mirror of
https://github.com/hanzoai/ANE.git
synced 2026-10-30 18:53:51 +00:00
zeekay
d861b8959b
- ane_universal.h: Auto-detecting runtime that selects InMem (M4+) or CoreML fallback (M1-M3) based on hardware capability probing - ane_probe_m1.m: Deep diagnostic found all 74 ANE classes present on M1 but _ANECompiler binary rejects all MIL programs (40+ format combos) - ane_m1_fix.m: Systematic MIL format fuzzer (program versions 1.0-1.3, opsets ios15-ios18, weight offsets, old/new syntax) — all fail on M1 - ane_coreml_bridge.m: CoreML ANE bridge showing 2.44x speedup at 256x256 - benchmark_apple_silicon.py: MLX vs MPS vs CoreML comparison on M1 Max - MLX: 6.72 TFLOPS peak (fp16), 62ms/step training - MPS: 4.12 TFLOPS peak, 84ms/step training - ANE via CoreML: ~1x (dispatch overhead negates hardware advantage) - demo_mlx_training.py: Working 4-layer transformer training, 4688 tok/s - Updated README with M1-M4 compatibility matrix and integration points - Added .gitignore for compiled binaries and .venv
114 lines
2.6 KiB
JSON
114 lines
2.6 KiB
JSON
[
|
|
{
|
|
"backend": "MLX",
|
|
"operation": "matmul",
|
|
"shape": "768x256 @ 256x768",
|
|
"ms_per_op": 6.083245840000018,
|
|
"tflops": 0.049642887356990184,
|
|
"notes": "Stories110M dim"
|
|
},
|
|
{
|
|
"backend": "MLX",
|
|
"operation": "matmul",
|
|
"shape": "768x256 @ 256x2048",
|
|
"ms_per_op": 3.1631091599956562,
|
|
"tflops": 0.25459329010355936,
|
|
"notes": "Stories110M FFN"
|
|
},
|
|
{
|
|
"backend": "MLX",
|
|
"operation": "matmul",
|
|
"shape": "4096x512 @ 512x4096",
|
|
"ms_per_op": 7.272356660005244,
|
|
"tflops": 2.362352396504658,
|
|
"notes": "Zen-4B dim"
|
|
},
|
|
{
|
|
"backend": "MLX",
|
|
"operation": "matmul",
|
|
"shape": "4096x512 @ 512x11008",
|
|
"ms_per_op": 6.866338339996219,
|
|
"tflops": 6.724238763921067,
|
|
"notes": "Zen-4B FFN"
|
|
},
|
|
{
|
|
"backend": "MLX",
|
|
"operation": "matmul",
|
|
"shape": "4096x2048 @ 2048x4096",
|
|
"ms_per_op": 10.98757249999835,
|
|
"tflops": 6.254291085315735,
|
|
"notes": "Zen-4B long ctx"
|
|
},
|
|
{
|
|
"backend": "MLX",
|
|
"operation": "transformer_fwd",
|
|
"shape": "dim=768 h=12 seq=256",
|
|
"ms_per_op": 7.699347950006086,
|
|
"tflops": 0.0,
|
|
"notes": "Stories110M"
|
|
},
|
|
{
|
|
"backend": "MLX",
|
|
"operation": "transformer_fwd",
|
|
"shape": "dim=4096 h=32 seq=512",
|
|
"ms_per_op": 63.143602100012686,
|
|
"tflops": 0.0,
|
|
"notes": "Zen-4B"
|
|
},
|
|
{
|
|
"backend": "MLX",
|
|
"operation": "train_step",
|
|
"shape": "1-layer dim=768 seq=256",
|
|
"ms_per_op": 62.13079170001947,
|
|
"tflops": 0.0,
|
|
"notes": "Stories110M"
|
|
},
|
|
{
|
|
"backend": "MPS",
|
|
"operation": "matmul",
|
|
"shape": "768x256 @ 256x768",
|
|
"ms_per_op": 7.707359159994667,
|
|
"tflops": 0.03918201834520567,
|
|
"notes": "Stories110M dim"
|
|
},
|
|
{
|
|
"backend": "MPS",
|
|
"operation": "matmul",
|
|
"shape": "768x256 @ 256x2048",
|
|
"ms_per_op": 7.709723320003832,
|
|
"tflops": 0.10445334217254372,
|
|
"notes": "Stories110M FFN"
|
|
},
|
|
{
|
|
"backend": "MPS",
|
|
"operation": "matmul",
|
|
"shape": "4096x512 @ 512x4096",
|
|
"ms_per_op": 11.832635000000664,
|
|
"tflops": 1.4519056139227684,
|
|
"notes": "Zen-4B dim"
|
|
},
|
|
{
|
|
"backend": "MPS",
|
|
"operation": "matmul",
|
|
"shape": "4096x512 @ 512x11008",
|
|
"ms_per_op": 14.450160839996897,
|
|
"tflops": 3.195182319646064,
|
|
"notes": "Zen-4B FFN"
|
|
},
|
|
{
|
|
"backend": "MPS",
|
|
"operation": "matmul",
|
|
"shape": "4096x2048 @ 2048x4096",
|
|
"ms_per_op": 16.676723340005992,
|
|
"tflops": 4.120682182881095,
|
|
"notes": "Zen-4B long ctx"
|
|
},
|
|
{
|
|
"backend": "MPS",
|
|
"operation": "train_step",
|
|
"shape": "1-layer dim=768 seq=256",
|
|
"ms_per_op": 83.85612920001222,
|
|
"tflops": 0.0,
|
|
"notes": "Stories110M"
|
|
}
|
|
] |