Normalize task queue priorities by workflow

.gitignore

@@ -2,3 +2,4 @@
 .DS_Store
 dist/
 iso/out/
+build-cache/

audit/internal/platform/benchmark.go

@@ -401,6 +401,7 @@ func (s *System) RunNvidiaBenchmark(ctx context.Context, baseDir string, opts Nv
 		serverLoadedW = serverLoadedWSum / float64(serverLoadedSamples)
 	}
 	result.ServerPower = characterizeServerPower(serverIdleW, serverLoadedW, gpuReportedSumW, serverIdleOK && serverLoadedOK)
+	result.Cooling = summarizeBenchmarkCooling(metricRows)
 
 	// Apply server-power penalty when IPMI reports the server delta is much
 	// lower than GPU-reported sum: GPU power telemetry is over-stated, making
@@ -739,7 +740,7 @@ func collectBenchmarkSamples(ctx context.Context, durationSec int, gpuIndices []
 		if ctx.Err() != nil {
 			return rows, ctx.Err()
 		}
-		samples, err := sampleGPUMetrics(gpuIndices)
+		samples, err := sampleBenchmarkTelemetry(gpuIndices)
 		if err == nil {
 			elapsed := time.Since(start).Seconds()
 			for i := range samples {
@@ -774,7 +775,7 @@ func runBenchmarkCommandWithMetrics(ctx context.Context, verboseLog, name string
 			case <-stopCh:
 				return
 			case <-ticker.C:
-				samples, err := sampleGPUMetrics(gpuIndices)
+				samples, err := sampleBenchmarkTelemetry(gpuIndices)
 				if err != nil {
 					continue
 				}
@@ -794,6 +795,37 @@ func runBenchmarkCommandWithMetrics(ctx context.Context, verboseLog, name string
 	return out, metricRows, err
 }
 
+type benchmarkCoolingSample struct {
+	AvgFanRPM             float64
+	AvgFanDutyCyclePct    float64
+	FanDutyCycleAvailable bool
+}
+
+func sampleBenchmarkTelemetry(gpuIndices []int) ([]GPUMetricRow, error) {
+	samples, err := sampleGPUMetrics(gpuIndices)
+	if err != nil {
+		return nil, err
+	}
+	fanSample := sampleBenchmarkCoolingSample()
+	for i := range samples {
+		samples[i].FanAvgRPM = fanSample.AvgFanRPM
+		samples[i].FanDutyCyclePct = fanSample.AvgFanDutyCyclePct
+		samples[i].FanDutyCycleAvailable = fanSample.FanDutyCycleAvailable
+	}
+	return samples, nil
+}
+
+func sampleBenchmarkCoolingSample() benchmarkCoolingSample {
+	fans, _ := sampleFanSpeeds()
+	avgRPM, _, _ := fanRPMStats(fans)
+	dutyPct, dutyAvailable := sampleFanDutyCyclePct()
+	return benchmarkCoolingSample{
+		AvgFanRPM:             avgRPM,
+		AvgFanDutyCyclePct:    dutyPct,
+		FanDutyCycleAvailable: dutyAvailable,
+	}
+}
+
 func annotateBenchmarkMetricRows(rows []GPUMetricRow, stage string, offset float64) []GPUMetricRow {
 	if len(rows) == 0 {
 		return nil
@@ -1022,6 +1054,37 @@ func summarizeBenchmarkTelemetry(rows []GPUMetricRow) BenchmarkTelemetrySummary
 	return summary
 }
 
+func summarizeBenchmarkCooling(rows []GPUMetricRow) *BenchmarkCoolingSummary {
+	if len(rows) == 0 {
+		return nil
+	}
+	var rpmValues []float64
+	var dutyValues []float64
+	for _, row := range rows {
+		if row.FanAvgRPM > 0 {
+			rpmValues = append(rpmValues, row.FanAvgRPM)
+		}
+		if row.FanDutyCycleAvailable {
+			dutyValues = append(dutyValues, row.FanDutyCyclePct)
+		}
+	}
+	if len(rpmValues) == 0 && len(dutyValues) == 0 {
+		return nil
+	}
+	summary := &BenchmarkCoolingSummary{
+		Available: true,
+		AvgFanRPM: benchmarkMean(rpmValues),
+	}
+	if len(dutyValues) > 0 {
+		summary.FanDutyCycleAvailable = true
+		summary.AvgFanDutyCyclePct = benchmarkMean(dutyValues)
+		summary.P95FanDutyCyclePct = benchmarkPercentile(dutyValues, 95)
+	} else {
+		summary.Notes = append(summary.Notes, "fan duty cycle unavailable on this host; RPM-only fan telemetry was collected")
+	}
+	return summary
+}
+
 func scoreBenchmarkGPUResult(gpu BenchmarkGPUResult) BenchmarkScorecard {
 	score := BenchmarkScorecard{}
 
