{ "artifacts": [ { "file": "sam2_tiny_mask_decoder_v2_fp16.tflite", "sha256": "80d668b19156c548f31a8c8eb3cc1da2127de36e24e96ef923467a1e54c99e76", "size_mb": 16.182 } ], "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": 398.045, "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": 7.908, "loads": true, "max_rel_diff": 0.026880677356045733, "output_match": false, "provenance": "measured", "runs": true } ], "demo_url": null, "sweep_source": "data/sweep/2.5.3/2026-08-13/sam2.1-hiera-tiny-mask-decoder__sam2_tiny_mask_decoder_v2_fp16.json" }, "conversion": { "command": "python conversion/convert_sam2_decoder.py", "quantization": "fp16 (float_casting weight quantization; from the artifact name and script)", "tool": "litert-torch (litert-samples interactive_segmentation conversion/convert_sam2_decoder.py)", "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=0.79666674->0.79666674 attempt=0" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": 10.548, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "iterations": 50, "latency_max_ms": 10.893, "latency_min_ms": 10.082, "load_ms": 1176.8 }, "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/sam2.1-hiera-tiny-mask-decoder__sam2_tiny_mask_decoder_v2_fp16__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=0.8033333->0.8 attempt=0" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": 8.045, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "iterations": 50, "jit_first_load_ms": 4352.9, "latency_max_ms": 8.227, "latency_min_ms": 7.963, "load_ms": 110.1, "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/sam2.1-hiera-tiny-mask-decoder__sam2_tiny_mask_decoder_v2_fp16__galaxy-s26.json" }, { "device": "pixel-8a", "run": { "accelerator": "cpu_xnnpack", "context_length": null, "date": "2026-08-28", "decode_tokens_per_s": null, "delegated_ops": 423, "env": { "device": "Pixel 8a", "machine_label": "pixel-8a-npubench", "os_build": "Android 16", "runtime": "litert", "runtime_version": "2.2.0", "soc": "Tensor G3", "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=0.6089151->0.61605835 attempt=0", "logcat: Replacing 423 out of 425 node(s) with delegate (TfLiteXNNPackDelegate) (largest subgraph captured by the driver)" ], "failure_class": null, "full_delegation": false, "latency_p50_ms": 222.813, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "iterations": 50, "latency_max_ms": 226.854, "latency_min_ms": 216.124, "load_ms": 47.5 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "total_ops": 425, "ttft_ms": null }, "source": "data/device_runs/2.2.0/2026-08-28/sam2.1-hiera-tiny-mask-decoder__sam2_tiny_mask_decoder_v2_fp16__pixel-8a.json" }, { "device": "pixel-8a", "run": { "accelerator": "gpu_mldrift", "context_length": null, "date": "2026-08-28", "decode_tokens_per_s": null, "delegated_ops": 425, "env": { "device": "Pixel 8a", "machine_label": "pixel-8a-npubench", "os_build": "Android 16", "runtime": "litert", "runtime_version": "2.2.0", "soc": "Tensor G3", "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=0.60405505->0.60405505 attempt=0", "logcat: Replacing 425 out of 425 node(s) with delegate (LITERT_CL) (largest subgraph captured by the driver)" ], "failure_class": null, "full_delegation": true, "latency_p50_ms": 30.875, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "iterations": 50, "latency_max_ms": 35.57, "latency_min_ms": 29.505, "load_ms": 1679.1 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "total_ops": 425, "ttft_ms": null }, "source": "data/device_runs/2.2.0/2026-08-28/sam2.1-hiera-tiny-mask-decoder__sam2_tiny_mask_decoder_v2_fp16__pixel-8a.