[richard-for-bugen]

Motivation

vllm-project#60024 ("[LoRA] Remove tensorizer", merged 2026-10-05) removed tensorizer_config_dict from Python's LoRARequest (vllm/lora/request.py). It is a msgspec array_like=True, omit_defaults=True struct, so it is encoded positionally and every field after the removed one moved down by one: the wire layout is now [lora_name, lora_int_id, lora_path, base_model_name, load_inplace, is_3d_lora_weight].

The Rust mirror LoraRequest (rust/src/engine-core-client/src/protocol/lora.rs) still carried tensorizer_config_dict: Option<OpaqueValue> at position 4. Rust therefore encoded seven positional fields and sent null where the engine now expects load_inplace: bool:

msgspec.ValidationError: Expected `bool`, got `null` - at `$[4]`

Concrete effect: with VLLM_USE_RUST_FRONTEND=1, vllm serve <model> --enable-lora --lora-modules Alice=... dies at startup (failed to load LoRA adapter Alice from --lora-modules), and /v1/load_lora_adapter, the gRPC LoadLora and per-request lora_request hit the same decode error, since all of them encode a LoraRequest.

This PR removes the field so the Rust struct matches Python's six-field layout. Rust never decodes a LoraRequest from Python in production (only the mock engine and tests decode what Rust encoded, so the derived Deserialize_tuple stays for them), and no other Rust struct mirrors anything else changed by vllm-project#60024 (the rest of it is engine-side tensorizer/PEFT code with no wire presence).

Ownership and layering

The consumer is the engine core's msgspec decoder of EngineCoreRequest.lora_request and the add_lora utility argument, so the owner is the wire struct in vllm-engine-core-client (protocol/lora.rs); the two construction sites (vllm-server LoraManager, a vllm-llm test) only drop the None initializer. No compatibility shim for pre-vllm-project#60024 engines: frontend and engine ship together.

Regression test

The existing python_msgpack_fixtures_match_rust_encoding mechanism (src/tests/python_compat.py) had no LoRA coverage, and its lora_request slot was an untyped object. This PR adds a Python LoRARequest mirror and an EngineCoreRequest fixture with every LoRA field non-default (so omit_defaults trims nothing and the full positional layout is pinned), and asserts both directions: Python wire -> Rust struct, and Rust struct -> wire for the lora_request slot of the request. The mirror's bytes were checked to equal the real vllm.lora.request.LoRARequest encoding of the same values.

Validation

• cargo fmt --check, cargo sort --check, cargo deny check bans, cargo clippy --workspace --all-targets --all-features --locked -D warnings: clean.
• cargo nextest run --workspace --all-features --locked --no-fail-fast: 2255 passed, 13 skipped.
• Red-green: restoring tensorizer_config_dict (with the test's expected value adjusted to compile) makes python_msgpack_fixtures_match_rust_encoding fail: the Rust decode of the Python fixture yields tensorizer_config_dict: Some(true), is_3d_lora_weight: false, i.e. fields shifted; with the fix it passes.
• End to end (1 GPU, --gpu-memory-utilization 0.12 via a pytest plugin because the GPU is shared), pytest entrypoints/scale_out/token_in_token_out/test_serving_tokens.py -k lora (test_generate_with_lora_adapter, server with --enable-lora --lora-modules Alice=... Bob=...):
  • before (base aa31fa038f): 1 error, server dies at startup with Expected bool, got null - at $[4].
  • after (this PR): 1 passed (Alice and Bob load at startup, /v1/models lists them, generation through the adapter succeeds). Python frontend on the same test: 1 passed.

Codex review

Verdict: approve. No blocking findings, no nits. It confirmed the problem against the Python decoder, that both Rust consumers (generation via EngineCoreRequest encoding, adapter loading via add_lora utility args) go through the changed serializer, that the new test fails on the base with the shifted fields and passes on HEAD (140 engine-core-client tests), and that the fixture bytes match the real LoRARequest for default and mixed-boolean values. Its only remark was that the E2E result and workspace-wide checks were author-reported, which this body now carries.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XTSySQLs9bSuTKD9ScKhDb

[BugenZhao]

Duplicated with vllm-project#63295