@@ -64,11 +64,11 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
6464 gomega .Expect (prefillPods ).Should (gomega .BeEmpty ())
6565 gomega .Expect (decodePods ).Should (gomega .HaveLen (1 ))
6666
67- nsHdr , podHdr , _ := runCompletion (simplePrompt , testutils . ModelName )
67+ nsHdr , podHdr , _ := runCompletion (simplePrompt )
6868 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
6969 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
7070
71- nsHdr , podHdr , _ = runChatCompletion (simplePrompt , testutils . ModelName )
71+ nsHdr , podHdr , _ = runChatCompletion (simplePrompt )
7272 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
7373 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
7474
@@ -98,32 +98,32 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
9898 gomega .Expect (prefillPods ).Should (gomega .HaveLen (prefillReplicas ))
9999 gomega .Expect (decodePods ).Should (gomega .HaveLen (decodeReplicas ))
100100
101- nsHdr , podHdrCompletion , _ := runCompletion (simplePrompt , testutils . ModelName )
101+ nsHdr , podHdrCompletion , _ := runCompletion (simplePrompt )
102102 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
103103 gomega .Expect (podHdrCompletion ).Should (gomega .BeElementOf (decodePods ))
104104
105- nsHdr , podHdrChat , _ := runChatCompletion (simplePrompt , testutils . ModelName )
105+ nsHdr , podHdrChat , _ := runChatCompletion (simplePrompt )
106106 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
107107 gomega .Expect (podHdrChat ).Should (gomega .BeElementOf (decodePods ))
108108
109109 // Do an extra completion call with a different prompt
110- nsHdr , podHdr , _ := runCompletion (extraPrompt , testutils . ModelName )
110+ nsHdr , podHdr , _ := runCompletion (extraPrompt )
111111 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
112112 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
113113
114114 // Run completion with the original prompt
115- nsHdr , podHdr , _ = runCompletion (simplePrompt , testutils . ModelName )
115+ nsHdr , podHdr , _ = runCompletion (simplePrompt )
116116 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
117117 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
118118 gomega .Expect (podHdr ).Should (gomega .Equal (podHdrCompletion ))
119119
120120 // Do an extra chat completion call with a different prompt
121- nsHdr , podHdr , _ = runChatCompletion (extraPrompt , testutils . ModelName )
121+ nsHdr , podHdr , _ = runChatCompletion (extraPrompt )
122122 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
123123 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
124124
125125 // Run chat completion with the original prompt
126- nsHdr , podHdr , _ = runChatCompletion (simplePrompt , testutils . ModelName )
126+ nsHdr , podHdr , _ = runChatCompletion (simplePrompt )
127127 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
128128 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
129129 gomega .Expect (podHdr ).Should (gomega .Equal (podHdrChat ))
@@ -170,22 +170,22 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
170170 gomega .Expect (decodePods ).Should (gomega .HaveLen (decodeReplicas ))
171171
172172 // Test regular completion request
173- nsHdr , podHdrCompletion , _ := runCompletion (simplePrompt , testutils . ModelName )
173+ nsHdr , podHdrCompletion , _ := runCompletion (simplePrompt )
174174 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
175175 gomega .Expect (podHdrCompletion ).Should (gomega .BeElementOf (decodePods ))
176176
177177 // Test regular chat completion request
178- nsHdr , podHdrChat , _ := runChatCompletion (simplePrompt , testutils . ModelName )
178+ nsHdr , podHdrChat , _ := runChatCompletion (simplePrompt )
179179 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
180180 gomega .Expect (podHdrChat ).Should (gomega .BeElementOf (decodePods ))
181181
182182 // Run completion with a different prompt
183- nsHdr , podHdr , _ := runCompletion (extraPrompt , testutils . ModelName )
183+ nsHdr , podHdr , _ := runCompletion (extraPrompt )
184184 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
185185 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
186186
187187 // Run completion with original prompt (should go to same pod due to prefix cache)
188- nsHdr , podHdr , _ = runCompletion (simplePrompt , testutils . ModelName )
188+ nsHdr , podHdr , _ = runCompletion (simplePrompt )
189189 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
190190 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
191191 gomega .Expect (podHdr ).Should (gomega .Equal (podHdrCompletion ))
@@ -347,32 +347,32 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
347347 gomega .Expect (prefillPods ).Should (gomega .HaveLen (prefillReplicas ))
348348 gomega .Expect (decodePods ).Should (gomega .HaveLen (decodeReplicas ))
349349
