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

@@ -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';