Experimental Strata engine support for Gyro-M on 12 GB VRAM and 32 GB RAM
Python
11
1 commits
updated Oct 14, 2026
Patched Strata for Gyro-M on Windows: 12 GB NVIDIA VRAM, 32 GB RAM, 256K text context. Experts run across VRAM, RAM and SSD.
Speeds depend heavily on your CPU, GPU bandwidth, RAM and SSD. Use this as a baseline; don't expect identical results.
Requires Windows x64, an AVX2 CPU, an NVIDIA GPU with compute capability 7.5+, 12 GB VRAM and a local SSD. Hardware details.
Prerequisites: Git, Python 3.12–3.14, VS 2022 C++ Build Tools with C++/Windows SDK, CUDA Toolkit 13.2, and NVIDIA driver R580+.
Allow 120 GiB for weights/pack, 4 GiB for engine files, plus prerequisites. Keep the engine folder in a permanent location.
Double-click: DOWNLOAD-MODEL.cmd → choose a folder → INSTALL.cmd → Start-Gyro-M-256K.cmd beside the weights.
PowerShell: Change the model path:
git clone https://github.com/Gitshop77/GyroStrata.git
cd GyroStrata
$ModelDir = "C:\Your Model Folder"
.\DOWNLOAD-MODEL.cmd "$ModelDir"
.\INSTALL.cmd --model-dir "$ModelDir"
& "$ModelDir\Start-Gyro-M-256K.cmd"
Already have the pinned GGUF? Skip DOWNLOAD-MODEL.cmd. Interrupted downloads resume in the same folder. Moving the engine or weights requires reinstalling with the new path.
The installer downloads source, applies patches, builds for your GPU and checks parity. Original weights stay unchanged.
API: http://127.0.0.1:8080/v1 · Model: qwen38-gyro-m · Stop: Ctrl+C.
No API key by default. Binding to 0.0.0.0 allows network access. Optional: install with --api-key; the key is saved in .gyro-recipe\api-key.txt. DELETE-API-KEY.cmd removes it; restart afterward.
Needs 14 GiB available RAM at startup. Runtime guards stop the model above min(11.5 GiB VRAM, reported VRAM minus 512 MiB) or below 512 MiB available RAM. Sampling can miss brief spikes. Q4 KV and BF16 pack conversions have precision tradeoffs.
Three runs on one 12 GB / 32 GB PC: synthetic Python continuation, thinking off, 256 tokens per reply. Prefill is the initial prompt; output rate covers its reply plus three prefix-reusing replies. Other workloads may be slower.
| Context | Prefill tok/s (mean, range) | Output tok/s (mean, range) |
|---|---|---|
| 8K | 953 (951–956) | 29.7 (28.0–30.8) |
| 32K | 1,552 (1,542–1,558) | 36.4 (35.7–36.8) |
| 64K | 1,725 (1,725–1,726) | 38.4 (36.9–40.6) |
| 128K | 1,791 (1,787–1,794) | 37.7 (36.9–38.3) |
| 192K | 1,785 (1,782–1,786) | 36.8 (35.6–37.8) |
| Near 256K | 1,755 (1,753–1,758) | 36.6 (35.4–37.8) |
Long-context retrieval, arithmetic and cache-return checks passed. The 32K GPU cache does not limit attention to 32K; leave room for output within 256K.
Patch attribution · Updates and rollback
Paste this prompt. It follows our tuning workflow; results on other PCs are not guaranteed.
Install and tune https://github.com/Gitshop77/GyroStrata for my Windows PC.
Read README.md, HARDWARE.md and PATCHES.md.
Inspect my GPU/VRAM, CPU, RAM speed, SSD performance, PCIe link and toolchain.
Reuse existing pinned Gyro-M weights; install missing requirements and run
setup/parity checks. Measure the baseline before changing settings.
Tune workers/affinity, task chunks, RAM expert placement, CPU/GPU scheduling,
VRAM reserve, prefill and SSD read merging using short, matched trials.
Keep stable gains within the memory guards, preserving requested context
and answer quality. Verify representative tasks, recall and cache reuse.
Run the bundled context grid once before and once after tuning. Repeat only
unclear or unstable results. Report prefill/output tok/s, first-token latency
and memory use. Reuse completed checks unless changes affect them.
Edit source only for a measured bottleneck; update patches/pins, rebuild
and check correctness. Preserve weights and rollback. Ask before driver
changes or destructive actions; leave other models untouched.
Leave one tuned launcher beside the weights and a short results table.
Stop test servers. Do not publish changes.
If performance is poor, open an issue with:
GPU and VRAM:
CPU:
RAM capacity and speed:
SSD model/type:
Windows, driver and engine version:
Context and launch settings:
Prefill and output tok/s:
Keep private prompts out of logs.
