Qwen3-Coder-Next (Gated Delta Net) load crash on RTX 5090 CUDA: shared object initialization failed - fixed by downgrading llama.cpp CUDA build

[logan683]

Environment

• GPU: NVIDIA GeForce RTX 5090 (Blackwell, compute capability 12.0)
• Driver: 610.74 (Windows)
• OS: Windows 11
• Runtime: LM Studio 4.20's bundled "CUDA 12 llama.cpp (Windows)" engine
• Broken version: v2.27.1
• Working version (downgrade fixes it): v2.25.2

Model

Qwen3-Coder-Next-Q4_K_M.gguf (unsloth quant). n_ctx=8192, n_parallel=4, kv_unified=true.

Symptom

Model consistently fails to load on runtime v2.27.1 and consistently succeeds on v2.25.2, same hardware/config/model file, no other changes. Ruled out:

• VRAM/contention: reproduces with zero other models loaded.
• Driver: 610.74 exceeds RTX 50-series minimum (570+), unchanged for ~1 month, and other CUDA models load fine on it.
• File corruption: same file loads fine once the engine is downgraded.

Logs (v2.27.1, fails)

resolve_fused_ops: layer 0 is assigned to device CPU but fused Gated Delta Net (chunked) is assigned to device CUDA0 (usually due to missing support)
resolve_fused_ops: fused Gated Delta Net (chunked) not supported, set to disabled
llm-engine\llama.cpp\ggml\src\ggml-cuda\ggml-cuda.cu:106: CUDA error
CUDA error: shared object initialization failed
  current device: 0, in function launch_fattn at llm-engine\llama.cpp\ggml\src\ggml-cuda\fattn-common.cuh:1113
  cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared)

A separate load attempt on the same runtime crashed instead in a different kernel, same CUDA error:

resolve_fused_ops: fused Gated Delta Net (chunked) not supported, set to disabled
llm-engine\llama.cpp\ggml\src\ggml-cuda\ggml-cuda.cu:106: CUDA error
ggml_cuda_compute_forward: MUL_MAT failed
CUDA error: shared object initialization failed
  current device: 0, in function ggml_cuda_compute_forward at llm-engine\llama.cpp\ggml\src\ggml-cuda\ggml-cuda.cu:2360

Both crashes immediately follow the "Gated Delta Net (chunked)...set to disabled" fallback, in whichever CUDA kernel executes next — suggesting the fallback/disabled code path itself is what's broken, not one specific kernel.

Possibly related

#24712 reports the same "Gated Delta Net (chunked) not supported" fallback warning on an RTX 5060, but only as a performance warning there, not a crash. This may indicate the fallback path is specifically broken on Blackwell/sm_120 in newer builds.

Happy to provide more logs/repro steps if useful.