Upload benchmark_matrix.json with huggingface_hub

benchmark_matrix.json

@@ -0,0 +1,386 @@
+{
+  "title": "dispatchAI Per-Chip Benchmark Matrix",
+  "description": "On-device inference benchmarks for mobile LLMs across chipsets",
+  "hardware": {
+    "snapdragon_865": {
+      "chipset": "Qualcomm Snapdragon 865",
+      "cpu": "Kryo 585 (1x2.84GHz + 3x2.42GHz + 4x1.80GHz)",
+      "gpu": "Adreno 650",
+      "npu": "Hexagon Tensor Accelerator (HTP)",
+      "process_nm": 7,
+      "devices": [
+        "Samsung Galaxy S20 FE 5G"
+      ]
+    },
+    "snapdragon_8_gen_2": {
+      "chipset": "Qualcomm Snapdragon 8 Gen 2",
+      "cpu": "Kryo CPU (1x3.2GHz + 2x2.8GHz + 2x3.0GHz + 3x2.0GHz)",
+      "gpu": "Adreno 740",
+      "npu": "Hexagon NPU ( Qualcomm AI Engine)",
+      "process_nm": 4,
+      "devices": [
+        "Samsung Galaxy S23",
+        "OnePlus 11"
+      ]
+    },
+    "apple_a17_pro": {
+      "chipset": "Apple A17 Pro",
+      "cpu": "2x3.78GHz + 4x2.11GHz",
+      "gpu": "Apple GPU 6-core",
+      "npu": "Neural Engine 16-core (35 TOPS)",
+      "process_nm": 3,
+      "devices": [
+        "iPhone 15 Pro"
+      ]
+    },
+    "mediatek_dimensity_9200": {
+      "chipset": "MediaTek Dimensity 9200",
+      "cpu": "1x3.05GHz + 3x2.85GHz + 4x1.8GHz",
+      "gpu": "Immortalis-G715",
+      "npu": "APU 690",
+      "process_nm": 4,
+      "devices": [
+        "Various 2023 flagships"
+      ]
+    }
+  },
+  "models": [
+    {
+      "model_id": "dispatchAI/Qwen2.5-0.5B-Instruct-mobile-int4",
+      "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+      "params_m": 500,
+      "quant": "int4",
+      "size_mb": 350,
+      "architecture": "Qwen2.5",
+      "task": "text-generation"
+    },
+    {
+      "model_id": "dispatchAI/Llama-3.2-1B-Instruct-mobile",
+      "base_model": "meta-llama/Llama-3.2-1B-Instruct",
+      "params_m": 1000,
+      "quant": "fp16",
+      "size_mb": 2000,
+      "architecture": "Llama-3.2",
+      "task": "text-generation"
+    },
+    {
+      "model_id": "dispatchAI/SmolLM2-360M-Instruct-mobile",
+      "base_model": "HuggingFaceTB/SmolLM2-360M-Instruct",
+      "params_m": 360,
+      "quant": "fp16",
+      "size_mb": 720,
+      "architecture": "SmolLM2",
+      "task": "text-generation"
+    },
+    {
+      "model_id": "dispatchAI/SmolLM2-135M-Instruct-mobile",
+      "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
+      "params_m": 135,
+      "quant": "fp16",
+      "size_mb": 270,
+      "architecture": "SmolLM2",
+      "task": "text-generation"
+    },
+    {
+      "model_id": "dispatchAI/TinyLlama-1.1B-Chat-Q5-mobile",
+      "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+      "params_m": 1100,
+      "quant": "Q5_K_M",
+      "size_mb": 450,
+      "architecture": "Llama-2",
+      "task": "text-generation"
+    },
+    {
+      "model_id": "dispatchAI/Qwen2.5-1.5B-Instruct-Q5-mobile",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "params_m": 1500,
+      "quant": "Q5_K_M",
+      "size_mb": 900,
+      "architecture": "Qwen2.5",
+      "task": "text-generation"
+    },
+    {
+      "model_id": "dispatchAI/Gemma-2-2B-IT-Q5-mobile",
+      "base_model": "google/gemma-2-2b-it",
+      "params_m": 2000,
+      "quant": "Q5_K_M",
+      "size_mb": 1300,
+      "architecture": "Gemma-2",
+      "task": "text-generation"
+    },
+    {
+      "model_id": "dispatchAI/Phi-3.5-mini-instruct-Q5-mobile",
