#2453 Windows/NVIDIA performance regression: v0.1.42 ~69 tok/s vs v0.1.44.1 ~55–58 tok/s on RTX 5060 Ti

Beschreibung

Hi, I am seeing a fairly large decode performance regression on Windows/NVIDIA after updating Strata.

## System

- OS: Windows 11
- CPU: Intel Core i9-14900K
- GPU: NVIDIA GeForce RTX 5060 Ti 16 GB
- RAM: 128 GB
- Model: Qwen3.8-Flash-Next
- Quant: IQ3_S
- Context: 64K
- KV cache: int8
- Images: OFF
- Experimental Speed Projection: ON

I have not manually changed the tuning parameters from the values generated by Strata setup.

## Performance

### Full v0.1.42 checkout

I cloned the `v0.1.42` tag into a separate directory and started it from there.

Observed:

- RAM usage: about 64 GB
- Decode speed: **69.2 tok/s**

The runtime log for this run shows:

- `session is up (engine 0.1.42)`
- `expert cache 4269 slots, 8.14 GiB of VRAM`
- `2078 generated in 30043 ms (69.2 tok/s)`
- `decode expert cache hit rate: 91.1%`
- `KV streaming: 99.93%`

I have attached this v0.1.42 runtime log.

### Full v0.1.44.1 checkout

After updating my normal Strata checkout to v0.1.44.1, I repeatedly observed:

- RAM usage: about 68 GB
- Decode speed: approximately **55–58.8 tok/s**

This was repeatable across multiple generations.

Before the update, v0.1.43.1 on this same machine was also running in roughly the same performance range as v0.1.42 (around 67–69 tok/s).

## Engine rollback test

I also tried:

`python setup.py --rollback-engine`

This restored engine v0.1.43.1 while keeping the newer Strata checkout.

In that configuration:

- RAM usage decreased slightly to about 66 GB
- Decode speed remained around 55–58 tok/s

So rolling back only the engine binary did not restore the previous performance.

When I instead ran the complete v0.1.42 checkout, RAM usage returned to about 64 GB and decode performance immediately returned to 69.2 tok/s.

This makes me wonder whether the regression may involve something outside the engine binary itself, such as generated configuration, expert/memory placement, or another change in the surrounding Strata tree.

## PCIe check

I also monitored the GPU PCIe link while generating.

The RTX 5060 Ti remained at **PCIe Gen5 x8 under load**, so I do not appear to be seeing the Gen1 link-drop behavior reported in another issue.

## Logs

Attached:

- `strata_full_0.1.42_69.2tps.log`
  - Full v0.1.42 checkout
  - 69.2 tok/s run

- `strata_newtree_0.1.44.1_rollback_0.1.43.1.log`
  - Contains the v0.1.44.1 runs at approximately 55–58.8 tok/s
  - Also contains the subsequent engine-only rollback to v0.1.43.1, which remained in the same slower performance range

[strata_v0.1.44.1_and_rollback_v0.1.43.1.log](https://github.com/user-attachments/files/33442461/strata_v0.1.44.1_and_rollback_v0.1.43.1.log)

↓strata_full_0.1.42_69.2tps.log
[strata-iq3_s.log](https://github.com/user-attachments/files/33442083/strata-iq3_s.log)