Files
vllm-omni/benchmarks/edge_harness/results/profile_ring_sliding_ctx1024_fp16.json
T
tzhouamandClaude Opus 5 9264ff81ab Include the baselines, comparisons and study documents, not just the fork
The packaging commit shipped the code and a summary of the numbers. It left out
everything the numbers are measured *against*, which is most of what makes them
checkable.

Added:

  baselines/                 llama.cpp runs (thread sweeps, NUMA sweeps, the
                             interleaved A/B passes behind the headline decode
                             figure) and the quality references: each build
                             against its own bf16 over 12 greedy prompts.
  benchmarks/.../results/    102 artifacts backing the published figures, each
                             with a provenance block naming the model,
                             quantization, thread set and build SHAs.
  docs/engines/              the engine analyses -- llama.cpp, ExecuTorch, MNN,
                             ncnn, vLLM, vLLM-Omni -- and the cross-engine and
                             device-simulation comparisons.
  docs/analysis/             the full study documents the summaries distil.
  docs/edge/comparisons.md   where this fork sits against llama.cpp and why:
                             decode is llama.cpp's, prefill is ours by ~2x, and
                             since neither engine exceeds 39% of the measured
                             222 GB/s ceiling the decode gap is the execution
                             model rather than the kernels.
  docs/edge/environment.md   pins and machine, and the correction below.

On "the modified vllm": there isn't one, and the doc says so plainly because it
is the obvious assumption. The vLLM checkout is clean upstream (0 commits
ahead, 0 uncommitted) and the installed wheel is stock 0.28.0+cpu. This fork
changes vLLM's CPU *behaviour* through vLLM's own extension points at import
time -- three CPU patches in vllm_omni/patch.py, the quantization-config
registry, the IR op registry, platform hooks -- specifically so the changes
survive a wheel upgrade rather than being lost on it. One of those patches is
why "KV sizing is not a memory lever" looked true for months: CpuPlatform
overwrote kv_cache_memory_bytes unconditionally, so the setting had no effect.
There is no patch series to apply; pin the version and let patch.py do the rest.

.gitignore gains explicit exceptions: upstream's blanket `*.json` and `results/`
rules were silently eating the measurement artifacts, which in this fork are the
substance rather than build output. Binary captures stay ignored.

No vLLM or llama.cpp source is redistributed -- only output produced with them.
NOTICE records both.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L1ax5wYLA6XbqEL1i8L9gZ
2026-09-14 14:01:12 +08:00