+      "base_model": "microsoft/Phi-3.5-mini-instruct",
+      "params_m": 2000,
+      "quant": "Q5_K_M",
+      "size_mb": 1300,
+      "architecture": "Phi-3.5",
+      "task": "text-generation"
+    }
+  ],
+  "results": [
+    {
+      "model_id": "dispatchAI/Qwen2.5-0.5B-Instruct-mobile-int4",
+      "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+      "architecture": "Qwen2.5",
+      "params_m": 500,
+      "quantization": "int4",
+      "size_mb": 350,
+      "task": "text-generation",
+      "device_serial": "R3CN509PLHA",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 47.5,
+      "generation_tps": 9.8,
+      "tokens_per_second": 9.8,
+      "notes": "Stable, good quality"
+    },
+    {
+      "model_id": "dispatchAI/Qwen2.5-0.5B-Instruct-mobile-int4",
+      "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+      "architecture": "Qwen2.5",
+      "params_m": 500,
+      "quantization": "int4",
+      "size_mb": 350,
+      "task": "text-generation",
+      "device_serial": "R3CN50A0HSV",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 45.2,
+      "generation_tps": 9.5,
+      "tokens_per_second": 9.5,
+      "notes": "Stable"
+    },
+    {
+      "model_id": "dispatchAI/Qwen2.5-0.5B-Instruct-mobile-int4",
+      "base_model": "Qwen/Qwen2.5-0.5B-Instruct",
+      "architecture": "Qwen2.5",
+      "params_m": 500,
+      "quantization": "int4",
+      "size_mb": 350,
+      "task": "text-generation",
+      "device_serial": "R3CN50AXHWE",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 43.8,
+      "generation_tps": 9.2,
+      "tokens_per_second": 9.2,
+      "notes": "Stable"
+    },
+    {
+      "model_id": "dispatchAI/Llama-3.2-1B-Instruct-mobile",
+      "base_model": "meta-llama/Llama-3.2-1B-Instruct",
+      "architecture": "Llama-3.2",
+      "params_m": 1000,
+      "quantization": "fp16",
+      "size_mb": 2000,
+      "task": "text-generation",
+      "device_serial": "R3CN509PLHA",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 28.0,
+      "generation_tps": 6.5,
+      "tokens_per_second": 6.5,
+      "notes": "Usable, moderate speed"
+    },
+    {
+      "model_id": "dispatchAI/Llama-3.2-1B-Instruct-mobile",
+      "base_model": "meta-llama/Llama-3.2-1B-Instruct",
+      "architecture": "Llama-3.2",
+      "params_m": 1000,
+      "quantization": "fp16",
+      "size_mb": 2000,
+      "task": "text-generation",
+      "device_serial": "R3CN50A0HSV",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 27.5,
+      "generation_tps": 6.3,
+      "tokens_per_second": 6.3,
+      "notes": "Usable"
+    },
+    {
+      "model_id": "dispatchAI/SmolLM2-360M-Instruct-mobile",
+      "base_model": "HuggingFaceTB/SmolLM2-360M-Instruct",
+      "architecture": "SmolLM2",
+      "params_m": 360,
+      "quantization": "fp16",
+      "size_mb": 720,
+      "task": "text-generation",
+      "device_serial": "R3CN509PLHA",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 55.0,
+      "generation_tps": 12.1,
+      "tokens_per_second": 12.1,
+      "notes": "Good speed/quality balance"
+    },
+    {
+      "model_id": "dispatchAI/SmolLM2-360M-Instruct-mobile",
+      "base_model": "HuggingFaceTB/SmolLM2-360M-Instruct",
+      "architecture": "SmolLM2",
+      "params_m": 360,
+      "quantization": "fp16",
+      "size_mb": 720,
+      "task": "text-generation",
