Conversation
There was a problem hiding this comment.
Does this change current set the names in Lemonade to be the exact flm name? Like qwen3:8b?
Need to think about how users will know this is an NPU model. In the past we had -NPU, -Hybrid, and -FLM suffixes so they were aware.
This is important because:
- People assume ollama-style model names mean llamacpp GGUFs on GPUs, which is not the case here
- The NPU has unique restrictions (only 1 LLM on NPU at a time) and performance limitations (less memory bandwidth than GPU)
So people should be aware of what they're loading. lemoande-server run qwen3:8b doesn't carry that information.
|
Also if we were to ever add a translation table for ollama model names for use with ollama API that would collide with the FLM model names. |
|
Another musing... it might finally be time to drop "recipes" at the top level. If we support Qwen3-8B on cpu, gpu, and npu, why not just show the model in the model list ONCE and just offer the user a choice of device? edit: one limitation of this strategy is that GPU vs. NPU is different model files. We dont want to trigger a lot of random giant downloads just because someone played with the backend menu. |
|
The problem with offering a list of devices is third party clients. Like think running in openwebui. So I think we really do want to keep suffixes. |
So convert their colons to dashes and append -flm? qwen3:8b becomes qwen3-8b-flm |
ya |
|
Reminder to self: ensure this doesn’t accidentally break the gpt-oss-20b filtering on windows. |
superm1
force-pushed
the
npu-json
branch
2 times, most recently
from
508bf59 to
73ff1c1
Compare
|
I've updated this branch OP to reflect an ongoing TODO. CI failures are expected until that list is all covered. |
superm1
force-pushed
the
npu-json
branch
3 times, most recently
from
60c63c7 to
1827a15
Compare
Deduplicate the backend row markup between BackendManager and ModelManager into a single BackendRow component. Also adds FLM discoverability: the model manager now shows recipe categories with install/upgrade prompts when the backend is supported but not yet installed. Co-Authored-By: Claude Opus 4.6 <[email protected]>
Consolidate FLM detection into 5 ordered states (unsupported, installable, update_required, action_required, installed) determined by system-info cache. Remove 4 redundant NPU checks, 2 version checkers, 2 validate wrappers, and 3 independent check-then-install dances. All code now reads FLM status from SystemInfoCache::get_flm_status() instead of discovering it independently. Key changes: - Add FlmStatus struct and invalidate_recipes() to SystemInfoCache - Split system-info cache into hardware (cached once) and recipes (invalidatable) - Fix is_recipe_installed() for FLM with correct ordered checks - Treat unknown FLM version as update_required (not action_required) - Remove ModelInvalidatedException, FLMCheckException, model invalidation logic - Simplify FastFlowLMServer: remove check(), validate(), check_npu_available() - Add LEMONADE_MOCK_FLM_PATH env var for testability - Add comprehensive FLM status tests (26 tests: 6 scenarios x 5 API actions) - Add flm-status to CI test matrix Co-Authored-By: Claude Opus 4.6 <[email protected]>
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
- Make FlmStatus::is_ready a method instead of a redundant field - Add FlmStatus::error_string() to eliminate 3x duplicated error formatting - Remove double build_recipes_info() call on first /system-info request - Consolidate triplicated #ifdef blocks in FLM branch of build_recipes_info() - Add _mock_flm() context manager to tests, removing 16x temp dir boilerplate Co-Authored-By: Claude Opus 4.6 <[email protected]>
Add allowlist-based path validation to prevent command injection via environment variables (LEMONADE_MOCK_FLM_PATH, PATH). Addresses CodeQL alert about user-derived input reaching popen(). Co-Authored-By: Claude Opus 4.6 <[email protected]>
# Conflicts: # src/app/src/renderer/BackendManager.tsx
Co-Authored-By: Claude Opus 4.6 <[email protected]>
…ogic - FLM install button opens setup instructions iframe instead of showing error - Server returns action URL from /install for backends requiring manual setup - Consolidate backend install/uninstall into shared useBackendInstall hook - Remove redundant recipes cache from BackendManager (SystemInfoCache is single source) - Invalidate model cache after backend install/uninstall so FLM models appear - Models list auto-refreshes on backendsUpdated event Co-Authored-By: Claude Opus 4.6 <[email protected]>
The local branch consolidated install/copy logic into ConnectedBackendRow and useBackendInstall hook, making the inline handlers from origin redundant. Co-Authored-By: Claude Opus 4.6 <[email protected]>
Export is_safe_executable_path() as a public API and call it at each popen site in model_manager.cpp and system_info.cpp. CodeQL traces data flow per compilation unit, so the sanitizer must be visible where the dangerous sink (popen) is called. Co-Authored-By: Claude Opus 4.6 <[email protected]>
Eliminates CodeQL taint-flow alerts by removing the test-specific LEMONADE_FLM_DIR env var from find_flm_executable(). Tests now mock FLM by placing scripts named `flm` in a temp dir prepended to PATH. Mock-FLM scenarios (update_required, action_required, installed) are skipped on Windows where SearchPathA can't be overridden via PATH. Co-Authored-By: Claude Opus 4.6 <[email protected]>
…nditions - Install tests: expect 200 with action URL on Linux (server returns setup page URL instead of auto-installing FLM) - unknown_version_load: accept "requires" as alternative to "unknown" in FLM hint (guards against inter-test race conditions) - Improve server lifecycle: wait for port release between tests to prevent stale server responses Co-Authored-By: Claude Opus 4.6 <[email protected]>
This information first lands upstream in 7.1. Effectively hide the NPU sensor if it's missing.
2 participants
A variety of things need to happen when FLM 0.9.35 releases. This branch tracks them all.
--jsonoutput from FLMRemaining things that we're waiting on to add before merging it.
flm validate --jsonneeds to return values for Windows with NPU driver. The Lemonade code that looks for driver version needs to be dropped and use the json on Windows