[Feature Request] Ability to disable mmproj via an argument and set max context checkpoints
Nobody has claimed this yet.
Assessment
- Difficulty
- 5/5
- Estimated time
- Over a week
- Newbie friendliness
- 35/100
Research direction
Start by reviewing the server launch arguments and --help output for model-loading options, then trace how mmproj weights and context checkpoints are handled. Confirm whether disabling vision/audio weights and limiting checkpoints are supported concepts, and define completion as documented, usable arguments that produce the intended memory behavior.
Written by the indexing model from the issue text.
Description
Currently (correct me if I'm wrong, please - but I can't find a way in docs, or via --help) there is no way to disable loading vision or audio weights (in case of Gemma 4 models) via parameter when launching the server. Since for Gemma 4 E4B those add up to nearly 2 GB and they aren't always used, it would be nice if we could just easily disable them.
I'm not quite sure about the second point since generally there is a very limited number of server params (so I'm not actually sure if the concept translates from llama.cpp) but I wanted to bring it up in case it's right:
Another potential saving, on llama.cpp you can set a maximum number of context checkpoints - I don't know how FLM handles this, but I would believe that there are a few equivalent checkpoints because the memory occupied by the model seems to keep growing over time.
Since I feel like NPU models are best used when loaded alongside another model via a different backend, those memory savings could be really useful. For me, currently I'm hovering around 6GB free memory when loading models for my usecase, and I could stay in that footprint when using two llama.cpp servers but when I use flm to serve the second model the memory use grows by 2 GB or so after use.
- Dominant language
- C++
- Stars
- 1.9k
- Forks
- 152
- Avg merge
- 4h 14m
- Merged PRs (30d)
- 11
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
More from ROCm/FastFlowLM
-
No valid checkpoint to restore + Max length reached on multi-turn tool calls (Qwen3.6-MoE, NPU) Open
Difficulty 2/5 1-3 hours Newbie friendliness 82/100
ROCm/FastFlowLM#744 · 1 comment ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
ROCm/FastFlowLM#741 · 1 comment ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 75/100
ROCm/FastFlowLM#730 ·
-
Download timeout Open
Difficulty 2/5 1-3 hours Newbie friendliness 74/100
ROCm/FastFlowLM#520 · 7 comments ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
ROCm/FastFlowLM#434 · 1 comment ·
Similar issues
-
Difficulty 2/5 1-3 hours Newbie friendliness 86/100
-
Sensor initialization takes very long when `--initial-sim-time` is set to current UNIX timestamp Open
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
gazebosim/gz-sensors#662 · 1 comment ·
-
enhancement
Difficulty 2/5 1-3 hours Newbie friendliness 76/100
-
comp-datalake
Difficulty 2/5 1-3 hours Newbie friendliness 88/100
ClickHouse/ClickHouse#121222 ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
LadybirdBrowser/ladybird#12123 ·