DSpark draft model fails to load on Windows Vulkan build: "invalid vector subscript"

[Zolikabacsi]

Summary

Loading the DSpark speculative-decoding drafter (Ternary-Bonsai-27B-dspark-Q4_1.gguf) via --model-draft fails on the Windows Vulkan prebuilt binary with invalid vector subscript. The same command works on the Windows CPU prebuilt binary from the same release.

Environment

• Release: prism-b9596-9fcaed7
• Binary: llama-bin-win-vulkan-x64.zip (reports build: b1-9fcaed7, MSVC 19.51)
• OS: Windows x64
• GPU: AMD Radeon RX 6800S (8 GB), Adrenalin driver, Vulkan backend
• Main model: Ternary-Bonsai-27B-Q2_0.gguf from prism-ml/Ternary-Bonsai-27B-gguf (loads and runs fine on the same Vulkan build without the drafter, ~17 tok/s)
• Draft model: Ternary-Bonsai-27B-dspark-Q4_1.gguf from the same repo (sha256 c4810091d244eddc61a0cc4966e584b0959f141e3c66c0d371a6652d9f647da9)

Reproduction

llama-cli.exe -m Ternary-Bonsai-27B-Q2_0.gguf ^
  --model-draft Ternary-Bonsai-27B-dspark-Q4_1.gguf ^
  --spec-type draft-dspark --spec-draft-n-max 4 ^
  -p "Hi" -n 8 --temp 0.7 --top-p 0.95 --top-k 20 -c 4096 -ngl 99

Observed (Vulkan build)

0.08.716.940 E llama_model_load: error loading model: invalid vector subscript
0.08.716.949 E llama_model_load_from_file_impl: failed to load model
0.08.717.101 E srv    load_model: failed to load draft model, 'Ternary-Bonsai-27B-dspark-Q4_1.gguf'

The failure happens while loading the draft model; the main model loads successfully on this build.

Expected / Control

The identical command using llama-bin-win-cpu-x64.zip from the same release (build b9596-9fcaed763, Clang 20.1.8) loads both models and generates correctly with speculative decoding enabled.

Notes

• The drafter's GGUF metadata reports general.architecture = dspark with tokenizer.ggml.model = none (tokenizer shared with the target), in case the draft-model load path on the Vulkan build handles this differently.
• Happy to provide full --verbose logs or test a patched build.