350- nsHdr , podHdrCompletion , _ := runCompletion (simplePrompt , testutils . ModelName )
350+ nsHdr , podHdrCompletion , _ := runCompletion (simplePrompt )
351351 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
352352 gomega .Expect (podHdrCompletion ).Should (gomega .BeElementOf (decodePods ))
353353
354- nsHdr , podHdrChat , _ := runChatCompletion (simplePrompt , testutils . ModelName )
354+ nsHdr , podHdrChat , _ := runChatCompletion (simplePrompt )
355355 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
356356 gomega .Expect (podHdrChat ).Should (gomega .BeElementOf (decodePods ))
357357
358358 // Do an extra completion call with a different prompt
359- nsHdr , podHdr , _ := runCompletion (extraPrompt , testutils . ModelName )
359+ nsHdr , podHdr , _ := runCompletion (extraPrompt )
360360 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
361361 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
362362
363363 // Run completion with the original prompt
364- nsHdr , podHdr , _ = runCompletion (simplePrompt , testutils . ModelName )
364+ nsHdr , podHdr , _ = runCompletion (simplePrompt )
365365 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
366366 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
367367 gomega .Expect (podHdr ).Should (gomega .Equal (podHdrCompletion ))
368368
369369 // Do an extra chat completion call with a different prompt
370- nsHdr , podHdr , _ = runChatCompletion (extraPrompt , testutils . ModelName )
370+ nsHdr , podHdr , _ = runChatCompletion (extraPrompt )
371371 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
372372 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
373373
374374 // Run chat completion with the original prompt
375- nsHdr , podHdr , _ = runChatCompletion (simplePrompt , testutils . ModelName )
375+ nsHdr , podHdr , _ = runChatCompletion (simplePrompt )
376376 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
377377 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
378378 gomega .Expect (podHdr ).Should (gomega .Equal (podHdrChat ))
@@ -408,11 +408,11 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
408408 gomega .Expect (prefillPods ).Should (gomega .BeEmpty ())
409409 gomega .Expect (decodePods ).Should (gomega .HaveLen (1 ))
410410
411- nsHdr , podHdr , _ := runCompletion (simplePrompt , testutils . ModelName )
411+ nsHdr , podHdr , _ := runCompletion (simplePrompt )
412412 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
413413 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
414414
415- nsHdr , podHdr , _ = runChatCompletion (simplePrompt , testutils . ModelName )
415+ nsHdr , podHdr , _ = runChatCompletion (simplePrompt )
416416 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
417417 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
418418
@@ -442,7 +442,7 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
442442 gomega .Expect (prefillDecodePods ).Should (gomega .HaveLen (decodeReplicas ))
443443
444444 // Text request: encode stage skipped, routed directly to a prefill-decode pod
445- nsHdr , podHdr , _ := runCompletion (simplePrompt , testutils . ModelName )
445+ nsHdr , podHdr , _ := runCompletion (simplePrompt )
446446 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
447447 gomega .Expect (podHdr ).Should (gomega .BeElementOf (prefillDecodePods ))
448448
@@ -518,7 +518,7 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
518518 gomega .Expect (decodePods ).Should (gomega .HaveLen (decodeReplicas ))
519519
520520 // Text request: encode stage skipped, prefill triggered by prefix-based-pd-decider
521- nsHdr , podHdr , _ := runCompletion (simplePrompt , testutils . ModelName )
521+ nsHdr , podHdr , _ := runCompletion (simplePrompt )
522522 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
523523 gomega .Expect (podHdr ).Should (gomega .BeElementOf (decodePods ))
524524
@@ -599,12 +599,12 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
599599 gomega .Expect (epdPods ).Should (gomega .HaveLen (replicas ))
600600
601601 // Text completion: encode skipped, routes to decode profile -> single deployment
602- nsHdr , podHdr , _ := runCompletion (simplePrompt , testutils . ModelName )
602+ nsHdr , podHdr , _ := runCompletion (simplePrompt )
603603 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
604604 gomega .Expect (podHdr ).Should (gomega .Equal (epdPods [0 ]))
605605
606606 // Text chat completion: same routing as above
607- nsHdr , podHdr , _ = runChatCompletion (simplePrompt , testutils . ModelName )
607+ nsHdr , podHdr , _ = runChatCompletion (simplePrompt )
608608 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
609609 gomega .Expect (podHdr ).Should (gomega .Equal (epdPods [0 ]))
610610
@@ -657,7 +657,7 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
657657 gomega .Expect (decodePods ).Should (gomega .HaveLen (1 ))
658658
659659 for range 5 {
