Diagnose NVIDIA L4 desktop WebGPU initialization crashes before enabling consumer GPU

[leehack]

Reproduced native desktop GPU crash

Downstream qualification for leehack/llamadart#506 bypassed the Dart platform gate using the low-level runtime client. Published v0.17.0-3 selects real NVIDIA L4 hardware, then crashes while initializing its WebGPU delegate. Removing the consumer gate is not safe yet.

Inputs and results

• Qwen3 0.6B: litert-community/Qwen3-0.6B, revision 8414150f2e9dcc82449bcc9c5abc404b399a4d06, SHA256 555579ff2f4fd13379abe69c1c3ab5200f7338bc92471557f1d6614a6e5ab0b4.
• Context 2048, GPU, speculative decoding disabled. Failure happens during litert_lm_engine_create, before conversation/template creation.
• Ubuntu 24.04 x64, GCP g2-standard-4/L4, NVIDIA 580.173.02: Vulkan identifies a discrete NVIDIA L4. Native libLiteRtWebGpuAccelerator.so segfaults. Qwen3.5 0.8B int8 also segfaults on this setup.
• Windows Server 2022 x64, GCP g2-standard-4/L4: native logs identify NVIDIA L4 ... backend=Direct3D 12, adapterType=Discrete GPU. DXC v1.9.2607 dxcompiler.dll and dxil.dll staged beside runtime DLLs. Access violation inside libLiteRtWebGpuAccelerator.dll+0x6e08b1.
• Direct Windows C API reproduction through C# P/Invoke, without Dart, also terminates with 0xC0000005 during engine creation. Thus the crash is not caused by Dart chat parsing or its GPU platform gate.
• Matched Windows v0.16.0-native.2 runtime, same Qwen3/model/settings/hardware and DXC dependencies: also selects L4/D3D12 and crashes during GPU initialization. Do not classify this as a new v0.17 regression. No old-runtime Linux GPU comparison was run.
• CPU chat works on Linux/Windows with downstream template correction PR507. macOS arm64 Metal and Pixel Vulkan evidence are separate and do not establish desktop x64 GPU support.

Linux setup pitfall

The accelerated image initially contained compute drivers only. Install matching libnvidia-gl-580-server and libnvidia-common-580-server plus libegl1; verify vulkaninfo --summary identifies NVIDIA. Prior software llvmpipe crashes were excluded from hardware conclusions. The hardware-only run used VK_DRIVER_FILES=/usr/share/vulkan/icd.d/nvidia_icd.json.

Acceptance

Identify the failing native/prebuilt delegate contract using a direct C API reproducer. Qualify a corrected artifact on real Linux Vulkan and Windows D3D12 hardware, with backend identity, generation/streaming/cancellation/disposal, and negative dependency/driver coverage. Then coordinate downstream gate changes in #506. Do not enable desktop GPU based on library registration or silently use a software/CPU adapter.

Temporary VMs were bounded to two hours; evidence was retrieved before deletion.

[leehack]

Current-artifact follow-up from leehack/llamadart#515: clean portable public-Dart bundle f1a13f350ce1adcb90191e755883e90e28f55c6d, LiteRT 0.17.0-5, context 1024, four threads, thinking disabled. Runtime archives and model files were checksum-verified. This is distinct from the earlier direct-C-API v0.17.0-3 crash evidence.

Model	Linux L4 / Vulkan	Windows L4 / D3D12
Gemma 4 E2B	18 implemented checks PASS; three release cases NOT_RUN	GPU engine creation fails: DynamicLib.Open: dxil.dll Windows Error: 87
Qwen3.5 0.8B int8	First tokenizer case hits 60-second deadline; process stopped at 240-second outer bound, no completed generation	D3D12 initialization reports the same DXIL error; partial errors followed by outer timeout

Linux used Ubuntu24.04 image common-cu129-ubuntu-2404-nvidia-580-v20260909, NVIDIA580.173.02, exact matching graphics libraries from Canonical snapshot 20260801T000000Z, and VK_DRIVER_FILES=/usr/share/vulkan/icd.d/nvidia_icd.json. vulkaninfo identified discrete NVIDIA L4 without llvmpipe. Gemma logs repeatedly select NVIDIA L4 ... backend=Vulkan, adapterType=Discrete GPU, create WebGPU environments, and initialize the serialized WebGPU kernels; public streaming/cancellation/reload assertions pass. The automated LiteRT placement verifier still remains unverified, and three unimplemented release obligations prevent qualification. This positive Gemma result means the older Linux failure must not be generalized to every current model/artifact combination. The Qwen timeout is not assumed to have the same cause as the older segfault.

Windows used Server2022 x64 and NVIDIA582.53. Official DXC v1.9.2607 archive dxc_2026_07_29.zip was SHA256-verified (a1dfb116ba3eeae6a1582291b53a8e7bf65ad760676bd3194685c8f7367cd241), and its bin/x64 directory was prepended to PATH. Both models selected the discrete NVIDIA L4 D3D12 adapter before the DXIL error. PATH availability does not prove Dawn's DLL loading contract; this run does not reproduce or attribute the earlier access violation. Gemma CPU passes18 implemented cases; Qwen CPU also times out, so that timeout is not established as GPU-specific.

Model locks:

• Gemma revision 6b78abd019e61a1ca4cbe3b212d2c9ce8ff38a94, SHA256 181938105e0eefd105961417e8da75903eacda102c4fce9ce90f50b97139a63c.
• Qwen revision c23b16e43ada6ead533b12593fe500bbe268014f, SHA256 684d4d34adf7176eb47f6026ff65c33d42584737254e5524a8d1ad62edc21b98.

Logs, manifests, exact prompts/configs and case records were retained before teardown. Both VMs, boot disks and temporary firewalls were deleted; final project instance/disk/task-firewall inventories are empty. No full desktop LiteRT GPU qualification is claimed.

[leehack]

Current triage separates three evidence tracks rather than assuming one desktop GPU cause. (1) Windows 0.17.0-5 DXIL initialization: published inventory omits dxil.dll/dxcompiler.dll; reviewed owning-repo candidate 8274222 bundles the pinned official DXC pair and licenses beside Dawn, but no PR/publication or actual Windows loader/GPU pass exists yet. Inventory tests are not binary compatibility evidence. (2) Current Linux Gemma GPU completes 18 implemented checks with native L4 selection; automatic placement/full release qualification remains separate. (3) Qwen Linux GPU and Windows CPU C02 timeouts include lazy engine creation: retained initialization logs progress beyond the 60-second case deadline; this does not establish a tokenizer hang or shared cause with historical access violations. Preserve historical 0.17.0-3 crashes as separate evidence. Next: exact target-hardware DXIL candidate load/generation and separately timed Qwen engine/tokenizer controls, without relaxing deadlines or claiming CPU fallback as GPU. Current priority P1; owner task has a reviewed packaging candidate but Windows GPU execution is still unverified.