Upload README.md with huggingface_hub

README.md

@@ -68,26 +68,36 @@ text tower offline and take the dot product on device.
 ## Conversion notes
 
 Converted with [litert-torch / ai-edge-torch](https://github.com/google-ai-edge/ai-edge-torch).
-Making a RoPE ViT image tower **fully GPU-compatible** required three verbatim
-(weights-exact, output corr ≈ 1.0) model-side rewrites:
+Making a RoPE ViT image tower **fully GPU-resident *and* numerically correct** on
+the ML Drift GPU delegate required four verbatim (weights-exact, output
+corr ≈ 1.0) model-side rewrites — the first three for residency, the last for
+on-device numerical correctness:
 
 1. **Fused-qkv → 4D manual attention** — the fused `qkv` reshape emits a 5D
-   head-split (`RESHAPE` rank 5) that the GPU delegate rejects; decompose into
-   separate q/k/v projections so attention stays 4D.
+   head-split the GPU delegate rejects; decompose into separate q/k/v projections.
+   Self-attention uses `scaled_dot_product_attention`, whose lowering keeps the
+   batch-matmul 3D with a materialized transpose (both required for residency).
 2. **Interleaved 2D-RoPE → rotate-half** — PE-Core's interleaved rotary uses a
    strided `x[..., ::2]` that lowers to `GATHER_ND` (GPU-banned). Bake an
    even→odd channel permutation into the q/k weights (preserves q·k exactly) and
-   apply the rotate-half form (slice + neg + concat) with constant cos/sin →
-   clean `MUL`/`ADD`/`SLICE`/`CONCAT`.
-3. **Attention-pool with constant query → constant-RHS matmul** — the pooling
-   query is a constant latent, so the `q·kᵀ` batch-matmul is `const @ non-const`,
-   which the delegate rejects; reorder as `k @ q_constᵀ` (constant RHS → the
-   fully-connected path). Self-attention uses `scaled_dot_product_attention`,
-   whose lowering keeps the batch-matmul 3D with a materialized transpose
-   (both required for GPU residency).
-
-Verified: zero banned ops, zero >4D tensors, TFLite-vs-PyTorch output
-correlation = 1.000000 (FP32 and FP16).
+   apply the rotate-half form with constant cos/sin → clean
+   `MUL`/`ADD`/`SLICE`/`CONCAT`.
+3. **Attention-pool single-query attention → broadcast-multiply + reduce-sum** —
+   the pooling query is a constant latent, so a batch-matmul there is
+   `const @ non-const` (rejected at compile, and the reordered `const-RHS` form is
+   mis-computed on device); expressing it as `(q·k).sum` + softmax + `(attn·v).sum`
+   is exact and GPU-correct.
+4. **Overflow-safe LayerNorm** — the delegate computes the LayerNorm variance
+   reduction in **fp16 even for an fp32 graph**; deep-ViT "massive activations"
+   (|x|~50+) make `sum((x-mean)²)` exceed fp16 max (65504), so the normalization
+   is wrong and the error compounds with depth (output correlation collapses to
+   ~0.28 over 12 blocks while *still reporting full GPU residency*). Scaling by
+   1/32 before squaring (undone after) keeps the running sum in range —
+   mathematically identical to `nn.LayerNorm`.
+
+Verified **on a Pixel 8a GPU**: zero banned ops, zero >4D tensors, full residency,
+and TFLite(GPU)-vs-PyTorch output correlation = 1.0 (the on-device GPU result —
+not just the host CPU result — matches the reference).
 
 ## Training data & PII