Add OpenVINO INT4 (8 steps) to benchmark; update results and model card

.gitattributes

@@ -53,3 +53,5 @@ results/images/avx512-Q4_K-r1.png filter=lfs diff=lfs merge=lfs -text
 results/images/avx512-Q4_K-r2.png filter=lfs diff=lfs merge=lfs -text
 results/images/stock-Q4_K-r1.png filter=lfs diff=lfs merge=lfs -text
 results/images/stock-Q4_K-r2.png filter=lfs diff=lfs merge=lfs -text
+results/images/ov-int4-8step-r1.png filter=lfs diff=lfs merge=lfs -text
+results/images/ov-int4-8step-r2.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -21,29 +21,36 @@ stable-diffusion.cpp 在 **純 CPU** 上跑 Z-Image-Turbo 的優化實驗。這
 
 ## 結論
 
-- 同一提示詞、同一 seed、512x512、8 steps。最快的組合是 **Q4_K + AMX 權重佈局 + FFN 融合**，總生成時間 **49.5 s**，相對 stock（56.8 s）快 **1.15 倍**。
-- 單獨啟用 AMX 權重佈局（`SD_CPU_EXTRA_BUFT=1`，Q4_K）相對 `amx_lto` 快約 15%（58.9 s → 49.8 s），是主要收益來源。
-- AVX512 相對 AVX2 約快 2 倍（118.4 s → 60.0 s）。
-- FFN gate/up 融合（w1 + w3 → w13）輸出與原版**逐位元相同**，但在這台機器上沒有可量測的加速，單獨使用沒有實際收益。
-- Q4_K 是此 CPU 上最快的量化格式。Q8_0 即使開 AMX 也比 Q4_K 慢約 15%；Q8_0 不開 AMX 時相對 stock 慢約 3 倍。
-- 峰值 RSS 在 7.7–11.1 GiB 之間，遠低於記憶體守衛上限。
+- 同一提示詞、同一 seed、512x512、**8 steps**。sd.cpp 最快的組合是 **Q4_K-ffn13 + AMX 權重佈局**（54.2 s），相對 stock（59.7 s）約快 1.10x。
+- **OpenVINO INT4**（同樣 8 steps）總生成 **37.5 s**，是本批次最快的一列（約 1.59x stock），峰值 RSS 12.6 GiB。但它是不同的權重與實作（optimum-intel + OpenVINO），像素與 sd.cpp 不可直接比較品質。
+- 單獨啟用 AMX 權重佈局（`SD_CPU_EXTRA_BUFT=1`，Q4_K）相對 `amx_lto` 約快 15–20%，是 sd.cpp 內最主要的收益來源。
+- AVX512 相對 AVX2 約快 2 倍（118.5 s → 58.1 s）。
+- FFN gate/up 融合（w1 + w3 → w13）輸出與原版逐位元相同，但在這台機器上沒有可量測的加速。
+- Q4_K 是 sd.cpp 在此 CPU 上最快的量化格式。Q8_0 即使開 AMX 也比 Q4_K 慢。
+- 峰值 RSS：sd.cpp 7.7–11.1 GiB，OpenVINO 12.6 GiB，皆遠低於記憶體守衛上限。
+
+**雜訊說明**：同一 stock 設定在不同批次的總生成時間為 56.8 s（前次）與 59.7 s（本次）。排名相近的列（例如 Q4_K-ffn13+AMXw 與 Q4_K+AMXw）差距在 ±5% 內，不宜過度解讀。
 
 ## 比較表（中位數，n = 2）
 
 | # | 版本 (variant) | 模型 | 額外環境變數 | 採樣 (s) | 總生成 (s) | 相對 stock | 峰值 RSS (GiB) | 與 stock 像素差 (mean/max) |
 |---|---|---|---|---|---|---|---|---|
-| 1 | stock | Q4_K | - | 46.4 | 56.8 | 1.00x | 7.86 | 基準 |
-| 2 | avx2 | Q4_K | - | 107.3 | 118.4 | 0.48x | 7.86 | 3.18 / 241 |
-| 3 | avx512 | Q4_K | - | 49.1 | 60.0 | 0.95x | 7.86 | 3.17 / 239 |
-| 4 | amx_lto | Q4_K | - | 48.2 | 58.9 | 0.96x | 7.86 | 0.00 / 0 |
-| 5 | amx_lto | Q4_K | `SD_CPU_EXTRA_BUFT=1` | 39.1 | 49.8 | 1.14x | 8.22 | 1.28 / 227 |
-| 6 | amx_lto | Q4_K-ffn13 | - | 45.9 | 56.4 | 1.01x | 7.86 | 0.00 / 0 |
-| 7 | amx_lto | Q4_K-ffn13 | `SD_CPU_EXTRA_BUFT=1` | 38.3 | **49.5** | **1.15x** | 8.18 | 1.28 / 227 |
-| 8 | amx_lto | Q4_0 | - | 174.4 | 187.2 | 0.30x | 7.69 | 22.84 / 252 |
-| 9 | amx_lto | Q8_0 | - | 162.9 | 173.6 | 0.33x | 10.38 | 12.51 / 249 |
-| 10 | amx_lto | Q8_0 | `SD_CPU_EXTRA_BUFT=1` | 46.7 | 57.0 | 1.00x | 11.11 | 12.33 / 249 |
+| 1 | stock | Q4_K | - | 49.2 | 59.7 | 1.00x | 7.86 | ref |
+| 2 | avx2 | Q4_K | - | 106.7 | 118.5 | 0.50x | 7.86 | 3.18 / 241 |
+| 3 | avx512 | Q4_K | - | 47.6 | 58.1 | 1.03x | 7.86 | 3.17 / 239 |
+| 4 | amx_lto | Q4_K | - | 52.4 | 63.7 | 0.94x | 7.86 | 0.00 / 0 |
+| 5 | amx_lto | Q4_K | SD_CPU_EXTRA_BUFT=1 | 42.0 | 52.7 | 1.13x | 8.16 | 1.28 / 227 |
+| 6 | amx_lto | Q4_K-ffn13 | - | 46.5 | 57.1 | 1.05x | 7.86 | 0.00 / 0 |
+| 7 | amx_lto | Q4_K-ffn13 | SD_CPU_EXTRA_BUFT=1 | 43.1 | 54.2 | 1.10x | 8.14 | 1.28 / 227 |
+| 8 | amx_lto | Q4_0 | - | 172.5 | 183.5 | 0.33x | 7.69 | 22.84 / 252 |
+| 9 | amx_lto | Q8_0 | - | 161.4 | 172.0 | 0.35x | 10.38 | 12.51 / 249 |
+| 10 | amx_lto | Q8_0 | SD_CPU_EXTRA_BUFT=1 | 55.9 | 66.4 | 0.90x | 11.11 | 12.33 / 249 |
+| 11 | openvino | OV-INT4 | - | 36.8 | 37.5 | 1.59x | 12.59 | 27.71 / 254 |
+
+n = 2 run(s) per row (median). Speedup = stock-Q4_K generate time / row generate time.
 
 - 「相對 stock」= stock 的總生成時間 / 該列的總生成時間，大於 1 表示更快。
+- OpenVINO 列為 **8 steps**（與 sd.cpp 相同步數）；它的「像素差」是與 stock 的差異，反映的是不同實作，不是品質差距。
 - 「像素差」只對**權重相同**的列有意義。第 6 列與第 1 列完全相同，表示 FFN 融合沒有改變輸出。第 5、7 列是 AMX 的數值差異（平均差 1.28），視覺上與 stock 相同。Q4_0、Q8_0 使用不同權重，差異大是預期的，不代表品質問題。
 - AVX2 與 AVX512 的輸出與 stock 不同，原因是浮點運算順序不同。
 
@@ -56,6 +63,8 @@ stable-diffusion.cpp 在 **純 CPU** 上跑 Z-Image-Turbo 的優化實驗。這
 | `avx512` | `GGML_NATIVE=OFF`，明確開 AVX512/VNNI/BF16/AMX，LTO | 與 native 版本效能接近（60.0 s vs 56.8 s，差距在雜訊附近） |
 | `amx_lto` | `GGML_NATIVE=ON`，AMX_TILE/INT8/BF16，LTO | 目前最佳的編譯選項 |
 
+| `openvino` | HelloSun/Z-Image-Turbo-OpenVINO-INT4，`tools/ov_bench.py`（optimum-intel 2.2.0 / OpenVINO 2026.4.0） | 不同推論框架��� INT4 權重，只作為對照，步數與 sd.cpp 相同（8） |
+
 執行期的選項：
 
 - `SD_CPU_EXTRA_BUFT=1`：擴散模型的 2 維權重改用 CPU 的 AMX 佈局（實驗性質，預設關閉）。

SmartTest.sh

@@ -80,57 +80,51 @@ echo "name,variant,model,env,rep,status,sampling_s,generate_s,wall_s,peak_rss_gi
     echo "gcc:         $(gcc -dumpfullversion 2>/dev/null || echo n/a)"
     echo "cmake:       $(cmake --version 2>/dev/null | head -1 || echo n/a)"
     echo "sd.cpp:      $(git -C "$WORK/stable-diffusion.cpp" rev-parse HEAD 2>/dev/null || echo n/a)"
+    echo "openvino env: $("$WORK/venv-ov/bin/python" -c 'import openvino, optimum.intel, diffusers; print(f"openvino {openvino.__version__}, optimum-intel {optimum.intel.__version__}, diffusers {diffusers.__version__}")' 2>/dev/null || echo n/a)"
+    echo "openvino model: HelloSun/Z-Image-Turbo-OpenVINO-INT4 (transformer/text_encoder INT4, VAE INT8)"
     echo "prompt:      $PROMPT"
     echo "negative:    $NEGATIVE"
     echo "size/steps/seed/cfg: ${WIDTH}x${HEIGHT} / $STEPS / $SEED / 1.0 (euler, simple scheduler)"
 } > "$OUT/hardware.txt"
 
