Repository navigation
Open
Labels
bugSomething isn't workingSomething isn't working
Description
Platform / backend
macOS (Metal)
Model tier
edge0-35b
Hardware / chip
NVIDIA GeForce RTX 5080, 16 GB VRAM
RAM / unified memory
16
OS version
Windows 11 Pro 25H2
Engine build
No response
How are you running it?
None
Routing / LoRA options
None
What happened?
Hi Edge0 team,
I'm trying to run Edge0-145B Q3_K on Windows 11 Pro with an RTX 5080 (16GB VRAM), 32GB RAM and CUDA 13.5.
I've successfully compiled the strata target from your source code:
[179/179] Linking CXX executable strata.exe
However, Windows Smart App Control is enabled and blocks the unsigned Edge0 executables.
I don't want to install Linux, WSL2, or another operating system.
Could you please clarify:
- Is there an officially signed Windows build of Edge0?
- Can I run Edge0-145B entirely on native Windows without disabling Smart App Control?
- Is there an officially supported alternative installation method for systems enforcing application control?
- Can you provide a signed
edge0-infer.exeandstrata.exe, or instructions for obtaining an authorized Windows build? - Is there any supported way to use the source-built engine with the official Edge0 inference launcher?
I have already completed the CUDA compilation successfully, so I would prefer not to repeat the build.
Thank you.
Steps to reproduce
# Title
π Windows 11 Smart App Control blocks unsigned Edge0 executables β request for supported installation method
### Platform / backend
Windows 11 x64 / NVIDIA CUDA 13.5
### Model tier
Edge0-145B Q3_K (intended model; inference not yet started)
### Hardware / chip
NVIDIA GeForce RTX 5080, 16 GB VRAM
### RAM / unified memory
32 GB system RAM
### OS version
Windows 11 Pro 25H2
### Engine build
- Edge0 release downloaded: v0.2.0 Windows x64
- Source repository cloned with submodules.
- Strata CUDA build completed successfully.
- Build output: `[179/179] Linking CXX executable strata.exe`
- `strata.exe` exists.
- Authenticode status: `NotSigned`.
- `edge0-infer info` could not be obtained because the packaged executable was blocked before it could run.
- Exact source commit and engine version: Not yet available.
### How are you running it?
Attempting a native Windows installation using the official Windows release and a source-built CUDA Strata engine.
No inference session has started.
### Routing / LoRA options
Not applicable. The engine has not started, so routing and LoRA options have not been tested.
### What happened?
I am attempting to run Edge0-145B Q3_K natively on Windows 11 using an RTX 5080.
The official Edge0 Windows v0.2.0 release contains unsigned executables. Windows Smart App Control is enabled and blocks `edge0-infer.exe`.
I subsequently cloned the official repository, downloaded the pinned llama.cpp dependency, configured CUDA and successfully compiled the Strata target.
However, the resulting `strata.exe` is also unsigned.
Windows Code Integrity logs show:
- Event ID 3077: executable blocked by Code Integrity.
- Event ID 3033: executable did not meet Enterprise signing-level requirements.
- Event ID 3118: Smart App Control block details.
- Policy ID: `{0283ac0f-fff1-49ae-ada1-8a933130cad6}`
I want to run Edge0 entirely on native Windows without disabling Smart App Control, installing another operating system, or using WSL2.
**Expected behaviour:**
An officially supported Windows installation method that works with Smart App Control enabled, ideally through publisher-signed binaries.
### Steps to reproduce
1. Install Windows 11 Pro 25H2 with Smart App Control enabled.
2. Download the official Edge0 v0.2.0 Windows x64 release.
3. Attempt to run `edge0-infer.exe`.
4. Observe the Windows Application Control block.
5. Clone the official Edge0 repository with submodules.
6. Download the pinned llama.cpp dependency.
7. Configure the CUDA build using CUDA 13.5.
8. Run `cmake --build build --target strata`.
9. Observe successful compilation and linking.
10. Check the resulting executable using `Get-AuthenticodeSignature`.
11. Observe `Status: NotSigned`.
### Logs / stack trace / output
Build output:
`[179/179] Linking CXX executable strata.exe`
Authenticode:
`Status: NotSigned`
Code Integrity:
- Event 3077
- Event 3033
- Event 3118
Smart App Control state:
`On`
### Questions for maintainers
1. Do you provide Authenticode-signed Windows executables?
2. Is there a supported native Windows installation method compatible with Smart App Control?
3. Can you provide signed builds of both `edge0-infer.exe` and `strata.exe`?
4. Is there an officially supported way to use the source-built Strata engine with the Edge0 inference launcher without disabling Windows security?
5. Are signed Windows releases planned for future versions?
I would appreciate specific installation guidance for this configuration.Logs / stack trace / output
Metadata
Metadata
Assignees
Labels
bugSomething isn't workingSomething isn't working
Activity
linyubupa commentedon Nov 14, 2026
Thanks for ruling out RAM and thermals β that is most of the triage already. sm_120 is not the problem: our 0.1.55 build with -DCMAKE_CUDA_ARCHITECTURES=120 runs this same 145B NVFP4 pack on a 5090 D (also sm_120) here, short prompt and all, in under a second. 120 is the right value; 120a/120f only matter for family-specific instructions we don't use.
What I think is happening: the -6 is abort(), not OOM, and it lands at the end of prefill rather than the start, so a CUDA error inside the FP4 path is the shape of it. The difference from our box is the stack β 595.84 / CUDA 13.2.1 / gcc 15.2 against our 580.178.04 / CUDA 13.0 / gcc 12.4. A KVM passthrough with 58 GB of VRAM in use and --pcie-mode dma is the other candidate, since a pinned-memory failure can abort before it prints the [pinned-memory-exhausted] hint.
To settle it I need the engine's stderr, which the delegate's temp log loses. Run the engine directly with the same arguments and 2>&1 | tee /tmp/engine.log, then send one request from another shell β the CUDA error text will name the call. If it aborts silently, gdb -batch -ex run -ex bt --args strata β¦ will give the frame.
Two things that would narrow it further: does the Q3_K pack do the same on that box, and is the GPU passed through whole (no vGPU or MIG)? And the three missing #include in the tests are real β we'll fix them.