LORA-RUNTIME work item W2 from .agents/specs/lora-adapter.md (§ "Non-overlapping work breakdown").
W1 (the CPU punica brick — LoRALayerWeights, BgmvShrink/BgmvExpand/BgmvExpandSlice, AddLoraLinear, single-slice LoRALinear) landed in 1d37f152, gated by test_punica_cpu. W2 is the next slice:
PackedLoRALayerWeights (vllm/lora/lora_weights.py:99-282) — pack + per-slice optimize + is_packed.- Multi-slice punica
add_shrink / add_expand with output_slices and offset_start (vllm/lora/punica_wrapper/punica_cpu.py:166-236), and the multi-slice add_lora_linear (:265-312).
- The LoRA-wrapped layer family (
vllm/lora/layers/): replicated, column-parallel, row-parallel, merged column-parallel (gate_up, 2 slices), QKV (1 slice) and merged QKV (3 slices), with TP slice_lora_a / slice_lora_b including the fully-sharded (S-LoRA) variants.
- Embedding LoRA (
vocal_parallel_embedding.py) via add_lora_embedding, and logits LoRA (logits_processor.py) via add_lora_logits + the vocab_size <= 258048 guard.
Tests: ports tests/lora/test_layers.py (pinned vLLM 555967922) — test_linear_replicated, test_linear_parallel (row/column x fully_shard), test_column_parallel_packed (repeats 1/2/3), test_embeddings, test_lm_head_logits_processor, test_lm_head_logits_processor_invalid_vocab_size.
Out of scope (later Ws in the same spec): metadata/mapping and sgmv segmentation (W3), adapter load/config (W4), the multi-adapter manager (W5), endpoints (W6), GPU punica kernels + fused-MoE packing and the model gate (W7).
Punch-list item 12 (ROAD-V1-D5, headline [H]) in .agents/specs/roadmap-v1-completion.md.
LORA-RUNTIMEwork item W2 from.agents/specs/lora-adapter.md(§ "Non-overlapping work breakdown").W1 (the CPU punica brick —
LoRALayerWeights,BgmvShrink/BgmvExpand/BgmvExpandSlice,AddLoraLinear, single-sliceLoRALinear) landed in1d37f152, gated bytest_punica_cpu. W2 is the next slice:PackedLoRALayerWeights(vllm/lora/lora_weights.py:99-282) —pack+ per-sliceoptimize+is_packed.add_shrink/add_expandwithoutput_slicesandoffset_start(vllm/lora/punica_wrapper/punica_cpu.py:166-236), and the multi-sliceadd_lora_linear(:265-312).vllm/lora/layers/): replicated, column-parallel, row-parallel, merged column-parallel (gate_up, 2 slices), QKV (1 slice) and merged QKV (3 slices), with TPslice_lora_a/slice_lora_bincluding the fully-sharded (S-LoRA) variants.vocal_parallel_embedding.py) viaadd_lora_embedding, and logits LoRA (logits_processor.py) viaadd_lora_logits+ thevocab_size <= 258048guard.Tests: ports
tests/lora/test_layers.py(pinned vLLM555967922) —test_linear_replicated,test_linear_parallel(row/column x fully_shard),test_column_parallel_packed(repeats 1/2/3),test_embeddings,test_lm_head_logits_processor,test_lm_head_logits_processor_invalid_vocab_size.Out of scope (later Ws in the same spec): metadata/mapping and sgmv segmentation (W3), adapter load/config (W4), the multi-adapter manager (W5), endpoints (W6), GPU punica kernels + fused-MoE packing and the model gate (W7).
Punch-list item 12 (
ROAD-V1-D5, headline[H]) in.agents/specs/roadmap-v1-completion.md.