-# ------------------------------------------------------------ run matrix
+# ------------------------------------------------------------ run cases
 REF_IMAGE=""
-for row in "${MATRIX[@]}"; do
-    IFS='|' read -r name variant mkey envspec <<< "$row"
-    if [[ -n "$ONLY" && "$name" != *"$ONLY"* ]]; then continue; fi
-
-    sdbin="$BIN/$variant/sd-cli"
-    model="$(model_file "$mkey")"
-    if [[ ! -x "$sdbin" || ! -f "$model" ]]; then
-        log "skip $name (missing binary or model)"
-        echo "$name,$variant,$mkey,$envspec,0,missing,,,,,,,," >> "$CSV"
-        continue
-    fi
-
-    envprefix=()
+OV_DIR="$MODELS/Z-Image-Turbo-OpenVINO-INT4"
+OV_PY="$WORK/venv-ov/bin/python"
+
+# run_case <name> <variant> <model-label> <env-spec> <rep-command...>
+# The rep-command is executed under taskset; "__IMG__" is replaced by the output image path.
+# It must print sd-cli style timing lines (sampling / generate_image) on stderr.
+run_case() {
+    local name="$1" variant="$2" mlabel="$3" envspec="$4"; shift 4
+    local envprefix=()
     [[ "$envspec" != "-" ]] && envprefix=(env "$envspec")
-
     for rep in $(seq 1 "$REPEAT"); do
-        tag="${name}-r${rep}"
+        local tag="${name}-r${rep}"
+        local log_file="$OUT/logs/$tag.log"
+        local img="$OUT/images/$tag.png"
         log "run $tag"
-        log_file="$OUT/logs/$tag.log"
-        img="$OUT/images/$tag.png"
+        local cmd=() a
+        for a in "$@"; do
+            if [[ "$a" == "__IMG__" ]]; then cmd+=("$img"); else cmd+=("$a"); fi
+        done
+        # stdout and stderr of the generator both go to the log; measure_rss appends its summary line too.
         set +e
-        "${envprefix[@]}" "$PY" "$ROOT/tools/measure_rss.py" \
-            taskset -c "$CPUSET" "$sdbin" \
-            --diffusion-model "$model" \
-            --vae "$MODELS/ae.safetensors" \
-            --llm "$MODELS/Qwen3-4B-Instruct-2507-Q4_K_M.gguf" \
-            -p "$PROMPT" -n "$NEGATIVE" \
-            -W "$WIDTH" -H "$HEIGHT" --steps "$STEPS" --cfg-scale 1.0 \
-            --seed "$SEED" --sampling-method euler --scheduler simple \
-            -t "$THREADS" -o "$img" > "$log_file" 2> "$log_file.rss"
-        rc=$?
+        "${envprefix[@]}" "$PY" "$ROOT/tools/measure_rss.py" taskset -c "$CPUSET" "${cmd[@]}" \
+            > "$log_file" 2>&1
+        local rc=$?
         set -e
-        rss_line="$(grep -o 'PEAK_RSS=[0-9.]*' "$log_file.rss" | cut -d= -f2 || true)"
-        sampling="$(grep -oE 'sampling completed, taking [0-9.]+s' "$log_file" | grep -oE '[0-9.]+' | tail -1 || true)"
-        generate="$(grep -oE 'generate_image completed in [0-9.]+s' "$log_file" | grep -oE '[0-9.]+' | tail -1 || true)"
-        wall="$(grep -oE 'WALL=[0-9.]+' "$log_file.rss" | cut -d= -f2 || true)"
+        local rss_line sampling generate wall
+        rss_line="$(grep -oE 'PEAK_RSS=[0-9.]+' "$log_file" | cut -d= -f2 | tail -1 || true)"
+        sampling="$(grep -oE 'sampling completed, taking [0-9.]+' "$log_file" | grep -oE '[0-9.]+' | tail -1 || true)"
+        generate="$(grep -oE 'generate_image completed in [0-9.]+' "$log_file" | grep -oE '[0-9.]+' | tail -1 || true)"
+        wall="$(grep -oE 'WALL=[0-9.]+' "$log_file" | cut -d= -f2 | tail -1 || true)"
 
-        status="ok"; diff_mean=""; diff_max=""
+        local status="ok" diff_mean="" diff_max=""
         if [[ $rc -ne 0 || ! -s "$img" ]]; then
             status="failed(rc=$rc)"
         else
-            if [[ -z "$REF_IMAGE" ]]; then
-                REF_IMAGE="$img"
-            fi
+            [[ -z "$REF_IMAGE" ]] && REF_IMAGE="$img"
             read -r diff_mean diff_max < <("$PY" - "$REF_IMAGE" "$img" <<'PY'
 import sys
 import numpy as np
@@ -142,9 +136,47 @@ print(f"{d.mean():.4f} {d.max()}")
 PY
 )
         fi
-        echo "$name,$variant,$mkey,$envspec,$rep,$status,$sampling,$generate,$wall,$rss_line,$diff_mean,$diff_max,images/$tag.png" >> "$CSV"
+        echo "$name,$variant,$mlabel,$envspec,$rep,$status,$sampling,$generate,$wall,$rss_line,$diff_mean,$diff_max,images/$tag.png" >> "$CSV"
         log "  $status sampling=${sampling:-?}s generate=${generate:-?}s rss=${rss_line:-?}GiB diff_mean=${diff_mean:-n/a}"
     done
+}
+
+# ---- sd.cpp matrix
+for row in "${MATRIX[@]}"; do
+    IFS='|' read -r name variant mkey envspec <<< "$row"
+    if [[ -n "$ONLY" && "$name" != *"$ONLY"* ]]; then continue; fi
+
+    sdbin="$BIN/$variant/sd-cli"
+    model="$(model_file "$mkey")"
+    if [[ ! -x "$sdbin" || ! -f "$model" ]]; then
+        log "skip $name (missing binary or model)"
+        echo "$name,$variant,$mkey,$envspec,0,missing,,,,,,,," >> "$CSV"
+        continue
+    fi
+    run_case "$name" "$variant" "$mkey" "$envspec" \
+        "$sdbin" \
+        --diffusion-model "$model" \
+        --vae "$MODELS/ae.safetensors" \
+        --llm "$MODELS/Qwen3-4B-Instruct-2507-Q4_K_M.gguf" \
+        -p "$PROMPT" -n "$NEGATIVE" \
+        -W "$WIDTH" -H "$HEIGHT" --steps "$STEPS" --cfg-scale 1.0 \
+        --seed "$SEED" --sampling-method euler --scheduler simple \
+        -t "$THREADS" -o "__IMG__"
+done
+
+# ---- OpenVINO INT4 (HelloSun/Z-Image-Turbo-OpenVINO-INT4), 8 steps to match sd.cpp; same prompt/seed/size
+for ov_row in "ov-int4-8step|8"; do
+    IFS='|' read -r name steps <<< "$ov_row"
+    if [[ -n "$ONLY" && "$name" != *"$ONLY"* ]]; then continue; fi
+    if [[ ! -x "$OV_PY" || ! -d "$OV_DIR" ]]; then
+        log "skip $name (missing OpenVINO venv or model; see build.sh)"
+        echo "$name,openvino,OV-INT4,-,0,missing,,,,,,,," >> "$CSV"
+        continue
+    fi
+    run_case "$name" "openvino" "OV-INT4" "-" \
+        "$OV_PY" "$ROOT/tools/ov_bench.py" \
+        --model-dir "$OV_DIR" -p "$PROMPT" \
+        -W "$WIDTH" -H "$HEIGHT" --steps "$steps" --seed "$SEED" -o "__IMG__"
 done
 
 # ------------------------------------------------------------ report (medians)
@@ -177,10 +209,12 @@ for i, (name, rs) in enumerate(groups.items(), 1):
     rss = med([r["peak_rss_gib"] for r in rs])
     dm = med([r["diff_mean_vs_ref"] for r in rs])
     dmx = med([r["diff_max_vs_ref"] for r in rs])
+    f1 = lambda v: "n/a" if v is None else f"{v:.1f}"
+    f2 = lambda v: "n/a" if v is None else f"{v:.2f}"
     speed = f"{ref / gen:.2f}x" if ref and gen else "n/a"
     diff = "ref" if name == "stock-Q4_K" else (f"{dm:.2f} / {dmx:.0f}" if dm is not None else "n/a")
     env = rs[0]["env"] if rs[0]["env"] != "-" else "-"
-    lines.append(f"| {i} | {rs[0]['variant']} | {rs[0]['model']} | {env} | {samp:.1f} | {gen:.1f} | {speed} | {rss:.2f} | {diff} |")
+    lines.append(f"| {i} | {rs[0]['variant']} | {rs[0]['model']} | {env} | {f1(samp)} | {f1(gen)} | {speed} | {f2(rss)} | {diff} |")
 lines.append("")
 lines.append(f"n = {max((len(v) for v in groups.values()), default=0)} run(s) per row (median). Speedup = stock-Q4_K generate time / row generate time.")
 open(sys.argv[2], "w").write("\n".join(lines) + "\n")

build.sh

@@ -5,6 +5,7 @@
 #   ./build.sh amx_lto avx512         # only the named variants
 #   ./build.sh --list                 # show variants and their CMake flags
 #   ./build.sh --skip-models amx_lto  # no model download
+#   ./build.sh --skip-openvino --all  # skip the OpenVINO INT4 env + model (about 5.5 GB)
 #
 # Environment overrides:
 #   WORK=./work                 working dir (source, binaries, models, venv)
@@ -52,7 +53,7 @@ usage() {
 }
 
 # ---------------------------------------------------------------- args
-DO_DEPS=1; DO_MODELS=1; SELECTED=()
+DO_DEPS=1; DO_MODELS=1; DO_OV=1; SELECTED=()
 for arg in "$@"; do
     case "$arg" in
         -h|--help) usage ;;
@@ -62,6 +63,7 @@ for arg in "$@"; do
         --all) SELECTED=("${ALL_VARIANTS[@]}") ;;
         --skip-deps) DO_DEPS=0 ;;
         --skip-models) DO_MODELS=0 ;;
+        --skip-openvino) DO_OV=0 ;;
         -*) die "unknown option: $arg" ;;
         *) variant_flags "$arg" >/dev/null || die "unknown variant: $arg (see --list)"; SELECTED+=("$arg") ;;
     esac
@@ -95,6 +97,19 @@ fi
 PY="$VENV/bin/python"
 [[ -x "$PY" ]] || PY="$(command -v python3)"
 
+OV_VENV="$WORK/venv-ov"
+OV_MODEL_REPO="${OV_MODEL_REPO:-HelloSun/Z-Image-Turbo-OpenVINO-INT4}"
+OV_MODEL_DIR="$MODELS/Z-Image-Turbo-OpenVINO-INT4"
+if [[ $DO_DEPS -eq 1 && $DO_OV -eq 1 ]]; then
+    if [[ ! -x "$OV_VENV/bin/python" ]]; then
+        log "creating OpenVINO venv at $OV_VENV (pinned: diffusers 0.37.1, optimum-intel 2.2.0, openvino 2026.4.0)"
+        python3 -m venv "$OV_VENV"
+        "$OV_VENV/bin/pip" install -q --upgrade pip
+        "$OV_VENV/bin/pip" install -q 'diffusers==0.37.1' 'transformers==4.57.6' 'tokenizers==0.22.0' \
+            'optimum==2.3.0' 'optimum-intel==2.2.0' 'openvino==2026.4.0' 'huggingface-hub==0.36.2' pillow numpy
+    fi
+fi
+
 # ---------------------------------------------------------------- source
 if [[ ${#SELECTED[@]} -gt 0 ]]; then
     if [[ ! -d "$SRC/.git" ]]; then
@@ -160,6 +175,21 @@ for repo, name in files:
     path = hf_hub_download(repo_id=repo, filename=name, local_dir=models_dir)
     print(f"  -> {path}")
 PY
+    if [[ $DO_OV -eq 1 ]]; then
+        if [[ ! -f "$OV_MODEL_DIR/transformer/openvino_model.bin" ]]; then
+            log "downloading OpenVINO INT4 model $OV_MODEL_REPO (about 5.5 GB)"
+            "$PY" - "$OV_MODEL_REPO" "$OV_MODEL_DIR" <<'PY'
+import sys
+from huggingface_hub import snapshot_download
+repo, dst = sys.argv[1:3]
+snapshot_download(repo_id=repo, local_dir=dst,
+                  ignore_patterns=["examples/*", "generated/*", "*.png", "README.md", "REPORT.md"])
+print("  ->", dst)
+PY
+        else
+            log "have OpenVINO model at $OV_MODEL_DIR"
+        fi
+    fi
     # FFN gate/up fusion at the weight level (w1 + w3 -> w13); see tools/fuse_ffn.py
     if [[ ! -f "$MODELS/z_image_turbo-Q4_K-ffn13.gguf" ]]; then
         log "fusing FFN w1/w3 -> w13 (Q4_K)"

