Inconsistent memory consumption?
Maintainers usually reply within 5 days
Nobody has claimed this yet.
Assessment
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Newbie friendliness
- 35/100
- Issue type
- Bug
- Clarity
- Needs clarification
- Activity status
- Active
- Domain
- ai, performance
Research direction
Start by reproducing the reported RAM measurements for the QWen3.5 2B, 4B, and 9B models with 4k and 8k contexts, recording usage at load and inference. Trace the inference and memory-management paths involved; done means determining whether the shared 27GB usage is expected or isolating a reproducible defect.
Written by the indexing model from the issue text.
Description
This might not be an issue but a misunderstanding.
I run FastFlowLM on a Ryzen AI 7 360 PRO NPU with 64GB VRAM.
The purpose is to run visual tasks on NPU and I am selecting a model for perf/RAM among the QWen3.5 family. I try 2B, 4B, and 9B models with 4k/8k context window size to concisely read short text from a 1080x720pix PNG image.
What is confusing: On model load, the consumed RAM is slightly different - OK, but when I run inference on the models they all end up with the same 27GB consumed RAM, which is very confusing. Also the context window size does not have an effect. I expected to see less RAM consumed by smaller models.
Here are RAM consumption details:
- QWen3.5-2B
- On load: 4.2GB
- On inference: 27GB
- QWen3.5-4B
- On load: 5.5GB
- On inference: 27GB
- QWen3.5-9B
- On load: 9.0GB
- On inference: 27GB
Am I supposed to expect the same RAM consumption, no matter the LLM size: 2B/4B/9B?
- Dominant language
- C++
- Stars
- 1.9k
- Forks
- 156
- Avg merge
- 5d 2h
- Merged PRs (30d)
- 12
Getting set up
- Ships a Dockerfile or Docker Compose file
- No pull request template
- Read the contributing 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
-
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
ROCm/FastFlowLM#757 ·
Maintainers usually reply within 5 days
-
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 · 2 comments ·
Maintainers usually reply within 5 days
-
Truncated tool call is emitted as a complete one (unterminated heredoc reaches the client)Possibly taken @Javinator9889 claimed this 17 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
ROCm/FastFlowLM#741 · 1 comment ·
Maintainers usually reply within 5 days
-
Download timeoutPossibly taken @stdevPavelmc claimed this 145 days ago. Open
Difficulty 2/5 1-3 hours Newbie friendliness 74/100
ROCm/FastFlowLM#520 · 7 comments ·
Maintainers usually reply within 5 days
-
Difficulty 2/5 1-3 hours Newbie friendliness 68/100
ROCm/FastFlowLM#434 · 1 comment ·
Maintainers usually reply within 5 days
Similar issues
-
Make Catch2 optional when `RDK_BUILD_CPP_TESTS=OFF`Possibly taken @pechersky claimed this today. Openbug
Difficulty 2/5 1-3 hours Newbie friendliness 82/100
Maintainers usually reply within 2 days
-
bug
Difficulty 2/5 1-3 hours Newbie friendliness 70/100
dice-group/dice-hash#111 ·
-
Difficulty 2/5 1-3 hours Newbie friendliness 64/100
-
Difficulty 2/5 1-3 hours Newbie friendliness 88/100
MerginMaps/mobile#4741 ·
Maintainers usually reply within 1 day
-
Difficulty 2/5 1-3 hours Newbie friendliness 78/100
ros-perception/image_pipeline#1198 ·