Add litertlm_manifest.json — machine-readable deployment manifest (variant selection, backend recommendations, measured performance)
Browse files- litertlm_manifest.json +134 -0
litertlm_manifest.json
ADDED
|
@@ -0,0 +1,134 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"manifest_schema": "0.1.0",
|
| 3 |
+
"repo": "litert-community/SmolLM3-3B",
|
| 4 |
+
"generated": "2026-08-24",
|
| 5 |
+
"generator": "make_manifest.py",
|
| 6 |
+
"model": {
|
| 7 |
+
"display_name": "SmolLM3-3B",
|
| 8 |
+
"base_model": "HuggingFaceTB/SmolLM3-3B",
|
| 9 |
+
"architecture": "Fully-open 3B decoder with GQA and a NoPE attention schedule (SmolLM3ForCausalLM, rotary disabled every 4th layer), multilingual, long-context trained",
|
| 10 |
+
"parameters_b": 3,
|
| 11 |
+
"license": "apache-2.0",
|
| 12 |
+
"context_length": 4096,
|
| 13 |
+
"capabilities": {
|
| 14 |
+
"vision": false,
|
| 15 |
+
"audio": false,
|
| 16 |
+
"thinking": {
|
| 17 |
+
"declared": false
|
| 18 |
+
}
|
| 19 |
+
}
|
| 20 |
+
},
|
| 21 |
+
"variants": [
|
| 22 |
+
{
|
| 23 |
+
"file": "SmolLM3-3B.litertlm",
|
| 24 |
+
"sha256": "a34da46f5697896c98a4d3b3dacddbc50487f8e35432b37b72e4fe38ed264bb4",
|
| 25 |
+
"size_bytes": 3108978688,
|
| 26 |
+
"sections": [
|
| 27 |
+
{
|
| 28 |
+
"type": "LlmMetadataProto",
|
| 29 |
+
"size_bytes": 268
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"type": "SP_Tokenizer",
|
| 33 |
+
"size_bytes": 2260840
|
| 34 |
+
},
|
| 35 |
+
{
|
| 36 |
+
"type": "TFLiteModel",
|
| 37 |
+
"size_bytes": 3106673904,
|
| 38 |
+
"model_type": "tf_lite_prefill_decode"
|
| 39 |
+
}
|
| 40 |
+
],
|
| 41 |
+
"quantization": "int4 (recipe not stated on card)",
|
| 42 |
+
"backends": [
|
| 43 |
+
"cpu"
|
| 44 |
+
],
|
| 45 |
+
"default_backend": "cpu",
|
| 46 |
+
"measured": [],
|
| 47 |
+
"known_issues": [
|
| 48 |
+
"File is present in the repo tree (3.11 GB) but not documented anywhere on the model card"
|
| 49 |
+
]
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"file": "SmolLM3-3B_q4_block32_ekv4096.litertlm",
|
| 53 |
+
"sha256": "38d7bf55e243f5a95d8b16c21b1f7ef7debe6ce6bb2ddb66eb54f4bc2be8e6eb",
|
| 54 |
+
"size_bytes": 2002257840,
|
| 55 |
+
"sections": [
|
| 56 |
+
{
|
| 57 |
+
"type": "LlmMetadataProto",
|
| 58 |
+
"size_bytes": 142
|
| 59 |
+
},
|
| 60 |
+
{
|
| 61 |
+
"type": "HF_Tokenizer_Zlib",
|
| 62 |
+
"size_bytes": 2608153
|
| 63 |
+
},
|
| 64 |
+
{
|
| 65 |
+
"type": "TFLiteModel",
|
| 66 |
+
"size_bytes": 1733852544,
|
| 67 |
+
"model_type": "tf_lite_prefill_decode"
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"type": "TFLiteModel",
|
| 71 |
+
"size_bytes": 265750448,
|
| 72 |
+
"model_type": "tf_lite_embedder"
|
| 73 |
+
}
|
| 74 |
+
],
|
| 75 |
+
"quantization": "int4 weights - blockwise (block 32) + OCTAV optimal-clipping, symmetric; embedding INT8",
|
| 76 |
+
"backends": [
|
| 77 |
+
"cpu",
|
| 78 |
+
"gpu"
|
| 79 |
+
],
|
| 80 |
+
"default_backend": "gpu",
|
| 81 |
+
"requirements": {
|
| 82 |
+
"platform_notes": [
|
| 83 |
+
"Gallery import needs package com.google.ai.edge.gallery 1.0.15+ (older 1.0.x builds reject .litertlm); Gallery v1.0.16+ can import litert-lm models directly from Hugging Face inside the app",
|
| 84 |
+
"Embedding externalized into its own bundle section so the main weights section stays under the iOS ~2 GiB single-mmap limit",
|
| 85 |
+
"On a Pixel 8a (Tensor G3, 8 GB) litert_lm_main runs the graph entirely on the OpenCL delegate (1476/1476 prefill, 1308/1308 decode nodes, zero rejected ops) and answers correctly"
|
| 86 |
+
]
|
| 87 |
+
},
|
| 88 |
+
"measured": [
|
| 89 |
+
{
|
| 90 |
+
"device": "Apple M4 Max",
|
| 91 |
+
"os": "macOS",
|
| 92 |
+
"backend": "cpu",
|
| 93 |
+
"runtime": "litert-lm benchmark (litert-lm 0.15.0)",
|
| 94 |
+
"prompt_tokens": 256,
|
| 95 |
+
"decode_tokens": 256,
|
| 96 |
+
"prefill_tps": 141,
|
| 97 |
+
"decode_tps": 24.1,
|
| 98 |
+
"ttft_s": 2.14,
|
| 99 |
+
"max_num_tokens": 4096,
|
| 100 |
+
"runs": 3,
|
| 101 |
+
"date": "2026-08-24",
|
| 102 |
+
"source": "model card Performance table (cardbench harness)"
|
| 103 |
+
},
|
| 104 |
+
{
|
| 105 |
+
"device": "Apple M4 Max",
|
| 106 |
+
"os": "macOS",
|
| 107 |
+
"backend": "gpu",
|
| 108 |
+
"runtime": "litert-lm benchmark (litert-lm 0.15.0)",
|
| 109 |
+
"prompt_tokens": 256,
|
| 110 |
+
"decode_tokens": 256,
|
| 111 |
+
"prefill_tps": 1354,
|
| 112 |
+
"decode_tps": 93.2,
|
| 113 |
+
"ttft_s": 0.21,
|
| 114 |
+
"max_num_tokens": 4096,
|
| 115 |
+
"runs": 3,
|
| 116 |
+
"date": "2026-08-24",
|
| 117 |
+
"source": "model card Performance table (cardbench harness)"
|
| 118 |
+
},
|
| 119 |
+
{
|
| 120 |
+
"device": "iPhone 17 Pro",
|
| 121 |
+
"os": "iOS 27.0",
|
| 122 |
+
"backend": "gpu",
|
| 123 |
+
"runtime": "LiteRTDemo harness (Metal GPU backend)",
|
| 124 |
+
"prefill_tps": 30.8,
|
| 125 |
+
"decode_tps": 22.5,
|
| 126 |
+
"ttft_s": 0.63,
|
| 127 |
+
"runs": 1,
|
| 128 |
+
"date": "2026-08-24",
|
| 129 |
+
"source": "model card Performance table (cardbench harness)"
|
| 130 |
+
}
|
| 131 |
+
]
|
| 132 |
+
}
|
| 133 |
+
]
|
| 134 |
+
}
|