results/hardware.txt

@@ -1,4 +1,4 @@
-date:        2026-10-28 22:07:08 UTC
+date:        2026-10-28 23:06:15 UTC
 os:          Debian GNU/Linux 13 (trixie)
 kernel:      6.18.48-107.148.amzn2023.x86_64
 cpu:         Intel(R) Xeon(R) Platinum 8559C
@@ -13,6 +13,8 @@ gpu:         Failed to initialize NVML: Unknown Error  (no GPU backend used; CPU
 gcc:         14.2.0
 cmake:       cmake version 3.31.6
 sd.cpp:      77b72a2efed78f55de26892c7e909cee001e8b83
+openvino env: n/a
+openvino model: HelloSun/Z-Image-Turbo-OpenVINO-INT4 (transformer/text_encoder INT4, VAE INT8)
 prompt:      cute cartoon cat and cute cartoon snake, best friends, hugging together, kawaii, simple clean background, children's book illustration, soft colors, adorable
 negative:    blurry, low quality, realistic, photo
 size/steps/seed/cfg: 512x512 / 8 / 42 / 1.0 (euler, simple scheduler)

results/logs/amx_lto-Q4_0-r1.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -   9.74GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.42s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -  10.00GB/s
+[I] loading tensors completed, taking 0.23s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.45s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 443/452 -  14.60GB/s
  |##################################################| 452/452 -   8.35GB/s
-[I] loading tensors completed, taking 0.41s (read: 0.22s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -  21.79s/it - ETA 02:32
  |============>                                     | 2/8 -  21.06s/it - ETA 02:08
  |==================>                               | 3/8 -  21.07s/it - ETA 01:46
  |=========================>                        | 4/8 -  21.23s/it - ETA 01:25
  |===============================>                  | 5/8 -  21.17s/it - ETA 01:03
  |=====================================>            | 6/8 -  21.09s/it - ETA 00:42
  |===========================================>      | 7/8 -  21.59s/it - ETA 00:21
  |==================================================| 8/8 -  21.20s/it
-[I] sampling completed, taking 170.21s --- image.cpp:953
-[I] generating 1 latent images completed, taking 170.21s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
+
  |#####                                             |  40/452 - 1008.41MB/s
  |##################################################| 452/452 -  16.57GB/s
+[I] loading tensors completed, taking 0.21s (read: 0.19s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -  21.68s/it - ETA 02:31
  |============>                                     | 2/8 -  21.66s/it - ETA 02:10
  |==================>                               | 3/8 -  21.39s/it - ETA 01:47
  |=========================>                        | 4/8 -  21.35s/it - ETA 01:26
  |===============================>                  | 5/8 -  21.17s/it - ETA 01:04
  |=====================================>            | 6/8 -  21.64s/it - ETA 00:42
  |===========================================>      | 7/8 -  21.60s/it - ETA 00:21
  |==================================================| 8/8 -  21.45s/it
+[I] sampling completed, taking 171.95s --- image.cpp:953
+[I] generating 1 latent images completed, taking 171.95s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  53/138 - 465.34MB/s
  |##################################################| 138/138 - 940.30MB/s
+
  |###################                               |  52/138 - 465.34MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.13s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.13s --- image.cpp:675
-[I] generate_image completed in 180.77s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.49s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.49s --- image.cpp:675
+[I] generate_image completed in 182.90s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_0-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.68 GiB WALL=183.8s

results/logs/amx_lto-Q4_0-r2.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  23/386 -  20.51MB/s
  |##################################################| 386/386 -   9.15GB/s
-[I] loading tensors completed, taking 0.25s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.43s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |#####                                             |  38/386 -   5.03GB/s
  |##################################################| 386/386 -   9.38GB/s
+[I] loading tensors completed, taking 0.24s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.44s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
+[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
 
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  16.98GB/s
-[I] loading tensors completed, taking 0.20s (read: 0.19s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -  21.45s/it - ETA 02:30
  |============>                                     | 2/8 -  21.47s/it - ETA 02:08
  |==================>                               | 3/8 -  21.37s/it - ETA 01:47
  |=========================>                        | 4/8 -  21.38s/it - ETA 01:25
  |===============================>                  | 5/8 -  22.24s/it - ETA 01:04
  |=====================================>            | 6/8 -  23.40s/it - ETA 00:43
  |===========================================>      | 7/8 -  23.29s/it - ETA 00:22
  |==================================================| 8/8 -  23.87s/it
-[I] sampling completed, taking 178.49s --- image.cpp:953
-[I] generating 1 latent images completed, taking 178.49s --- image.cpp:965
+[I] loading tensors completed, taking 0.20s (read: 0.17s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -  21.91s/it - ETA 02:33
  |============>                                     | 2/8 -  21.41s/it - ETA 02:09
  |==================>                               | 3/8 -  21.40s/it - ETA 01:47
  |=========================>                        | 4/8 -  22.07s/it - ETA 01:26
  |===============================>                  | 5/8 -  21.33s/it - ETA 01:04
  |=====================================>            | 6/8 -  22.26s/it - ETA 00:43
  |===========================================>      | 7/8 -  21.35s/it - ETA 00:21
  |==================================================| 8/8 -  21.27s/it
+[I] sampling completed, taking 173.02s --- image.cpp:953
+[I] generating 1 latent images completed, taking 173.02s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 935.65MB/s
+[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |####################                              |  54/138 - 232.92MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 14.69s --- image.cpp:671
-[I] decode_first_stage completed, taking 14.69s --- image.cpp:675
-[I] generate_image completed in 193.62s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.69s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.69s --- image.cpp:675
+[I] generate_image completed in 184.15s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_0-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.69 GiB WALL=185.0s

results/logs/amx_lto-Q4_K+AMXw-r1.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  10.42MB/s
  |##################################################| 386/386 -   9.96GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.85s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |######                                            |  45/386 -   4.23GB/s
  |##################################################| 386/386 -   9.58GB/s
+[I] loading tensors completed, taking 0.24s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.43s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 13 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 -  77.39MB/s
  |######################                            | 197/452 -   6.64GB/s
  |##################################################| 452/452 -   8.14GB/s
+[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  34/452 - 232.18MB/s
  |#########################                         | 221/452 -   7.52GB/s
  |##################################################| 452/452 -   8.16GB/s
 [I] loading tensors completed, taking 0.44s (read: 0.05s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.33s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   5.03s/it - ETA 00:35
  |============>                                     | 2/8 -   5.26s/it - ETA 00:30
  |==================>                               | 3/8 -   4.75s/it - ETA 00:25
  |=========================>                        | 4/8 -   4.62s/it - ETA 00:19
  |===============================>                  | 5/8 -   4.75s/it - ETA 00:14
  |=====================================>            | 6/8 -   4.84s/it - ETA 00:09
  |===========================================>      | 7/8 -   5.18s/it - ETA 00:04
  |==================================================| 8/8 -   4.87s/it
-[I] sampling completed, taking 39.32s --- image.cpp:953
-[I] generating 1 latent images completed, taking 39.32s --- image.cpp:965
+
  |======>                                           | 1/8 -   6.14s/it - ETA 00:42
  |============>                                     | 2/8 -   4.95s/it - ETA 00:33
  |==================>                               | 3/8 -   5.51s/it - ETA 00:27
  |=========================>                        | 4/8 -   6.91s/it - ETA 00:23
  |===============================>                  | 5/8 -   5.19s/it - ETA 00:17
  |=====================================>            | 6/8 -   4.81s/it - ETA 00:11
  |===========================================>      | 7/8 -   4.49s/it - ETA 00:05
  |==================================================| 8/8 -   4.73s/it
+[I] sampling completed, taking 42.74s --- image.cpp:953
+[I] generating 1 latent images completed, taking 42.74s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 935.65MB/s
+[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.09s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.09s --- image.cpp:675
-[I] generate_image completed in 50.27s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.13s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.13s --- image.cpp:675
+[I] generate_image completed in 53.30s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_K+AMXw-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=8.16 GiB WALL=54.2s

results/logs/amx_lto-Q4_K+AMXw-r2.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -  10.23GB/s
-[I] loading tensors completed, taking 0.22s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.40s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |########                                          |  60/386 -   8.22GB/s
  |##################################################| 386/386 -  10.47GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.11s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.41s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 232.18MB/s
  |##########################                        | 234/452 -   7.99GB/s
  |##################################################| 452/452 -   8.74GB/s
-[I] loading tensors completed, taking 0.41s (read: 0.05s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.29s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   5.19s/it - ETA 00:36
  |============>                                     | 2/8 -   4.95s/it - ETA 00:30
  |==================>                               | 3/8 -   4.98s/it - ETA 00:25
  |=========================>                        | 4/8 -   4.75s/it - ETA 00:19
  |===============================>                  | 5/8 -   4.50s/it - ETA 00:14
  |=====================================>            | 6/8 -   5.37s/it - ETA 00:09
  |===========================================>      | 7/8 -   4.69s/it - ETA 00:04
  |==================================================| 8/8 -   4.54s/it
-[I] sampling completed, taking 38.98s --- image.cpp:953
-[I] generating 1 latent images completed, taking 38.98s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  34/452 - 116.09MB/s
  |##########################                        | 234/452 -   7.95GB/s
  |##################################################| 452/452 -   8.67GB/s
+[I] loading tensors completed, taking 0.41s (read: 0.05s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.30s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   5.10s/it - ETA 00:35
  |============>                                     | 2/8 -   4.72s/it - ETA 00:29
  |==================>                               | 3/8 -   4.89s/it - ETA 00:24
  |=========================>                        | 4/8 -   5.31s/it - ETA 00:20
  |===============================>                  | 5/8 -   5.13s/it - ETA 00:15
  |=====================================>            | 6/8 -   5.04s/it - ETA 00:10
  |===========================================>      | 7/8 -   5.75s/it - ETA 00:05
  |==================================================| 8/8 -   5.40s/it
+[I] sampling completed, taking 41.34s --- image.cpp:953
+[I] generating 1 latent images completed, taking 41.34s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 935.65MB/s
+
  |####################                              |  53/138 - 340.34MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 9.98s --- image.cpp:671
-[I] decode_first_stage completed, taking 9.98s --- image.cpp:675
-[I] generate_image completed in 49.37s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.24s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.24s --- image.cpp:675
+[I] generate_image completed in 52.01s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_K+AMXw-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=8.16 GiB WALL=53.0s