@@ -1601,7 +1664,10 @@ func maxInt(a, b int) int {
 // queryIPMIServerPowerW reads the current server power draw via ipmitool dcmi.
 // Returns 0 and an error if IPMI is unavailable or the output cannot be parsed.
 func queryIPMIServerPowerW() (float64, error) {
-	out, err := satExecCommand("ipmitool", "dcmi", "power", "reading").Output()
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	cmd := exec.CommandContext(ctx, "ipmitool", "dcmi", "power", "reading")
+	out, err := cmd.Output()
 	if err != nil {
 		return 0, fmt.Errorf("ipmitool dcmi power reading: %w", err)
 	}
@@ -1620,6 +1686,7 @@ func sampleIPMIPowerSeries(ctx context.Context, durationSec int) (meanW float64,
 	}
 	deadline := time.Now().Add(time.Duration(durationSec) * time.Second)
 	var samples []float64
+loop:
 	for {
 		if w, err := queryIPMIServerPowerW(); err == nil {
 			samples = append(samples, w)
@@ -1629,7 +1696,7 @@ func sampleIPMIPowerSeries(ctx context.Context, durationSec int) (meanW float64,
 		}
 		select {
 		case <-ctx.Done():
-			break
+			break loop
 		case <-time.After(2 * time.Second):
 		}
 	}

audit/internal/platform/benchmark_report.go

@@ -290,6 +290,31 @@ func renderBenchmarkReportWithCharts(result NvidiaBenchmarkResult) string {
 		}
 	}
 
+	// ── Cooling ───────────────────────────────────────────────────────────────
+	if cooling := result.Cooling; cooling != nil {
+		b.WriteString("## Cooling\n\n")
+		if cooling.Available {
+			b.WriteString("| Metric | Value |\n|--------|-------|\n")
+			fmt.Fprintf(&b, "| Average fan speed | %.0f RPM |\n", cooling.AvgFanRPM)
+			if cooling.FanDutyCycleAvailable {
+				fmt.Fprintf(&b, "| Average fan duty cycle | %.1f%% |\n", cooling.AvgFanDutyCyclePct)
+				fmt.Fprintf(&b, "| P95 fan duty cycle | %.1f%% |\n", cooling.P95FanDutyCyclePct)
+			} else {
+				b.WriteString("| Average fan duty cycle | N/A |\n")
+				b.WriteString("| P95 fan duty cycle | N/A |\n")
+			}
+			b.WriteString("\n")
+		} else {
+			b.WriteString("Cooling telemetry unavailable.\n\n")
+		}
+		for _, note := range cooling.Notes {
+			fmt.Fprintf(&b, "- %s\n", note)
+		}
+		if len(cooling.Notes) > 0 {
+			b.WriteString("\n")
+		}
+	}
+
 	// ── Raw files ─────────────────────────────────────────────────────────────
 	b.WriteString("## Raw Files\n\n")
 	b.WriteString("- `result.json`\n- `report.md`\n- `summary.txt`\n- `verbose.log`\n")

audit/internal/platform/benchmark_test.go

@@ -131,6 +131,13 @@ func TestRenderBenchmarkReportIncludesFindingsAndScores(t *testing.T) {
 				DegradationReasons: []string{"power_capped"},
 			},
 		},
+		Cooling: &BenchmarkCoolingSummary{
+			Available:             true,
+			AvgFanRPM:             9200,
+			FanDutyCycleAvailable: true,
+			AvgFanDutyCyclePct:    47.5,
+			P95FanDutyCyclePct:    62.0,
+		},
 	}
 
 	report := renderBenchmarkReport(result)
@@ -140,6 +147,9 @@ func TestRenderBenchmarkReportIncludesFindingsAndScores(t *testing.T) {
 		"1176.00",
 		"fp16_tensor",
 		"700.00",
+		"Cooling",
+		"Average fan duty cycle",
+		"47.5%",
 	} {
 		if !strings.Contains(report, needle) {
 			t.Fatalf("report missing %q\n%s", needle, report)

audit/internal/platform/benchmark_types.go

@@ -25,6 +25,17 @@ type BenchmarkCPULoad struct {
 	Note   string `json:"note,omitempty"`
 }
 
