestRouter() (types.Router, error) {
return... | pkg/plugins/gateway/algorithms/least_request.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [feat] add distribute-dp api server least_request route | 1,866 | vllm-project/aibrix | 2,572 | Go | paranoidRick | |
contentTypeKey = "content-type"
)
func (s *Server) HandleRequestHeaders(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest) (*extProcPb.ProcessingResponse, utils.User, int64, *types.RoutingContext) {
var username, requestPath string
var user utils.User
var rpm int64
var err error
var errRe... | No issues found. | contentTypeKey = "content-type"
)
func (s *Server) HandleRequestHeaders(ctx context.Context, requestID string, req *extProcPb.ProcessingRequest) (*extProcPb.ProcessingResponse, utils.User, int64, *types.RoutingContext) {
var username, requestPath string
var user utils.User
var rpm int64
var err error
var errRe... | @@ -106,14 +106,10 @@ func (s *Server) HandleRequestHeaders(ctx context.Context, requestID string, req
},
})
- switch requestPath {
- case "/v1/image/generations":
- headers = buildEnvoyProxyHeaders(headers, ":path", "/generate")
- case "/v1/video/generations":
- headers = buildEnvoyProxyHeaders(headers, ":pat... | pkg/plugins/gateway/gateway_req_headers.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Support OpenAI compatible image generation endpoint | 1,867 | vllm-project/aibrix | 2,572 | Go | Jeffwan | |
Kind: "Deployment",
},
ScalingStrategy: autoscalingv1alpha1.HPA,
MetricsSources: []autoscalingv1alpha1.MetricSource{
{
MetricSourceType: autoscalingv1alpha1.RESOURCE,
TargetMetric: "cpu",
TargetValue: "50",
},
},
},
},
expectError: false,
}... | No issues found. | Kind: "Deployment",
},
ScalingStrategy: autoscalingv1alpha1.HPA,
MetricsSources: []autoscalingv1alpha1.MetricSource{
{
MetricSourceType: autoscalingv1alpha1.RESOURCE,
TargetMetric: "cpu",
TargetValue: "50",
},
},
},
},
expectError: false,
}... | @@ -0,0 +1,130 @@
+/*
+Copyright 2026 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 a... | pkg/webhook/podautoscaler_webhook_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bug] Prevent div by zero in APA | 1,879 | vllm-project/aibrix | 2,572 | Go | alpe | |
err := recover()
if err != nil {
klog.Errorf("slowStartBatch panic: %v", err)
}
}()
defer wg.Done()
if err := fn(index); err != nil {
errCh <- err
}
}(finished + i)
}
wg.Wait()
curSuccesses := batchSize - len(errCh)
successes += curSuccesses
if len(errCh) > 0 {
... | No issues found. | err := recover()
if err != nil {
klog.Errorf("slowStartBatch panic: %v", err)
}
}()
defer wg.Done()
if err := fn(index); err != nil {
errCh <- err
}
}(finished + i)
}
wg.Wait()
curSuccesses := batchSize - len(errCh)
successes += curSuccesses
if len(errCh) > 0 {
... | @@ -268,7 +268,7 @@ func MinInt(a, b int) int {
return b
}
-func EnsurePodGroupExist(ctx context.Context, dc dynamic.Interface, podGroup client.Object, name, namespace string) (created bool, err error) {
+func EnsurePodGroup(ctx context.Context, dc dynamic.Interface, podGroup client.Object, name, namespace string)... | pkg/controller/util/orchestration/util.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Feat]: update PodGroup when RoleSet.SchedulingStrategy changed | 1,889 | vllm-project/aibrix | 2,572 | Go | fungaren | |
) -> str:
"""
Download from GCS.
Credentials dict should contain:
- gcp_service_account_json: Service account JSON string
"""
try:
from google.cloud import storage
from google.oauth2 import service_account
except ImportError:
... | No issues found. | ) -> str:
"""
Download from GCS.
Credentials dict should contain:
- gcp_service_account_json: Service account JSON string
"""
try:
from google.cloud import storage
from google.oauth2 import service_account
except ImportError:
... | @@ -14,6 +14,8 @@
"""Artifact downloaders for different storage backends."""
+import asyncio
+import functools
import os
from abc import ABC, abstractmethod
from pathlib import Path
@@ -58,6 +60,18 @@ class S3ArtifactDownloader(ArtifactDownloader):
async def download(
self, source_url: str, loca... | python/aibrix/aibrix/runtime/downloaders.py | 0 | Python | 1 | none | 16 | 51 | 51 | true | [Bug] Refactor downloader and artifact_service to be non-blocking | 1,895 | vllm-project/aibrix | 2,572 | Go | xieus | |
self.download_calls = []
fake_tos.last_client = self
def list_objects_type2(self, bucket_name: str, **kwargs):
prefix = kwargs.get("prefix", "")
keys = [
prefix,
f"{prefix}.git/",
f"{prefix}sub/",
f"... | No issues found. | self.download_calls = []
fake_tos.last_client = self
def list_objects_type2(self, bucket_name: str, **kwargs):
prefix = kwargs.get("prefix", "")
keys = [
prefix,
f"{prefix}.git/",
f"{prefix}sub/",
f"... | @@ -0,0 +1,131 @@
+# 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/tests/runtime/test_runtime_downloaders_tos.py | 0 | Python | 1 | none | 16 | 51 | 51 | true | Feat: delegate artifacts preparation to lora downloader | 1,898 | vllm-project/aibrix | 2,572 | Go | scarlet25151 | |
t.Fatalf("computeMetricBasedReplicas returned error: %v", err)
}
if res == nil {
t.Fatalf("expected non-nil result")
}
if as.lastRequest == nil {
t.Fatalf("autoscaler did not receive request")
}
if as.lastRequest.CurrentReplicas != currentReplicas {
t.Fatalf("CurrentReplicas mismatch: got=%d want=%d", as.... | No issues found. | t.Fatalf("computeMetricBasedReplicas returned error: %v", err)
}
if res == nil {
t.Fatalf("expected non-nil result")
}
if as.lastRequest == nil {
t.Fatalf("autoscaler did not receive request")
}
if as.lastRequest.CurrentReplicas != currentReplicas {
t.Fatalf("CurrentReplicas mismatch: got=%d want=%d", as.... | @@ -29,6 +29,7 @@ import (
"k8s.io/apimachinery/pkg/runtime"
"k8s.io/apimachinery/pkg/selection"
"k8s.io/client-go/kubernetes/scheme"
+ "k8s.io/utils/ptr"
"sigs.k8s.io/controller-runtime/pkg/client/fake"
autoscalingv1alpha1 "github.com/vllm-project/aibrix/api/autoscaling/v1alpha1"
@@ -111,6 +112,28 @@ func ... | pkg/controller/podautoscaler/podautoscaler_controller_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | [Bugfix] add APA support to stormService without podGroupSize (#1913) | 1,917 | vllm-project/aibrix | 2,572 | Go | rayne-Li | |
}
if len(prefixHashes) > 0 && targetPrefillPod != nil {
r.enqueuePrefixUpdate(prefixHashes, routingCtx.Model, targetPrefillPod.Name)
}
r.countersMu.Lock()
r.selectionCounts[targetPrefillPod.Name]++
r.selectionCounts[targetDecodePod.Name]++
r.countersMu.Unlock()
metrics.EmitCounterMetric(routingCtx, targetPr... | No issues found. | }
if len(prefixHashes) > 0 && targetPrefillPod != nil {
r.enqueuePrefixUpdate(prefixHashes, routingCtx.Model, targetPrefillPod.Name)
}
r.countersMu.Lock()
r.selectionCounts[targetPrefillPod.Name]++
r.selectionCounts[targetDecodePod.Name]++
r.countersMu.Unlock()
metrics.EmitCounterMetric(routingCtx, targetPr... | @@ -127,7 +127,7 @@ func NewPDRouter() (types.Router, error) {
pdRouter := pdRouter{
cache: c,
tokenizer: tokenizerObj,
- prefixCacheIndexer: prefixcacheindexer.NewPrefixHashTable(),
+ prefixCacheIndexer: prefixcacheindexer.GetSharedPrefixHashTable(),
prefillRequestTracker... | pkg/plugins/gateway/algorithms/pd_disaggregation.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | fix: use a shared prefix cache indexer across routing-strategies | 1,939 | vllm-project/aibrix | 2,572 | Go | varungup90 | |
}
}
/*
* Template arguments:
* - TOnload: True if offload_kv_cache to kv_cache.