results/logs/amx_lto-Q4_K-ffn13+AMXw-r1.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  27/386 -   1.60GB/s
  |##################################################| 386/386 -   9.83GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.41s --- image.cpp:580
+
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -  10.18GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.55s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
 [I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |#####                                             |  34/418 - 116.33MB/s
  |############################                      | 226/418 -   7.81GB/s
  |##################################################| 418/418 -   8.53GB/s
-[I] loading tensors completed, taking 0.42s (read: 0.04s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.30s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   5.08s/it - ETA 00:35
  |============>                                     | 2/8 -   4.99s/it - ETA 00:30
  |==================>                               | 3/8 -   5.11s/it - ETA 00:25
  |=========================>                        | 4/8 -   5.00s/it - ETA 00:20
  |===============================>                  | 5/8 -   4.46s/it - ETA 00:14
  |=====================================>            | 6/8 -   4.46s/it - ETA 00:09
  |===========================================>      | 7/8 -   4.54s/it - ETA 00:04
  |==================================================| 8/8 -   4.50s/it
-[I] sampling completed, taking 38.15s --- image.cpp:953
-[I] generating 1 latent images completed, taking 38.15s --- image.cpp:965
+
  |#####                                             |  37/418 - 261.96MB/s
  |##########################                        | 217/418 -   7.44GB/s
  |##################################################| 418/418 -   8.16GB/s
+[I] loading tensors completed, taking 0.44s (read: 0.04s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.29s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   5.73s/it - ETA 00:40
  |============>                                     | 2/8 -   6.07s/it - ETA 00:35
  |==================>                               | 3/8 -   5.57s/it - ETA 00:28
  |=========================>                        | 4/8 -   4.98s/it - ETA 00:22
  |===============================>                  | 5/8 -   4.92s/it - ETA 00:16
  |=====================================>            | 6/8 -   4.87s/it - ETA 00:10
  |===========================================>      | 7/8 -   5.68s/it - ETA 00:05
  |==================================================| 8/8 -   4.69s/it
+[I] sampling completed, taking 42.52s --- image.cpp:953
+[I] generating 1 latent images completed, taking 42.52s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
+[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
+
  |####################                              |  54/138 - 232.92MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.24s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.24s --- image.cpp:675
-[I] generate_image completed in 48.81s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.84s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.84s --- image.cpp:675
+[I] generate_image completed in 53.92s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_K-ffn13+AMXw-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=8.18 GiB WALL=54.8s

results/logs/amx_lto-Q4_K-ffn13+AMXw-r2.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  10.42MB/s
  |##################################################| 386/386 -   9.19GB/s
-[I] loading tensors completed, taking 0.25s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 1.90s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  24/386 -   1.08GB/s
  |##################################################| 386/386 -   9.96GB/s
+[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.42s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |#####                                             |  37/418 - 261.96MB/s
  |###########################                       | 224/418 -   8.03GB/s
  |##################################################| 418/418 -   8.37GB/s
-[I] loading tensors completed, taking 0.43s (read: 0.04s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.29s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   4.97s/it - ETA 00:34
  |============>                                     | 2/8 -   5.05s/it - ETA 00:30
  |==================>                               | 3/8 -   4.88s/it - ETA 00:24
  |=========================>                        | 4/8 -   5.19s/it - ETA 00:20
  |===============================>                  | 5/8 -   4.44s/it - ETA 00:14
  |=====================================>            | 6/8 -   4.55s/it - ETA 00:09
  |===========================================>      | 7/8 -   4.81s/it - ETA 00:04
  |==================================================| 8/8 -   4.48s/it
-[I] sampling completed, taking 38.38s --- image.cpp:953
-[I] generating 1 latent images completed, taking 38.38s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |#####                                             |  36/418 - 123.66MB/s
  |###########################                       | 222/418 -   7.69GB/s
  |##################################################| 418/418 -   8.51GB/s
+[I] loading tensors completed, taking 0.42s (read: 0.04s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.29s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   5.52s/it - ETA 00:38
  |============>                                     | 2/8 -   5.37s/it - ETA 00:32
  |==================>                               | 3/8 -   4.91s/it - ETA 00:26
  |=========================>                        | 4/8 -   5.03s/it - ETA 00:20
  |===============================>                  | 5/8 -   5.35s/it - ETA 00:15
  |=====================================>            | 6/8 -   5.84s/it - ETA 00:10
  |===========================================>      | 7/8 -   6.16s/it - ETA 00:05
  |==================================================| 8/8 -   5.45s/it
+[I] sampling completed, taking 43.63s --- image.cpp:953
+[I] generating 1 latent images completed, taking 43.63s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  53/138 - 465.34MB/s
  |##################################################| 138/138 - 935.65MB/s
+
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 9.95s --- image.cpp:671
-[I] decode_first_stage completed, taking 9.95s --- image.cpp:675
-[I] generate_image completed in 50.24s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.38s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.38s --- image.cpp:675
+[I] generate_image completed in 54.44s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_K-ffn13+AMXw-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=8.11 GiB WALL=55.4s

results/logs/amx_lto-Q4_K-ffn13-r1.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -  10.05GB/s
-[I] loading tensors completed, taking 0.22s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -  10.51GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.11s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
 [I] get_learned_condition completed, taking 0.41s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
 [I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |#####                                             |  37/418 -  52.39MB/s
  |##################################################| 418/418 -  15.58GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.20s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.07s/it - ETA 00:42
  |============>                                     | 2/8 -   5.77s/it - ETA 00:35
  |==================>                               | 3/8 -   5.85s/it - ETA 00:29
  |=========================>                        | 4/8 -   5.70s/it - ETA 00:23
  |===============================>                  | 5/8 -   5.58s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.60s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.70s/it - ETA 00:05
  |==================================================| 8/8 -   5.69s/it
-[I] sampling completed, taking 45.98s --- image.cpp:953
-[I] generating 1 latent images completed, taking 45.98s --- image.cpp:965
+
  |#####                                             |  37/418 - 130.98MB/s
  |##################################################| 418/418 -  17.82GB/s
+[I] loading tensors completed, taking 0.20s (read: 0.17s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   6.26s/it - ETA 00:43
  |============>                                     | 2/8 -   5.88s/it - ETA 00:36
  |==================>                               | 3/8 -   5.84s/it - ETA 00:29
  |=========================>                        | 4/8 -   5.75s/it - ETA 00:23
  |===============================>                  | 5/8 -   5.64s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.49s/it - ETA 00:11
  |===========================================>      | 7/8 -   6.08s/it - ETA 00:05
  |==================================================| 8/8 -   5.66s/it
+[I] sampling completed, taking 46.61s --- image.cpp:953
+[I] generating 1 latent images completed, taking 46.61s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 935.65MB/s
+[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.22s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.23s --- image.cpp:675
-[I] generate_image completed in 56.62s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.00s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.00s --- image.cpp:675
+[I] generate_image completed in 57.03s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_K-ffn13-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=57.9s

results/logs/amx_lto-Q4_K-ffn13-r2.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  20/386 -  15.62MB/s
  |##################################################| 386/386 -   9.87GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.42s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |######                                            |  42/386 -   3.25GB/s
  |##################################################| 386/386 -  10.61GB/s
+[I] loading tensors completed, taking 0.21s (read: 0.11s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.41s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
 [I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |#####                                             |  37/418 - 130.98MB/s
  |##################################################| 418/418 -  14.22GB/s
-[I] loading tensors completed, taking 0.25s (read: 0.20s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.06s/it - ETA 00:42
  |============>                                     | 2/8 -   5.58s/it - ETA 00:34
  |==================>                               | 3/8 -   5.58s/it - ETA 00:28
  |=========================>                        | 4/8 -   6.03s/it - ETA 00:23
  |===============================>                  | 5/8 -   5.66s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.70s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.66s/it - ETA 00:05
  |==================================================| 8/8 -   5.61s/it
-[I] sampling completed, taking 45.89s --- image.cpp:953
-[I] generating 1 latent images completed, taking 45.89s --- image.cpp:965
+
  |#####                                             |  37/418 - 261.96MB/s
  |##################################################| 418/418 -  17.82GB/s
+[I] loading tensors completed, taking 0.20s (read: 0.17s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   5.83s/it - ETA 00:40
  |============>                                     | 2/8 -   5.61s/it - ETA 00:34
  |==================>                               | 3/8 -   5.73s/it - ETA 00:28
  |=========================>                        | 4/8 -   5.87s/it - ETA 00:23
  |===============================>                  | 5/8 -   5.86s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.79s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.71s/it - ETA 00:05
  |==================================================| 8/8 -   5.97s/it
+[I] sampling completed, taking 46.38s --- image.cpp:953
+[I] generating 1 latent images completed, taking 46.38s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
+
  |####################                              |  53/138 - 170.17MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 9.92s --- image.cpp:671
-[I] decode_first_stage completed, taking 9.92s --- image.cpp:675
-[I] generate_image completed in 56.23s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.44s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.44s --- image.cpp:675
+[I] generate_image completed in 57.24s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_K-ffn13-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=58.2s

results/logs/amx_lto-Q4_K-r1.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |#######                                           |  54/386 -   5.76GB/s
  |##################################################| 386/386 -  10.51GB/s
-[I] loading tensors completed, taking 0.22s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.47s --- image.cpp:580
+
  |#######                                           |  50/386 -   5.04GB/s
  |##################################################| 386/386 -   9.46GB/s
+[I] loading tensors completed, taking 0.24s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.43s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  33/452 - 115.84MB/s
  |##################################################| 452/452 -  17.30GB/s
-[I] loading tensors completed, taking 0.21s (read: 0.19s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.25s/it - ETA 00:43
  |============>                                     | 2/8 -   6.24s/it - ETA 00:37
  |==================>                               | 3/8 -   6.20s/it - ETA 00:31
  |=========================>                        | 4/8 -   6.04s/it - ETA 00:24
  |===============================>                  | 5/8 -   6.14s/it - ETA 00:18
  |=====================================>            | 6/8 -   6.06s/it - ETA 00:12
  |===========================================>      | 7/8 -   6.24s/it - ETA 00:06
  |==================================================| 8/8 -   6.13s/it
-[I] sampling completed, taking 49.29s --- image.cpp:953
-[I] generating 1 latent images completed, taking 49.29s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  16.58GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.20s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   6.12s/it - ETA 00:42
  |============>                                     | 2/8 -   6.06s/it - ETA 00:36
  |==================>                               | 3/8 -   5.87s/it - ETA 00:30
  |=========================>                        | 4/8 -   6.10s/it - ETA 00:24
  |===============================>                  | 5/8 -   6.02s/it - ETA 00:18
  |=====================================>            | 6/8 -   6.91s/it - ETA 00:12
  |===========================================>      | 7/8 -   7.30s/it - ETA 00:06
  |==================================================| 8/8 -   6.11s/it
+[I] sampling completed, taking 50.50s --- image.cpp:953
+[I] generating 1 latent images completed, taking 50.50s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###################                               |  52/138 - 340.34MB/s
  |##################################################| 138/138 - 940.30MB/s
+
  |####################                              |  54/138 - 232.92MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.42s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.42s --- image.cpp:675
-[I] generate_image completed in 60.19s --- image.cpp:1140
+[I] latent 1 decoded, taking 11.28s --- image.cpp:671
+[I] decode_first_stage completed, taking 11.28s --- image.cpp:675
+[I] generate_image completed in 62.21s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_K-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=63.3s

