GPU-clean conversion
A conversion is done when three things hold, in this order:
- it converts,
- every node runs on the GPU,
- the on-device output matches the source model.
Step 3 is not implied by step 2. The delegate can report full residency and still return wrong numbers — that is the failure mode most of the rewrites below exist for. Never call a model done without a numerical check against CPU or PyTorch on the actual device.
Loop
1. Convert plain first. No patches. This tells you what the model actually needs rather than what you assumed.
2. Verify through the CompiledModel API before touching a device.
pythonfrom litert_gpu_toolkit import check_gpu_compatibility, print_report print_report(check_gpu_compatibility("model.tflite"))
It compiles the model for the GPU accelerator, runs every signature on random inputs, and compares the outputs against a CPU-compiled reference. A failed GPU compile names the offending op — route it through the table below. A pass with a CPU-fallback warning means some ops fell back to the CPU; treat them the same way if you need full residency. This exercises the host GPU, so step 5 on the actual device is still required.
3. Map each symptom to a rewrite. The rewrites live in
utilities/litert_gpu_toolkit — plain Python, no build step. Outside this
repo, clone litert-samples and put utilities/ on PYTHONPATH, or vendor
the directory:
| What you see | Rewrite |
|---|---|
GATHER_ND named in the compile error | Find its source. Stride-2 slicing (x[:, ::2], Focus stems, patch merging), grid_sample, MaxPool padding, bicubic interpolate, and reflect-mode F.pad all lower to it. patch_grid_sample, patch_maxpool_zeropad, patch_interpolate, patch_patch_merging |
| A rank-5+ tensor named in the compile error | PixelShuffle (rank-6 reshape), windowed attention, einops.rearrange, packed-QKV attention head splits. pixelshuffle_to_conv_transpose, patch_window_attention, patch_einops |
TRANSPOSE_CONV rejected | Version skew, not a missing op. ZeroStuffConvT1d / ZeroStuffConvT2d — zero-stuff plus a plain conv, exact to ~1e-7 |
SELECT / SELECT_V2 | PReLU, ELU, in-place index assignment, and torch.where masking. Replace with arithmetic: x*(1-m) + v*m. Keep v finite in fp16: −1e9 is −inf there, and (1 − m) × −1e9 is 0 × −inf = NaN on every valid position; −30000 is exact in fp16 and gives the same softmax in fp32 when the logits are bounded |
BROADCAST_TO | Two cases. On a compile-time constant: an outer product or .expand that did not fold — bake the result as a constant at its target shape. On a runtime tensor the GPU delegate rejects it outright, even at rank 4 — the canonical case is GQA's repeat_kv (x[:,:,None].expand(...), which is also rank-5, so two walls in one line). Exact rewrite: torch.cat([x[:, i:i+1].expand(b, n_rep, s, d) for i in range(n_kv)], dim=1) — same head order, bit-exact. Tracked upstream: google-ai-edge/LiteRT#9191 |
| Masked attention wrong only on device: token 0 bit-exact, every later token wrong | Broadcast ADD whose LHS is a BATCH_MATMUL result (the scores + mask[1,1,S,S] idiom) silently miscomputed on ai-edge-litert 2.1.3 and 2.1.5 and is fixed in 2.1.6 (head-axis size-1 broadcast only; google-ai-edge/LiteRT#8593, closed as fixed; the 2.1.6 release notes list GPU broadcast fixes). The signature mimics broken RoPE — tap the rope output before blaming it. Rewrites on the older runtimes: pre-expand the mask to [1,H,S,S], or materialize the BMM as a second output |
An ADD result that is both a graph output and consumed downstream comes back wrong | Output aliasing: the returned tensor holds an operand, not the sum — ADD with two runtime operands (SUB/MUL exact, x + 1.0 exact). This is the shape of every explicit state update in a streaming/recurrent graph. Workaround: emit acc * one where one is a runtime input holding 1.0 — a constant folds straight back into the pattern. Tracked upstream: google-ai-edge/LiteRT#8599 |
RELU_0_TO_1 rejected by the GPU delegate | Emitted by hard-sigmoid / nn.Hardtanh(0,1). Accepted in litert 2.1.3, rejected from 2.1.5 on — a model at full residency on an older runtime hard-fails CompiledModel creation after an upgrade. Rewrite: relu(x) - relu(x-1), exact. Tracked upstream: google-ai-edge/LiteRT#8598 |
DIV: No support of few identical inputs / Expected 1 const input tensor(s), device only | The delegate declines an op whose two inputs are the same tensor, and ops whose inputs are all constants — together these split a perceiver-style block (softmax over a length-1 axis of a constant latent bank) into several partitions. Fixes: special-case the degenerate axis (a softmax over a length-1 axis is identically 1), or make one input non-constant. Note the sibling LayerNorm-over-a-constant pattern no longer reaches the delegate at all — the converter folds it to a single MUL. Tracked upstream: google-ai-edge/LiteRT#9192 |
NHWC node rewriter not found: amax | x.amax(...)/x.max(dim) in stable-softmax, adaptive norms, qk-norm. Rewrite channel reduce-max as max_pool2d(x.reshape(N,1,C,H*W), kernel=(C,1)) — numerically identical — or drop the norm to 3D |
Lowering not found: aten._fft_r2c / aten.complex | torch.stft/istft and complex views have no lowering (fails before any GPU check). A DFT is a fixed linear map: windowed-DFT as Conv1d with the cos/sin basis baked into kernels (stride = hop), iSTFT as inverse-DFT matmul + overlap-add via zero-stuffed conv-transpose — exact. Model-selection corollary: prefer time-domain vocoder branches over iSTFT-based ones. ⚠ Library STFT-as-conv stacks (torchlibrosa-style) have numerically mis-converted (corr 0.83) while the op check looks clean — verify the spectrogram numerically or compute log-mel host-side |
