Add ramp-up mode to NVIDIA GPU benchmark

Adds a new checkbox (enabled by default) in the benchmark section.
In ramp-up mode N tasks are spawned simultaneously: 1 GPU, then 2,
then 3, up to all selected GPUs — each step runs its GPUs in parallel.
NCCL runs only on the final step.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

audit/internal/webui/api.go

@@ -12,6 +12,7 @@ import (
 	"path/filepath"
 	"regexp"
 	"sort"
+	"strconv"
 	"strings"
 	"sync/atomic"
 	"syscall"
@@ -209,6 +210,14 @@ func joinTaskIndices(indices []int) string {
 	return strings.Join(parts, ",")
 }
 
+func formatGPUIndexList(indices []int) string {
+	parts := make([]string, len(indices))
+	for i, idx := range indices {
+		parts[i] = strconv.Itoa(idx)
+	}
+	return strings.Join(parts, ",")
+}
+
 func formatSplitTaskName(baseName, selectionLabel string) string {
 	baseName = strings.TrimSpace(baseName)
 	selectionLabel = strings.TrimSpace(selectionLabel)
@@ -540,6 +549,7 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 		ExcludeGPUIndices []int  `json:"exclude_gpu_indices"`
 		RunNCCL           *bool  `json:"run_nccl"`
 		ParallelGPUs      *bool  `json:"parallel_gpus"`
+		RampUp            *bool  `json:"ramp_up"`
 		DisplayName       string `json:"display_name"`
 	}
 	if r.Body != nil {
@@ -557,10 +567,63 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 	if body.ParallelGPUs != nil {
 		parallelGPUs = *body.ParallelGPUs
 	}
+	rampUp := false
+	if body.RampUp != nil {
+		rampUp = *body.RampUp
+	}
 	name := taskDisplayName("nvidia-benchmark", "", "")
 	if strings.TrimSpace(body.DisplayName) != "" {
 		name = body.DisplayName
 	}
+
+	if rampUp && len(body.GPUIndices) > 1 {
+		// Ramp-up mode: resolve GPU list, then create one task per prefix
+		// [gpu0], [gpu0,gpu1], ..., [gpu0,...,gpuN-1], each running in parallel.
+		gpus, err := apiListNvidiaGPUs(h.opts.App)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, err.Error())
+			return
+		}
+		resolved, err := expandSelectedGPUIndices(gpus, body.GPUIndices, body.ExcludeGPUIndices)
+		if err != nil {
+			writeError(w, http.StatusBadRequest, err.Error())
+			return
+		}
+		if len(resolved) < 2 {
+			// Fall through to normal single-task path.
+			rampUp = false
+		} else {
+			now := time.Now()
+			var allTasks []*Task
+			for step := 1; step <= len(resolved); step++ {
+				subset := resolved[:step]
+				stepName := fmt.Sprintf("%s [ramp %d/%d: GPU %s]", name, step, len(resolved), formatGPUIndexList(subset))
+				t := &Task{
+					ID:        newJobID("benchmark-nvidia"),
+					Name:      stepName,
+					Target:    "nvidia-benchmark",
+					Priority:  15,
+					Status:    TaskPending,
+					CreatedAt: now,
+					params: taskParams{
+						GPUIndices:       append([]int(nil), subset...),
+						SizeMB:           body.SizeMB,
+						BenchmarkProfile: body.Profile,
+						RunNCCL:          runNCCL && step == len(resolved),
+						ParallelGPUs:     true,
+						DisplayName:      stepName,
+					},
+				}
+				allTasks = append(allTasks, t)
+			}
+			for _, t := range allTasks {
+				globalQueue.enqueue(t)
+			}
+			writeTaskRunResponse(w, allTasks)
+			return
+		}
+	}
+
 	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", 15, time.Now(), taskParams{
 		GPUIndices:        body.GPUIndices,
 		ExcludeGPUIndices: body.ExcludeGPUIndices,

