Motivation

Several LoRA loading bugs affect correctness and usability:

1. Partial MoE rank loading — when a LoRA adapter has fewer ranks than max_lora_rank, the A-buffer stacked components are placed at wrong offsets, causing the MoE kernel's [:max_r] / [max_r:2*max_r] slicing to read garbage.
2. VL model lm_head LoRA — vision-language models with should_apply_lora patterns skip embed_tokens/lm_head modules because the gate fires before they are handled.
3. Loose LoRA loading — weight name mismatches between the adapter and the model silently drop weights, making debugging difficult. No validation that loaded weights actually match the model's target modules.
4. DeepSeek on-demand loading — the "all" target module sentinel is expanded too early in server_args, before the model is loaded, preventing model-aware resolution of which modules to target.

Modifications

Partial MoE rank fix (mem_pool.py):

• Place stacked A-buffer components at max_rank-spaced positions instead of contiguous lora_rank * c slicing.
• Zero B-buffer beyond the loaded rank so the MoE kernel reads correct padding.

VL model lm_head fix (lora_manager.py, mem_pool.py):

• Move should_apply_lora gate to after embed_tokens/lm_head handling so VL models' patterns don't skip these modules.
• Add lora_lm_head_module is not None guards and PP-stage fallback assertion for non-last ranks.

Strict LoRA loading (server_args.py, lora_manager.py, mem_pool.py):

• Add --lora-strict-loading flag.
• Pre-validate all adapter weight names against target modules before GPU buffer mutation. Log matched/skipped modules, raise on mismatch when strict.
• Keep "all" as a sentinel in server_args and resolve it model-aware in lora_manager using auto_detect_lora_target_modules.

Accuracy Tests

Changes are bug fixes to weight loading — they restore correct behavior rather than changing model outputs.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.