{ "artifacts": [ { "file": "granite-4.0-h-1b_int8.litertlm", "sha256": "d04280eded8a133cc053df460e131b7102098295f97d550c10796d7bd98f8cbf", "size_mb": 1553.615 } ], "benchmarks": [], "conversion": { "command": "git -C litert-torch-granite checkout 115a136 && git -C litert-torch-granite apply granite_hybrid_litert_torch.patch && PYTHONPATH=litert-torch-granite python convert_granite4h.py ibm-granite/granite-4.0-h-1b out_granite_1b", "quantization": "int8 dynamic on linears + embedding; convs and the selective scan stay float (ship commit ea409da; HF card recipe table)", "tool": "litert-torch, pinned base + granite hybrid patch (hf-to-litertlm granite_work/convert_granite4h.py; patch adds export-cache layers for linear_attention, state-continuation tracing, prefill-pad guard)", "tool_version": "editable checkout 115a136 + granite_hybrid_litert_torch.patch (hf-to-litertlm granite_work/README.md)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "galaxy-s26", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-05", "decode_tokens_per_s": 29.39, "delegated_ops": 4729, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-cold-cache-cooled", "os_build": "Android 16", "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 4034 out of 4346 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 384 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 4034 out of 4346 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 384 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 3998 out of 4310 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 384 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 4176 out of 4488 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 384 partitions for subgraph 3 (prefill_128).", "results block: prefill=178.72 decode=29.39 tokens/s" ], "failure_class": null, "full_delegation": false, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 240.0, "init_s": 2.97487, "prefill_tokens": 224.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 178.72, "provenance": "measured", "runs": true, "total_ops": 4962, "ttft_ms": 1290.0 }, "source": "data/device_runs/0.16.0/2026-09-05/granite-4.0-h-1b__galaxy-s26.json" }, { "device": "galaxy-s26", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-24", "decode_tokens_per_s": 24.06, "delegated_ops": 4488, "env": { "device": "Galaxy S26 (SM-S942Q)", "machine_label": "galaxy-s26-cl-pinned", "os_build": "Android 16", "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Qualcomm SM8850", "vendor_sdk": null }, "error": null, "evidence": [ "VERBOSE: Replacing 4346 out of 4346 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).", "VERBOSE: Replacing 4346 out of 4346 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).", "VERBOSE: Replacing 4310 out of 4310 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).", "VERBOSE: Replacing 4488 out of 4488 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).", "results block: prefill=232.64 decode=24.06 tokens/s" ], "failure_class": null, "full_delegation": true, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 285.0, "init_s": 24.58452, "prefill_tokens": 225.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 232.64, "provenance": "measured", "runs": true, "total_ops": 4488, "ttft_ms": 1010.0 }, "source": "data/device_runs/0.16.0/2026-08-24/granite-4.0-h-1b__galaxy-s26.json" }, { "device": "mac-studio-m4-max", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-08-12", "decode_tokens_per_s": null, "delegated_ops": null, "env": { "device": "Mac Studio (M4 Max)", "machine_label": "mac-studio-m4-max", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Apple M4 Max", "vendor_sdk": null }, "error": null, "evidence": [ "compat_check status ok (runtime /Users/USER/code/litertlm-convert/.qa-venvs/litert-lm-0.16.0/bin/litert-lm 0.16.0); fixed-question answer: '17 + 25 = 42'" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "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/0.16.0/2026-08-12/granite-4.0-h-1b__mac-studio-m4-max.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 4.38, "delegated_ops": null, "env": { "device": "Raspberry Pi 5 Model B Rev 1.1", "machine_label": "raspberry-pi-5-cooled-52c", "os_build": "Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41", "runtime": "litert-lm", "runtime_version": "0.16.1", "soc": "Broadcom BCM2712", "vendor_sdk": null }, "error": null, "evidence": [ "pi5 LLM sweep row: `litert-lm benchmark --backend cpu --cpu-thread-count 4 -p 256 -d 256 --runs 1 --cache memory` (wave-2 driver pi5_llm_bench.py; --cache memory rather than the house --cache no, which OOM-kills every >=1.2B file on the 8 GB Pi — equivalence measured on granite-350m int8, +2-3%), 3 invocations per file with cool-down to <=52 C between them, vcgencmd measure_temp + get_throttled logged per invocation, peak RSS polled from /proc; throughput = median of the three invocations (spread in metrics); a row counts as measured only when the real-generation gate (`litert-lm run`, degenerate-output check) passed and every invocation exited 0 with get_throttled 0x0", "versions: cpu=Raspberry Pi 5 Model B Rev 1.1, litert-lm=0.16.1, litert-lm-api=0.16.1, platform=Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41, python=3.13.5", "cache mode 'memory'; -p 256 -d 256 --runs 1 --cpu-thread-count 4", "gate ('What is 17 plus 26? Answer with the number only.'): status pass, exit 0, wall 73.7 s, output head '43'", "invocation 0: exit=0 wall_s=170.1 temp 50.5->53.2C throttled=0x0 prefill_tps=42.56 decode_tps=4.38 ttft_s=6.243 init_s=42.4893 peak_rss_mb=2599", "invocation 1: exit=0 wall_s=172.1 temp 49.9->54.3C throttled=0x0 prefill_tps=42.3 decode_tps=4.38 ttft_s=6.2801 init_s=42.1156 peak_rss_mb=2599", "invocation 2: exit=0 wall_s=176.1 temp 51.6->55.4C throttled=0x0 prefill_tps=41.65 decode_tps=4.25 ttft_s=6.382 init_s=42.1878 peak_rss_mb=2599" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 256.0, "decode_tps_max": 4.38, "decode_tps_min": 4.25, "init_s": 42.1878, "init_s_max": 42.4893, "init_s_min": 42.1156, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 42.56, "prefill_tps_min": 41.65, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 6.382, "ttft_s_min": 6.243 }, "output_match": null, "peak_mem_mb": 2599.0, "prefill_tokens_per_s": 42.3, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 6280.1 }, "source": "data/device_runs/0.16.1/2026-09-01/granite-4.0-h-1b__raspberry-pi-5.json" } ] }, "model": { "family": "granite4h", "id": "granite-4.0-h-1b", "license": "apache-2.0", "source_url": "https://huggingface.co/ibm-granite/granite-4.0-h-1b", "task": "text-generation" }, "pitfalls": [ "Mamba2 conv/SSM state buffers bind through litert-lm >= 0.15's generalized state binding (ExecutorMetadata) — the 0.14 engine cannot bind them (HF card; 350m draft card wording).", "This 2026-08-04 artifact runs the selective scan as generic float ops on CPU only (stated plainly on the card of record); the GPU story belongs to the 2026-08-13 folded-scan re-ship, a different file (see header note).", "Correctness of record: FP parity corr 1.000000 at all 8 checked decode positions vs HF; 8Q sanity 8/8 == the HF reference; first-token sweep clean at prompt lengths 12-60 against the engine's real chunk plans (ship commit ea409da).", "Family BOS trap: the 350m sibling's greedy decoding flips with a BOS token (LlmMetadata start_token root cause, commit 5f3176d) — the 1b is BOS-robust, but verify start_token when reusing metadata across granite sizes.", "The 2026-08-12 compat_check loadability record on this card measured exactly these bytes (sha-verified against HF revision history); records dated after 2026-08-13 measure the re-shipped file instead." ], "schema_version": "1.2" }