audit/internal/webui/pages.go

@@ -1342,11 +1342,11 @@ func renderValidate(opts HandlerOptions) string {
 
 <div class="grid3">
 ` + renderSATCard("nvidia", "NVIDIA GPU", "runNvidiaValidateSet('nvidia')", "", renderValidateCardBody(
-			inv.NVIDIA,
-			`Runs NVIDIA diagnostics and board inventory checks.`,
-			`<code>nvidia-smi</code>, <code>dmidecode</code>, <code>dcgmi diag</code>`,
-			`Level 2 in Validate, Level 3 in Stress. Runs one GPU at a time on the selected NVIDIA GPUs.`,
-		)) +
+		inv.NVIDIA,
+		`Runs NVIDIA diagnostics and board inventory checks.`,
+		`<code>nvidia-smi</code>, <code>dmidecode</code>, <code>dcgmi diag</code>`,
+		`Level 2 in Validate, Level 3 in Stress. Runs one GPU at a time on the selected NVIDIA GPUs.`,
+	)) +
 		`<div id="sat-card-nvidia-targeted-stress">` +
 		renderSATCard("nvidia-targeted-stress", "NVIDIA GPU Targeted Stress", "runNvidiaValidateSet('nvidia-targeted-stress')", "", renderValidateCardBody(
 			inv.NVIDIA,
@@ -1966,9 +1966,13 @@ func renderBenchmark(opts HandlerOptions) string {
           <p style="color:var(--muted);font-size:13px">Loading NVIDIA GPUs...</p>
         </div>
       </div>
+      <label class="benchmark-cb-row">
+        <input type="checkbox" id="benchmark-ramp-up" checked onchange="benchmarkUpdateSelectionNote()">
+        <span>Ramp-up mode: run 1 GPU, then 2, then 3… up to all selected GPUs (each step is a separate task)</span>
+      </label>
       <label class="benchmark-cb-row">
         <input type="checkbox" id="benchmark-parallel-gpus">
-        <span>Run all selected GPUs simultaneously (parallel mode)</span>
+        <span>Run all selected GPUs simultaneously (parallel mode, ignored in ramp-up)</span>
       </label>
       <label class="benchmark-cb-row">
         <input type="checkbox" id="benchmark-run-nccl" checked>
@@ -2036,11 +2040,16 @@ function benchmarkUpdateSelectionNote() {
     return;
   }
   btn.disabled = false;
-  note.textContent = 'Selected GPUs: ' + selected.join(', ') + '.';
-  if (nccl && nccl.checked && selected.length < 2) {
-    note.textContent += ' NCCL will be skipped because fewer than 2 GPUs are selected.';
-  } else if (nccl && nccl.checked) {
-    note.textContent += ' NCCL interconnect will use only these GPUs.';
+  const rampUp = selected.length > 1 && !!document.getElementById('benchmark-ramp-up').checked;
+  if (rampUp) {
+    note.textContent = 'Ramp-up: will spawn ' + selected.length + ' tasks (1 GPU → ' + selected.length + ' GPUs). NCCL runs on the final step only.';
+  } else {
+    note.textContent = 'Selected GPUs: ' + selected.join(', ') + '.';
+    if (nccl && nccl.checked && selected.length < 2) {
+      note.textContent += ' NCCL will be skipped because fewer than 2 GPUs are selected.';
+    } else if (nccl && nccl.checked) {
+      note.textContent += ' NCCL interconnect will use only these GPUs.';
+    }
   }
 }
 
@@ -2095,12 +2104,14 @@ function runNvidiaBenchmark() {
     return;
   }
   if (benchmarkES) { benchmarkES.close(); benchmarkES = null; }
-  const parallelGPUs = !!document.getElementById('benchmark-parallel-gpus').checked;
+  const rampUp = selected.length > 1 && !!document.getElementById('benchmark-ramp-up').checked;
+  const parallelGPUs = !rampUp && !!document.getElementById('benchmark-parallel-gpus').checked;
   const body = {
     profile: document.getElementById('benchmark-profile').value || 'standard',
     gpu_indices: selected,
     run_nccl: !!document.getElementById('benchmark-run-nccl').checked,
     parallel_gpus: parallelGPUs,
+    ramp_up: rampUp,
     display_name: 'NVIDIA Benchmark'
   };
   document.getElementById('benchmark-output').style.display = 'block';
@@ -2373,7 +2384,7 @@ func renderBurn() string {
 	    <p id="burn-selection-note" style="font-size:12px;color:var(--muted);margin:10px 0 0">Select at least one NVIDIA GPU to enable NVIDIA burn recipes.</p>
 	    <label class="cb-row" style="margin-top:10px">
 	      <input type="checkbox" id="burn-stagger-nvidia">
-	      <span>Ramp selected NVIDIA GPUs one by one before full-load hold. Uses a 3-minute stabilization window per GPU, then keeps all selected GPUs under load for the chosen Burn Profile duration.</span>
+	      <span>Ramp selected NVIDIA GPUs one by one before the full-load hold. Smoke: +2 min per GPU, then 5 min with all selected GPUs under load. Acceptance: +10 min per GPU, then at least 1 hour with all selected GPUs under load. Overnight: +1 hour per GPU, then at least 1 hour with all selected GPUs under load, capped at 10 hours total.</span>
 	    </label>
 	  </div>
 	</div>
@@ -3108,7 +3119,6 @@ usbRefresh();
 </script>`
 }
 
