fix(tune): skip empty-buffer LoRA layers in save_peft_safetensors (GDN slots)

[ohdearquant]

Bug

save_peft_safetensors aborts with InvalidTensorView(F32, [rank, d_in], 0) when an adapter contains a LoRA layer with empty A/B factor buffers. This happens for GDN-attention slots when only the GQA layers are trained: those slots leave q_proj/v_proj factors empty, but the save loop still pushed a zero-byte byte buffer paired with the non-zero shape [rank, d_in] / [d_out, rank], which TensorView::new rejects. The whole save then fails — so a real LoRA adapter trained on a GDN model (e.g. Qwen3.5-0.8B, 18 GDN + 6 GQA layers) cannot be persisted at all.

Fix

Skip any layer whose A or B factor buffer is empty rather than emit a zero-byte tensor with a non-zero shape. The trained layers round-trip intact; the untrained GDN slots are dropped from the saved adapter.

for ((layer_idx, module), layer) in &adapter.layers {
    // A LoRA layer with an empty factor buffer means "this module was not trained"
    // (e.g. GDN-attention slots leave q_proj/v_proj empty). Skip it rather than
    // emit an InvalidTensorView for a zero-byte tensor with non-zero shape.
    if layer.a.is_empty() || layer.b.is_empty() {
        continue;
    }
    ...
}

Provenance & why a fresh PR

This is the safetensors.rs library slice of the GDN-save fix originally developed on feat/lora-backward-training (#193). That branch predates #191 / #212/#217 / #261 and has substantially diverged from main (60 commits behind, 20 add/add conflicts), so it can't be cleanly rebased without a scope decision. Applying just this guard onto current main preserves the fixes the branch would otherwise regress:

• hardening(tune): LoRA safetensors transpose panics on shape/element mismatch; f32 reader drops trailing bytes #212/fix(tune): validate LoRA safetensors payload length + tensor shape (#212) #217 — load_peft_safetensors shape/length validation (branch removed it)
• fix(tune): load LoRA adapter alpha from safetensors header (was hardcoded to rank) #261 — alpha-from-header behavior (branch reverts to alpha = rank)

Verification

• Regression proof (TDD): new test_save_skips_empty_buffer_layers builds an adapter with one real GQA layer + one empty GDN-slot layer. With the guard disabled the test fails with exactly InvalidTensorView(F32, [8, 1024], 0); with the guard restored it passes (real layer present + bit-exact, empty layer absent).
• cargo test -p lattice-tune --lib --features safetensors — 25/25 green, including the preserved test_alpha_metadata_round_trips, test_rejects_shape_product_overflow, test_transpose_rejects_shape_element_mismatch.
• cargo clippy -p lattice-tune --features safetensors --lib — clean (-D warnings).
• Diff is purely additive: +82 lines, one file.

🤖 Generated with Claude Code