* - TLayout: The layout of offload_kv_cache.
* Args:
* - offload_kv_cache: Supports LCND and NCLD layouts.
* LCND: [num_blocks, num_layers, 2, block_size, dim]
* NCLD: [num_blocks, block_siz... | No issues found. | }
}
/*
* Template arguments:
* - TOnload: True if offload_kv_cache to kv_cache.
* - TLayout: The layout of offload_kv_cache.
* Args:
* - offload_kv_cache: Supports LCND and NCLD layouts.
* LCND: [num_blocks, num_layers, 2, block_size, dim]
* NCLD: [num_blocks, block_siz... | @@ -52,15 +52,23 @@ torch::Tensor get_device_ptrs(const std::vector<TTensor> &tensors) {
return gpu_data_ptrs;
}
-__device__ __forceinline__ int64_t
-get_kv_cache_offset(const int64_t kv_type, const int64_t num_blocks,
- const int64_t block_size, const int64_t embed_dim,
- co... | python/aibrix_kvcache/csrc/cache_kernels.cu | 0 | CUDA | 1 | none | 16 | 51 | 51 | true | [Feature] KVCache support block-first layout | 1,947 | vllm-project/aibrix | 2,572 | Go | DwyaneShi | |
"gateway_pod",
}, labelNames)
require.Equal(t, []string{
"ns1",
"llm-pod-1",
"m1",
"vllm",
"rs1",
"role1",
"0",
"gw-pod-1",
}, labelValues)
}
func TestEmitMetricToPrometheus_GaugeAndCounter(t *testing.T) {
var gaugeCalls []struct {
name string
value float64
}
originalGaugeFn := metrics.Se... | No issues found. | "gateway_pod",
}, labelNames)
require.Equal(t, []string{
"ns1",
"llm-pod-1",
"m1",
"vllm",
"rs1",
"role1",
"0",
"gw-pod-1",
}, labelValues)
}
func TestEmitMetricToPrometheus_GaugeAndCounter(t *testing.T) {
var gaugeCalls []struct {
name string
value float64
}
originalGaugeFn := metrics.Se... | @@ -15,6 +15,11 @@ limitations under the License.
package cache
import (
+ "encoding/base64"
+ "fmt"
+ "net/http"
+ "net/http/httptest"
+ "sync"
"testing"
"time"
@@ -23,6 +28,7 @@ import (
"github.com/vllm-project/aibrix/pkg/metrics"
v1 "k8s.io/api/core/v1"
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
... | pkg/cache/cache_metrics_test.go | 0 | Go | 1 | none | 16 | 51 | 51 | true | Feat: load prom auth from secret | 1,949 | vllm-project/aibrix | 2,572 | Go | scarlet25151 | |
# ruff: noqa
import json
import random
import string
from vllm import LLM
from vllm.sampling_params import SamplingParams
# This script is an offline demo for running Pixtral.
#
# If you want to run a server/client setup, please follow this code:
#
# - Server:
#
# ```bash
# vllm serve mistralai/Mistral-7B-Instruct-v0... | Example file of how to use Mistral models with function calling | # ruff: noqa
import json
import random
import string
from vllm import LLM
from vllm.sampling_params import SamplingParams
# This script is an offline demo for function calling
#
# If you want to run a server/client setup, please follow this code:
#
# - Server:
#
# ```bash
# vllm serve mistralai/Mistral-7B-Instruct-v0... | @@ -0,0 +1,138 @@
+# ruff: noqa | examples/offline_chat_with_tools.py | 1 | Python | 0.357 | suggestion | 63 | 26 | 26 | false | [Model] Add mistral function calling format to all models loaded with "mistral" format | 8,515 | vllm-project/vllm | 71,895 | Python | patrickvonplaten | patrickvonplaten |
return {}
def encode(self, prompt: str) -> List[int]:
# `encode` should only be used for prompt completion
# it should never be used for chat_completion.
# For chat completion use `apply_chat_template`
return self.tokenizer.encode(prompt, bos=True, eos=False)
def apply_... | ```suggestion
return "".join(
t for t in tokens
if t not in self.tokenizer._all_special_tokens
)
```
`str.join` works on iterables so we can avoid iterating through it twice. | return {}
def encode(self, prompt: str) -> List[int]:
# `encode` should only be used for prompt completion
# it should never be used for chat_completion.
# For chat completion use `apply_chat_template`
return self.tokenizer.encode(prompt, bos=True, eos=False)
def apply_... | @@ -165,18 +165,20 @@ def apply_chat_template(self,
messages: List["ChatCompletionMessageParam"],
tools: Optional[Dict[str, Any]] = None,
**kwargs) -> List[int]:
- assert tools is None, "`tools` are not yet supported."
- ... | vllm/transformers_utils/tokenizers/mistral.py | 26 | Python | 1 | suggestion | 238 | 51 | 49 | false | [Model] Add mistral function calling format to all models loaded with "mistral" format | 8,515 | vllm-project/vllm | 71,895 | Python | DarkLight1337 | patrickvonplaten |
) as mistral_format_model:
mistral_format_outputs = mistral_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_format_outputs,
outputs_1_lst=mistral_format_outputs,
name_0="hf",
name_1="mis... | > @DarkLight1337 seems like mistral-nemo is too large for the decoder-only tests:
We can set `max_model_len` here to reduce the memory usage. | load_format="mistral",
config_format="mistral",
) as mistral_format_model:
mistral_format_outputs = mistral_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_format_outputs,
output... | @@ -81,3 +127,22 @@ def test_mistral_format(
name_0="hf",
name_1="mistral",
)
+
+
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("model", MODELS[1:]) # v1 can't do func calling
+def test_mistral_function_calling(
+ vllm_runner,
+ model: str,
+ dtype: str,
+) ... | tests/models/decoder_only/language/test_mistral.py | 26 | Python | 0.786 | suggestion | 144 | 30 | 32 | false | [Model] Add mistral function calling format to all models loaded with "mistral" format | 8,515 | vllm-project/vllm | 71,895 | Python | DarkLight1337 | patrickvonplaten |