+// BenchmarkCoolingSummary captures fan telemetry averaged across the full
+// benchmark run.
+type BenchmarkCoolingSummary struct {
+	Available             bool     `json:"available"`
+	AvgFanRPM             float64  `json:"avg_fan_rpm,omitempty"`
+	FanDutyCycleAvailable bool     `json:"fan_duty_cycle_available,omitempty"`
+	AvgFanDutyCyclePct    float64  `json:"avg_fan_duty_cycle_pct,omitempty"`
+	P95FanDutyCyclePct    float64  `json:"p95_fan_duty_cycle_pct,omitempty"`
+	Notes                 []string `json:"notes,omitempty"`
+}
+
 const (
 	NvidiaBenchmarkProfileStandard  = "standard"
 	NvidiaBenchmarkProfileStability = "stability"
@@ -61,6 +72,7 @@ type NvidiaBenchmarkResult struct {
 	Normalization      BenchmarkNormalization       `json:"normalization"`
 	HostConfig         *BenchmarkHostConfig         `json:"host_config,omitempty"`
 	CPULoad            *BenchmarkCPULoad            `json:"cpu_load,omitempty"`
+	Cooling            *BenchmarkCoolingSummary     `json:"cooling,omitempty"`
 	GPUs               []BenchmarkGPUResult         `json:"gpus"`
 	Interconnect       *BenchmarkInterconnectResult `json:"interconnect,omitempty"`
 	ServerPower        *BenchmarkServerPower        `json:"server_power,omitempty"`

audit/internal/platform/gpu_metrics.go

@@ -13,15 +13,18 @@ import (
 
 // GPUMetricRow is one telemetry sample from nvidia-smi during a stress test.
 type GPUMetricRow struct {
-	Stage       string  `json:"stage,omitempty"`
-	ElapsedSec  float64 `json:"elapsed_sec"`
-	GPUIndex    int     `json:"index"`
-	TempC       float64 `json:"temp_c"`
-	UsagePct    float64 `json:"usage_pct"`
-	MemUsagePct float64 `json:"mem_usage_pct"`
-	PowerW      float64 `json:"power_w"`
-	ClockMHz    float64 `json:"clock_mhz"`
-	MemClockMHz float64 `json:"mem_clock_mhz"`
+	Stage                 string  `json:"stage,omitempty"`
+	ElapsedSec            float64 `json:"elapsed_sec"`
+	GPUIndex              int     `json:"index"`
+	TempC                 float64 `json:"temp_c"`
+	UsagePct              float64 `json:"usage_pct"`
+	MemUsagePct           float64 `json:"mem_usage_pct"`
+	PowerW                float64 `json:"power_w"`
+	ClockMHz              float64 `json:"clock_mhz"`
+	MemClockMHz           float64 `json:"mem_clock_mhz"`
+	FanAvgRPM             float64 `json:"fan_avg_rpm,omitempty"`
+	FanDutyCyclePct       float64 `json:"fan_duty_cycle_pct,omitempty"`
+	FanDutyCycleAvailable bool    `json:"fan_duty_cycle_available,omitempty"`
 }
 
 // sampleGPUMetrics runs nvidia-smi once and returns current metrics for each GPU.
@@ -142,10 +145,14 @@ func sampleAMDGPUMetrics() ([]GPUMetricRow, error) {
 // WriteGPUMetricsCSV writes collected rows as a CSV file.
 func WriteGPUMetricsCSV(path string, rows []GPUMetricRow) error {
 	var b bytes.Buffer
-	b.WriteString("stage,elapsed_sec,gpu_index,temperature_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz\n")
+	b.WriteString("stage,elapsed_sec,gpu_index,temperature_c,usage_pct,mem_usage_pct,power_w,clock_mhz,mem_clock_mhz,fan_avg_rpm,fan_duty_cycle_pct,fan_duty_cycle_available\n")
 	for _, r := range rows {
-		fmt.Fprintf(&b, "%s,%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f\n",
-			strconv.Quote(strings.TrimSpace(r.Stage)), r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.MemUsagePct, r.PowerW, r.ClockMHz, r.MemClockMHz)
+		dutyAvail := 0
+		if r.FanDutyCycleAvailable {
+			dutyAvail = 1
+		}
+		fmt.Fprintf(&b, "%s,%.1f,%d,%.1f,%.1f,%.1f,%.1f,%.0f,%.0f,%.0f,%.1f,%d\n",
+			strconv.Quote(strings.TrimSpace(r.Stage)), r.ElapsedSec, r.GPUIndex, r.TempC, r.UsagePct, r.MemUsagePct, r.PowerW, r.ClockMHz, r.MemClockMHz, r.FanAvgRPM, r.FanDutyCyclePct, dutyAvail)
 	}
 	return os.WriteFile(path, b.Bytes(), 0644)
 }