-
 func renderNvidiaSelfHealInline() string {
 	return `<p style="font-size:13px;color:var(--muted);margin-bottom:12px">Inspect NVIDIA GPU health, restart the bee-nvidia driver service, and issue a per-GPU reset when the driver reports reset required.</p>
 <div style="display:flex;gap:8px;flex-wrap:wrap;margin-bottom:12px">

audit/internal/webui/tasks.go

@@ -152,6 +152,12 @@ type burnPreset struct {
 	DurationSec int
 }
 
+type nvidiaRampSpec struct {
+	DurationSec      int
+	StaggerSeconds   int
+	TotalDurationSec int
+}
+
 func resolveBurnPreset(profile string) burnPreset {
 	switch profile {
 	case "overnight":
@@ -163,11 +169,43 @@ func resolveBurnPreset(profile string) burnPreset {
 	}
 }
 
-func boolToNvidiaStaggerSeconds(enabled bool, selected []int) int {
-	if enabled && len(selected) > 1 {
-		return 180
+func resolveNvidiaRampPlan(profile string, enabled bool, selected []int) (nvidiaRampSpec, error) {
+	base := resolveBurnPreset(profile).DurationSec
+	plan := nvidiaRampSpec{
+		DurationSec:      base,
+		TotalDurationSec: base,
 	}
-	return 0
+	if !enabled {
+		return plan, nil
+	}
+	count := len(selected)
+	if count == 0 {
+		return nvidiaRampSpec{}, fmt.Errorf("staggered NVIDIA burn requires explicit GPU selection")
+	}
+	if count == 1 {
+		return plan, nil
+	}
+
+	switch profile {
+	case "acceptance":
+		plan.StaggerSeconds = 10 * 60
+		plan.TotalDurationSec = plan.DurationSec + plan.StaggerSeconds*(count-1)
+	case "overnight":
+		plan.StaggerSeconds = 60 * 60
+		plan.TotalDurationSec = 8 * 60 * 60
+		minTotal := count * 60 * 60
+		if plan.TotalDurationSec < minTotal {
+			plan.TotalDurationSec = minTotal
+		}
+		if plan.TotalDurationSec > 10*60*60 {
+			return nvidiaRampSpec{}, fmt.Errorf("overnight staggered NVIDIA burn supports at most 10 GPUs")
+		}
+		plan.DurationSec = plan.TotalDurationSec - plan.StaggerSeconds*(count-1)
+	default:
+		plan.StaggerSeconds = 2 * 60
+		plan.TotalDurationSec = plan.DurationSec + plan.StaggerSeconds*(count-1)
+	}
+	return plan, nil
 }
 
 func resolvePlatformStressPreset(profile string) platform.PlatformStressOptions {
@@ -600,7 +638,7 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 			RunNCCL:           t.params.RunNCCL,
 			ParallelGPUs:      t.params.ParallelGPUs,
 		}, j.append)
-		case "nvidia-compute":
+	case "nvidia-compute":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
 			break
@@ -609,11 +647,18 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		if t.params.BurnProfile != "" && dur <= 0 {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
-			staggerSec := boolToNvidiaStaggerSeconds(t.params.StaggerGPUStart, t.params.GPUIndices)
-			if staggerSec > 0 {
-				j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU", staggerSec))
-			}
-			archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, staggerSec, j.append)
+		rampPlan, planErr := resolveNvidiaRampPlan(t.params.BurnProfile, t.params.StaggerGPUStart, t.params.GPUIndices)
+		if planErr != nil {
+			err = planErr
+			break
+		}
+		if t.params.BurnProfile != "" && t.params.StaggerGPUStart && dur <= 0 {
+			dur = rampPlan.DurationSec
+		}
+		if rampPlan.StaggerSeconds > 0 {
+			j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU; post-ramp hold: %ds; total runtime: %ds", rampPlan.StaggerSeconds, dur, rampPlan.TotalDurationSec))
+		}
+		archive, err = a.RunNvidiaOfficialComputePack(ctx, "", dur, t.params.GPUIndices, rampPlan.StaggerSeconds, j.append)
 	case "nvidia-targeted-power":
 		if a == nil {
 			err = fmt.Errorf("app not configured")
@@ -663,13 +708,24 @@ func (q *taskQueue) runTask(t *Task, j *jobState, ctx context.Context) {
 		if t.params.BurnProfile != "" && dur <= 0 {
 			dur = resolveBurnPreset(t.params.BurnProfile).DurationSec
 		}
-			archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
-				DurationSec:       dur,
-				Loader:            t.params.Loader,
-				GPUIndices:        t.params.GPUIndices,
-				ExcludeGPUIndices: t.params.ExcludeGPUIndices,
-				StaggerSeconds:    boolToNvidiaStaggerSeconds(t.params.StaggerGPUStart, t.params.GPUIndices),
-			}, j.append)
+		rampPlan, planErr := resolveNvidiaRampPlan(t.params.BurnProfile, t.params.StaggerGPUStart, t.params.GPUIndices)
+		if planErr != nil {
+			err = planErr
+			break
+		}
+		if t.params.BurnProfile != "" && t.params.StaggerGPUStart && dur <= 0 {
+			dur = rampPlan.DurationSec
+		}
+		if rampPlan.StaggerSeconds > 0 {
+			j.append(fmt.Sprintf("NVIDIA staggered ramp-up enabled: %ds per GPU; post-ramp hold: %ds; total runtime: %ds", rampPlan.StaggerSeconds, dur, rampPlan.TotalDurationSec))
+		}
+		archive, err = runNvidiaStressPackCtx(a, ctx, "", platform.NvidiaStressOptions{
+			DurationSec:       dur,
+			Loader:            t.params.Loader,
+			GPUIndices:        t.params.GPUIndices,
+			ExcludeGPUIndices: t.params.ExcludeGPUIndices,
+			StaggerSeconds:    rampPlan.StaggerSeconds,
+		}, j.append)
 	case "memory":
 		if a == nil {
 			err = fmt.Errorf("app not configured")