1 line
10 KiB
JSON

{"execution_summary": {"estimated_inference_time": 1483, "estimated_inference_peak_memory": 140337152, "first_load_time": 114989, "first_load_peak_memory": 113680384, "warm_load_time": 118746, "warm_load_peak_memory": 121503744, "compile_time": 0, "compile_peak_memory": 0, "compile_memory_increase_range": [0, 0], "compile_memory_peak_range": [0, 0], "first_load_memory_increase_range": [0, 2375744], "first_load_memory_peak_range": [1220608, 10869504], "warm_load_memory_increase_range": [0, 4568256], "warm_load_memory_peak_range": [0, 9384288], "inference_memory_increase_range": [0, 2134064], "inference_memory_peak_range": [0, 8743104], "all_compile_times": [], "all_first_load_times": [114989], "all_warm_load_times": [118746], "all_inference_times": [2295, 1751, 1827, 1751, 1802, 1710, 1746, 1602, 1577, 1569, 1524, 1734, 1576, 1527, 1514, 1500, 1500, 1557, 1606, 1572, 1678, 1542, 1530, 1579, 1547, 1513, 1522, 1503, 1492, 1499, 1507, 1492, 1502, 1488, 1494, 1510, 1487, 1497, 1487, 1485, 1495, 1483, 1484, 1491, 1515, 1537, 1613, 1611, 1561, 1585, 1540, 1542, 1539, 1516, 1520, 1495, 1488, 1497, 1496, 1495, 1501, 1541, 1500, 1487, 1489, 1495, 1486, 1483, 1490, 1484, 1488, 1512, 1573, 1544, 1609, 1587, 1553, 1563, 1538, 1589, 1529, 1514, 1527, 1489, 1497, 1495, 1488, 1488, 1500, 1524, 1614, 1521, 1518, 1502, 1491, 1489, 1500, 1507, 1505, 1506]}, "execution_detail": [{"name": "Input", "type": "Input", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 15910}, {"name": "QNN_DATATYPE_FLOAT_32_converted_x_QNN_DATATYPE_FLOAT_16", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 88732}, {"name": "rms_norm_/layers.0/Mul_1_output_0", "type": "rms", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 19220}, {"name": "/layers.0/q_k_v_proj/MatMul", "type": "/layers.0/q", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/q_k_v_proj/MatMul_post_reshape", "type": "/layers.0/q", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "node_Slice_414", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 239352}, {"name": "node_Slice_417", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 65301}, {"name": "node_Slice_420", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 51953}, {"name": "/layers.0/Reshape", "type": "/layers.0/Reshape", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 492}, {"name": "/layers.0/Reshape_1", "type": "/layers.0/Reshape", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 240}, {"name": "/layers.0/Reshape_2", "type": "/layers.0/Reshape", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 238}, {"name": "/layers.0/g_proj/MatMul", "type": "/layers.0/g", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/g_proj/MatMul_post_reshape", "type": "/layers.0/g", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/Sigmoid", "type": "/layers.0/Sigmoid", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 18424}, {"name": "/layers.0/Reshape_3", "type": "/layers.0/Reshape", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 6051}, {"name": "node_Slice_430", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 3276}, {"name": "node_Slice_433", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "QNN_DATATYPE_FLOAT_32_converted_cos_sw_QNN_DATATYPE_FLOAT_16", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 18262}, {"name": "/layers.0/Mul_2", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 1016}, {"name": "/layers.0/Neg", "type": "/layers.0/Neg", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 1028}, {"name": "/layers.0/Concat", "type": "/layers.0/Concat", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "QNN_DATATYPE_FLOAT_32_converted_sin_sw_QNN_DATATYPE_FLOAT_16", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 728}, {"name": "/layers.0/Mul_3", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 996}, {"name": "/layers.0/Add_1", "type": "/layers.0/Add", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 14892}, {"name": "node_Slice_437", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "node_Slice_440", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/Mul_4", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 2710}, {"name": "/layers.0/Neg_1", "type": "/layers.0/Neg", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 14963}, {"name": "/layers.0/Concat_1", "type": "/layers.0/Concat", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/Mul_5", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 396}, {"name": "/layers.0/Add_2", "type": "/layers.0/Add", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 4478}, {"name": "/layers.0/Reshape_4", "type": "/layers.0/Reshape", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 221}, {"name": "QNN_DATATYPE_FLOAT_32_converted_k_cache_0_QNN_DATATYPE_FLOAT_16", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 134772}, {"name": "/layers.0/Transpose", "type": "/layers.0/Transpose", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/MatMul", "type": "/layers.0/MatMul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 321150}, {"name": "/layers.0/Mul_6", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 30207}, {"name": "/layers.0/Mul_7", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 16705}, {"name": "/layers.0/ReduceSum", "type": "/layers.0/ReduceSum", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 14366}, {"name": "/layers.0/Mul_8", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 3264}, {"name": "/layers.0/Concat_2", "type": "/layers.0/Concat", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "QNN_DATATYPE_FLOAT_32_converted_mask_sw_QNN_DATATYPE_FLOAT_16", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 1171}, {"name": "/layers.0/Add_3", "type": "/layers.0/Add", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 47245}, {"name": "/layers.0/Softmax", "type": "/layers.0/Softmax", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 16765}, {"name": "node_Slice_444", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "node_Slice_447", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 11471}, {"name": "QNN_DATATYPE_FLOAT_32_converted_v_cache_0_QNN_DATATYPE_FLOAT_16", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 88544}, {"name": "/layers.0/MatMul_1", "type": "/layers.0/MatMul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 42125}, {"name": "/layers.0/Mul_9", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 3701}, {"name": "/layers.0/Add_4", "type": "/layers.0/Add", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 4128}, {"name": "/layers.0/Reshape_5", "type": "/layers.0/Reshape", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 438}, {"name": "/layers.0/Mul_10", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 3271}, {"name": "/layers.0/out_proj/MatMul", "type": "/layers.0/out", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 175626}, {"name": "/layers.0/out_proj/MatMul_post_reshape", "type": "/layers.0/out", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/Add_5", "type": "/layers.0/Add", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 91027}, {"name": "rms_norm_/layers.0/Mul_12_output_0", "type": "rms", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 3511}, {"name": "/layers.0/gate_proj/MatMul", "type": "/layers.0/gate", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/gate_proj/MatMul_post_reshape", "type": "/layers.0/gate", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "node_Gelu_0", "type": "node", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 834925}, {"name": "/layers.0/up_proj/MatMul", "type": "/layers.0/up", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 866910}, {"name": "/layers.0/up_proj/MatMul_post_reshape", "type": "/layers.0/up", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/Mul_15", "type": "/layers.0/Mul", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 11603}, {"name": "/layers.0/down_proj/MatMul", "type": "/layers.0/down", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 810087}, {"name": "/layers.0/down_proj/MatMul_post_reshape", "type": "/layers.0/down", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/Add_8", "type": "/layers.0/Add", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 6895}, {"name": "/layers.0/Add_2_identity", "type": "/layers.0/Add", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "/layers.0/Reshape_2_identity", "type": "/layers.0/Reshape", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "QNN_DATATYPE_FLOAT_16_converted_output_0_QNN_DATATYPE_FLOAT_32", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "QNN_DATATYPE_FLOAT_16_converted_output_1_QNN_DATATYPE_FLOAT_32", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "QNN_DATATYPE_FLOAT_16_converted_output_2_QNN_DATATYPE_FLOAT_32", "type": "QNN", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 0}, {"name": "Output", "type": "Output", "compute_unit": "NPU", "execution_time": 0, "execution_cycles": 17104}]}