audit/internal/platform/sat_fan_stress.go

@@ -426,6 +426,101 @@ func sampleFanSpeedsViaSensorsJSON() ([]FanReading, error) {
 	return fans, nil
 }
 
+// sampleFanDutyCyclePct reads fan PWM/duty-cycle controls from lm-sensors.
+// Returns the average duty cycle across all exposed PWM controls.
+func sampleFanDutyCyclePct() (float64, bool) {
+	out, err := exec.Command("sensors", "-j").Output()
+	if err != nil || len(out) == 0 {
+		return 0, false
+	}
+	return parseFanDutyCyclePctSensorsJSON(out)
+}
+
+func parseFanDutyCyclePctSensorsJSON(raw []byte) (float64, bool) {
+	var doc map[string]map[string]any
+	if err := json.Unmarshal(raw, &doc); err != nil {
+		return 0, false
+	}
+	var samples []float64
+	for _, features := range doc {
+		for name, feature := range features {
+			if strings.EqualFold(name, "Adapter") {
+				continue
+			}
+			featureMap, ok := feature.(map[string]any)
+			if !ok {
+				continue
+			}
+			if duty, ok := firstFanDutyValue(name, featureMap); ok {
+				samples = append(samples, duty)
+			}
+		}
+	}
+	if len(samples) == 0 {
+		return 0, false
+	}
+	return benchmarkMean(samples), true
+}
+
+func firstFanDutyValue(featureName string, feature map[string]any) (float64, bool) {
+	featureName = strings.ToLower(strings.TrimSpace(featureName))
+	if strings.Contains(featureName, "enable") || strings.Contains(featureName, "mode") || strings.Contains(featureName, "alarm") {
+		return 0, false
+	}
+	if strings.Contains(featureName, "pwm") {
+		for _, key := range []string{"input", "value", "current"} {
+			if value, ok := feature[key]; ok {
+				if duty, parsed := parseFanDutyValue(value); parsed {
+					return duty, true
+				}
+			}
+		}
+	}
+	keys := make([]string, 0, len(feature))
+	for key := range feature {
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	for _, key := range keys {
+		lower := strings.ToLower(key)
+		if !strings.Contains(lower, "pwm") {
+			continue
+		}
+		if strings.Contains(lower, "enable") || strings.Contains(lower, "mode") || strings.Contains(lower, "alarm") {
+			continue
+		}
+		if duty, parsed := parseFanDutyValue(feature[key]); parsed {
+			return duty, true
+		}
+	}
+	return 0, false
+}
+
+func parseFanDutyValue(value any) (float64, bool) {
+	switch v := value.(type) {
+	case float64:
+		return normalizePWMAsDutyPct(v)
+	case string:
+		if f, err := strconv.ParseFloat(strings.TrimSpace(v), 64); err == nil {
+			return normalizePWMAsDutyPct(f)
+		}
+	}
+	return 0, false
+}
+
+func normalizePWMAsDutyPct(raw float64) (float64, bool) {
+	if raw < 0 {
+		return 0, false
+	}
+	if raw <= 100 {
+		return raw, true
+	}
+	if raw <= 255 {
+		return raw / 255.0 * 100.0, true
+	}
+	return 0, false
+}
+
 func firstFanInputValue(feature map[string]any) (float64, bool) {
 	keys := make([]string, 0, len(feature))
 	for key := range feature {

