port(LORA-RUNTIME): ExpandPackedLora refuses a packed group with an unadapted member, which is partial LoRA on GatedDeltaNet

Row: LORA-RUNTIME (.agents/engine-matrix.md:256, re-resolved by line number before filing)

Found by wave PORTQ-6 (#2718) re-deriving PORT-NOW entry [205], upstream c2e7242ab7 vllm#47640, of 5559679229..e126687a9a. Nothing was executed. This is a static reading of source.

What upstream does

vllm/lora/layers/column_parallel_linear.py widens expand_packed_lora's signature to list[torch.Tensor | None] and inserts a if b_i is None: guard at the top of the loop. An unadapted packed-group member emits self.n_slices - start_idx None placeholders, advances start_idx, and continues, so the remaining slices stay aligned and are left at base weights. Before the fix the first statement of the loop body was b_rows, cu_rows, covered = b_i.shape[0], 0, 0, which dereferences None. The case is partial LoRA on a Qwen3.5/3.6 GatedDeltaNet.

The pre-fix shape is at the pin: git show 5559679229:vllm/lora/layers/column_parallel_linear.py:266-299. This fix is post-pin and is owed by the pin advance.

What is here

ExpandPackedLora is ported, and it is exactly the pre-commit shape.

// src/vllm/lora/layers.cpp:354-355
for (size_t g = 0; g < lora_b.size(); ++g) {
  const int64_t b_rows = lora_b[g].rows;

There is no emptiness guard. An absent sub-module is an empty LoRAMat — include/vllm/lora/layers.h:66-78 gives rows = 0 and empty() as rows == 0 || cols == 0, and :79-80 documents MatList as upstream's list[torch.Tensor | None] with "an EMPTY LoRAMat standing for None". So b_rows reads 0; the inner loop at layers.cpp:357-363 adds a positive output_sizes_[i] before every comparison and can never reach cu_rows == 0; covered stays 0 and :364-369 throws std::invalid_argument.

The asymmetry is the point. The None tolerance upstream's docstring cites as already present is present here, on both of the other two paths: MergedColumnParallelLinearWithLoRA::SliceLoraB (src/vllm/lora/layers.cpp:333-343, if (lora_b[ii].empty()) continue;) and the stacking loop in BaseLinearLayerWithLoRA::SetLora (:252-263). ExpandPackedLora is the one that was never given it, and it is entered from MergedColumnParallelLinearWithLoRA::SetLora at layers.cpp:389-392 whenever lora_b.size() != n_slices_, which is the partial-group case.

Consequence: a refusal, not a crash and not silent wrongness

std::invalid_argument("LoRA: cannot split lora_b with 0 rows into N slices starting at S").

It is not user-visible today. include/vllm/lora/layers.h is included by src/vllm/lora/layers.cpp and nothing else, and the only constructors of MergedColumnParallelLinear{,VariableSlice}WithLoRA are in tests/vllm/lora/test_lora_layers.cpp. The LoRA layer family is a staged unreached slice, which the LORA-RUNTIME row already records ("the CPU brick exists but nothing reaches it", .agents/feature-matrix.md:39). The gap becomes reachable the day the seam is wired, which is the argument for fixing it in the same change rather than after.

Size

~15 product lines in src/vllm/lora/layers.cpp: an if (lora_b[g].empty()) branch that appends n_slices_ - start_idx empty LoRAMats to both lists, advances start_idx and continues. ~35 test lines in tests/vllm/lora/test_lora_layers.cpp, red-first.

Probe note

The queue text asserts "our layers.h ports expand_packed_lora". That was verified, not assumed: ExpandPackedLora returns 4 hits over src/, include/ and tests/. The secondary zero — no existing test drives an empty member through ExpandPackedLora — carries its control: test_lora_layers.cpp:290 asserts CHECK_FALSE(packed.subloras[1].has_value()); // the "None" sub-module, so the None concept is tested in that same file, just not on this path.

Start in src/vllm/lora/layers.cpp at ExpandPackedLora and read LoRAMat's empty representation in include/vllm/lora/layers.h. Compare the existing empty-member handling in SliceLoraB and SetLora, then add coverage in tests/vllm/lora/test_lora_layers.cpp for a partial packed group. Done means empty members are preserved as empty slices and the test passes without the invalid_argument refusal.