# ruff: noqa
import json
import random
import string
from vllm import LLM
from vllm.sampling_params import SamplingParams
# This script is an offline demo for running Pixtral.
#
# If you want to run a server/client setup, please follow this code:
#
# - Server:
#
# ```bash
# vllm serve mistralai/Mistral-7B-Instruct-v0... | Note that this is not a demo for Pixtral | # ruff: noqa
import json
import random
import string
from vllm import LLM
from vllm.sampling_params import SamplingParams
# This script is an offline demo for function calling
#
# If you want to run a server/client setup, please follow this code:
#
# - Server:
#
# ```bash
# vllm serve mistralai/Mistral-7B-Instruct-v0... | @@ -0,0 +1,138 @@
+# ruff: noqa
+import json
+import random
+import string
+
+from vllm import LLM
+from vllm.sampling_params import SamplingParams
+
+# This script is an offline demo for running Pixtral. | examples/offline_chat_with_tools.py | 9 | Python | 0.071 | suggestion | 40 | 34 | 34 | false | [Model] Add mistral function calling format to all models loaded with "mistral" format | 8,515 | vllm-project/vllm | 71,895 | Python | mgoin | patrickvonplaten |
assert (prepare_object_to_dump(CustomClass(
1, "b")) == "CustomClass(a=1, b='b')")
def test_request_logger_log_outputs():
"""Test the new log_outputs functionality."""
# Create a mock logger to capture log calls
mock_logger = MagicMock()
with patch("vllm.entrypoints.logger.logger", mock_... | Please remove the commented-out code | def test_request_logger_log_outputs():
"""Test the new log_outputs functionality."""
# Create a mock logger to capture log calls
mock_logger = MagicMock()
with patch("vllm.entrypoints.logger.logger", mock_logger):
request_logger = RequestLogger(max_log_len=None)
# Test basic output log... | @@ -252,4 +259,253 @@ class CustomClass:
b: str
assert (prepare_object_to_dump(CustomClass(
- 1, 'b')) == "CustomClass(a=1, b='b')")
+ 1, "b")) == "CustomClass(a=1, b='b')")
+
+
+def test_request_logger_log_outputs():
+ """Test the new log_outputs functionality."""
+ # Create a mock ... | tests/test_logger.py | 26 | Python | 0.286 | suggestion | 36 | 51 | 51 | false | feat: Add --enable-log-outputs flag for logging model generations | 20,707 | vllm-project/vllm | 71,895 | Python | DarkLight1337 | mizadri |
help="The file path to the SSL cert file.",
)
parser.add_argument(
"--ssl-ca-certs",
type=optional_type(str),
default=None,
help="The CA certificates file.",
)
parser.add_argument(
"--enable-ssl-refresh",
action="store_true",
default=False,... | Please revert the unnecessary formatting changes | # Special case: Middleware needs append action
frontend_kwargs["middleware"]["action"] = "append"
frontend_kwargs["middleware"]["type"] = str
if "nargs" in frontend_kwargs["middleware"]:
del frontend_kwargs["middleware"]["nargs"]
frontend_kwargs["middleware"]["default... | @@ -92,102 +92,128 @@ def make_arg_parser(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
"--uvicorn-log-level",
type=str,
default="info",
- choices=['debug', 'info', 'warning', 'error', 'critical', 'trace'],
- help="Log level for uvicorn.")
- parser.add_argument("... | vllm/entrypoints/openai/cli_args.py | 26 | Python | 0.286 | style | 48 | 51 | 51 | false | feat: Add --enable-log-outputs flag for logging model generations | 20,707 | vllm-project/vllm | 71,895 | Python | DarkLight1337 | mizadri |
extract_reasoning_content_streaming(
previous_text,
current_text,
delta_text,
previous_token_ids,
... | When `Stream=True`, this line causes an error by repeatedly accumulating the previous content.
|
# when only reasoning
elif self.reasoning_parser:
delta_message = (reasoning_parser.
extract_reasoning_content_streaming(
previous_text,
... | @@ -808,6 +812,10 @@ async def chat_completion_stream_generator(
assert all_previous_token_ids is not None
previous_texts[i] = current_text
all_previous_token_ids[i] = current_token_ids
+ else:
+ # Updat... | vllm/entrypoints/openai/serving_chat.py | 26 | Python | 0.857 | suggestion | 96 | 51 | 51 | false | feat: Add --enable-log-outputs flag for logging model generations | 20,707 | vllm-project/vllm | 71,895 | Python | chaunceyjiang | mizadri |
lora_config: LoRAConfig,
model_config: Optional[PretrainedConfig] = None,
) -> LogitsProcessorWithLoRA:
ret = LogitsProcessorWithLoRA(layer, lm_head.embedding_dim,
lm_head.weight.dtype, lm_head.weight.device,
lm_head.get_sharded_to_full_map... | Could we move this warning into `vllm/lora/models.py`?
If this helper function gets used in other places this warning might not be relevant | model_config: Optional[PretrainedConfig] = None,
) -> LogitsProcessorWithLoRA:
ret = LogitsProcessorWithLoRA(layer, lm_head.embedding_dim,
lm_head.weight.dtype, lm_head.weight.device,
lm_head.get_sharded_to_full_mapping())
ret.create_lora_w... | @@ -97,6 +98,17 @@ def replace_submodule(model: nn.Module, module_name: str,
return new_module
+def is_moe_model(model: nn.Module) -> bool:
+ """Checks if the model contains FusedMoE layers and warns the user."""
+ if any(isinstance(module, FusedMoE) for module in model.modules()):
+ logger.warnin... | vllm/lora/utils.py | 26 | Python | 0.714 | refactor | 142 | 51 | 51 | false | [Misc] Qwen MoE model supports LoRA | 20,932 | vllm-project/vllm | 71,895 | Python | hmellor | jeejeelee |
q_i, k_i, v_i = (
rearrange(x, "b s h d -> b h s d") for x in [q_i, k_i, v_i]
)
output_i = F.scaled_dot_product_attention(q_i, k_i, v_i, dropout_p=0.0)
output_i = rearrange(output_i, "b h s d -> b s h d ")
outputs.append... | qq: what happens to other dimensions? does dynamism on other dimensions trigger a recompile? | context_layer = einops.rearrange(
context_layer, "b s h d -> s b (h d)"
).contiguous()
elif self.attn_backend == _Backend.XFORMERS:
context_layer = vit_xformers_attn_wrapper(q, k, v, seqlens)
output, _ = self.proj(context_layer)
return output
... | @@ -462,6 +465,10 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
return x
+@set_model_tag("Qwen2_5_VisionPatchMerger")
+@support_torch_compile(dynamic_arg_dims={
+ "x": 0, | vllm/model_executor/models/qwen2_5_vl.py | 26 | Python | 0.143 | question | 92 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | miladm | Lucaskabela |
self.use_upstream_fa = use_upstream_fa
self.attn_backend, self.flash_attn_varlen_func = (
maybe_get_vit_flash_attn_backend(
self.attn_backend,
self.use_upstream_fa,
)
)
self.is_flash_attn_backend = self.attn_backend in {
... | NOTE: We push all the compile breaking things here (such as .item/.tolist). This way, we can specify this as a custom op and compile around it | output_size=embed_dim,
quant_config=quant_config,
prefix=f"{prefix}.proj",
disable_tp=use_data_parallel,
)
self.attn_backend = attn_backend
self.use_upstream_fa = use_upstream_fa
self.attn_backend, self.flash_attn_varlen_func = (
... | @@ -356,62 +359,101 @@ def forward(
q = apply_rotary_pos_emb_vision(q, rotary_pos_emb)
k = apply_rotary_pos_emb_vision(k, rotary_pos_emb)
- if self.is_flash_attn_backend:
- if self.attn_backend == _Backend.ROCM_AITER_FA:
- from aiter import flash_attn_varlen_... | vllm/model_executor/models/qwen2_5_vl.py | 26 | Python | 0.286 | suggestion | 143 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | Lucaskabela | Lucaskabela |
output, "(b s) h d -> s b (h d)", b=batch_size
).contiguous()
elif self.attn_backend == _Backend.TORCH_SDPA:
# Execute attention entry by entry for speed & less VRAM.
from vllm.platforms import current_platform
if current_platform.is_rocm():
... | Other major change here; don't use optional primitives that are difficult to trace. Instead use tensor which works here | )
elif self.attn_backend == _Backend.TORCH_SDPA:
# Execute attention entry by entry for speed & less VRAM.
outputs = []
for i in range(1, len(cu_seqlens)):
start_idx = cu_seqlens[i - 1]
end_idx = cu_seqlens[i]
q_i = ... | @@ -450,8 +492,8 @@ def forward(
x: torch.Tensor,
cu_seqlens: torch.Tensor,
rotary_pos_emb: torch.Tensor,
- max_seqlen: Optional[int] = None, # Only used for Flash Attention | vllm/model_executor/models/qwen2_5_vl.py | 26 | Python | 0.5 | suggestion | 120 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | Lucaskabela | Lucaskabela |
seq_len, bs, _ = qkv.shape
if self.tp_size > 1:
qkv = all_gather_interleave(qkv, self.qkv.hidden_size, self.tp_size)
# [s, b, 3 * head * head_dim] -> 3 * [s, b, head * head_dim]
q, k, v = qkv.chunk(3, dim=2)
# 3 * [s, b, head * head_dim]
if self.tp_size > 1:... | Is there a reason why we're doing this at top level instead of per-attention backend?
We have already changed the logic here so that there's no more repeative code on deciding attention backend - it'll be just populated from `Qwen2.5VLVisionTransformer`
https://github.com/vllm-project/vllm/blob/2e4fe48c370e833350ea... | _Backend.FLASH_ATTN,
_Backend.ROCM_AITER_FA,
}
def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]:
# [s, b, 3 * head * head_dim]
seq_len, bs, _ = qkv.shape
if self.tp_size > 1:
qkv = all_gather_interleave(qkv, self.qkv.hidden_size,... | @@ -366,67 +369,104 @@ def forward(
qk_rotated = apply_rotary_pos_emb_vision(qk_concat, rotary_pos_emb)
q, k = torch.chunk(qk_rotated, 2, dim=0)
- if self.is_flash_attn_backend:
- if self.attn_backend == _Backend.ROCM_AITER_FA:
- from aiter import flash_attn_... | vllm/model_executor/models/qwen2_5_vl.py | 26 | Python | 0.857 | question | 579 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | ywang96 | Lucaskabela |
video_token_id = hf_config.video_token_id
vision_start_token_id = hf_config.vision_start_token_id
spatial_merge_size = hf_config.vision_config.spatial_merge_size
tokens_per_second = getattr(hf_config.vision_config, "tokens_per_second", 1.0)
input_tokens_tensor = torch.tensor(inp... | Do we not need to `set_forward_context` for the DP ViT case? | image_grid_thw: list[list[int]] | torch.Tensor,
video_grid_thw: list[list[int]] | torch.Tensor,
second_per_grid_ts: list[float],
context_len: int = 0,
seq_len: int | None = None,
audio_feature_lengths: torch.Tensor | None = None,
use_audio_in_video: bool = False,
... | @@ -1079,8 +1130,9 @@ def _process_image_input(
grid_thw_list,
rope_type="rope_3d")
else:
- image_embeds = self.visual(pixel_values,
- ... | vllm/model_executor/models/qwen2_5_vl.py | 26 | Python | 0.357 | question | 60 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | ywang96 | Lucaskabela |
outputs = []
for i in range(1, len(cu_seqlens)):
start_idx = cu_seqlens[i - 1]
end_idx = cu_seqlens[i]
q_i = q[:, start_idx:end_idx]
k_i = k[:, start_idx:end_idx]
v_i = v[:, start_idx:end_idx]
q_i, k_... | @Lucaskabela to check: previously we had an invariant that all dynamic dimensions were actually "the same". After this PR, that is no longer the case, right? (cc @laithsakka who wanted to make use of this information)
This is not a problem, just a FYI | q_i, k_i, v_i = (
einops.rearrange(x, "b s h d -> b h s d") for x in [q_i, k_i, v_i]
)
output_i = F.scaled_dot_product_attention(q_i, k_i, v_i, dropout_p=0.0)
output_i = einops.rearrange(output_i, "b h s d -> b s h d ")
... | @@ -440,24 +440,21 @@ def forward(
context_layer, "b s h d -> s b (h d)"
).contiguous()
elif self.attn_backend == _Backend.XFORMERS:
- from xformers import ops as xops
- from xformers.ops.fmha.attn_bias import BlockDiagonalMask
-
- attn_bias = Bloc... | vllm/model_executor/models/qwen2_5_vl.py | 26 | Python | 0.857 | style | 254 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | zou3519 | Lucaskabela |
elif self.attn_backend == _Backend.TORCH_SDPA:
# Execute attention entry by entry for speed & less VRAM.
from vllm.platforms import current_platform
if current_platform.is_rocm():
q = q.contiguous()
k = k.contiguous()
v = v.con... | Does this turn on CUDAGraphs capture too? How does that interaction work? | # Execute attention entry by entry for speed & less VRAM.
outputs = []
for i in range(1, len(cu_seqlens)):
start_idx = cu_seqlens[i - 1]
end_idx = cu_seqlens[i]
q_i = q[:, start_idx:end_idx]
k_i = k[:, start_idx:end_idx]... | @@ -440,24 +440,21 @@ def forward(
context_layer, "b s h d -> s b (h d)"
).contiguous()
elif self.attn_backend == _Backend.XFORMERS:
- from xformers import ops as xops
- from xformers.ops.fmha.attn_bias import BlockDiagonalMask
-
- attn_bias = Bloc... | vllm/model_executor/models/qwen2_5_vl.py | 26 | Python | 0.143 | question | 73 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | zou3519 | Lucaskabela |
# support decorating multiple times
return cls
# take care of method resolution order
# make sure super().__init__ is called on the base class
# other than TorchCompileWrapperWithCustomDispatcher
cls.__bases__ = cls.__bases__ + (TorchCompileWrapperWithCustomDispatcher,)
old_init =... | this is kind of sketchy, do you have an example? I don't know where `old_init` came from |
def _verify_source_unchanged(source_info, vllm_config) -> None:
from .caching import _compute_code_hash, _compute_code_hash_with_content
file_contents = {}
for source in source_info.inlined_sources:
module = sys.modules[source.module]
file = inspect.getfile(module)
vllm_config.com... | @@ -228,8 +233,18 @@ def _support_torch_compile(
setattr(cls, IGNORE_COMPILE_KEY, False)
- def __init__(self, *, vllm_config: VllmConfig, prefix: str = "", **kwargs):
- old_init(self, vllm_config=vllm_config, prefix=prefix, **kwargs)
+ def __init__(self, **kwargs):
+ # NOTE: to support mult... | vllm/compilation/decorators.py | 26 | Python | 0.357 | suggestion | 88 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | zou3519 | Lucaskabela |
@overload
def support_torch_compile(
*,
enable_if: Callable[[VllmConfig], bool] | None = None,
) -> Callable[[_T], _T]: ...