Experimental Strata engine support for Gyro-M on 12 GB VRAM and 32 GB RAM
Python
11
1 commits
updated Oct 14, 2026
Patched Strata for Gyro-M on Windows: 12 GB NVIDIA VRAM, 32 GB RAM, 256K text context. Experts run across VRAM, RAM and SSD.
Speeds depend heavily on your CPU, GPU bandwidth, RAM and SSD. Use this as a baseline; don't expect identical results.
Requires Windows x64, an AVX2 CPU, an NVIDIA GPU with compute capability 7.5+, 12 GB VRAM and a local SSD. Hardware details.
Prerequisites: Git, Python 3.12–3.14, VS 2022 C++ Build Tools with C++/Windows SDK, CUDA Toolkit 13.2, and NVIDIA driver R580+.
Allow 120 GiB for weights/pack, 4 GiB for engine files, plus prerequisites. Keep the engine folder in a permanent location.
Double-click: DOWNLOAD-MODEL.cmd → choose a folder → INSTALL.cmd → Start-Gyro-M-256K.cmd beside the weights.
PowerShell: Change the model path:
git clone https://github.com/Gitshop77/GyroStrata.git
cd GyroStrata
$ModelDir = "C:\Your Model Folder"
.\DOWNLOAD-MODEL.cmd "$ModelDir"
.\INSTALL.cmd --model-dir "$ModelDir"
& "$ModelDir\Start-Gyro-M-256K.cmd"
Already have the pinned GGUF? Skip DOWNLOAD-MODEL.cmd. Interrupted downloads resume in the same folder. Moving the engine or weights requires reinstalling with the new path.
The installer downloads source, applies patches, builds for your GPU and checks parity. Original weights stay unchanged.
API: http://127.0.0.1:8080/v1 · Model: qwen38-gyro-m · Stop: Ctrl+C.
No API key by default. Binding to 0.0.0.0 allows network access. Optional: install with --api-key; the key is saved in .gyro-recipe\api-key.txt. DELETE-API-KEY.cmd removes it; restart afterward.
Needs 14 GiB available RAM at startup. Runtime guards stop the model above min(11.5 GiB VRAM, reported VRAM minus 512 MiB) or below 512 MiB available RAM. Sampling can miss brief spikes. Q4 KV and BF16 pack conversions have precision tradeoffs.
Three runs on one 12 GB / 32 GB PC: synthetic Python continuation, thinking off, 256 tokens per reply. Prefill is the initial prompt; output rate covers its reply plus three prefix-reusing replies. Other workloads may be slower.
| Context | Prefill tok/s (mean, range) | Output tok/s (mean, range) |
|---|---|---|
| 8K | 953 (951–956) | 29.7 (28.0–30.8) |
| 32K | 1,552 (1,542–1,558) | 36.4 (35.7–36.8) |
| 64K | 1,725 (1,725–1,726) | 38.4 (36.9–40.6) |
| 128K | 1,791 (1,787–1,794) | 37.7 (36.9–38.3) |
| 192K | 1,785 (1,782–1,786) | 36.8 (35.6–37.8) |
| Near 256K | 1,755 (1,753–1,758) | 36.6 (35.4–37.8) |
Long-context retrieval, arithmetic and cache-return checks passed. The 32K GPU cache does not limit attention to 32K; leave room for output within 256K.
Patch attribution · Updates and rollback
Paste this prompt. It follows our tuning workflow; results on other PCs are not guaranteed.
Install and tune https://github.com/Gitshop77/GyroStrata for my Windows PC.
Read README.md, HARDWARE.md and PATCHES.md.
Inspect my GPU/VRAM, CPU, RAM speed, SSD performance, PCIe link and toolchain.
Reuse existing pinned Gyro-M weights; install missing requirements and run
setup/parity checks. Measure the baseline before changing settings.
Tune workers/affinity, task chunks, RAM expert placement, CPU/GPU scheduling,
VRAM reserve, prefill and SSD read merging using short, matched trials.
Keep stable gains within the memory guards, preserving requested context
and answer quality. Verify representative tasks, recall and cache reuse.
Run the bundled context grid once before and once after tuning. Repeat only
unclear or unstable results. Report prefill/output tok/s, first-token latency
and memory use. Reuse completed checks unless changes affect them.
Edit source only for a measured bottleneck; update patches/pins, rebuild
and check correctness. Preserve weights and rollback. Ask before driver
changes or destructive actions; leave other models untouched.
Leave one tuned launcher beside the weights and a short results table.
Stop test servers. Do not publish changes.
If performance is poor, open an issue with:
GPU and VRAM:
CPU:
RAM capacity and speed:
SSD model/type:
Windows, driver and engine version:
Context and launch settings:
Prefill and output tok/s:
Keep private prompts out of logs.