[Bug] flm serve crashes with Access Violation (0xc0000005) at start of token generation on Ryzen AI NPU (XDNA2)

Hi, I encounter a systematic crash when trying to stream model answer through ip:port access, from a third party app, that I tried to diagnostic with the help of gemini. Here is the report :

Describe the bug

When running flm serve on a Windows system with a Ryzen AI (XDNA2) NPU, the server crashes with a segmentation fault/access violation (0xc0000005) immediately when it starts generating tokens. This happens for both streaming and non-streaming requests, and has been confirmed to affect multiple models (gemma4-it:e4b and llama3.2:1b), crashing at the exact same fault offset.

Environment

• OS: Windows 11
• FLM Version: 0.9.44
• NPU Architecture: XDNA2
• NPU Driver Version: 32.0.20102.3930
• Models tested:
  • gemma4-it:e4b
  • llama3.2:1b

Steps to Reproduce

1. Start the FLM server:

   "C:\Program Files\flm\flm.exe" serve --host 0.0.0.0 --port 11434 --cors 0
   

2. Send a POST request to /api/chat:

   {
       "model": "llama3.2:1b",
       "messages": [
           {
               "role": "user",
               "content": "Write a one-sentence greeting."
           }
       ],
       "stream": true
   }
   

3. The server locks the NPU, loads the model, begins generation, and then crashes/terminates immediately.

Console Logs (Llama 3.2 1B)

[🟢 ]  NPU Locked!
[FLM]  Loading model: C:\Users\Adminlocal\.flm\models\Llama-3.2-1B-NPU2
[FLM]  Start generating...
[FLM]  Prefill chunk 1/1 with 42 tokens
<Crash / Exit code 1>

Windows Event Viewer Log (Same for both Gemma 4 and Llama 3.2)

Faulting application name: flm.exe, version: 0.0.0.0
Faulting module name: flm.exe, version: 0.0.0.0
Exception code: 0xc0000005 (Access Violation)
Fault offset: 0x00000000002c6823
Path: C:\Program Files\flm\flm.exe

Additional context

flm validate output:

[Windows]  NPU: XDNA2
[Windows]  NPU driver version: 32.0.20102.3930

Start with the flm serve entry point and reproduce the crash by sending the documented POST request to /api/chat on the specified Ryzen AI XDNA2 setup. Investigate the access violation reported at offset 0x2c6823; done means both streaming and non-streaming generation complete without the server terminating.