audit/internal/platform/sat_fan_stress_test.go

@@ -29,6 +29,27 @@ func TestFirstFanInputValue(t *testing.T) {
 	}
 }
 
+func TestParseFanDutyCyclePctSensorsJSON(t *testing.T) {
+	raw := []byte(`{
+		"chip0": {
+			"fan1": {"input": 9000},
+			"pwm1": {"input": 128},
+			"pwm1_enable": {"input": 1}
+		},
+		"chip1": {
+			"pwm2": {"input": 64}
+		}
+	}`)
+
+	got, ok := parseFanDutyCyclePctSensorsJSON(raw)
+	if !ok {
+		t.Fatalf("expected duty cycle telemetry to be parsed")
+	}
+	if got < 57 || got > 58 {
+		t.Fatalf("got=%v want ~57.1", got)
+	}
+}
+
 func TestParseDCMIPowerReading(t *testing.T) {
 	raw := `
 Instantaneous power reading:                   512 Watts

audit/internal/webui/api.go

@@ -36,6 +36,16 @@ var apiListNvidiaGPUStatuses = func(a *app.App) ([]platform.NvidiaGPUStatus, err
 	return a.ListNvidiaGPUStatuses()
 }
 
+const (
+	taskPriorityBenchmark      = 10
+	taskPriorityBurn           = 20
+	taskPriorityValidateStress = 30
+	taskPriorityValidate       = 40
+	taskPriorityAudit          = 50
+	taskPriorityInstallToRAM   = 60
+	taskPriorityInstall        = 70
+)
+
 // ── Job ID counter ────────────────────────────────────────────────────────────
 
 var jobCounter atomic.Uint64
@@ -109,6 +119,30 @@ func shouldSplitHomogeneousNvidiaTarget(target string) bool {
 	}
 }
 
+func defaultTaskPriority(target string, params taskParams) int {
+	switch strings.TrimSpace(target) {
+	case "install":
+		return taskPriorityInstall
+	case "install-to-ram":
+		return taskPriorityInstallToRAM
+	case "audit":
+		return taskPriorityAudit
+	case "nvidia-benchmark":
+		return taskPriorityBenchmark
+	case "nvidia-stress", "amd-stress", "memory-stress", "sat-stress", "platform-stress", "nvidia-compute":
+		return taskPriorityBurn
+	case "nvidia", "nvidia-targeted-stress", "nvidia-targeted-power", "nvidia-pulse",
+		"nvidia-interconnect", "nvidia-bandwidth", "memory", "storage", "cpu",
+		"amd", "amd-mem", "amd-bandwidth":
+		if params.StressMode {
+			return taskPriorityValidateStress
+		}
+		return taskPriorityValidate
+	default:
+		return 0
+	}
+}
+
 func expandHomogeneousNvidiaSelections(gpus []platform.NvidiaGPU, include, exclude []int) ([]nvidiaTaskSelection, error) {
 	if len(gpus) == 0 {
 		return nil, fmt.Errorf("no NVIDIA GPUs detected")
@@ -458,6 +492,7 @@ func (h *handler) handleAPIAuditRun(w http.ResponseWriter, _ *http.Request) {
 		ID:        newJobID("audit"),
 		Name:      "Audit",
 		Target:    "audit",
+		Priority:  defaultTaskPriority("audit", taskParams{}),
 		Status:    TaskPending,
 		CreatedAt: time.Now(),
 	}
@@ -526,7 +561,7 @@ func (h *handler) handleAPISATRun(target string) http.HandlerFunc {
 			DisplayName:        body.DisplayName,
 			PlatformComponents: body.PlatformComponents,
 		}
-		tasks, err := buildNvidiaTaskSet(target, 0, time.Now(), params, name, h.opts.App, "sat-"+target)
+		tasks, err := buildNvidiaTaskSet(target, defaultTaskPriority(target, params), time.Now(), params, name, h.opts.App, "sat-"+target)
 		if err != nil {
 			writeError(w, http.StatusBadRequest, err.Error())
 			return
@@ -613,7 +648,7 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 					ID:        newJobID("benchmark-nvidia"),
 					Name:      stepName,
 					Target:    "nvidia-benchmark",
-					Priority:  15,
+					Priority:  defaultTaskPriority("nvidia-benchmark", taskParams{}),
 					Status:    TaskPending,
 					CreatedAt: now,
 					params: taskParams{
@@ -645,7 +680,7 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 		name = fmt.Sprintf("%s · sequential", name)
 	}
 
