Add mixed precision support for LoRA expand & shrink kernels

[chaojun-zhang]

… and inputs

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.

Purpose

Add mixed precision support for LoRA expand & shrink kernels (used in vllm ) with float32 inputs and float16 weights.

Test Plan

pytest -s -v tests/test_lora_ops::test_kernels_mixed_precision
vllm side: pytest -s -v tests/lora/test_layers.py

Test Result

(Optional) Documentation Update

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)

[Copilot]

Pull request overview

Adds mixed-precision support to the XPU LoRA BGMV kernels so they can accept float32 inputs with fp16/bf16 LoRA weights, and introduces a new test variant to exercise this path.

Changes:

• Extend bgmv_shrink / bgmv_expand_slice kernels to dispatch input and weight dtypes independently (allowing float32 inputs with fp16/bf16 weights).
• Update test data generation utilities to allow separate input_dtype vs weight_dtype.
• Add a new test_kernels_mixed_precision test matrix and mini-profiler params.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 6 comments.

File	Description
tests/test_lora_ops.py	Adds mixed-precision test coverage and threads input_dtype through kernel checks.
tests/lora/utils.py	Allows generating inputs with a different dtype than the LoRA weights.
csrc/xpu/lora/lora_shrink.cpp	Enables mixed dtype dispatch for shrink (independent input/weight types) and updates validation.
csrc/xpu/lora/lora_expand.cpp	Enables mixed dtype dispatch for expand-slice and updates validation.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

This PR’s description template is still unfilled (missing Purpose and Test Result details). Please update the PR description so reviewers can understand intent and validate changes more quickly.

[Copilot]

This file now uses std::min to compute vec_size, but it does not include <algorithm>. Please add the direct header include to avoid relying on transitive includes (which can break builds under different toolchains).

[Copilot]

The new dispatch logic uses std::remove_const_t / std::remove_pointer_t but this file does not include <type_traits>. Add the explicit include to avoid build fragility from transitive includes.

[Copilot]

This file now uses std::min to compute vec_size, but it does not include <algorithm>. Please add the direct header include to avoid relying on transitive includes (which can break builds under different toolchains).

[Copilot]

The new dispatch logic uses std::remove_const_t / std::remove_pointer_t but this file does not include <type_traits>. Add the explicit include to avoid build fragility from transitive includes.

[Copilot]

The new mixed-precision test sets input_dtype=torch.float32, but the Torch reference path (tests/lora/torch_ops.py) casts inputs to output_tensor.dtype (fp16/bf16) before computing. That means this test may not actually validate the float32-input execution path (it may just compare against a fp16/bf16 reference). Consider updating the reference calculation (or adding a special-case in this test) so the reference keeps float32 inputs and only casts the final result to the output dtype as needed.