@overload
def support_torch_compile(
*,
dynamic_arg_dims: dict[str, int | list[int]] | None,
) -> Callable[[_T], _T]: ...
@overload
def support_torch_compile(cls: _T) -> _T: ... | This arg is used with `dynamic_arg_dims` for `Qwen2_5_VisionBlock`. This overload does not provide a `dynamic_arg_dims` arg and my type checker complains about this. Should this be merged into overload above? | def _should_ignore_torch_compile(cls) -> bool:
"""
Check if the class should be ignored for torch.compile.
"""
return getattr(cls, IGNORE_COMPILE_KEY, False)
@overload
def support_torch_compile(
*,
enable_if: Callable[[VllmConfig], bool] | None = None,
) -> Callable[[_T], _T]: ...
@overload
... | @@ -73,6 +78,13 @@ def support_torch_compile(
) -> Callable[[_T], _T]: ...
+@overload
+def support_torch_compile(
+ *,
+ mark_unbacked_dims: dict[str, int | list[int]] | None,
+) -> Callable[[_T], _T]: ... | vllm/compilation/decorators.py | 26 | Python | 1 | question | 208 | 51 | 51 | false | [Misc][qwen2_5_vl][torch.compile] Enable `supports_torch_compile` on generic nn.Module and demonstrate speedup on Qwen Vision model | 23,207 | vllm-project/vllm | 71,895 | Python | yanyongyu | Lucaskabela |
mode 3, i.e. the backend must be either eager or inductor. Furthermore,
compilation is only piecewise if splitting ops is set accordingly and
use_inductor_graph_partition is off. Note that the default options for
splitting ops are sufficient for piecewise compilation.
"""
custom_ops: list[str] =... | This config has been placed between `splitting_ops` and its docstring.
And the docstring for `compile_mm_encoder` has been placed above it. | (it sees a part of the graph). The backend can not be custom for compilation
mode 3, i.e. the backend must be either eager or inductor. Furthermore,
compilation is only piecewise if splitting ops is set accordingly and
use_inductor_graph_partition is off. Note that the default options for
splitting ... | @@ -250,6 +251,13 @@ class CompilationConfig:
disabled when running with Inductor: mode>=VLLM_COMPILE and use_inductor=True.
Inductor generates (fused) Triton kernels for disabled custom ops."""
splitting_ops: list[str] | None = None
+
+ """
+ Provide control over whether to compile the multimodal ... | vllm/config/compilation.py | 26 | Python | 0.5 | suggestion | 142 | 51 | 51 | false | [Multimodal][torch.compile] Add compilation config field for turning off ViT/MM compile | 28,242 | vllm-project/vllm | 71,895 | Python | hmellor | Lucaskabela |
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Copyright 2024 The vLLM 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
#
#... | This is not where this should live. This file is for the Transformers backend. | return f"{prefix}{tip}"
def can_enable_torch_compile(vllm_config: "VllmConfig") -> bool:
"""
Callable to be passed to `@support_torch_compile`'s `enable_if` argument.
Defaults to `True` but is disabled in the following situations:
- The model uses dynamic rope scaling.
"""
text_config = ... | @@ -205,3 +205,14 @@ def can_enable_torch_compile(vllm_config: "VllmConfig") -> bool:
# Dynamic rope scaling is not compatible with torch.compile
rope_scaling: dict = getattr(text_config, "rope_scaling", None) or {}
return rope_scaling.get("rope_type") != "dynamic"
+
+
+def should_torch_compile_mm_vit(vl... | vllm/model_executor/models/transformers/utils.py | 26 | Python | 0.357 | suggestion | 78 | 208 | 27 | false | [Multimodal][torch.compile] Add compilation config field for turning off ViT/MM compile | 28,242 | vllm-project/vllm | 71,895 | Python | hmellor | Lucaskabela |
IS_AITER_FOUND and _______________|
cls._AITER_ENABLED -----> Check by the logic in `is_enabled()`
Example:
from vllm._aiter_ops import rocm_aiter_ops
# Check if aiter is enabled before using operations
if rocm_aiter_ops.is_enabled():
... | Please reuse `VLLM_ROCM_USE_AITER_MOE` for this feature and not introduce another FLAG `VLLM_ROCM_USE_CK_MXFP4_MOE` . It should be dispatched automatically based on the model config. | def _rocm_aiter_rmsnorm_with_add_fp8_group_quant_fake(
x: torch.Tensor,
residual: torch.Tensor,
weight: torch.Tensor,
variance_epsilon: float,
group_size: int,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
M, N = x.shape
scale_shape = (M, (N + group_size - 1) // group_size)
retur... | @@ -708,6 +708,7 @@ class rocm_aiter_ops:
_LINEAR_ENABLED = envs.VLLM_ROCM_USE_AITER_LINEAR
_RMSNORM_ENABLED = envs.VLLM_ROCM_USE_AITER_RMSNORM
_FMOE_ENABLED = envs.VLLM_ROCM_USE_AITER_MOE
+ _CK_MXFP4_MOE = envs.VLLM_ROCM_USE_CK_MXFP4_MOE | vllm/_aiter_ops.py | 26 | Python | 0.714 | suggestion | 182 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | tjtanaa | maleksan85 |
# See also vllm/config/profiler.py and `--profiler-config` argument
VLLM_TORCH_CUDA_PROFILE: str | None = None
VLLM_TORCH_PROFILER_DIR: str | None = None
VLLM_TORCH_PROFILER_RECORD_SHAPES: str | None = None
VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY: str | None = None
VLLM_TORCH_PROFILER_DISABLE_AS... | Let use the VLLM_ROCM_USE_AITER_MOE flag. The conditions whether to use the a16w4 should be determined based in the quantization config. | VLLM_FORCE_AOT_LOAD: bool = False
VLLM_USE_MEGA_AOT_ARTIFACT: bool = False
VLLM_USE_TRITON_AWQ: bool = False
VLLM_ALLOW_RUNTIME_LORA_UPDATING: bool = False
VLLM_SKIP_P2P_CHECK: bool = False
VLLM_DISABLED_KERNELS: list[str] = []
VLLM_DISABLE_PYNCCL: bool = False
VLLM_USE_OINK_OPS: bool = ... | @@ -115,6 +115,7 @@
VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
VLLM_ROCM_USE_AITER_LINEAR: bool = True
VLLM_ROCM_USE_AITER_MOE: bool = True
+ VLLM_ROCM_USE_AITER_FUSED_MOE_A16W4: bool = True | vllm/envs.py | 26 | Python | 0.5 | suggestion | 136 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | tjtanaa | maleksan85 |
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
prepare_fp8_moe_layer_for_marlin,
)
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import _swizzle_mxfp4
from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import (
OCP_MX_BLOCK_SIZE,
OCP_MX_Scheme,
)
f... | No need to add these two. Both are supported under `QuarkOCP_MX_MoEMethod`.
