[CLI] Catalog returns zero non-CPU (GPU/NPU) model variants on Snapdragon X Elite despite both being detected and EPs loading successfully

[steamedeo]

Problem

On a Snapdragon X Elite (X1E80100) laptop running Windows 11, foundry model list and foundry model list --variants show 43 available variants — all 43 are CPU-only. Zero GPU or NPU variants appear anywhere in the catalog, despite:

• foundry status correctly detecting both the Adreno X1-85 GPU and the Hexagon NPU.
• foundry status printing a GPU-specific warning ("Qualcomm Adreno GPU detected. If acceleration fails, try a CPU model variant or update the GPU driver.") — implying GPU variants are expected to be attempted, yet none exist in the catalog to attempt.
• foundry server start explicitly downloading and reporting success for QNNExecutionProvider.
• Attempting to run a known NPU model ID directly (foundry run phi-3.5-mini-instruct-qnn-npu) fails with "No model matching 'phi-3.5-mini-instruct-qnn-npu' was found in the catalog" — confirming the entries are absent from the catalog response itself, not just hidden from the default table view.

Ruled out: eastus regional throttling (#757)

Since foundry config has no exposed key for ModelRegistryRegion (only port, cache-directory, idle-timeout-minutes, log-level), I tested this directly via the Python SDK, pinning the region explicitly and running each test in a fresh process to avoid the FoundryLocalManager singleton carrying over the first config:

config = Configuration(
    app_name=f"probe_{region}",
    additional_settings={"ModelRegistryRegion": region},
    log_level="Warning",
)
FoundryLocalManager.initialize(config)
models = FoundryLocalManager.instance.catalog.list_models()

Results across 5 regions (NPU check):

Region	Total models	NPU-tagged	GPU-tagged
westeurope	43	0	—
westus2	43	0	—
northeurope	43	0	—
eastus2	43	0	—
southeastasia	43	0	—
eastus (default)	43	0	0

Every region returns an identical count with zero non-CPU entries. This rules out eastus throttling as the cause — results are identical regardless of region, and the CPU-only pattern is consistent across the board, not a partial/intermittent fetch failure.

Environment

• CLI version: 0.10.2
• Hardware: Snapdragon(R) X Elite - X1E80100, Qualcomm(R) Hexagon(TM) NPU, Qualcomm(R) Adreno(TM) X1-85 GPU
• OS: Windows 11 (reports as build 10.0.26200)
• RAM: 15.6 GB total
• QNNExecutionProvider: downloads and registers successfully on foundry server start

Ask

1. Is this catalog snapshot genuinely missing all non-CPU variants for Snapdragon X Elite hardware right now, or is there a hardware-profile-matching bug preventing GPU/NPU entries from being returned even though they exist server-side?
2. Is there a catalog endpoint/flag (beyond ModelRegistryRegion) relevant to hardware-profile matching that might need adjusting?
3. Could foundry config eventually expose ModelRegistryRegion for CLI-only users, since the SDK is currently the only way to attempt this workaround at all?

Happy to provide foundry report output or further diagnostics if useful.

Diagnostics

System

• OS: Microsoft Windows 10.0.26200
• Architecture: Arm64
• CPU: Snapdragon(R) X 12-core X1E80100 @ 3.40 GHz (12 logical cores)
• GPU: Qualcomm Incorporated Qualcomm(R) Adreno(TM) X1-85 GPU (—)
• NPU: Qualcomm Technologies, Inc. Snapdragon(R) X Elite - X1E80100 - Qualcomm(R) Hexagon(TM) NPU
• RAM: 7.5 GB available / 15.6 GB total
• Disk: 707.8 GB free / 952.6 GB total (C:)
• .NET: .NET 9.0.18

Service

• State: Ready
• Web URLs: http://127.0.0.1:50990
• CLI version: 0.10.2

Models

• Available variants: 43
• Loaded: 0
• Cached: 0

Connectivity

• Local service: Reachable
• Model registry: Reachable

Warnings

• Qualcomm Adreno GPU detected. If acceleration fails, try a CPU model variant or update the GPU driver.

[natke]

Thank you for reporting this @steamedeo

Can you please check the driver version of your Qualcomm NPU ie

And also check the foundry local logs in ~/.foundry and add them here (with any confidential information redacted

[steamedeo]

Thanks for looking into this! Here's the requested info:

NPU Driver Version

• Driver package: qcnspmcdm8380.inf
• File Version: 30.0.0220.3000
• Provider: Qualcomm Technologies, Inc.
• Digital Signer: Microsoft Windows Hardware Compatibility
• Device: Snapdragon(R) X Elite - X1E80100 - Qualcomm(R) Hexagon(TM) NPU

foundrylocald logs

Startup/EP registration sequence (redacted SID) — no errors appear during EP download/registration or catalog fetch, the daemon reports "Ready" cleanly each time:

2026-07-15T08:10:59.467Z [INF] Foundry.Daemon.Host: Daemon configured with --port 5273.
2026-07-15T08:10:59.469Z [INF] Foundry.Daemon.Startup: Binding IPC server
2026-07-15T08:10:59.469Z [INF] Foundry.Daemon.Ipc.IpcServer: IPC server listening at foundry-cli-[REDACTED-SID]
2026-07-15T08:10:59.469Z [INF] Foundry.Daemon.Startup: Starting heartbeat publisher (cadence=00:00:05)
2026-07-15T08:10:59.469Z [INF] Foundry.Daemon.Startup: Activating noise suppressor
2026-07-15T08:10:59.469Z [INF] Foundry.Daemon.Startup: Calling FoundryLocalManager.CreateAsync
2026-07-15T08:10:59.708Z [INF] Foundry.Daemon.Startup: Downloading and registering EPs
2026-07-15T08:10:59.710Z [INF] Foundry.Daemon.Startup: Starting SDK web service
2026-07-15T08:10:59.733Z [INF] Foundry.Daemon.Startup: Writing daemon discovery file (1 URLs)
2026-07-15T08:10:59.735Z [INF] Foundry.Daemon.Startup: Daemon Ready

This pattern repeats identically across every daemon restart in the log (multiple sessions), with no visible errors or warnings around EP registration or the model catalog fetch — the NPU-catalog gap doesn't appear to throw any exception locally, which suggests it's failing silently upstream (this is consistent with the region-pinning test I ran earlier: 5 different ModelRegistryRegion values all returned identical TOTAL=43/NPU=0 results).

Separate, likely-unrelated observation

The log also contains two System.IO.IOException: Pipe is broken errors during chat.complete IPC calls, e.g.:

2026-07-15T07:03:06.354Z [ERR] Foundry.Daemon.Ipc.IpcDispatcher: Handler for op 'chat.complete' threw.
System.IO.IOException: Pipe is broken.
   at System.IO.Pipes.PipeStream.WriteAsyncCore(...)
   ...
   at Microsoft.AI.Foundry.Local.OpenAIChatClient.<ChatStreamingImplAsync>d__13.MoveNext()

These occurred during separate interactive foundry chat CLI sessions (likely from an abrupt session exit), not during catalog/model-list operations, so probably unrelated to the NPU issue — flagging in case it's useful context, happy to open a separate issue for it if helpful.

Let me know if there's any other diagnostic info I can provide.

[baijumeswani]

I'll investigate. My naive thought is that something might be crashing the daemon. I have a similar machine. Will try to reproduce on my end. Thanks for reporting.

[fil-bol]

I am experiencing a similar problem with a Vitis NPU: the CLI reports the EP as loaded, however the debug log reports

2026-07-17 15:29:01.284 +02:00 [INF] DownloadAndRegisterEpsAsync [3/3]: starting VitisAIExecutionProvider.
2026-07-17 15:29:01.284 +02:00 [INF] WinML EP VitisAIExecutionProvider: Starting download and registration (state: "NotPresent").
2026-07-17 15:29:08.999 +02:00 [WRN] WinML EP VitisAIExecutionProvider: Download failed (result: "Failure", error: The product is not applicable or cannot be found.

The product is not applicable or cannot be found.
). Will retry on next attempt.
2026-07-17 15:29:08.999 +02:00 [WRN] DownloadAndRegisterEpsAsync [3/3]: VitisAIExecutionProvider failed.
2026-07-17 15:29:08.999 +02:00 [INF] EP registration partially complete: 2 succeeded, 1 failed (VitisAIExecutionProvider) in 16907ms. Available EPs: CPUExecutionProvider, WebGpuExecutionProvider
2026-07-17 15:29:08.999 +02:00 [INF] [Telemetry] AppName:foundry UserAgent:FoundryLocalCore Command:AutoRegisterCertifiedEps Status:Failure Direct:True Time:16907ms
2026-07-17 15:29:08.999 +02:00 [INF] [Telemetry] AppName:foundry UserAgent:CoreNativeInterop Command:AutoRegisterCertifiedEps Status:Failure Direct:True Time:16907ms

The execution provider seems to be preset on my machine, as

Get-AppxPackage -AllUsers "*.EP.*"  | Select-Object -ExpandProperty PackageFullName

reports

WindowsWorkload.EP.AMD.VitisAI.1.8_1.8.58.0_x64__8wekyb3d8bbwe
WindowsWorkload.EP.AMD.VitisAI.Framework.1.8_1.8.60.0_x64__8wekyb3d8bbwe

[ErikMoreau]

Can repro on multiple Surface Snapdragon devices, no NPU models are showing up Elite and Plus processors
The QNNExecutionProvider takes like only 1 second to load, seems it's not really loading