MemForge

Make your GPUs work smarter.

MemForge profiles GPU workloads, explains where performance is being wasted, and suggests safer, faster ways to run them. The Community edition measures and diagnoses; it never modifies your model or framework.

memforge analyze -- python inference.py

MemForge found:

⚠ This looks like Tensor Parallelism across GPUs with no direct peer access (PCIe).
    why        : TP synchronizes between GPUs nearly every layer; over PCIe that can dominate.
    suggestion : Test Pipeline Parallelism. Run 'memforge p2p-test' to measure your link.
    what changes: runtime launch configuration only
    risk       : SAFE

Model changed: NO     Framework source changed: NO

What it is

An automatic performance engineer for GPUs. Point it at a PyTorch script, a vLLM or llama.cpp command, or a CUDA program. It profiles, finds where performance is wasted, and tells you - in plain English - what you can safely do about it.

30-second start

git clone https://github.com/elmehdi-eljair/memforge.git
cd memforge && ./install.sh
memforge doctor
memforge analyze -- python inference.py

What MemForge measures

GPU hardware & topology · memory bandwidth/latency · P2P / multi-GPU transfers · CUPTI profiling · traffic amplification (how much more memory a kernel moves than it needs) · cache behavior · launch overhead · bottleneck detection.

Safety / NOOP philosophy

Every finding states what would change, what would not, the risk, the expected benefit, how it is verified, and how to roll back. And:

MemForge never keeps an optimization because it looks faster. It keeps it only after proving it is correct and faster end to end.

NOOP is a successful MemForge decision. Refusing a change that would not pay back is a feature, not a gap.

Real measured results

On 2× RTX 5060 Ti (no NVLink), MemForge Pro produced:

workload	before	after	speedup
Qwen3-0.6B decode	108.6 tok/s	184.9 tok/s	1.70× (fused kernels, identical output)
Qwen3.6-35B-A3B decode	27.0 tok/s	80.4 tok/s	2.98× (parallelism placement + CUDA graphs)

These are measured results on specific hardware/workloads and are not promised gains for every workload. MemForge measures every optimization rather than assuming it generalizes.

Community vs Pro

Community (this repo, MPL-2.0) profiles, understands and diagnoses. MemForge Pro adds autonomous optimization: it decides exactly what to do, generates CUDA, composes strategies, verifies, benchmarks, and deploys or rolls back. See docs/COMMUNITY_VS_PRO.md.

Docs

docs/QUICKSTART.md · docs/CLI.md · docs/PRODUCT_OVERVIEW.md · docs/COMMUNITY_VS_PRO.md

License

MPL-2.0 (see LICENSE). Trademarks: see TRADEMARKS.md. MemForge Pro/Enterprise are separately licensed and not covered by this license.