Both wmxfp4_amxfp4 and wmxfp4_afp8 are covered under `QuarkOCP_MX_MoEMethod`. | from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
prepare_fp8_moe_layer_for_marlin,
)
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import _swizzle_mxfp4
from vllm.model_executor.layers.quantization.utils.ocp_mx_utils import (
OCP_MX_BLOCK_SIZE,
OCP_MX_Scheme,
)
f... | @@ -48,7 +50,13 @@
logger = init_logger(__name__)
-__all__ = ["QuarkMoEMethod", "QuarkW8A8Fp8MoEMethod", "QuarkOCP_MX_MoEMethod"]
+__all__ = [
+ "QuarkMoEMethod",
+ "QuarkW8A8Fp8MoEMethod",
+ "QuarkW4MXFp4MoEMethod_OSS",
+ "QuarkW4MXFp4MoEMethod", | vllm/model_executor/layers/quantization/quark/quark_moe.py | 26 | Python | 0.5 | suggestion | 154 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 |
else:
return ocp_mx_moe_quant_config(
quant_dtype=self.input_dtype,
weight_dtype=self.weight_dtype,
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
w1_bias=layer.w13_bias,
w2_bias=layer.w... | For wmxfp4_afp8, just need to add the triton kernel dispatch at here.
If there are some triton kernel specific setups needed like pre-shuffling, kwargs setting etc, please do it in process_weights_after_loading. | return mxfp4_w4a8_moe_quant_config(
w1_scale=layer.w13_weight_scale,
w2_scale=layer.w2_weight_scale,
a1_scale=layer.w13_input_scale,
a2_scale=layer.w2_input_scale,
w1_bias=layer.w13_bias,
w2_bias=layer.w2_bias,
... | vllm/model_executor/layers/quantization/quark/quark_moe.py | 26 | Python | 0.357 | suggestion | 212 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 | |
FnSpecs("swiglu", triton_kernels.swiglu.swiglu_fn, ("alpha", "limit")),
(swiglu_alpha, swiglu_limit),
2,
)
intermediate_cache1 = matmul_ogs(
hidden_states,
w1,
quant_config.w1_bias,
routing_data,
gather_indx=... | why is it `use_mxfp4_w4a4` in config but `a8w4` kernel | # Shape check, only check non-mxfp4
assert hidden_states.shape[-1] == w1.shape[-2]
assert w2.shape[-1] == w1.shape[1]
E, _, N = w1.shape
if global_num_experts == -1:
global_num_experts = E
gammas = routing_data.gate_scal if routing_data else None
from aiter.ops.triton.moe_op_gemm... | @@ -299,6 +299,182 @@ def triton_kernel_fused_experts(
return output_tensor
+def triton_kernel_moe_oss_forward(
+ hidden_states: torch.Tensor,
+ w1, # Tensor or triton_kernels.Tensor
+ w2, # Tensor or triton_kernels.Tensor
+ gating_output: torch.Tensor,
+ topk: int,
+ renormalize: bool,
+ ... | vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py | 26 | Python | 0.571 | suggestion | 54 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 |
x,
layer.w13_weight,
layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
activation=layer.activation,
quant_config=self.moe_quant_config,
expert_map=layer.expert_map,
)
... | I think this can be merged with `QuarkOCP_MX_MoEMethod`. But probably it's okay to keep separate in this PR.
There will be a larger refactor to abstract out kernel dependent logic for `process_weights_after_loading` and align with modular kernel refactors. That will do the merge anyways. | layer.w2_weight,
topk_weights=topk_weights,
topk_ids=topk_ids,
activation=layer.activation,
quant_config=self.moe_quant_config,
expert_map=layer.expert_map,
)
else:
from vllm.model_executor.la... | @@ -1031,3 +1047,133 @@ def apply(
expert_map=layer.expert_map,
quant_config=self.moe_quant_config,
)
+
+
+class QuarkOCP_MX_MoEMethod_OSS(QuarkOCP_MX_MoEMethod): | vllm/model_executor/layers/quantization/quark/quark_moe.py | 26 | Python | 0.786 | refactor | 289 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 |
fused_activation=act,
)
intermediate_cache3 = matmul_ogs(
intermediate_cache1,
w2,
quant_config.w2_bias,
routing_data,
scatter_indx=scatter_indx,
precision_config=quant_config.w2_precision,
gammas=None if app... | ah i see how input scale is passed, through within w1_precision.
I think by using `mxfp4_w4a8_moe_quant_config`, we can use `quant_config.a1_scale` here instead | assert quant_config.w2_precision is not None, (
"w2_precision in quant config can't be None"
)
hidden_states = downcast_to_static_fp8(
hidden_states, quant_config.w1_precision.flex_ctx.lhs_data.scale
)
intermediate_cache1 = moe_gemm_a8w4(
hidden_states,
w1.storage.d... | @@ -304,6 +304,182 @@ def triton_kernel_fused_experts(
return output_tensor
+def triton_kernel_moe_oss_forward(
+ hidden_states: torch.Tensor,
+ w1, # Tensor or triton_kernels.Tensor
+ w2, # Tensor or triton_kernels.Tensor
+ gating_output: torch.Tensor,
+ topk: int,
+ renormalize: bool,
+ ... | vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py | 26 | Python | 0.714 | suggestion | 161 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 |
# Note: get_quant_method will look at the layer's local_num_experts
# for heuristic purposes, so it must be initialized first.
self.quant_method: FusedMoEMethodBase = _get_quant_method()
if not self.moe_config.is_act_and_mul and not current_platform.is_cuda_alike():
raise No... | fyi: no action for this pr, i have a PR #34285 refactoring how hidden_size is padded. we can land this first and I'll pull in the update. |