660- nsHdr , podHdr , _ := runCompletion (simplePrompt , testutils . ModelName )
660+ nsHdr , podHdr , _ := runCompletion (simplePrompt )
661661 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
662662 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
663663 }
@@ -681,18 +681,18 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
681681 gomega .Expect (decodePods ).Should (gomega .HaveLen (1 ))
682682
683683 // Test completions
684- nsHdr , podHdr , _ := runCompletion (simplePrompt , testutils . ModelName )
684+ nsHdr , podHdr , _ := runCompletion (simplePrompt )
685685 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
686686 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
687687
688688 // Test chat completions
689- nsHdr , podHdr , _ = runChatCompletion (simplePrompt , testutils . ModelName )
689+ nsHdr , podHdr , _ = runChatCompletion (simplePrompt )
690690 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
691691 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
692692
693693 // Repeat to verify prefix cache affinity with pre-tokenized prompts
694694 for range 3 {
695- nsHdr , podHdr , _ = runCompletion (simplePrompt , testutils . ModelName )
695+ nsHdr , podHdr , _ = runCompletion (simplePrompt )
696696 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
697697 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
698698 }
@@ -716,7 +716,7 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
716716
717717 var nsHdr , podHdr string
718718 for range 5 {
719- nsHdr , podHdr , _ = runCompletion (simplePrompt , testutils . ModelName )
719+ nsHdr , podHdr , _ = runCompletion (simplePrompt )
720720 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
721721 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
722722 }
@@ -730,7 +730,7 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
730730 var scaledNsHdr , scaledPodHdr string
731731 // Run inference multiple times until one is scheduled on the new pod
732732 for range 30 {
733- scaledNsHdr , scaledPodHdr , _ = runCompletion (extraPrompt , testutils . ModelName )
733+ scaledNsHdr , scaledPodHdr , _ = runCompletion (extraPrompt )
734734 gomega .Expect (scaledNsHdr ).Should (gomega .Equal (nsName ))
735735 gomega .Expect (scaledPodHdr ).Should (gomega .BeElementOf (scaledUpDecodePods ))
736736 if scaledPodHdr != podHdr {
@@ -748,7 +748,7 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
748748
749749 // Run multiple times and insure that they are scheduled on the remaining pod
750750 for range 5 {
751- nsHdr , podHdr , _ = runCompletion (simplePrompt , testutils . ModelName )
751+ nsHdr , podHdr , _ = runCompletion (simplePrompt )
752752 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
753753 gomega .Expect (podHdr ).Should (gomega .Equal (scaledDownDecodePods [0 ]))
754754 }
@@ -770,15 +770,15 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
770770 gomega .Expect (prefillPods ).Should (gomega .BeEmpty ())
771771 gomega .Expect (decodePods ).Should (gomega .HaveLen (1 ))
772772
773- nsHdr , podHdr , portHdr := runCompletion (simplePrompt , testutils . ModelName )
773+ nsHdr , podHdr , portHdr := runCompletion (simplePrompt )
774774 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
775775 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
776776
777777 var parallelNsHdr , parallelPodHdr , parallelPortHdr string
778778
779779 // Run inference multiple times until one is scheduled on the other port
780780 for range 30 {
781- parallelNsHdr , parallelPodHdr , parallelPortHdr = runCompletion (extraPrompt , testutils . ModelName )
781+ parallelNsHdr , parallelPodHdr , parallelPortHdr = runCompletion (extraPrompt )
782782 gomega .Expect (parallelNsHdr ).Should (gomega .Equal (nsName ))
783783 gomega .Expect (parallelPodHdr ).Should (gomega .Equal (decodePods [0 ]))
784784 if parallelPortHdr != portHdr {
@@ -787,13 +787,13 @@ var _ = ginkgo.Describe("Run end to end tests", ginkgo.Ordered, func() {
787787 }
788788 gomega .Expect (parallelPortHdr ).ShouldNot (gomega .Equal (portHdr ))
789789
790- nsHdr , podHdr , portHdr = runChatCompletion (simplePrompt , testutils . ModelName )
790+ nsHdr , podHdr , portHdr = runChatCompletion (simplePrompt )
791791 gomega .Expect (nsHdr ).Should (gomega .Equal (nsName ))
792792 gomega .Expect (podHdr ).Should (gomega .Equal (decodePods [0 ]))
793793
794794 // Run inference multiple times until one is scheduled on the other port
795795 for range 30 {
796- parallelNsHdr , parallelPodHdr , parallelPortHdr = runChatCompletion (extraPrompt , testutils . ModelName )
796+ parallelNsHdr , parallelPodHdr , parallelPortHdr = runChatCompletion (extraPrompt )
797797 gomega .Expect (parallelNsHdr ).Should (gomega .Equal (nsName ))
798798 gomega .Expect (parallelPodHdr ).Should (gomega .Equal (decodePods [0 ]))
799799 if parallelPortHdr != portHdr {
0 commit comments