results/logs/amx_lto-Q4_K-r2.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -   8.60GB/s
-[I] loading tensors completed, taking 0.26s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.50s --- image.cpp:580
+
  |###                                               |  20/386 -  10.42MB/s
  |##################################################| 386/386 -   9.46GB/s
+[I] loading tensors completed, taking 0.24s (read: 0.14s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.75s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
 [I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 -  33.17MB/s
  |##################################################| 452/452 -  15.58GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.21s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.40s/it - ETA 00:44
  |============>                                     | 2/8 -   6.20s/it - ETA 00:37
  |==================>                               | 3/8 -   6.01s/it - ETA 00:31
  |=========================>                        | 4/8 -   5.87s/it - ETA 00:24
  |===============================>                  | 5/8 -   5.61s/it - ETA 00:18
  |=====================================>            | 6/8 -   5.60s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.73s/it - ETA 00:05
  |==================================================| 8/8 -   5.60s/it
-[I] sampling completed, taking 47.03s --- image.cpp:953
-[I] generating 1 latent images completed, taking 47.03s --- image.cpp:965
+
  |####                                              |  34/452 -  77.39MB/s
  |################################################  | 430/452 -  14.75GB/s
  |##################################################| 452/452 -   8.89GB/s
+[I] loading tensors completed, taking 0.41s (read: 0.23s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   7.70s/it - ETA 00:53
  |============>                                     | 2/8 -   7.18s/it - ETA 00:44
  |==================>                               | 3/8 -   6.77s/it - ETA 00:36
  |=========================>                        | 4/8 -   7.29s/it - ETA 00:28
  |===============================>                  | 5/8 -   7.60s/it - ETA 00:21
  |=====================================>            | 6/8 -   6.09s/it - ETA 00:14
  |===========================================>      | 7/8 -   5.82s/it - ETA 00:06
  |==================================================| 8/8 -   5.90s/it
+[I] sampling completed, taking 54.36s --- image.cpp:953
+[I] generating 1 latent images completed, taking 54.36s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
+
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.00s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.00s --- image.cpp:675
-[I] generate_image completed in 57.54s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.13s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.13s --- image.cpp:675
+[I] generate_image completed in 65.26s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q4_K-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=66.2s

results/logs/amx_lto-Q8_0+AMXw-r1.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  31.25MB/s
  |##################################################| 386/386 -   9.58GB/s
-[I] loading tensors completed, taking 0.24s (read: 0.11s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.44s --- image.cpp:580
+
  |###                                               |  23/386 - 348.63MB/s
  |##################################################| 386/386 -   9.83GB/s
+[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.48s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 232.18MB/s
  |################                                  | 144/452 -   7.17GB/s
  |###############################                   | 279/452 -   8.17GB/s
  |##############################################    | 413/452 -   8.57GB/s
  |##################################################| 452/452 -   7.64GB/s
-[I] loading tensors completed, taking 0.80s (read: 0.09s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.53s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.32s/it - ETA 00:44
  |============>                                     | 2/8 -   5.94s/it - ETA 00:36
  |==================>                               | 3/8 -   5.67s/it - ETA 00:29
  |=========================>                        | 4/8 -   5.57s/it - ETA 00:23
  |===============================>                  | 5/8 -   6.16s/it - ETA 00:17
  |=====================================>            | 6/8 -   6.14s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.79s/it - ETA 00:05
  |==================================================| 8/8 -   5.56s/it
-[I] sampling completed, taking 47.17s --- image.cpp:953
-[I] generating 1 latent images completed, taking 47.17s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  34/452 - 232.18MB/s
  |#################                                 | 153/452 -   7.61GB/s
  |################################                  | 288/452 -   8.59GB/s
  |################################################# | 435/452 -   9.08GB/s
  |##################################################| 452/452 -   7.64GB/s
+[I] loading tensors completed, taking 0.80s (read: 0.09s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.49s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   6.93s/it - ETA 00:48
  |============>                                     | 2/8 -   9.07s/it - ETA 00:47
  |==================>                               | 3/8 -   9.59s/it - ETA 00:42
  |=========================>                        | 4/8 -   9.83s/it - ETA 00:35
  |===============================>                  | 5/8 -   8.50s/it - ETA 00:26
  |=====================================>            | 6/8 -   7.59s/it - ETA 00:17
  |===========================================>      | 7/8 -   5.94s/it - ETA 00:08
  |==================================================| 8/8 -   5.56s/it
+[I] sampling completed, taking 63.02s --- image.cpp:953
+[I] generating 1 latent images completed, taking 63.02s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
+
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 9.85s --- image.cpp:671
-[I] decode_first_stage completed, taking 9.85s --- image.cpp:675
-[I] generate_image completed in 57.46s --- image.cpp:1140
+[I] latent 1 decoded, taking 9.95s --- image.cpp:671
+[I] decode_first_stage completed, taking 9.95s --- image.cpp:675
+[I] generate_image completed in 73.46s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q8_0+AMXw-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=11.11 GiB WALL=74.6s

results/logs/amx_lto-Q8_0+AMXw-r2.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  20/386 -  15.62MB/s
  |##################################################| 386/386 -  10.82GB/s
-[I] loading tensors completed, taking 0.21s (read: 0.11s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.39s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |###                                               |  20/386 -  15.62MB/s
  |##################################################| 386/386 -  10.23GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.42s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 232.18MB/s
  |##################                                | 156/452 -   7.95GB/s
  |################################                  | 288/452 -   8.59GB/s
  |###############################################   | 422/452 -   8.78GB/s
  |##################################################| 452/452 -   7.64GB/s
-[I] loading tensors completed, taking 0.80s (read: 0.09s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.51s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.50s/it - ETA 00:45
  |============>                                     | 2/8 -   5.87s/it - ETA 00:37
  |==================>                               | 3/8 -   6.01s/it - ETA 00:30
  |=========================>                        | 4/8 -   5.68s/it - ETA 00:24
  |===============================>                  | 5/8 -   5.73s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.58s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.43s/it - ETA 00:05
  |==================================================| 8/8 -   5.42s/it
-[I] sampling completed, taking 46.24s --- image.cpp:953
-[I] generating 1 latent images completed, taking 46.24s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  34/452 - 116.09MB/s
  |#################                                 | 145/452 -   7.35GB/s
  |###############################                   | 274/452 -   7.99GB/s
  |################################################  | 430/452 -   8.89GB/s
  |##################################################| 452/452 -   7.61GB/s
+[I] loading tensors completed, taking 0.81s (read: 0.09s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.49s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   6.71s/it - ETA 00:46
  |============>                                     | 2/8 -   5.97s/it - ETA 00:38
  |==================>                               | 3/8 -   6.59s/it - ETA 00:32
  |=========================>                        | 4/8 -   5.63s/it - ETA 00:24
  |===============================>                  | 5/8 -   6.69s/it - ETA 00:18
  |=====================================>            | 6/8 -   5.89s/it - ETA 00:12
  |===========================================>      | 7/8 -   5.72s/it - ETA 00:06
  |==================================================| 8/8 -   5.62s/it
+[I] sampling completed, taking 48.82s --- image.cpp:953
+[I] generating 1 latent images completed, taking 48.82s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
+[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |####################                              |  55/138 - 232.92MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 9.91s --- image.cpp:671
-[I] decode_first_stage completed, taking 9.91s --- image.cpp:675
-[I] generate_image completed in 56.54s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.07s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.07s --- image.cpp:675
+[I] generate_image completed in 59.32s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q8_0+AMXw-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=11.11 GiB WALL=60.4s

results/logs/amx_lto-Q8_0-r1.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -  10.00GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.42s --- image.cpp:580
+
  |###                                               |  20/386 -  15.38MB/s
  |##################################################| 386/386 -  10.14GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.60s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 116.09MB/s
  |############################                      | 248/452 -  13.23GB/s
  |##################################################| 452/452 -  14.87GB/s
-[I] loading tensors completed, taking 0.41s (read: 0.39s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -  21.19s/it - ETA 02:28
  |============>                                     | 2/8 -  20.88s/it - ETA 02:06
  |==================>                               | 3/8 -  20.99s/it - ETA 01:45
  |=========================>                        | 4/8 -  20.49s/it - ETA 01:23
  |===============================>                  | 5/8 -  20.11s/it - ETA 01:02
  |=====================================>            | 6/8 -  20.37s/it - ETA 00:41
  |===========================================>      | 7/8 -  20.67s/it - ETA 00:20
  |==================================================| 8/8 -  20.25s/it
-[I] sampling completed, taking 164.95s --- image.cpp:953
-[I] generating 1 latent images completed, taking 164.95s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  34/452 - 116.09MB/s
  |################################                  | 288/452 -  15.83GB/s
  |##################################################| 452/452 -  14.87GB/s
+[I] loading tensors completed, taking 0.41s (read: 0.35s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -  20.99s/it - ETA 02:26
  |============>                                     | 2/8 -  20.33s/it - ETA 02:03
  |==================>                               | 3/8 -  20.04s/it - ETA 01:42
  |=========================>                        | 4/8 -  20.12s/it - ETA 01:21
  |===============================>                  | 5/8 -  20.52s/it - ETA 01:01
  |=====================================>            | 6/8 -  20.40s/it - ETA 00:40
  |===========================================>      | 7/8 -  20.04s/it - ETA 00:20
  |==================================================| 8/8 -  19.77s/it
+[I] sampling completed, taking 162.21s --- image.cpp:953
+[I] generating 1 latent images completed, taking 162.21s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |################                                  |  44/138 - 230.83MB/s
  |##################################################| 138/138 - 875.00MB/s
-[I] loading tensors completed, taking 0.22s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.19s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.19s --- image.cpp:675
-[I] generate_image completed in 175.58s --- image.cpp:1140
+
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 931.04MB/s
+[I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] latent 1 decoded, taking 10.05s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.05s --- image.cpp:675
+[I] generate_image completed in 172.87s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q8_0-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=10.38 GiB WALL=173.9s

results/logs/amx_lto-Q8_0-r2.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |#####                                             |  32/386 -   3.88GB/s
  |##################################################| 386/386 -   9.83GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |###                                               |  20/386 -  15.62MB/s
  |##################################################| 386/386 -  10.05GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.14s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
 [I] get_learned_condition completed, taking 0.42s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
 [I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 -  58.04MB/s
  |################################                  | 286/452 -  16.19GB/s
  |##################################################| 452/452 -  15.16GB/s
-[I] loading tensors completed, taking 0.40s (read: 0.35s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -  20.85s/it - ETA 02:25
  |============>                                     | 2/8 -  20.21s/it - ETA 02:03
  |==================>                               | 3/8 -  20.20s/it - ETA 01:42
  |=========================>                        | 4/8 -  20.17s/it - ETA 01:21
  |===============================>                  | 5/8 -  19.87s/it - ETA 01:00
  |=====================================>            | 6/8 -  19.73s/it - ETA 00:40
  |===========================================>      | 7/8 -  20.01s/it - ETA 00:20
  |==================================================| 8/8 -  19.82s/it
-[I] sampling completed, taking 160.86s --- image.cpp:953
-[I] generating 1 latent images completed, taking 160.86s --- image.cpp:965
+
  |####                                              |  34/452 - 116.09MB/s
  |#################################                 | 292/452 -  16.17GB/s
  |##################################################| 452/452 -  14.94GB/s
+[I] loading tensors completed, taking 0.41s (read: 0.34s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -  20.87s/it - ETA 02:26
  |============>                                     | 2/8 -  20.12s/it - ETA 02:02
  |==================>                               | 3/8 -  19.96s/it - ETA 01:41
  |=========================>                        | 4/8 -  19.96s/it - ETA 01:20
  |===============================>                  | 5/8 -  19.82s/it - ETA 01:00
  |=====================================>            | 6/8 -  20.01s/it - ETA 00:40
  |===========================================>      | 7/8 -  20.06s/it - ETA 00:20
  |==================================================| 8/8 -  19.74s/it
+[I] sampling completed, taking 160.54s --- image.cpp:953
+[I] generating 1 latent images completed, taking 160.54s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
+[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
+
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.43s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.43s --- image.cpp:675
-[I] generate_image completed in 171.71s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.10s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.10s --- image.cpp:675
+[I] generate_image completed in 171.07s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/amx_lto-Q8_0-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=10.38 GiB WALL=172.1s