# Note: get_quant_method will look at the layer's local_num_experts
# for heuristic purposes, so it must be initialized first.
self.quant_method: FusedMoEMethodBase = _get_quant_method()
if not self.moe_config.is_act_and_mul and not current_platform.is_cuda_alike():
raise N... | @@ -609,6 +611,7 @@ def _get_quant_method() -> FusedMoEMethodBase:
moe_quant_params = {
"num_experts": self.local_num_experts,
"hidden_size": hidden_size,
+ "unpadded_hidden_size": unpadded_hidden_size, | vllm/model_executor/layers/fused_moe/layer.py | 26 | Python | 0.5 | refactor | 137 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 |
elif (
self.quant_config.get_name() == "quark" and self.model_type == "gpt_oss"
):
# When self._is_mxfp4 is true, model_dtype must be gpt_oss
expert_data = param.data[expert_id]
if "input_scale" in weight_name:
a... | are these needed? I thought Xuebin's earlier PR that enables w4a8 emulation should support all the necessary scales loading? |
quant_method_name = self.quant_method.__class__.__name__
global_expert_id = expert_id
expert_id = self._map_global_expert_id_to_local_expert_id(global_expert_id)
use_global_sf = (
getattr(self.quant_method, "use_global_sf", False)
and "input_scale" in weight_nam... | @@ -1060,16 +1063,43 @@ def weight_loader(
expert_id: int,
return_success: bool = False,
) -> bool | None:
- if self.quant_config and self.quant_config.get_name() == "mxfp4":
- # (FIXME) for gpt-oss all experts are combined
- if "bias" in weight_name:
- ... | vllm/model_executor/layers/fused_moe/layer.py | 26 | Python | 0.5 | question | 124 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 |
layer_quant_config = quant_config._find_matched_config(layer_name, module)
if layer_quant_config.get("output_tensors") or layer_quant_config.get("bias"):
raise NotImplementedError(
"Currently, Quark models with "
"output_tensors and bias "
"qu... | best to avoid conditioning on model type, ideally something like `if self.ocp_mx_scheme == "w_mxfp4_a_fp8" and not emulate:` |
if layer_quant_config.get("output_tensors") or layer_quant_config.get("bias"):
raise NotImplementedError(
"Currently, Quark models with "
"output_tensors and bias "
"quantized are not supported"
)
weight_config = layer_quant_confi... | @@ -71,14 +77,31 @@ def get_moe_method(
"output_tensors and bias "
"quantized are not supported"
)
+
+ emulate = not current_platform.supports_mx() or not (
+ rocm_aiter_ops.is_fused_moe_enabled()
+ )
weight_config = layer_quant_config.get... | vllm/model_executor/layers/quantization/quark/quark_moe.py | 26 | Python | 1 | suggestion | 124 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 |
layer.w2_input_scale = torch.nn.Parameter(
layer.w2_input_scale.max().to(torch.float32), requires_grad=False
)
from triton_kernels.numerics import InFlexData
lhs_data13 = InFlexData(scale=layer.w13_input_scale)
lhs_data2 = InFlexData(scale=la... | as commented elsewhere, suggest return `mxfp4_w4a8_moe_quant_config` | )
from triton_kernels.numerics import InFlexData
lhs_data13 = InFlexData(scale=layer.w13_input_scale)
lhs_data2 = InFlexData(scale=layer.w2_input_scale)
self.w13_precision_config = PrecisionConfig(
weight_scale=w13_scale,
fle... | @@ -1031,3 +1047,133 @@ def apply(
expert_map=layer.expert_map,
quant_config=self.moe_quant_config,
)
+
+
+class QuarkOCP_MX_MoEMethod_OSS(QuarkOCP_MX_MoEMethod):
+ def __init__(
+ self,
+ weight_config: dict[str, Any],
+ input_config: dict[str, Any... | vllm/model_executor/layers/quantization/quark/quark_moe.py | 26 | Python | 0.571 | suggestion | 68 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | BowenBao | maleksan85 |
hidden_states,
w1,
quant_config.w1_bias,
routing_data,
gather_indx=gather_indx,
precision_config=quant_config.w1_precision,
gammas=gammas if apply_router_weight_on_input else None,
fused_activation=act,
y=intermediate_cache,
)
matmul_ogs(
... | can we unify this with [`triton_kernel_moe_forward`](https://github.com/ROCm/vllm/blob/0b21db603eae8077c9a25ce00cb2af922fae19a9/vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py#L169)?
Currently for this function, we always have `quant_config.use_mxfp4_w4a8 = True`; but the w4a16 section here would ac... | apply_router_weight_on_input: bool = False,
global_num_experts: int = -1,
expert_map: torch.Tensor | None = None,
intermediate_cache: torch.Tensor | None = None,
a1q_scale: torch.Tensor | None = None,
) -> torch.Tensor:
"""Triton implementation of fused expert computation using OAI kernels."""
... | @@ -304,6 +304,182 @@ def triton_kernel_fused_experts(
return output_tensor
+def triton_kernel_moe_oss_forward( | vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py | 26 | Python | 0.857 | bug | 421 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | Rohan138 | maleksan85 |
routing_data, gather_idx, scatter_idx = aiter_routing(
gating_output, topk, sm_first=not renormalize
)
return triton_kernel_fused_oss_experts(
None,
hidden_states,
w1,
w2,
routing_data,
gather_idx,
scatter_idx,
activation=a... | Same comment here, can we reuse [triton_kernel_fused_experts](https://github.com/ROCm/vllm/blob/0b21db603eae8077c9a25ce00cb2af922fae19a9/vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py#L206) and add an `elif quant_config.use_mxfp4_w4a8 and rocm_aiter_ops.is_enabled()` for aiter triton with w4a8 moe? | else FusedActivation(
FnSpecs("swiglu", triton_kernels.swiglu.swiglu_fn, ("alpha", "limit")),
(swiglu_alpha, swiglu_limit),
2,
)
)
gammas = routing_data.gate_scal if routing_data else None
matmul_ogs(
hidden_states,
w1,
quant_confi... | @@ -304,6 +304,182 @@ def triton_kernel_fused_experts(
return output_tensor
+def triton_kernel_moe_oss_forward(
+ hidden_states: torch.Tensor,
+ w1, # Tensor or triton_kernels.Tensor
+ w2, # Tensor or triton_kernels.Tensor
+ gating_output: torch.Tensor,
+ topk: int,
+ renormalize: bool,
+ ... | vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py | 26 | Python | 0.857 | question | 319 | 51 | 51 | false | [ROCm][Quantization] GPT OSS Upstream MoE wmxfp4_afp8 with static scales | 30,357 | vllm-project/vllm | 71,895 | Python | Rohan138 | maleksan85 |
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import subprocess
import sys
import regex as re
# --------------------------------------------------------------------------- #
# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
# ---------------... | The hook doesn't need to work out which files it should run on, pre-commit already does that for us. Please look at `tools/pre_commit/check_pickle_imports.py` to see how this should be done | # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import sys
import regex as re
# --------------------------------------------------------------------------- #
# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
# ---------------------------------... | tools/pre_commit/check_torch_cuda.py | 1 | Python | 0.714 | suggestion | 190 | 26 | 26 | false | [Hardware] Replace `torch.cuda.empty_cache` with `torch.accelerator.empty_cache` | 30,681 | vllm-project/vllm | 71,895 | Python | hmellor | jikunshang | |
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import sys
import regex as re
# --------------------------------------------------------------------------- #
# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
# ---------------------------------... | ```suggestion
with open(path, encoding="utf-8") as f:
content = f.read()
for match in re.finditer(_TORCH_CUDA_RE, content, re.MULTILINE):
# Calculate line number from match position
line_num = content[: match.start() + 1].count("\n") + 1
print(
f"{path}:{line_num}: "
... | # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import sys
import regex as re
# --------------------------------------------------------------------------- #
# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
# ---------------------------------... | @@ -0,0 +1,39 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import sys
+
+import regex as re
+
+# --------------------------------------------------------------------------- #
+# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
+# -------... | tools/pre_commit/check_torch_cuda.py | 25 | Python | 1 | suggestion | 440 | 40 | 44 | false | [Hardware] Replace `torch.cuda.empty_cache` with `torch.accelerator.empty_cache` | 30,681 | vllm-project/vllm | 71,895 | Python | hmellor | jikunshang |
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import sys
import regex as re
# --------------------------------------------------------------------------- #
# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
# ---------------------------------... | ```suggestion
_TORCH_CUDA_RE = r"\btorch\.cuda\.empty_cache\b"
``` | # SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import sys
import regex as re
# --------------------------------------------------------------------------- #
# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
# ---------------------------------... | @@ -0,0 +1,39 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import sys
+
+import regex as re
+
+# --------------------------------------------------------------------------- #
+# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
+# -------... | tools/pre_commit/check_torch_cuda.py | 10 | Python | 0.857 | performance | 66 | 35 | 35 | false | [Hardware] Replace `torch.cuda.empty_cache` with `torch.accelerator.empty_cache` | 30,681 | vllm-project/vllm | 71,895 | Python | hmellor | jikunshang |
import regex as re
# --------------------------------------------------------------------------- #
# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
# --------------------------------------------------------------------------- #
_TORCH_CUDA_RE = r"\btorch\.cuda\.empty_cache\b"
ALLOWED_FILES = {"vllm/... | ```suggestion
with open(path, encoding="utf-8") as f:
content = f.read()
return_code = 0
for match in re.finditer(_TORCH_CUDA_RE, content, re.MULTILINE):
# Calculate line number from match position
line_num = content[: match.start() + 1].count("\n") + 1
print(
f"{... | import regex as re
# --------------------------------------------------------------------------- #
# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
# --------------------------------------------------------------------------- #
_TORCH_CUDA_PATTERNS = [
r"\btorch\.cuda\.empty_cache\b",
]
ALLOWED_F... | @@ -0,0 +1,41 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import sys
+
+import regex as re
+
+# --------------------------------------------------------------------------- #
+# Regex: match `torch.cuda.xxx` but allow `torch.accelerator.xxx`
+# -------... | tools/pre_commit/check_torch_cuda.py | 26 | Python | 1 | suggestion | 490 | 38 | 40 | false | [Hardware] Replace `torch.cuda.empty_cache` with `torch.accelerator.empty_cache` | 30,681 | vllm-project/vllm | 71,895 | Python | hmellor | jikunshang |
if isinstance(pg, ProcessGroup):
world_size = torch.distributed.get_world_size(group=pg)
else:
world_size = pg.world_size
if world_size == 1:
return 1
# Build node assignment map
node_assignment = [0] * world_size # rank -> node_id
next_node_id = 0
for current_ran... | This block was not confitional on `if not current_platform.is_cpu()` before?
