straylight: pin amdgpu DPM high and enable Laguna DFlash

Keep the iGPU out of powersave while llama-server runs, and pair Laguna
S 2.1 with Poolside's DFlash draft (f16 draft KV, n-max 4).

hw/straylight/default.nix

@@ -39,8 +39,16 @@ let
 
     ; Poolside Laguna S 2.1 (118B total, 8B active): primary coding agent.
     ; ~69 GB of weights, so it only loads once the raised TTM/GTT limit is active (reboot).
+    ; DFlash draft (~2.2 GB BF16) from poolside/Laguna-S-2.1-GGUF. Keep draft KV in
+    ; f16; q8 draft KV collapses acceptance. spec-draft-n-max 4 is the gfx1151
+    ; sweet spot (n=15 is the trained cap and is slower here).
     [laguna-s-2.1]
     model = ${llamaModelsDir}/laguna-s-2.1-UD-Q4_K_S/Laguna-S-2.1-UD-Q4_K_S-00001-of-00003.gguf
+    model-draft = ${llamaModelsDir}/laguna-s-2.1-UD-Q4_K_S/laguna-s-2.1-DFlash-BF16.gguf
+    spec-type = draft-dflash
+    spec-draft-n-max = 4
+    cache-type-k-draft = f16
+    cache-type-v-draft = f16
     parallel = 1
     ctx-size = 131072
 
@@ -143,6 +151,15 @@ let
   whisperTranscribe = pkgs.writeShellScriptBin "whisper-transcribe" ''
     exec ${whisperCpp}/bin/whisper-cli -m ${whisperModel} --language auto -f "$@"
   '';
+  amdgpuDpmHigh = pkgs.writeShellScript "amdgpu-dpm-high" ''
+    set -eu
+    for f in /sys/class/drm/card*/device/power_dpm_force_performance_level; do
+      vendor="$(dirname "$f")/vendor"
+      [ -r "$vendor" ] || continue
+      grep -qx '0x1002' "$vendor" || continue
+      echo high > "$f"
+    done
+  '';
 in
 {
   imports = [
@@ -159,6 +176,19 @@ in
     # amdgpu.gttsize is deprecated on recent kernels; ttm.* is the supported knob.
     boot.kernelParams = [ "ttm.pages_limit=27262976" "ttm.page_pool_size=27262976" ];
 
+    # amdgpu defaults to powersave/auto and will clock the iGPU down under a
+    # desktop compositor. llama.cpp on gfx1151 wants the high DPM profile.
+    systemd.services.amdgpu-dpm-high = {
+      description = "Set amdgpu DPM performance level to high";
+      wantedBy = [ "multi-user.target" ];
+      after = [ "systemd-udev-settle.service" ];
+      serviceConfig = {
+        Type = "oneshot";
+        RemainAfterExit = true;
+        ExecStart = amdgpuDpmHigh;
+      };
+    };
+
     # llama.cpp server (router mode) with the Vulkan backend for gfx1151 (Strix Halo).
     # Models and per-model settings live in llamaModelsIni above.
     # NOTE: BIOS "UMA Frame Buffer Size" must be ≥32GB for a 23GB model to fit on-GPU.
@@ -182,6 +212,8 @@ in
         # Bind all interfaces so localhost and the tailnet can both reach the
         # OpenAI-compatible API. LAN access is still blocked: 11434 is opened
         # only on tailscale0, not in the global allowedTCPPorts list.
+        # "+" runs as root so we can pin amdgpu DPM before the server starts.
+        ExecStartPre = "+${amdgpuDpmHigh}";
         ExecStart = ''
           ${llama-cpp}/bin/llama-server \
             --host 0.0.0.0 \