audit/internal/webui/tasks_test.go

@@ -491,6 +491,83 @@ func TestResolveBurnPreset(t *testing.T) {
 	}
 }
 
+func TestResolveNvidiaRampPlan(t *testing.T) {
+	tests := []struct {
+		name     string
+		profile  string
+		enabled  bool
+		selected []int
+		want     nvidiaRampSpec
+		wantErr  string
+	}{
+		{
+			name:     "disabled uses base preset",
+			profile:  "acceptance",
+			selected: []int{0, 1},
+			want:     nvidiaRampSpec{DurationSec: 60 * 60, TotalDurationSec: 60 * 60},
+		},
+		{
+			name:     "smoke ramp uses two minute steps",
+			profile:  "smoke",
+			enabled:  true,
+			selected: []int{0, 1, 2},
+			want:     nvidiaRampSpec{DurationSec: 5 * 60, StaggerSeconds: 2 * 60, TotalDurationSec: 9 * 60},
+		},
+		{
+			name:     "acceptance ramp uses ten minute steps",
+			profile:  "acceptance",
+			enabled:  true,
+			selected: []int{0, 1, 2},
+			want:     nvidiaRampSpec{DurationSec: 60 * 60, StaggerSeconds: 10 * 60, TotalDurationSec: 80 * 60},
+		},
+		{
+			name:     "overnight stays at eight hours when possible",
+			profile:  "overnight",
+			enabled:  true,
+			selected: []int{0, 1, 2},
+			want:     nvidiaRampSpec{DurationSec: 6 * 60 * 60, StaggerSeconds: 60 * 60, TotalDurationSec: 8 * 60 * 60},
+		},
+		{
+			name:     "overnight extends to keep one hour after final gpu",
+			profile:  "overnight",
+			enabled:  true,
+			selected: []int{0, 1, 2, 3, 4, 5, 6, 7, 8},
+			want:     nvidiaRampSpec{DurationSec: 60 * 60, StaggerSeconds: 60 * 60, TotalDurationSec: 9 * 60 * 60},
+		},
+		{
+			name:     "overnight rejects impossible gpu count",
+			profile:  "overnight",
+			enabled:  true,
+			selected: []int{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10},
+			wantErr:  "at most 10 GPUs",
+		},
+		{
+			name:    "enabled requires explicit selection",
+			profile: "smoke",
+			enabled: true,
+			wantErr: "requires explicit GPU selection",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := resolveNvidiaRampPlan(tc.profile, tc.enabled, tc.selected)
+			if tc.wantErr != "" {
+				if err == nil || !strings.Contains(err.Error(), tc.wantErr) {
+					t.Fatalf("err=%v want substring %q", err, tc.wantErr)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("resolveNvidiaRampPlan error: %v", err)
+			}
+			if got != tc.want {
+				t.Fatalf("resolveNvidiaRampPlan(%q, %t, %v)=%+v want %+v", tc.profile, tc.enabled, tc.selected, got, tc.want)
+			}
+		})
+	}
+}
+
 func TestTaskDisplayNameUsesNvidiaStressLoader(t *testing.T) {
 	tests := []struct {
 		loader string