-	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", 15, time.Now(), taskParams{
+	params := taskParams{
 		GPUIndices:        body.GPUIndices,
 		ExcludeGPUIndices: body.ExcludeGPUIndices,
 		SizeMB:            body.SizeMB,
@@ -653,7 +688,8 @@ func (h *handler) handleAPIBenchmarkNvidiaRun(w http.ResponseWriter, r *http.Req
 		RunNCCL:           runNCCL,
 		ParallelGPUs:      parallelGPUs,
 		DisplayName:       body.DisplayName,
-	}, name, h.opts.App, "benchmark-nvidia")
+	}
+	tasks, err := buildNvidiaTaskSet("nvidia-benchmark", defaultTaskPriority("nvidia-benchmark", params), time.Now(), params, name, h.opts.App, "benchmark-nvidia")
 	if err != nil {
 		writeError(w, http.StatusBadRequest, err.Error())
 		return
@@ -1054,7 +1090,7 @@ func (h *handler) handleAPIInstallToRAM(w http.ResponseWriter, r *http.Request)
 		ID:        newJobID("install-to-ram"),
 		Name:      "Install to RAM",
 		Target:    "install-to-ram",
-		Priority:  10,
+		Priority:  defaultTaskPriority("install-to-ram", taskParams{}),
 		Status:    TaskPending,
 		CreatedAt: time.Now(),
 	}
@@ -1169,7 +1205,7 @@ func (h *handler) handleAPIInstallRun(w http.ResponseWriter, r *http.Request) {
 		ID:        newJobID("install"),
 		Name:      "Install to Disk",
 		Target:    "install",
-		Priority:  20,
+		Priority:  defaultTaskPriority("install", taskParams{}),
 		Status:    TaskPending,
 		CreatedAt: time.Now(),
 		params: taskParams{
@@ -1461,4 +1497,3 @@ func (h *handler) rollbackPendingNetworkChange() error {
 	}
 	return nil
 }
-

audit/internal/webui/api_test.go

@@ -39,6 +39,9 @@ func TestHandleAPISATRunDecodesBodyWithoutContentLength(t *testing.T) {
 	if got := globalQueue.tasks[0].params.BurnProfile; got != "smoke" {
 		t.Fatalf("burn profile=%q want smoke", got)
 	}
+	if got := globalQueue.tasks[0].Priority; got != taskPriorityValidate {
+		t.Fatalf("priority=%d want %d", got, taskPriorityValidate)
+	}
 }
 
 func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
@@ -84,6 +87,9 @@ func TestHandleAPIBenchmarkNvidiaRunQueuesSelectedGPUs(t *testing.T) {
 	if task.params.RunNCCL {
 		t.Fatal("RunNCCL should reflect explicit false from request")
 	}
+	if task.Priority != taskPriorityBenchmark {
+		t.Fatalf("priority=%d want %d", task.Priority, taskPriorityBenchmark)
+	}
 }
 
 func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
@@ -133,6 +139,12 @@ func TestHandleAPIBenchmarkNvidiaRunSplitsMixedGPUModels(t *testing.T) {
 	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
 		t.Fatalf("task[1] gpu indices=%v want [2]", got)
 	}
+	if got := globalQueue.tasks[0].Priority; got != taskPriorityBenchmark {
+		t.Fatalf("task[0] priority=%d want %d", got, taskPriorityBenchmark)
+	}
+	if got := globalQueue.tasks[1].Priority; got != taskPriorityBenchmark {
+		t.Fatalf("task[1] priority=%d want %d", got, taskPriorityBenchmark)
+	}
 }
 
 func TestHandleAPISATRunSplitsMixedNvidiaTaskSet(t *testing.T) {
@@ -175,6 +187,39 @@ func TestHandleAPISATRunSplitsMixedNvidiaTaskSet(t *testing.T) {
 	if got := globalQueue.tasks[1].params.GPUIndices; len(got) != 1 || got[0] != 2 {
 		t.Fatalf("task[1] gpu indices=%v want [2]", got)
 	}
+	if got := globalQueue.tasks[0].Priority; got != taskPriorityValidate {
+		t.Fatalf("task[0] priority=%d want %d", got, taskPriorityValidate)
+	}
+	if got := globalQueue.tasks[1].Priority; got != taskPriorityValidate {
+		t.Fatalf("task[1] priority=%d want %d", got, taskPriorityValidate)
+	}
+}
+
+func TestDefaultTaskPriorityOrder(t *testing.T) {
+	got := []int{
+		defaultTaskPriority("install-to-ram", taskParams{}),
+		defaultTaskPriority("audit", taskParams{}),
+		defaultTaskPriority("cpu", taskParams{}),
+		defaultTaskPriority("cpu", taskParams{StressMode: true}),
+		defaultTaskPriority("nvidia-stress", taskParams{}),
+		defaultTaskPriority("nvidia-benchmark", taskParams{}),
+	}
+	want := []int{
+		taskPriorityInstallToRAM,
+		taskPriorityAudit,
+		taskPriorityValidate,
+		taskPriorityValidateStress,
+		taskPriorityBurn,
+		taskPriorityBenchmark,
+	}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Fatalf("priority[%d]=%d want %d", i, got[i], want[i])
+		}
+	}
+	if !(got[0] > got[1] && got[1] > got[2] && got[2] > got[3] && got[3] > got[4] && got[4] > got[5]) {
+		t.Fatalf("priority order=%v", got)
+	}
 }
 
