[TRTLLM-6683][feat] Support LoRA reload CPU cache evicted adapter#6510

Merged

shaharmor98 merged 9 commits intoNVIDIA:mainfrom

amitz-nv:dev-support-torch-lora-reload-adapter-after-cpu-cache-eviction

Aug 7, 2025

Commits on Aug 6, 2025

PeftCacheManager.add_request_peft loads LoRA adapter from filesystem when it's not in LoRA CPU cache and it doesn't have weights attached, worker.py _enqueue_request now always passes the LoRA conf…
amitz-nv
committed
Fix CPP tests to call LlmRequest constructor with the new loraPath argument
amitz-nv
committed
Fix inline comment in microBatchSchdulerTest.cpp to match convention
amitz-nv
committed
Fix more CPP tests to call LlmRequest constructor with the new loraPath argument
amitz-nv
committed
Validate LoraConfig with weights also has config tensor, fix loraConfigTest accordingly
amitz-nv
committed
Move lora path to be a pytorch only field in Request and LlmRequest classes, revert lora path changes in CPP
amitz-nv
committed
Restored accidentally erased bindings for LlmRequest::removeLoraTensors, added a test that verifies the flow it's required in
amitz-nv
committed
PeftCacheManager.add_request_peft - Using unsqueeze instead of reshape
amitz-nv
committed
Fix TestResourceManager::_create_request function to create wrapper LlmRequest instance instead of bindings.internal.batch_manager.LlmRequest
amitz-nv
committed