{
  "artifacts": [
    {
      "file": "siglip2_base_224_fp16.tflite",
      "sha256": "a30ebb7b3ee15eaa68a18f9ab6a2ed740c15c343d25d898dc482317473320854",
      "size_mb": 176.847
    }
  ],
  "benchmarks": [],
  "browser": {
    "backends": [
      {
        "backend": "wasm_xnnpack",
        "date": "2026-08-13",
        "env": {
          "browser": "chromium",
          "browser_version": "151.0.7922.34",
          "headless": true,
          "jspi": true,
          "litertjs_core_version": "2.5.3",
          "machine_label": "mac-studio-m4-max",
          "os": "macOS",
          "os_version": "27.0.0",
          "webgpu_adapter": {
            "architecture": "metal-3",
            "description": "",
            "device": "",
            "vendor": "apple"
          }
        },
        "full_delegation": null,
        "latency_p50_ms": 2124.707,
        "loads": true,
        "max_rel_diff": null,
        "output_match": null,
        "provenance": "measured",
        "runs": true
      },
      {
        "backend": "webgpu_mldrift",
        "date": "2026-08-13",
        "env": {
          "browser": "chromium",
          "browser_version": "151.0.7922.34",
          "headless": true,
          "jspi": true,
          "litertjs_core_version": "2.5.3",
          "machine_label": "mac-studio-m4-max",
          "os": "macOS",
          "os_version": "27.0.0",
          "webgpu_adapter": {
            "architecture": "metal-3",
            "description": "",
            "device": "",
            "vendor": "apple"
          }
        },
        "full_delegation": true,
        "latency_p50_ms": 19.635,
        "loads": true,
        "max_rel_diff": 0.0008081583214905942,
        "output_match": true,
        "provenance": "measured",
        "runs": true
      }
    ],
    "demo_url": null,
    "sweep_source": "data/sweep/2.5.3/2026-08-13/siglip2-base-patch16-224.json"
  },
  "conversion": {
    "command": "python scripts/convert_siglip2.py",
    "quantization": "fp16 (float_casting weight quantization; single-graph 185 MB file)",
    "tool": "litert-torch (litertlm-convert scripts/convert_siglip2.py; timm image tower)",
    "tool_version": "0.10.0 (editable dev checkout 115a136 + local patches)"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu_mldrift",
          "context_length": null,
          "date": "2026-08-26",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-npubench",
            "os_build": "Android 16",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "npubench sweep row: N=50 median, 1 backend = 1 process, accepted only with thermal NONE->NONE; NPU rows additionally required qnn_partition delegate evidence in logcat (a stock model asked for on the NPU can silently land on XNNPACK and report a CPU number)",
            "mode=jit thermal=NONE->NONE headroom=0.6066667->0.6066667 attempt=0"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 13.704,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "iterations": 50,
            "latency_max_ms": 18.062,
            "latency_min_ms": 13.427,
            "load_ms": 1872.4
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-08-26/siglip2-base-patch16-224__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "npu_qnn",
          "context_length": null,
          "date": "2026-08-26",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-npubench",
            "os_build": "Android 16",
            "runtime": "litert",
            "runtime_version": "2.2.0",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": "QAIRT (Hexagon, JIT on-device)"
          },
          "error": null,
          "evidence": [
            "npubench sweep row: N=50 median, 1 backend = 1 process, accepted only with thermal NONE->NONE; NPU rows additionally required qnn_partition delegate evidence in logcat (a stock model asked for on the NPU can silently land on XNNPACK and report a CPU number)",
            "mode=jit thermal=NONE->NONE headroom=0.59000003->0.59000003 attempt=0"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": 6.943,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "iterations": 50,
            "jit_first_load_ms": 6454.4,
            "latency_max_ms": 7.14,
            "latency_min_ms": 6.88,
            "load_ms": 183.9,
            "qnn_partition_lines": 57
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/2.2.0/2026-08-26/siglip2-base-patch16-224__galaxy-s26.json"
      }
    ]
  },
  "model": {
    "family": "siglip",
    "id": "siglip2-base-patch16-224",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/SigLIP2-base-patch16-224",
    "task": "image-feature-extraction"
  },
  "pitfalls": [
    "Input is NCHW [1,3,224,224] normalized to [-1,1] ((x/255-0.5)/0.5) BY THE CALLER; output [1,768] is already L2-normalized (HF card).",
    "Re-authoring (script docstring): fused qkv 5-D head-split decomposed to separate q/k/v with 4-D SDPA; the attention-pool's const-latent batch-matmul expressed as broadcast-multiply + reduce-sum (const@non-const BMM is rejected/mis-computed); overflow-safe LayerNorm because the delegate reduces variance in fp16 even for an fp32 graph (deep-ViT activations overflow 65504).",
    "Full GPU residency on Pixel 8a: 809/809 nodes LITERT_CL, no Flex ops, GPU output corr ~1.0 vs PyTorch (HF card).",
    "Zero-shot classification needs the TEXT tower host-side: open_clip ViT-B-16-SigLIP2 embeddings (prompt 'This is a photo of {label}.'), dot product on device (HF card)."
  ],
  "schema_version": "1.2"
}