[AMD NPU] models fail to load with "invalid unordered_map<K,T> key" on Foundry Local 0.8.113

[MedAnd]

Description

On AMD Ryzen AI systems, all AMD Vitis NPU models fail during the load step in Foundry Local, even though the models download successfully and appear in the local cache. The error occurs consistently across multiple models (Phi-4-mini, Qwen2.5-Coder, etc.).

Environment

Foundry Local version:
0.8.113+70167233c5

AMD NPU Execution Provider (Windows ML):
Name: MicrosoftCorporationII.WinML.AMD.NPU.EP.1.8
Version: 1.8.26.0
OS: Windows 11 24H2 (Copilot+ PC)

Repro Steps

 foundry model run qwen2.5-coder-7b-instruct-vitis-npu

Expected Behavior

Model should load successfully into Foundry Local service and be available via local OpenAI-compatible endpoint.

Actual Behavior

Fails during EP initialization with:

invalid unordered_map<K,T> key"

Additional Notes

• CPU variants load and run fine.
• NPU EP and Foundry Local are both latest versions.
• This appears identical to Issue [AMD NPU] qwen2.5-0.5b-instruct-vitis-npu fails to load with "invalid unordered_map<K, T> key" #327

[natke]

@MedAnd we have not been able to reproduce this error and continue to investigate as to why you are seeing this error. Can you please share your AMD NPU driver version? And also run again with Debug logging:

foundry service set --default-log-level Debug and share relevent logs from ~/.foundry/logs

[OrenKohavi]

I'm not the OP but am running into the same issue on version 0.8.113+70167233c5 and windows 25H2 (Build 26200.7171)
CPU is a Ryzen 7840S (not Copilot+, but has an AMD NPU)

Here's the log produced after running the relevant commands:

foundry20251206.log

[MedAnd]

@natke - as requested please find logs: foundry20251208.log

AMD NPU Driver version 32.0.203.304

[natke]

Thank you both. We are investigating the issue

[TPollmeier]

running into same issue

• foundry --version
  0.8.117
• NPU Driver:
  32.00.0203.314
• log
  foundry20260115.log

[xxyk]

I think maybe XDNA1 doesn't support LLM(OGA)...It only supports CNN INT8...

[nicholasdbrady]

Hitting a similar issue but with a different error signature and newer EP version — adding details in case it's useful.

foundry20260323.log

Environment:

• Foundry Local version: 0.8.119
• AMD NPU EP version: MicrosoftCorporationII.WinML.AMD.NPU.EP.1.8 → 1.8.55.0
• CPU/NPU: AMD Ryzen AI 7 PRO 350 (XDNA2 architecture)
• OS: Windows 11
• Model attempted: Mistral-7B-Instruct-v0-2-vitis-npu:1

Steps to reproduce:

foundry model run Mistral-7B-Instruct-v0-2-vitis-npu:1

Error:

[ERR] Failed loading model: Mistral-7B-Instruct-v0-2-vitis-npu:1
Exception: Failed: Loading model Mistral-7B-Instruct-v0-2-vitis-npu:1 from http://127.0.0.1:52251/openai/load/Mistral-7B-Instruct-v0-2-vitis-npu:1?ttl=600
Internal Server Error
Exception during initialization: [C:\Users\z1aiebuild\dod\src\ops\llm_ops\mladfrmsnorm\mladfrmsnorm.cpp:254]
[ERROR] No supported shapes with matching txn_fname_prefix found

Notes:

• Model was found in local cache — download is not the issue
• This error is different from the invalid unordered_map<K,T> key others have reported — specifically failing at the mladfrmsnorm RMS norm op during NPU kernel dispatch
• Running on XDNA2 (Ryzen AI 7 PRO 350), EP version 1.8.55.0 — newer than what prior reporters had (1.8.26.0), so upgrading the EP does not resolve this
• CPU variant runs fine as a workaround

[leestott]

Hardware validation

Windows 11 Enterprise 10.0.26220 Build 26220
HP EliteBook 8 G1a
AMD Ryzen AI 7 Pro 350w/ Radeon 860M
Memory 64GB
AMD driver
32.0.22050.1003

Update driver at www.amd.com/en/supporr/download/drivers.html

[MedAnd]

Hi @natke - it's been several months and wondering you can share an update on this issue? Thx.

[AgentEpsilon]

Adding that I'm seeing the same issue. Is there an expected fix for this (or at least better detection of unsupported hardware, if that is the issue)? CPU and GPU models work fine, but that doesn't change the fact that Foundry Local still auto-detects NPU support by default.

> foundry model run phi-4-mini
Model phi-4-mini-instruct-vitis-npu:2 was found in the local cache.
🕗 Loading model... [11:33:05 ERR] Failed loading model:phi-4-mini-instruct-vitis-npu:2
Exception: Failed: Loading model phi-4-mini-instruct-vitis-npu:2 from http://127.0.0.1:60274/openai/load/phi-4-mini-instruct-vitis-npu:2?ttl=600
Internal Server Error
Failed loading model phi-4-mini-instruct-vitis-npu:2
Exception during initialization: invalid unordered_map<K, T> key

foundry20260413.log

Foundry Local version: 0.8.119
Hardware: AMD Ryzen 7 PRO 7840U w/ Radeon 780M Graphics
NPU Driver version: 32.0.203.280

[MedAnd]

Hi @natke / @samuel100 - any update on this issue, it's been close to five months and affecting multiple customers / users as per above...?

[samuel100]

@MedAnd Please use the latest CLI https://github.com/microsoft/Foundry-Local/releases/tag/cli-preview-0.10.0 that uses the latest EP/Hardware discovery APIs.

In the CLI you can list by different devices:

You can also, see the different variants of an alias:

You can then plug the full model ID into foundry chat qwen2.5-0.5b-instruct-generic-gpu:4 to force a specific variant.

I am closing the issue but please re-open if you feel this does not address the issue.