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu_xnnpack", "context_length": null, "date": "2026-08-31", "decode_tokens_per_s": null, "delegated_ops": null, "env": { "device": "Raspberry Pi 5 Model B Rev 1.1", "machine_label": "raspberry-pi-5", "os_build": "Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41", "runtime": "litert", "runtime_version": "2.2.0.dev20260804", "soc": null, "vendor_sdk": null }, "error": null, "evidence": [ "pi5 sweep row: LiteRT benchmark_model, CPU/XNNPACK at --num_threads=4, 3 invocations per file of 10 warm-up + 50 timed runs (the tool caps a phase at 150 s, so very slow graphs run fewer); latency = median of the three per-invocation medians over the timed phase; a row counts as measured only when every invocation exited 0 with XNNPACK engaged and vcgencmd get_throttled 0x0 before and after", "versions: ai-edge-litert-nightly=2.2.0.dev20260804, benchmark_model_sha256=babc9275addd8612caf0842294cb5b3f301c3cce46e6c5979cd2803f215f6bee, cpu=Raspberry Pi 5 Model B Rev 1.1, litert-cli-nightly=0.2.0.dev20260805, platform=Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41, python=3.13.5", "invocation 0: exit=0 wall_s=9.7 temp 50.5->58.2C throttled=0x0 xnnpack=True median_us=158888.0 runs=50 footprint_peak_mb=135.48", "invocation 1: exit=0 wall_s=9.6 temp 51.0->59.8C throttled=0x0 xnnpack=True median_us=158270.0 runs=50 footprint_peak_mb=135.48", "invocation 2: exit=0 wall_s=9.6 temp 50.5->58.7C throttled=0x0 xnnpack=True median_us=158169.0 runs=50 footprint_peak_mb=135.48" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": 158.27, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "init_ms": 17.667, "invocations": 3, "iterations": 150, "latency_max_ms": 161.422, "latency_min_ms": 157.109, "threads": 4, "warmup_runs_per_invocation": 10 }, "output_match": null, "peak_mem_mb": 135.48, "prefill_tokens_per_s": null, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": null }, "source": "data/device_runs/2.2.0.dev20260804/2026-08-31/sam2.1-hiera-tiny-mask-decoder__sam2_tiny_mask_decoder_v2_fp16__raspberry-pi-5.json" } ] }, "model": { "family": "sam2", "id": "sam2.1-hiera-tiny-mask-decoder__sam2_tiny_mask_decoder_v2_fp16", "license": "apache-2.0", "source_url": "https://huggingface.co/litert-community/SAM2.1-Hiera-Tiny-Mask-Decoder", "task": "mask-generation" }, "pitfalls": [ "Why v2 exists: the v1 build's attention was written with the batch dim collapsed (q/k/v [heads,N,d], rank 3) — it compiles, fully delegates (358/358 LITERT_CL) and matches PyTorch on desktop, yet returns silently wrong masks on the Pixel 8a GPU (corr 0.265 vs CPU; not fp16 — fp32 GPU compute still 0.473). v2 keeps the leading batch dim ([1,heads,N,d], rank-4 SDPA): corr 0.9998 / binary-IoU 0.999 vs CPU and ~20% faster (6.8 vs 8.5 ms/tap) (HF card).", "The batchless miscompute is not Android-specific: on Mac Metal 2.1.6 v1 is rel 8.1e-02 vs v2 4.2e-07, and the browser sweep shows the same (v1 max_abs 4.81) — isolation probes cleared every op family individually, so it is a fusion/partition-level interaction on the batchless layout (rewrite-reach-survey 2026-08-13 addendum).", "Drop-in replacement: v2 inputs/outputs are identical to v1 (image_embeddings [1,256,64,64], sparse_prompt [1,2,256], feat_s1 [1,64,128,128], feat_s0 [1,32,256,256] -> masks [1,3,256,256] + iou [1,3]) at +45 RESHAPE +2 TRANSPOSE (HF card; survey addendum).", "Re-authoring in the script: ConvTranspose2d -> zero-stuff Conv2d; SafeLayerNorm (scale-before-square, fp16-overflow-safe); image positional embeddings + no-mask dense prompt baked constant; multimask path as a static slice (no argmax/gather) (convert_sam2_decoder.py docstring).", "The point prompt encoder runs host-side (sin/cos in Kotlin); pair with litert-community/SAM2.1-Hiera-Tiny-Image-Encoder run once per image (HF card)." ], "schema_version": "1.2" }