problem

[3879171-cn]

Environment

• GPU: NVIDIA GeForce RTX 5090 Laptop GPU (24GB VRAM)
• Compute capability: 12.0 (Blackwell)
• OS: WSL2
• llama.cpp branch: turboquant (feature/turboquant-kv-cache)
• Model: Qwen3.6-35B-A3B-Q4_K_M.gguf
  Steps to reproduce

1. Build llama-server with CUDA support
2. Run: ./llama-server -m Qwen3.6-35B-A3B-Q4_K_M.gguf --port 1234 --host 0.0.0.0 --ctx-size 2048
   Expected behavior
   Server starts and responds to API requests
   Actual behavior

• Model loads successfully
• Server hangs during warmup phase
• All API requests return HTTP 503 "Loading model"
• Server process eventually crashes/exits
  Key warnings from logs
• "llama_kv_cache: upstream attention rotation disabled (TurboQuant uses kernel-level WHT)"
• "sched_reserve: fused Gated Delta Net (autoregressive) enabled"
  Notes
• Same model works fine in LM Studio on Windows 11
• Issue occurs even with -ngl 0 (CPU-only mode)
• Issue occurs even with --ctx-size 2048 (small context)