results/logs/avx2-Q4_K-r1.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  20/386 -  15.62MB/s
  |##################################################| 386/386 -  10.42GB/s
-[I] loading tensors completed, taking 0.22s (read: 0.11s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.76s --- image.cpp:580
+
  |###                                               |  22/386 -  15.62MB/s
  |##################################################| 386/386 -   9.83GB/s
+[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 1.79s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
 [I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 232.18MB/s
  |##################################################| 452/452 -  17.90GB/s
-[I] loading tensors completed, taking 0.20s (read: 0.17s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -  13.55s/it - ETA 01:34
  |============>                                     | 2/8 -  13.32s/it - ETA 01:20
  |==================>                               | 3/8 -  13.33s/it - ETA 01:06
  |=========================>                        | 4/8 -  13.10s/it - ETA 00:53
  |===============================>                  | 5/8 -  13.03s/it - ETA 00:39
  |=====================================>            | 6/8 -  13.94s/it - ETA 00:26
  |===========================================>      | 7/8 -  13.28s/it - ETA 00:13
  |==================================================| 8/8 -  13.13s/it
-[I] sampling completed, taking 106.68s --- image.cpp:953
-[I] generating 1 latent images completed, taking 106.68s --- image.cpp:965
+
  |####                                              |  34/452 - 232.18MB/s
  |##################################################| 452/452 -  16.74GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.20s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -  13.28s/it - ETA 01:32
  |============>                                     | 2/8 -  12.77s/it - ETA 01:18
  |==================>                               | 3/8 -  13.21s/it - ETA 01:05
  |=========================>                        | 4/8 -  13.94s/it - ETA 00:53
  |===============================>                  | 5/8 -  12.66s/it - ETA 00:39
  |=====================================>            | 6/8 -  12.45s/it - ETA 00:26
  |===========================================>      | 7/8 -  12.66s/it - ETA 00:12
  |==================================================| 8/8 -  12.73s/it
+[I] sampling completed, taking 103.71s --- image.cpp:953
+[I] generating 1 latent images completed, taking 103.71s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |#################                                 |  45/138 -  56.16MB/s
  |##################################################| 138/138 - 940.30MB/s
+
  |################                                  |  42/138 -  54.70MB/s
  |##################################################| 138/138 - 940.30MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.38s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.38s --- image.cpp:675
-[I] generate_image completed in 117.82s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.45s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.45s --- image.cpp:675
+[I] generate_image completed in 115.96s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/avx2-Q4_K-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=117.0s

results/logs/avx2-Q4_K-r2.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |#######                                           |  52/386 -   5.67GB/s
  |##################################################| 386/386 -  10.05GB/s
-[I] loading tensors completed, taking 0.22s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.65s --- image.cpp:580
+
  |#####                                             |  38/386 -   3.85GB/s
  |##################################################| 386/386 -  10.32GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 1.05s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  17.06GB/s
-[I] loading tensors completed, taking 0.21s (read: 0.19s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -  14.08s/it - ETA 01:38
  |============>                                     | 2/8 -  13.74s/it - ETA 01:23
  |==================>                               | 3/8 -  13.15s/it - ETA 01:08
  |=========================>                        | 4/8 -  13.21s/it - ETA 00:54
  |===============================>                  | 5/8 -  13.39s/it - ETA 00:40
  |=====================================>            | 6/8 -  13.55s/it - ETA 00:27
  |===========================================>      | 7/8 -  13.21s/it - ETA 00:13
  |==================================================| 8/8 -  13.50s/it
-[I] sampling completed, taking 107.83s --- image.cpp:953
-[I] generating 1 latent images completed, taking 107.83s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  16.66GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.20s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -  12.98s/it - ETA 01:30
  |============>                                     | 2/8 -  13.11s/it - ETA 01:18
  |==================>                               | 3/8 -  13.27s/it - ETA 01:05
  |=========================>                        | 4/8 -  13.45s/it - ETA 00:52
  |===============================>                  | 5/8 -  14.45s/it - ETA 00:40
  |=====================================>            | 6/8 -  14.71s/it - ETA 00:27
  |===========================================>      | 7/8 -  14.10s/it - ETA 00:13
  |==================================================| 8/8 -  13.63s/it
+[I] sampling completed, taking 109.71s --- image.cpp:953
+[I] generating 1 latent images completed, taking 109.71s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |################                                  |  43/138 -  55.19MB/s
  |##################################################| 138/138 - 935.65MB/s
+[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
+
  |#################                                 |  46/138 -  56.65MB/s
  |##################################################| 138/138 - 940.30MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.53s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.53s --- image.cpp:675
-[I] generate_image completed in 119.02s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.27s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.27s --- image.cpp:675
+[I] generate_image completed in 121.03s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/avx2-Q4_K-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=121.9s

results/logs/avx512-Q4_K-r1.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  23/386 -  13.67MB/s
  |##################################################| 386/386 -  10.05GB/s
-[I] loading tensors completed, taking 0.22s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.41s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  30/386 -   1.36GB/s
  |##################################################| 386/386 -  10.00GB/s
+[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.43s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
+[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
 
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  16.28GB/s
 [I] loading tensors completed, taking 0.22s (read: 0.20s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.13s/it - ETA 00:42
  |============>                                     | 2/8 -   5.96s/it - ETA 00:36
  |==================>                               | 3/8 -   6.06s/it - ETA 00:30
  |=========================>                        | 4/8 -   6.22s/it - ETA 00:24
  |===============================>                  | 5/8 -   6.28s/it - ETA 00:18
  |=====================================>            | 6/8 -   5.96s/it - ETA 00:12
  |===========================================>      | 7/8 -   6.21s/it - ETA 00:06
  |==================================================| 8/8 -   6.12s/it
-[I] sampling completed, taking 48.96s --- image.cpp:953
-[I] generating 1 latent images completed, taking 48.96s --- image.cpp:965
+
  |======>                                           | 1/8 -   6.04s/it - ETA 00:42
  |============>                                     | 2/8 -   5.80s/it - ETA 00:35
  |==================>                               | 3/8 -   5.90s/it - ETA 00:29
  |=========================>                        | 4/8 -   5.81s/it - ETA 00:23
  |===============================>                  | 5/8 -   5.85s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.84s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.86s/it - ETA 00:05
  |==================================================| 8/8 -   6.20s/it
+[I] sampling completed, taking 47.32s --- image.cpp:953
+[I] generating 1 latent images completed, taking 47.32s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |###################                               |  52/138 - 340.34MB/s
  |##################################################| 138/138 - 940.30MB/s
+[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.47s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.47s --- image.cpp:675
-[I] generate_image completed in 59.85s --- image.cpp:1140
+[I] latent 1 decoded, taking 9.96s --- image.cpp:671
+[I] decode_first_stage completed, taking 9.96s --- image.cpp:675
+[I] generate_image completed in 57.71s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/avx512-Q4_K-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=58.6s

results/logs/avx512-Q4_K-r2.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  20/386 -  15.62MB/s
  |##################################################| 386/386 -   9.34GB/s
-[I] loading tensors completed, taking 0.24s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.45s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |###                                               |  20/386 -  15.62MB/s
  |##################################################| 386/386 -  10.71GB/s
+[I] loading tensors completed, taking 0.21s (read: 0.11s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.40s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
-[I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  15.92GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.21s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.28s/it - ETA 00:43
  |============>                                     | 2/8 -   6.30s/it - ETA 00:37
  |==================>                               | 3/8 -   6.04s/it - ETA 00:31
  |=========================>                        | 4/8 -   6.27s/it - ETA 00:24
  |===============================>                  | 5/8 -   6.07s/it - ETA 00:18
  |=====================================>            | 6/8 -   6.25s/it - ETA 00:12
  |===========================================>      | 7/8 -   6.06s/it - ETA 00:06
  |==================================================| 8/8 -   5.97s/it
-[I] sampling completed, taking 49.25s --- image.cpp:953
-[I] generating 1 latent images completed, taking 49.25s --- image.cpp:965
+[I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
+
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  17.82GB/s
+[I] loading tensors completed, taking 0.20s (read: 0.18s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   5.89s/it - ETA 00:41
  |============>                                     | 2/8 -   6.11s/it - ETA 00:36
  |==================>                               | 3/8 -   6.15s/it - ETA 00:30
  |=========================>                        | 4/8 -   5.94s/it - ETA 00:24
  |===============================>                  | 5/8 -   5.80s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.79s/it - ETA 00:11
  |===========================================>      | 7/8 -   6.05s/it - ETA 00:05
  |==================================================| 8/8 -   6.11s/it
+[I] sampling completed, taking 47.87s --- image.cpp:953
+[I] generating 1 latent images completed, taking 47.87s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
-[I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  53/138 - 465.34MB/s
  |##################################################| 138/138 - 940.30MB/s
-[I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 10.52s --- image.cpp:671
-[I] decode_first_stage completed, taking 10.52s --- image.cpp:675
-[I] generate_image completed in 60.23s --- image.cpp:1140
+[I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
+
  |#############################################     | 124/138 -   2.93GB/s
  |##################################################| 138/138 - 904.31MB/s
+[I] loading tensors completed, taking 0.21s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] latent 1 decoded, taking 10.30s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.30s --- image.cpp:675
+[I] generate_image completed in 58.57s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/avx512-Q4_K-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=59.4s

results/logs/ov-int4-8step-r1.log

@@ -0,0 +1,18 @@
+Multiple distributions found for package optimum. Picked distribution: optimum
+/root/app/zimage/verify_work/venv-ov/lib/python3.12/site-packages/torch/cuda/__init__.py:1264: UserWarning: Can't initialize NVML
+  raw_cnt = _raw_device_count_nvml()
+[I] load completed in 6.61 s
+
+
  0%|          | 0/8 [00:00<?, ?it/s]
+
 12%|█▎        | 1/8 [00:05<00:40,  5.81s/it]
+
 25%|██▌       | 2/8 [00:09<00:26,  4.35s/it]
+
 38%|███▊      | 3/8 [00:12<00:20,  4.00s/it]
+
 50%|█████     | 4/8 [00:16<00:16,  4.03s/it]
+
 62%|██████▎   | 5/8 [00:20<00:12,  4.07s/it]
+
 75%|███████▌  | 6/8 [00:24<00:08,  4.06s/it]
+
 88%|████████▊ | 7/8 [00:29<00:04,  4.06s/it]
+
+[I] sampling completed, taking 35.45 s
+[I] generate_image completed in 36.13 s
+[I] saved /root/app/zimage/hf_upload/results/images/ov-int4-8step-r1.png
+EXIT=0 PEAK_RSS=12.54 GiB WALL=50.0s