```suggestion
try:
torch._C._host_emptyCache()
except AttributeError:
logger.warning("torch._C._host_emptyCache() only available in Pytorch >=2.5")
``` | torch.distributed.destroy_process_group()
def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
# Reset environment variable cache
envs.disable_envs_cache()
# Ensure all objects are not frozen before cleanup
gc.unfreeze()
destroy_model_parallel()
destroy_distributed_environment... | @@ -1732,14 +1732,14 @@ def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
gc.collect()
from vllm.platforms import current_platform
- empty_cache = current_platform.empty_cache
- if empty_cache is not None:
- empty_cache()
- try:
- if not current_platform.is_cpu():
+ if n... | vllm/distributed/parallel_state.py | 26 | Python | 1 | performance | 253 | 28 | 51 | false | [Hardware] Replace `torch.cuda.empty_cache` with `torch.accelerator.empty_cache` | 30,681 | vllm-project/vllm | 71,895 | Python | hmellor | jikunshang |
Args:
pg: The process group to analyze
Returns:
int: The total number of nodes
"""
if isinstance(pg, ProcessGroup):
world_size = torch.distributed.get_world_size(group=pg)
else:
world_size = pg.world_size
if world_size == 1:
return 1
# Build node as... | If the call is no longer CUDA specific does it need to be guarded?
```suggestion
torch.accelerator.empty_cache()
``` | global _WORLD, _NODE_COUNT
if _WORLD:
_WORLD.destroy()
_WORLD = None
_NODE_COUNT = None
if torch.distributed.is_initialized():
torch.distributed.destroy_process_group()
def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
# Reset environment variable cache
envs.disa... | @@ -1732,14 +1732,14 @@ def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
gc.collect()
from vllm.platforms import current_platform
- empty_cache = current_platform.empty_cache
- if empty_cache is not None:
- empty_cache()
- try:
- if not current_platform.is_cpu():
+ if n... | vllm/distributed/parallel_state.py | 26 | Python | 1 | performance | 121 | 34 | 51 | false | [Hardware] Replace `torch.cuda.empty_cache` with `torch.accelerator.empty_cache` | 30,681 | vllm-project/vllm | 71,895 | Python | hmellor | jikunshang |
if isinstance(pg, ProcessGroup):
world_size = torch.distributed.get_world_size(group=pg)
else:
world_size = pg.world_size
if world_size == 1:
return 1
# Build node assignment map
node_assignment = [0] * world_size # rank -> node_id
next_node_id = 0
for current_ran... | @hmellor
before code logic is:
```
call empty_cache for all platform (if possible, actually do nothing on cpu)
then for non-cpu platform, call torch._C._host_emptyCache()
```
after:
```
for non-cpu platform,
call empty_cache
then call torch._C._host_emptyCache()
(do nothing here for cpu platform)
```
I think... | torch.distributed.destroy_process_group()
def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
# Reset environment variable cache
envs.disable_envs_cache()
# Ensure all objects are not frozen before cleanup
gc.unfreeze()
destroy_model_parallel()
destroy_distributed_environment... | @@ -1732,14 +1732,14 @@ def cleanup_dist_env_and_memory(shutdown_ray: bool = False):
gc.collect()
from vllm.platforms import current_platform
- empty_cache = current_platform.empty_cache
- if empty_cache is not None:
- empty_cache()
- try:
- if not current_platform.is_cpu():
+ if n... | vllm/distributed/parallel_state.py | 26 | Python | 0.929 | performance | 354 | 28 | 51 | false | [Hardware] Replace `torch.cuda.empty_cache` with `torch.accelerator.empty_cache` | 30,681 | vllm-project/vllm | 71,895 | Python | jikunshang | jikunshang |
self.quant_config = quant_config
self.vocab_size = config.vocab_size
if get_pp_group().is_first_rank or (
config.tie_word_embeddings and get_pp_group().is_last_rank
):
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
c... | This is needed to avoid a recompile (discovered via tlparse) | config = vllm_config.model_config.hf_config
quant_config = vllm_config.quant_config
self.config = config
self.quant_config = quant_config
self.vocab_size = config.vocab_size
if get_pp_group().is_first_rank or (
config.tie_word_embeddings and get_pp_group().... | @@ -407,6 +407,9 @@ def __init__(
)
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+ # Need explicit mark here to avoid recompile from 0/1 spec
+ # since VocabEmbedding uses a different torch.compile decorator
+ torch._dynamo.decorators.mark_unbacked(input_ids, ... | vllm/model_executor/models/llama.py | 26 | Python | 0.357 | suggestion | 60 | 51 | 51 | false | [Misc][LLaMa4] Compile LLaMa Vision Encoder | 30,709 | vllm-project/vllm | 71,895 | Python | Lucaskabela | Lucaskabela |
if modality.startswith("image"):
return "<|image|>"
raise ValueError("Only image modality is supported")
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
config = vllm_config.model_config.hf_config
quant_config = vllm_config.q... | I think we try to avoid lazy import, it can be moved at the top with others | "gate_up_proj": ["gate_proj", "up_proj"],
}
supports_encoder_tp_data = True
@classmethod
def get_placeholder_str(cls, modality: str, i: int) -> str | None:
if modality.startswith("image"):
return "<|image|>"
raise ValueError("Only image modality is supported")
... | @@ -762,18 +769,28 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
multimodal_config = vllm_config.model_config.multimodal_config
self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"
+ self.vllm_config = vllm_config
self.config = config
... | vllm/model_executor/models/mllama4.py | 26 | Python | 0.357 | suggestion | 75 | 51 | 51 | false | [Misc][LLaMa4] Compile LLaMa Vision Encoder | 30,709 | vllm-project/vllm | 71,895 | Python | jeremyteboul | Lucaskabela |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.