gpu_mldrift @ litert 0.0.0-example — 40.0% of 5 ops delegated · 1 partition · 3 blocking ops
model:  data/examples/model_mixed_example.tflite (sha256 9af4f555f718…, 1 subgraph(s))
matrix: data/examples/matrix_example.json (generated 2026-08-10)

Non-delegated ops:
  s0/n0  FULLY_CONNECTED  dtypes=int8  fallback  provenance=example
      reason: matched entry specific to op, dtypes, and constraints
  s0/n1  CUSTOM  dtypes=float32  unknown  provenance=-
      reason: custom op 'ExampleCustomOp' — not a builtin, outside the matrix vocabulary
  s0/n4  TRANSPOSE_CONV  dtypes=float32  fallback  provenance=example
      reason: matched entry specific to op, dtypes, and constraints
      condition: rejected by some mobile GPU drivers regardless of rank

Correctness warnings (delegated but numerics outside tolerance):
  s0/n3  SOFTMAX  dtypes=float16 — numerics exceed tolerance vs CPU reference for large logits

Partition map (each sync is a host<->device transfer):
  subgraph 0:
    (host)      node 0 FULLY_CONNECTED (fallback)
    (host)      node 1 CUSTOM (unknown)
    ----- host<->device sync -----
    [device #0] nodes 2-3 (2 ops: CONV_2D, SOFTMAX)
    ----- host<->device sync -----
    (host)      node 4 TRANSPOSE_CONV (fallback)

Blocking ops, ranked (impact_score is a heuristic (output element count as a FLOPs proxy, adjacent-partition cost split across each CPU island) — a ranking aid, not a measurement):
  1. s0/n4 TRANSPOSE_CONV (fallback) — impact 128.0
  2. s0/n0 FULLY_CONNECTED (fallback) — impact 64.0
  3. s0/n1 CUSTOM (unknown) — impact 64.0

Rewrite suggestions (verbatim from the matrix, never generated):
  s0/n0 FULLY_CONNECTED:
    symptom: partition split at FULLY_CONNECTED with dynamic batch
    rewrite: export with a fixed batch dimension
    expected effect: single contiguous GPU partition

Custom ops (outside the matrix vocabulary; not probeable):
  s0/n1 CUSTOM 'ExampleCustomOp'

Next steps:
  fallback / incorrect / crash: apply the rewrite hint recorded above, or try the recorded transforms with `edge-fix run <model> --matrix <same snapshot> --rules data/transforms` (a dry run unless --apply is given).
  unknown / needs probe: no measured entry for gpu_mldrift @ litert 0.0.0-example — generate probe fixtures with `edge-compat probe gen` and measure them on your device, or look for a newer snapshot at https://john-rocky.github.io/edge-compat/data/matrix/
  verdicts hold for gpu_mldrift @ litert 0.0.0-example only; delegate support changes between runtime versions.