results/logs/ov-int4-8step-r2.log

@@ -0,0 +1,18 @@
+Multiple distributions found for package optimum. Picked distribution: optimum
+/root/app/zimage/verify_work/venv-ov/lib/python3.12/site-packages/torch/cuda/__init__.py:1264: UserWarning: Can't initialize NVML
+  raw_cnt = _raw_device_count_nvml()
+[I] load completed in 6.93 s
+
+
  0%|          | 0/8 [00:00<?, ?it/s]
+
 12%|█▎        | 1/8 [00:06<00:47,  6.75s/it]
+
 25%|██▌       | 2/8 [00:10<00:31,  5.21s/it]
+
 38%|███▊      | 3/8 [00:15<00:23,  4.73s/it]
+
 50%|█████     | 4/8 [00:19<00:18,  4.54s/it]
+
 62%|██████▎   | 5/8 [00:23<00:13,  4.55s/it]
+
 75%|███████▌  | 6/8 [00:27<00:08,  4.40s/it]
+
 88%|████████▊ | 7/8 [00:31<00:04,  4.11s/it]
+
+[I] sampling completed, taking 38.11 s
+[I] generate_image completed in 38.84 s
+[I] saved /root/app/zimage/hf_upload/results/images/ov-int4-8step-r2.png
+EXIT=0 PEAK_RSS=12.63 GiB WALL=53.2s

results/logs/stock-Q4_K-r1.log

@@ -21,22 +21,23 @@
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
 [I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  23/386 - 361.82MB/s
  |##################################################| 386/386 -  10.05GB/s
-[I] loading tensors completed, taking 0.23s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.71s --- image.cpp:580
+
  |###                                               |  21/386 -  10.42MB/s
  |##################################################| 386/386 -  10.14GB/s
+[I] loading tensors completed, taking 0.22s (read: 0.13s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.54s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
 [I] z_image build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 232.18MB/s
  |##################################################| 452/452 -  16.74GB/s
+
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  16.14GB/s
 [I] loading tensors completed, taking 0.22s (read: 0.20s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   6.29s/it - ETA 00:44
  |============>                                     | 2/8 -   5.77s/it - ETA 00:36
  |==================>                               | 3/8 -   6.00s/it - ETA 00:30
  |=========================>                        | 4/8 -   5.84s/it - ETA 00:23
  |===============================>                  | 5/8 -   5.73s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.65s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.73s/it - ETA 00:05
  |==================================================| 8/8 -   5.65s/it
-[I] sampling completed, taking 46.67s --- image.cpp:953
-[I] generating 1 latent images completed, taking 46.67s --- image.cpp:965
+
  |======>                                           | 1/8 -   6.53s/it - ETA 00:45
  |============>                                     | 2/8 -   6.52s/it - ETA 00:39
  |==================>                               | 3/8 -   6.25s/it - ETA 00:32
  |=========================>                        | 4/8 -   5.88s/it - ETA 00:25
  |===============================>                  | 5/8 -   7.13s/it - ETA 00:19
  |=====================================>            | 6/8 -   6.32s/it - ETA 00:12
  |===========================================>      | 7/8 -   5.72s/it - ETA 00:06
  |==================================================| 8/8 -   5.87s/it
+[I] sampling completed, taking 50.23s --- image.cpp:953
+[I] generating 1 latent images completed, taking 50.23s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  53/138 - 465.34MB/s
  |##################################################| 138/138 - 921.96MB/s
-[I] loading tensors completed, taking 0.21s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 9.75s --- image.cpp:671
-[I] decode_first_stage completed, taking 9.75s --- image.cpp:675
-[I] generate_image completed in 57.15s --- image.cpp:1140
+
  |####################                              |  55/138 - 465.83MB/s
  |##################################################| 138/138 - 940.30MB/s
+[I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] latent 1 decoded, taking 9.86s --- image.cpp:671
+[I] decode_first_stage completed, taking 9.86s --- image.cpp:675
+[I] generate_image completed in 60.64s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/stock-Q4_K-r1.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=61.6s

results/logs/stock-Q4_K-r2.log

@@ -20,23 +20,24 @@
 [I] generate_image 512x512 --- image.cpp:869
 [I] get_sigmas with Simple scheduler --- denoiser.hpp:1157
 [I] sampling using Euler method --- request.cpp:477
-[I] qwen3 build cached graph cut plan done (taking 2 ms) --- ggml_graph_cut.cpp:1032
-
  |###                                               |  21/386 -  15.62MB/s
  |##################################################| 386/386 -  10.05GB/s
-[I] loading tensors completed, taking 0.22s (read: 0.12s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] get_learned_condition completed, taking 0.42s --- image.cpp:580
+[I] qwen3 build cached graph cut plan done (taking 1 ms) --- ggml_graph_cut.cpp:1032
+
  |###                                               |  21/386 -  31.25MB/s
  |##################################################| 386/386 -  10.71GB/s
+[I] loading tensors completed, taking 0.21s (read: 0.11s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+[I] get_learned_condition completed, taking 0.41s --- image.cpp:580
 [I] generating image: 1/1 - seed 42 --- image.cpp:920
 [I] z_image build cached graph cut plan done (taking 3 ms) --- ggml_graph_cut.cpp:1032
-
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  16.07GB/s
-[I] loading tensors completed, taking 0.22s (read: 0.21s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-
  |======>                                           | 1/8 -   5.96s/it - ETA 00:41
  |============>                                     | 2/8 -   5.71s/it - ETA 00:35
  |==================>                               | 3/8 -   5.86s/it - ETA 00:29
  |=========================>                        | 4/8 -   5.86s/it - ETA 00:23
  |===============================>                  | 5/8 -   5.74s/it - ETA 00:17
  |=====================================>            | 6/8 -   5.68s/it - ETA 00:11
  |===========================================>      | 7/8 -   5.63s/it - ETA 00:05
  |==================================================| 8/8 -   5.63s/it
-[I] sampling completed, taking 46.08s --- image.cpp:953
-[I] generating 1 latent images completed, taking 46.08s --- image.cpp:965
+
  |####                                              |  34/452 - 116.09MB/s
  |##################################################| 452/452 -  17.82GB/s
+[I] loading tensors completed, taking 0.20s (read: 0.17s, memcpy: 0.00s, convert: 0.00s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
+
  |======>                                           | 1/8 -   6.33s/it - ETA 00:44
  |============>                                     | 2/8 -   5.91s/it - ETA 00:36
  |==================>                               | 3/8 -   6.03s/it - ETA 00:30
  |=========================>                        | 4/8 -   6.09s/it - ETA 00:24
  |===============================>                  | 5/8 -   6.07s/it - ETA 00:18
  |=====================================>            | 6/8 -   5.78s/it - ETA 00:12
  |===========================================>      | 7/8 -   5.80s/it - ETA 00:06
  |==================================================| 8/8 -   6.12s/it
+[I] sampling completed, taking 48.14s --- image.cpp:953
+[I] generating 1 latent images completed, taking 48.14s --- image.cpp:965
 [I] decoding 1 latents --- image.cpp:605
 [I] vae build cached graph cut plan done (taking 0 ms) --- ggml_graph_cut.cpp:1032
-
  |####################                              |  54/138 - 465.83MB/s
  |##################################################| 138/138 - 935.65MB/s
+
  |####################                              |  54/138 - 232.92MB/s
  |##################################################| 138/138 - 935.65MB/s
 [I] loading tensors completed, taking 0.20s (read: 0.00s, memcpy: 0.00s, convert: 0.01s, copy_to_backend: 0.00s) --- model_loader.cpp:1407
-[I] latent 1 decoded, taking 9.87s --- image.cpp:671
-[I] decode_first_stage completed, taking 9.87s --- image.cpp:675
-[I] generate_image completed in 56.37s --- image.cpp:1140
+[I] latent 1 decoded, taking 10.28s --- image.cpp:671
+[I] decode_first_stage completed, taking 10.28s --- image.cpp:675
+[I] generate_image completed in 58.83s --- image.cpp:1140
 [I] save result image 0 to '/root/app/zimage/hf_upload/results/images/stock-Q4_K-r2.png' (success) --- main.cpp:495
 [I] 1/1 images saved --- main.cpp:571
+EXIT=0 PEAK_RSS=7.86 GiB WALL=59.7s