+      "device_serial": "R3CN50A0HSV",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 53.2,
+      "generation_tps": 11.8,
+      "tokens_per_second": 11.8,
+      "notes": "Good balance"
+    },
+    {
+      "model_id": "dispatchAI/SmolLM2-135M-Instruct-mobile",
+      "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
+      "architecture": "SmolLM2",
+      "params_m": 135,
+      "quantization": "fp16",
+      "size_mb": 270,
+      "task": "text-generation",
+      "device_serial": "R3CN509PLHA",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 85.0,
+      "generation_tps": 18.3,
+      "tokens_per_second": 18.3,
+      "notes": "Very fast, limited quality"
+    },
+    {
+      "model_id": "dispatchAI/SmolLM2-135M-Instruct-mobile",
+      "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct",
+      "architecture": "SmolLM2",
+      "params_m": 135,
+      "quantization": "fp16",
+      "size_mb": 270,
+      "task": "text-generation",
+      "device_serial": "R3CN50A0HSV",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 82.5,
+      "generation_tps": 17.8,
+      "tokens_per_second": 17.8,
+      "notes": "Very fast"
+    },
+    {
+      "model_id": "dispatchAI/TinyLlama-1.1B-Chat-Q5-mobile",
+      "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+      "architecture": "Llama-2",
+      "params_m": 1100,
+      "quantization": "Q5_K_M",
+      "size_mb": 450,
+      "task": "text-generation",
+      "device_serial": "R3CN509PLHA",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 26.5,
+      "generation_tps": 6.2,
+      "tokens_per_second": 6.2,
+      "notes": "Q5 quant, stable"
+    },
+    {
+      "model_id": "dispatchAI/Qwen2.5-1.5B-Instruct-Q5-mobile",
+      "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+      "architecture": "Qwen2.5",
+      "params_m": 1500,
+      "quantization": "Q5_K_M",
+      "size_mb": 900,
+      "task": "text-generation",
+      "device_serial": "R3CN509PLHA",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 18.0,
+      "generation_tps": 4.2,
+      "tokens_per_second": 4.2,
+      "notes": "Slower but higher quality"
+    },
+    {
+      "model_id": "dispatchAI/Gemma-2-2B-IT-Q5-mobile",
+      "base_model": "google/gemma-2-2b-it",
+      "architecture": "Gemma-2",
+      "params_m": 2000,
+      "quantization": "Q5_K_M",
+      "size_mb": 1300,
+      "task": "text-generation",
+      "device_serial": "R3CN509PLHA",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 12.5,
+      "generation_tps": 3.0,
+      "tokens_per_second": 3.0,
+      "notes": "Slow but best quality"
+    },
+    {
+      "model_id": "dispatchAI/Phi-3.5-mini-instruct-Q5-mobile",
+      "base_model": "microsoft/Phi-3.5-mini-instruct",
+      "architecture": "Phi-3.5",
+      "params_m": 2000,
+      "quantization": "Q5_K_M",
+      "size_mb": 1300,
+      "task": "text-generation",
+      "device_serial": "R3CN509PLHA",
+      "device_model": "SM-G981V",
+      "chipset": "Snapdragon 865",
+      "chipset_family": "Qualcomm",
+      "cpu_cores": 8,
+      "ram_gb": 8,
+      "android_version": 13,
+      "prompt_tps": 13.0,
+      "generation_tps": 3.1,
+      "tokens_per_second": 3.1,
+      "notes": "Comparable to Gemma"
+    }
+  ],
+  "methodology": "Models run via llama-cli on Samsung S20 FE 5G (Snapdragon 865, 8GB RAM, Android 13). Prompt: 'The capital of France is', max_tokens=20, threads=4, context=512.",
+  "cost": "$0",
+  "date": "2026-06-26"
+}