fix(lora): sync with vLLM 0.18.0 and update LoRA tests

[junstar92]

🚀 Summary of Changes

This PR focuses on LoRA test coverage and validation for the RBLN torch-compile path.

• Updated vllm_rbln/lora/layer.py to realign the RBLN LoRA embedding path with the current upstream implementation while preserving the RBLN-specific input-shape handling.
• Added a basic LoRA E2E smoke test at tests/torch_compile/e2e/v1/lora/test_basic_lora.py using a single SQL prompt.
• Updated the experimental LoRA example at examples/experimental/run_lora_test.py to use:
  • meta-llama/Llama-3.2-3B-Instruct
  • jeeejeee/llama32-3b-text2sql-spider
• Applied the same runtime configuration in tests via monkey patching:
  • VLLM_RBLN_ENFORCE_MODEL_FP32=1
  • VLLM_RBLN_ENABLE_WARM_UP=0
  • VLLM_RBLN_USE_VLLM_MODEL=1
  • VLLM_DISABLE_COMPILE_CACHE=0
• Relaxed TorchDynamo recompile limits in the LoRA unit tests to reduce excessive recompilation and remove fallbacks during test runs.

---

📌 Related Issues / Tickets

• Resolves #
• Related to #

---

✅ Type of Change

• 🚀 Release (release)
• ✨ Feature (feature)
• 🧠 Model support (model)
• 🧬 Core engine changes (core)
• 🛠 Bug fix (fix)
• ⚙️ Performance improvement (perf)
• 🔁 Refactor or code cleanup (refactor)
• 📄 Documentation (docs)
• ❓ Other (other): please describe

---

🧪 How to Test

1. Run unit tests: pytest tests/torch_compile/unit/v1/lora
2. Run the new E2E smoke test: pytest tests/torch_compile/e2e/v1/lora/test_basic_lora.py
3. Run the example script: python examples/experimental/run_lora_test.py

Verify output:

• The basic E2E test should produce non-empty outputs for both base and LoRA generation, and the outputs should differ.
• The example script should run with the updated model/LoRA pair.
• Known local unit-test result on this branch: 25 failed, 111 passed.

Known failures:

• test_embeddings fails in test_layers.py.
  • The same math appears to pass in eager mode.
  • The failure seems to happen only after torch.compile(..., backend="rbln").
  • This suggests a compile/lowering issue rather than a pure PyTorch math issue.
• test_lora_functions.py fails locally with:
  • RuntimeError: RBLNRuntimeError: RBLN_DEVICES environment variable changed at runtime. Initial value: , Current value: 0
  • This does not currently look like a test-logic issue.
  • It may be environment-specific and may only be reproduciable on my side.

---

📸 Screenshots / Logs (if applicable)

---

📋 Checklist

• PR title follows Conventional Commits format
• This PR is linked to an existing issue
• The test method is described, and the expected result is clearly stated
• Relevant documentation has been updated (if applicable)

---

💬 Notes

• The example still shows broken or corrupted output in compiled mode.
• The same example behaves normally in eager mode, so this also appears to be a compile-specific issue.
• The exact failure point in the compiled path is still unclear and has been difficult to debug from our side.
• Review and investigation from the Rebellions compiler/runtime side is likely required.

---

[rebel-jinhwan]

https://github.com/rebellions-sw/fsw-inference/issues/144
https://github.com/rebellions-sw/fsw-inference/issues/165

[rebel-jiwoopark]

lgtm

[rebel-jiwoopark]

@rebel-jinhwan Could you check if any additional review is needed?