results/results.csv

@@ -1,21 +1,23 @@
-name,variant,model,env,rep,status,sampling_s,generate_s,wall_s,peak_rss_gib,diff_mean_vs_ref,diff_max_vs_ref,image
-stock-Q4_K,stock,Q4_K,-,1,ok,46.67,57.15,58.0,7.86,0.0000,0,images/stock-Q4_K-r1.png
-stock-Q4_K,stock,Q4_K,-,2,ok,46.08,56.37,57.2,7.86,0.0000,0,images/stock-Q4_K-r2.png
-avx2-Q4_K,avx2,Q4_K,-,1,ok,106.68,117.82,118.7,7.86,3.1796,241,images/avx2-Q4_K-r1.png
-avx2-Q4_K,avx2,Q4_K,-,2,ok,107.83,119.02,119.9,7.86,3.1796,241,images/avx2-Q4_K-r2.png
-avx512-Q4_K,avx512,Q4_K,-,1,ok,48.96,59.85,60.7,7.86,3.1690,239,images/avx512-Q4_K-r1.png
-avx512-Q4_K,avx512,Q4_K,-,2,ok,49.25,60.23,61.1,7.86,3.1690,239,images/avx512-Q4_K-r2.png
-amx_lto-Q4_K,amx_lto,Q4_K,-,1,ok,49.29,60.19,61.1,7.86,0.0000,0,images/amx_lto-Q4_K-r1.png
-amx_lto-Q4_K,amx_lto,Q4_K,-,2,ok,47.03,57.54,58.4,7.86,0.0000,0,images/amx_lto-Q4_K-r2.png
-amx_lto-Q4_K+AMXw,amx_lto,Q4_K,SD_CPU_EXTRA_BUFT=1,1,ok,39.32,50.27,51.2,8.24,1.2844,227,images/amx_lto-Q4_K+AMXw-r1.png
-amx_lto-Q4_K+AMXw,amx_lto,Q4_K,SD_CPU_EXTRA_BUFT=1,2,ok,38.98,49.37,50.3,8.20,1.2844,227,images/amx_lto-Q4_K+AMXw-r2.png
-amx_lto-Q4_K-ffn13,amx_lto,Q4_K-ffn13,-,1,ok,45.98,56.62,57.6,7.86,0.0000,0,images/amx_lto-Q4_K-ffn13-r1.png
-amx_lto-Q4_K-ffn13,amx_lto,Q4_K-ffn13,-,2,ok,45.89,56.23,57.1,7.86,0.0000,0,images/amx_lto-Q4_K-ffn13-r2.png
-amx_lto-Q4_K-ffn13+AMXw,amx_lto,Q4_K-ffn13,SD_CPU_EXTRA_BUFT=1,1,ok,38.15,48.81,49.7,8.19,1.2844,227,images/amx_lto-Q4_K-ffn13+AMXw-r1.png
-amx_lto-Q4_K-ffn13+AMXw,amx_lto,Q4_K-ffn13,SD_CPU_EXTRA_BUFT=1,2,ok,38.38,50.24,51.1,8.16,1.2844,227,images/amx_lto-Q4_K-ffn13+AMXw-r2.png
-amx_lto-Q4_0,amx_lto,Q4_0,-,1,ok,170.21,180.77,181.6,7.69,22.8399,252,images/amx_lto-Q4_0-r1.png
-amx_lto-Q4_0,amx_lto,Q4_0,-,2,ok,178.49,193.62,194.5,7.68,22.8399,252,images/amx_lto-Q4_0-r2.png
-amx_lto-Q8_0,amx_lto,Q8_0,-,1,ok,164.95,175.58,176.7,10.38,12.5065,249,images/amx_lto-Q8_0-r1.png
-amx_lto-Q8_0,amx_lto,Q8_0,-,2,ok,160.86,171.71,172.7,10.38,12.5065,249,images/amx_lto-Q8_0-r2.png
-amx_lto-Q8_0+AMXw,amx_lto,Q8_0,SD_CPU_EXTRA_BUFT=1,1,ok,47.17,57.46,58.6,11.11,12.3329,249,images/amx_lto-Q8_0+AMXw-r1.png
-amx_lto-Q8_0+AMXw,amx_lto,Q8_0,SD_CPU_EXTRA_BUFT=1,2,ok,46.24,56.54,57.7,11.11,12.3329,249,images/amx_lto-Q8_0+AMXw-r2.png
+name,variant,model,env,rep,status,sampling_s,generate_s,wall_s,peak_rss_gib,diff_mean_vs_ref,diff_max_vs_ref,image
+stock-Q4_K,stock,Q4_K,-,1,ok,50.23,60.64,61.6,7.86,0.0000,0,images/stock-Q4_K-r1.png
+stock-Q4_K,stock,Q4_K,-,2,ok,48.14,58.83,59.7,7.86,0.0000,0,images/stock-Q4_K-r2.png
+avx2-Q4_K,avx2,Q4_K,-,1,ok,103.71,115.96,117.0,7.86,3.1796,241,images/avx2-Q4_K-r1.png
+avx2-Q4_K,avx2,Q4_K,-,2,ok,109.71,121.03,121.9,7.86,3.1796,241,images/avx2-Q4_K-r2.png
+avx512-Q4_K,avx512,Q4_K,-,1,ok,47.32,57.71,58.6,7.86,3.1690,239,images/avx512-Q4_K-r1.png
+avx512-Q4_K,avx512,Q4_K,-,2,ok,47.87,58.57,59.4,7.86,3.1690,239,images/avx512-Q4_K-r2.png
+amx_lto-Q4_K,amx_lto,Q4_K,-,1,ok,50.50,62.21,63.3,7.86,0.0000,0,images/amx_lto-Q4_K-r1.png
+amx_lto-Q4_K,amx_lto,Q4_K,-,2,ok,54.36,65.26,66.2,7.86,0.0000,0,images/amx_lto-Q4_K-r2.png
+amx_lto-Q4_K+AMXw,amx_lto,Q4_K,SD_CPU_EXTRA_BUFT=1,1,ok,42.74,53.30,54.2,8.16,1.2844,227,images/amx_lto-Q4_K+AMXw-r1.png
+amx_lto-Q4_K+AMXw,amx_lto,Q4_K,SD_CPU_EXTRA_BUFT=1,2,ok,41.34,52.01,53.0,8.16,1.2844,227,images/amx_lto-Q4_K+AMXw-r2.png
+amx_lto-Q4_K-ffn13,amx_lto,Q4_K-ffn13,-,1,ok,46.61,57.03,57.9,7.86,0.0000,0,images/amx_lto-Q4_K-ffn13-r1.png
+amx_lto-Q4_K-ffn13,amx_lto,Q4_K-ffn13,-,2,ok,46.38,57.24,58.2,7.86,0.0000,0,images/amx_lto-Q4_K-ffn13-r2.png
+amx_lto-Q4_K-ffn13+AMXw,amx_lto,Q4_K-ffn13,SD_CPU_EXTRA_BUFT=1,1,ok,42.52,53.92,54.8,8.18,1.2844,227,images/amx_lto-Q4_K-ffn13+AMXw-r1.png
+amx_lto-Q4_K-ffn13+AMXw,amx_lto,Q4_K-ffn13,SD_CPU_EXTRA_BUFT=1,2,ok,43.63,54.44,55.4,8.11,1.2844,227,images/amx_lto-Q4_K-ffn13+AMXw-r2.png
+amx_lto-Q4_0,amx_lto,Q4_0,-,1,ok,171.95,182.90,183.8,7.68,22.8399,252,images/amx_lto-Q4_0-r1.png
+amx_lto-Q4_0,amx_lto,Q4_0,-,2,ok,173.02,184.15,185.0,7.69,22.8399,252,images/amx_lto-Q4_0-r2.png
+amx_lto-Q8_0,amx_lto,Q8_0,-,1,ok,162.21,172.87,173.9,10.38,12.5065,249,images/amx_lto-Q8_0-r1.png
+amx_lto-Q8_0,amx_lto,Q8_0,-,2,ok,160.54,171.07,172.1,10.38,12.5065,249,images/amx_lto-Q8_0-r2.png
+amx_lto-Q8_0+AMXw,amx_lto,Q8_0,SD_CPU_EXTRA_BUFT=1,1,ok,63.02,73.46,74.6,11.11,12.3329,249,images/amx_lto-Q8_0+AMXw-r1.png
+amx_lto-Q8_0+AMXw,amx_lto,Q8_0,SD_CPU_EXTRA_BUFT=1,2,ok,48.82,59.32,60.4,11.11,12.3329,249,images/amx_lto-Q8_0+AMXw-r2.png
+ov-int4-8step,openvino,OV-INT4,-,1,ok,35.45,36.13,50.0,12.54,27.7079,254,images/ov-int4-8step-r1.png
+ov-int4-8step,openvino,OV-INT4,-,2,ok,38.11,38.84,53.2,12.63,27.7079,254,images/ov-int4-8step-r2.png

results/results.md

@@ -1,14 +1,15 @@
 | # | 版本 (variant) | 模型 | 額外環境變數 | 採樣 (s) | 總生成 (s) | 相對 stock | 峰值 RSS (GiB) | 與 stock 像素差 (mean/max) |
 |---|---|---|---|---|---|---|---|---|
-| 1 | stock | Q4_K | - | 46.4 | 56.8 | 1.00x | 7.86 | ref |
-| 2 | avx2 | Q4_K | - | 107.3 | 118.4 | 0.48x | 7.86 | 3.18 / 241 |
-| 3 | avx512 | Q4_K | - | 49.1 | 60.0 | 0.95x | 7.86 | 3.17 / 239 |
-| 4 | amx_lto | Q4_K | - | 48.2 | 58.9 | 0.96x | 7.86 | 0.00 / 0 |
-| 5 | amx_lto | Q4_K | SD_CPU_EXTRA_BUFT=1 | 39.1 | 49.8 | 1.14x | 8.22 | 1.28 / 227 |
-| 6 | amx_lto | Q4_K-ffn13 | - | 45.9 | 56.4 | 1.01x | 7.86 | 0.00 / 0 |
-| 7 | amx_lto | Q4_K-ffn13 | SD_CPU_EXTRA_BUFT=1 | 38.3 | 49.5 | 1.15x | 8.18 | 1.28 / 227 |
-| 8 | amx_lto | Q4_0 | - | 174.4 | 187.2 | 0.30x | 7.69 | 22.84 / 252 |
-| 9 | amx_lto | Q8_0 | - | 162.9 | 173.6 | 0.33x | 10.38 | 12.51 / 249 |
-| 10 | amx_lto | Q8_0 | SD_CPU_EXTRA_BUFT=1 | 46.7 | 57.0 | 1.00x | 11.11 | 12.33 / 249 |
+| 1 | stock | Q4_K | - | 49.2 | 59.7 | 1.00x | 7.86 | ref |
+| 2 | avx2 | Q4_K | - | 106.7 | 118.5 | 0.50x | 7.86 | 3.18 / 241 |
+| 3 | avx512 | Q4_K | - | 47.6 | 58.1 | 1.03x | 7.86 | 3.17 / 239 |
+| 4 | amx_lto | Q4_K | - | 52.4 | 63.7 | 0.94x | 7.86 | 0.00 / 0 |
+| 5 | amx_lto | Q4_K | SD_CPU_EXTRA_BUFT=1 | 42.0 | 52.7 | 1.13x | 8.16 | 1.28 / 227 |
+| 6 | amx_lto | Q4_K-ffn13 | - | 46.5 | 57.1 | 1.05x | 7.86 | 0.00 / 0 |
+| 7 | amx_lto | Q4_K-ffn13 | SD_CPU_EXTRA_BUFT=1 | 43.1 | 54.2 | 1.10x | 8.14 | 1.28 / 227 |
+| 8 | amx_lto | Q4_0 | - | 172.5 | 183.5 | 0.33x | 7.69 | 22.84 / 252 |
+| 9 | amx_lto | Q8_0 | - | 161.4 | 172.0 | 0.35x | 10.38 | 12.51 / 249 |
+| 10 | amx_lto | Q8_0 | SD_CPU_EXTRA_BUFT=1 | 55.9 | 66.4 | 0.90x | 11.11 | 12.33 / 249 |
+| 11 | openvino | OV-INT4 | - | 36.8 | 37.5 | 1.59x | 12.59 | 27.71 / 254 |
 
 n = 2 run(s) per row (median). Speedup = stock-Q4_K generate time / row generate time.

tools/ov_bench.py

@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+"""Generate one image with the OpenVINO INT4 Z-Image-Turbo pipeline (optimum-intel) and print
+timing lines in the same format as sd-cli, so SmartTest.sh can parse both.
+
+usage: ov_bench.py --model-dir <dir> -p <prompt> -o <out.png> [-W 512] [-H 512] [--steps 8] [--seed 42]
+
+Printed lines (stderr):
+  [I] load completed in <s>
+  [I] sampling completed, taking <s> s   (sum of per-step wall times, from callback_on_step_end)
+  [I] generate_image completed in <s> s  (full pipeline call: text encode + denoise + VAE decode)
+"""
+import argparse
+import sys
+import time
+
+import torch
+from optimum.intel import OVDiffusionPipeline
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--model-dir", required=True)
+    ap.add_argument("-p", "--prompt", required=True)
+    ap.add_argument("-o", "--output", required=True)
+    ap.add_argument("-W", "--width", type=int, default=512)
+    ap.add_argument("-H", "--height", type=int, default=512)
+    ap.add_argument("--steps", type=int, default=8)
+    ap.add_argument("--seed", type=int, default=42)
+    args = ap.parse_args()
+
+    t0 = time.time()
+    pipe = OVDiffusionPipeline.from_pretrained(args.model_dir, compile=True)
+    print(f"[I] load completed in {time.time() - t0:.2f} s", file=sys.stderr)
+
+    step_marks = []
+
+    def on_step_end(pipeline, step, timestep, callback_kwargs):
+        step_marks.append(time.time())
+        return callback_kwargs
+
+    gen = torch.Generator().manual_seed(args.seed)
+    t_gen = time.time()
+    image = pipe(
+        prompt=args.prompt,
+        height=args.height,
+        width=args.width,
+        num_inference_steps=args.steps,
+        guidance_scale=0.0,  # Turbo: no CFG, one transformer forward per step
+        generator=gen,
+        callback_on_step_end=on_step_end,
+    ).images[0]
+    t_end = time.time()
+
+    # sampling = time spent between pipeline start and the last denoising step
+    sampling = (step_marks[-1] - t_gen) if step_marks else float("nan")
+    print(f"[I] sampling completed, taking {sampling:.2f} s", file=sys.stderr)
+    print(f"[I] generate_image completed in {t_end - t_gen:.2f} s", file=sys.stderr)
+
+    image.save(args.output)
+    print(f"[I] saved {args.output}", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())