 func TestPushFanRingsTracksByNameAndCarriesForwardMissingSamples(t *testing.T) {

iso/builder/VERSIONS

@@ -6,7 +6,7 @@ NCCL_CUDA_VERSION=13.0
 NCCL_SHA256=2e6faafd2c19cffc7738d9283976a3200ea9db9895907f337f0c7e5a25563186
 NCCL_TESTS_VERSION=2.13.10
 NVCC_VERSION=12.8
-CUBLAS_VERSION=13.0.2.14-1
+CUBLAS_VERSION=13.1.1.3-1
 CUDA_USERSPACE_VERSION=13.0.96-1
 DCGM_VERSION=4.5.3-1
 JOHN_JUMBO_COMMIT=67fcf9fe5a

iso/builder/bee-gpu-stress.c

@@ -33,7 +33,6 @@ typedef void *CUstream;
 #define CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR 75
 #define CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR 76
 #define MAX_STRESS_STREAMS 16
-#define MAX_CUBLAS_PROFILES 5
 #define MIN_PROFILE_BUDGET_BYTES ((size_t)4u * 1024u * 1024u)
 #define MIN_STREAM_BUDGET_BYTES ((size_t)64u * 1024u * 1024u)
 
@@ -689,6 +688,8 @@ static const struct profile_desc k_profiles[] = {
 #endif
 };
 
+#define PROFILE_COUNT ((int)(sizeof(k_profiles) / sizeof(k_profiles[0])))
+
 static int load_cublaslt(struct cublaslt_api *api) {
     memset(api, 0, sizeof(*api));
     api->lib = dlopen("libcublasLt.so.13", RTLD_NOW | RTLD_LOCAL);
@@ -1124,7 +1125,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
                                const char *precision_filter,
                                struct stress_report *report) {
     struct cublaslt_api cublas;
-    struct prepared_profile prepared[MAX_STRESS_STREAMS * MAX_CUBLAS_PROFILES];
+    struct prepared_profile prepared[MAX_STRESS_STREAMS * PROFILE_COUNT];
     cublasLtHandle_t handle = NULL;
     CUcontext ctx = NULL;
     CUstream streams[MAX_STRESS_STREAMS] = {0};
@@ -1134,7 +1135,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
     int active = 0;
     int mp_count = 0;
     int stream_count = 1;
-    int profile_count = (int)(sizeof(k_profiles) / sizeof(k_profiles[0]));
+    int profile_count = PROFILE_COUNT;
     int prepared_count = 0;
     size_t requested_budget = 0;
     size_t total_budget = 0;
@@ -1159,6 +1160,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         return 0;
     }
 
+    /* Count profiles matching the filter (for deciding what to run). */
     for (size_t i = 0; i < sizeof(k_profiles) / sizeof(k_profiles[0]); i++) {
         if (k_profiles[i].enabled && cc >= k_profiles[i].min_cc &&
             (precision_filter == NULL || strcmp(k_profiles[i].block_label, precision_filter) == 0)) {
@@ -1172,18 +1174,31 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         return 0;
     }
 