| Compiles, runs, output is wrong or NaN | The fp16 reduction family — and the trigger is the fp16 accumulator passing 65504, so a plain single-axis mean/sum over enough elements overflows just like variance does. patch_safe_layernorm, patch_rmsnorm, patch_instance_norm, hierarchical_mean. Caveats: at extreme magnitudes (|x| in the thousands) even the adaptive safe-LN form overflows when it reconstructs the large variance — the robust form stays entirely in the down-scaled domain (xs = x/S, normalize xs, never multiply the variance back by S²); hierarchical_mean is exact only for power-of-two spatial dims (for arbitrary dims, cascade /2 avg-pools with ceil_mode so each stage averages ≤~49 elements). A static form keeps fp32 bit-exact and puts no reduce-max in the graph: scale each norm's input by a power of two 2^-k chosen per site from a calibration over the whole check set (headroom for the row's sum of squares and for the largest single x² — both overflowed in the case on record, the mean of squares did not), and eps by 4^-k. When eps·4^-k drops out of fp16's range (eps 1e-6 at k = 3), an all-zero row (padding) becomes rsqrt(0) = inf and 0 × inf = NaN: clamp the rsqrt at 2^k/√eps — the zero row's own value, never reached in fp32. Diagnostic split: all-zero/all-blank output = an overflow in one block; a result that starts near-correct and degrades with depth = precision compounding, which no overflow patch (and no fp32-precision flag) fixes |
| Head outputs exactly zero | RMSNorm Σx² overflowed fp16 to inf. patch_rmsnorm |
GPU compile fails only at fp16 precision with a kernel-source error (write_imageh on Adreno and Mali OpenCL, write on Metal, at the same source line), while fp32 precision compiles | A boolean value on the way to a graph output. In the case on record an output mask computed as (mask > 0.5) and cast back to float made all three GPUs fail; the same output computed from the float input mask by float ops (a strided slice) compiled at fp16 everywhere and stayed bit-exact in fp32. Small graphs with the same GREATER → CAST pair compiled, so bisect on the real graph. Keep masks float (1 = valid) from input to output |
4. Re-convert and re-verify. Repeat 2–3 until the check passes.
5. Verify on device, numerically. Run the same input through the source
model and the on-device graph and compare. Correlation on the output tensor,
plus a task-level check where one exists (argmax match, IoU, mask foreground
count). Record the device and the residency line (N/N nodes) alongside the
number — a result without both is not reproducible.
If residency is full and the output is wrong, bisect by intermediate: dump tensors at block boundaries and find the first one that diverges. The cause is usually a reduction, not the op you suspect.
Watch for
- fp16 is used even for an fp32 graph. The delegate reduces in fp16
regardless of tensor dtype by default (the GPU options carry a precision
setting —
CompiledModel.GpuOptions(precision = …)in Kotlin — that can request fp32 at a speed cost). Anything that sums many large values — variance,Σx², multi-axis mean — is a candidate. - Approximation choices are not free. Substituting a GELU or Swish approximation changes numerics. A head with a wide output range can lose real accuracy to the coarser form.
inplace=Trueactivations mutate a residual.out = self.act(x)thenx + outaddsrelu(x), notx. Swapping the activation silently changes the math.- A silent CPU fallback looks like success. If the GPU output is
bit-identical to CPU fp32, it probably did not run on the GPU. Genuine fp16
execution drifts in the last digits. The runtime's own count is the line
N operations will run on the GPU, and the remaining M operations will run on the CPU.; a node the delegate declines while building (TopK requires src tensor C dimension to be divisible by 4) leaves only that log behind — the checker's report can read fully-on-GPU while the op ran on the CPU (ai-edge-litert2.2.0 and 2.3.0). - Reflect padding routes through
F.padeven atpadding=0— every conv withpadding_mode='reflect'is affected, not just padded ones. Slice+concat reflect pads are cheap for small pad widths. - Channel-attention through a
Linearconfuses layout handling at convert time (tfl.mul operands don't have broadcast-compatible shapes) — express channel attention as a 1×1 conv. .chunk()lowers toSPLIT(GPU-rejected) — slice directly instead; bit-exact.- Constant folding can explode the file: a frozen-param × constant product materializes at full size per block (fp16 casting skips non-weight constants, so it cannot rescue it). Feed one factor as a runtime input — param × input never folds.
Output layout
Ship the result as a model recipe, not a one-off script:
models/<family>/<model>/ README.md what the model is, what was produced converted/ README.md environments, pipeline, verification results export_*.py one script per graph dump_*_ref.py reference dumps from the source model verify_*.py parity check against those references
Keep the export, the reference dump, and the verification separate so each can be re-run alone. State the verification numbers in the README with the device they came from. Weights are not committed.
The tree above is this repo's convention. In your own project the directory names matter less than the split — keep the same three separately-runnable pieces wherever your models live.
Do the model first. A demo can follow, and reusable inference code matters more in it than a full UI.
