File size: 1,786 Bytes
3da88e3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
{
  "queue": [
    {
      "id": "head-layout",
      "class": "exact-or-numerical",
      "mechanism": "Contiguous per-head Q/K/V might remove strided memory access; count copy costs. Reject no operator/block win or quality gate."
    },
    {
      "id": "m5-fc2-tiles",
      "class": "exact-or-numerical",
      "mechanism": "Native FC2 K=14336 selects BK64; test MPP BK128/256/512 and cache walk. FP32 accumulation, BF16 epilogue. Full trajectory validates reduction changes."
    },
    {
      "id": "vector-unpack",
      "class": "exact",
      "mechanism": "Reconstruct 2/4/8 lossless BF16 lanes per thread, fewer redundant nibble loads. Decode+GEMM and full packed block must improve."
    },
    {
      "id": "sparse-embedding",
      "class": "exact",
      "mechanism": "Read only actual token rows, avoiding loading 1.45GiB vocabulary on cold encoding; measure phase-specific peak."
    },
    {
      "id": "exact-qk-rope",
      "class": "numerical-until-proven",
      "mechanism": "Replicate native RMS reduction and cast order in one head-major producer. Retain native fallback."
    }
  ],
  "deprioritized": [
    {
      "mechanism": "CoreML CPU/GPU/ANE/all VAE",
      "reason": "Already measured complete images in v2. All-GPU ~MLX speed, ANE slower and b8 corrupted tiles; VAE <=3% total latency limits payoff."
    },
    {
      "mechanism": "Unscaled FP16 GEMM, skipped flow heads, approximate Q4/Q8",
      "reason": "Actual full trajectories previously failed finite or visual accuracy gates."
    },
    {
      "mechanism": "MPS/CPU dense, whole DiT CoreML export",
      "reason": "Prior real operator tests did not identify a throughput advantage; huge export/compile and memory cost. Revisit only if a new operator mechanism qualifies."
    }
  ]
}