+    /* Count all profiles active on this GPU regardless of filter.
+     * Used as the budget divisor so matrix sizes stay consistent whether
+     * running all precisions together or a single-precision phase. */
+    int planned_total = 0;
+    for (size_t i = 0; i < sizeof(k_profiles) / sizeof(k_profiles[0]); i++) {
+        if (k_profiles[i].enabled && cc >= k_profiles[i].min_cc) {
+            planned_total++;
+        }
+    }
+    if (planned_total < planned) {
+        planned_total = planned;
+    }
+
     requested_budget = (size_t)size_mb * 1024u * 1024u;
-    if (requested_budget < (size_t)planned * MIN_PROFILE_BUDGET_BYTES) {
-        requested_budget = (size_t)planned * MIN_PROFILE_BUDGET_BYTES;
+    if (requested_budget < (size_t)planned_total * MIN_PROFILE_BUDGET_BYTES) {
+        requested_budget = (size_t)planned_total * MIN_PROFILE_BUDGET_BYTES;
     }
     total_budget = clamp_budget_to_free_memory(cuda, requested_budget);
-    if (total_budget < (size_t)planned * MIN_PROFILE_BUDGET_BYTES) {
-        total_budget = (size_t)planned * MIN_PROFILE_BUDGET_BYTES;
+    if (total_budget < (size_t)planned_total * MIN_PROFILE_BUDGET_BYTES) {
+        total_budget = (size_t)planned_total * MIN_PROFILE_BUDGET_BYTES;
     }
     if (query_multiprocessor_count(cuda, dev, &mp_count) &&
         cuda->cuStreamCreate &&
         cuda->cuStreamDestroy) {
-        stream_count = choose_stream_count(mp_count, planned, total_budget, 1);
+        stream_count = choose_stream_count(mp_count, planned_total, total_budget, 1);
     }
     if (stream_count > 1) {
         int created = 0;
@@ -1196,7 +1211,7 @@ static int run_cublaslt_stress(struct cuda_api *cuda,
         }
     }
     report->stream_count = stream_count;
-    per_profile_budget = total_budget / ((size_t)planned * (size_t)stream_count);
+    per_profile_budget = total_budget / ((size_t)planned_total * (size_t)stream_count);
     if (per_profile_budget < MIN_PROFILE_BUDGET_BYTES) {
         per_profile_budget = MIN_PROFILE_BUDGET_BYTES;
     }
@@ -1424,7 +1439,17 @@ int main(int argc, char **argv) {
     ok = run_cublaslt_stress(&cuda, dev, name, cc_major, cc_minor, seconds, size_mb, precision_filter, &report);
 #endif
     if (!ok) {
-        if (!run_ptx_fallback(&cuda, dev, name, cc_major, cc_minor, seconds, size_mb, &report)) {
+        if (precision_filter != NULL) {
+            fprintf(stderr,
+                    "requested precision path unavailable: precision=%s device=%s cc=%d.%d\n",
+                    precision_filter,
+                    name,
+                    cc_major,
+                    cc_minor);
+            return 1;
+        }
+        int ptx_mb = size_mb;
+        if (!run_ptx_fallback(&cuda, dev, name, cc_major, cc_minor, seconds, ptx_mb, &report)) {
             return 1;
         }
     }

iso/builder/build.sh

@@ -873,6 +873,22 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
 
     CUBLAS_CACHE="${DIST_DIR}/cublas-${CUBLAS_VERSION}+cuda${NCCL_CUDA_VERSION}"
 
+    echo "=== bee-gpu-burn FP4 header probe ==="
+    fp4_type_match="$(grep -Rsnm 1 'CUDA_R_4F_E2M1' "${CUBLAS_CACHE}/include" 2>/dev/null || true)"
+    fp4_scale_match="$(grep -Rsnm 1 'CUBLASLT_MATMUL_MATRIX_SCALE_VEC16_UE4M3' "${CUBLAS_CACHE}/include" 2>/dev/null || true)"
+    if [ -n "$fp4_type_match" ]; then
+        echo "fp4_header_symbol=present"
+        echo "$fp4_type_match"
+    else
+        echo "fp4_header_symbol=missing"
+    fi
+    if [ -n "$fp4_scale_match" ]; then
+        echo "fp4_scale_mode_symbol=present"
+        echo "$fp4_scale_match"
+    else
+        echo "fp4_scale_mode_symbol=missing"
+    fi
+
     GPU_STRESS_NEED_BUILD=1
     if [ -f "$GPU_BURN_WORKER_BIN" ]; then
         GPU_STRESS_NEED_BUILD=0
@@ -901,6 +917,12 @@ if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     else
         echo "=== bee-gpu-burn worker up to date, skipping build ==="
     fi
+    echo "=== bee-gpu-burn compiled profile probe ==="
+    if grep -aq 'fp4_e2m1' "$GPU_BURN_WORKER_BIN"; then
+        echo "fp4_profile_string=present"
+    else
+        echo "fp4_profile_string=missing"
+    fi
 fi
 
 echo "=== preparing staged overlay (${BUILD_VARIANT}) ==="