Eval bug: Qwen3.8-27B (qwen35 hybrid) llama-server crashes silently at ~520K prefill tokens with YaRN rope-scale 4 (just under 2x yarn-orig-ctx), b10434 CUDA

[asaffulks]

Name and Version

llama-server, build b10434 (commit range including #26623), Windows x64 CUDA 13.3 release binaries (llama-b10434-bin-win-cuda-13.3-x64.zip). Also reproduced on b10430.

Operating systems

Windows

GGML backends

CUDA

Hardware

1× NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96GB, compute 12.0), driver-side CUDA 13.x, Windows 11.

Models

unsloth/Qwen3.8-27B-GGUF → Qwen3.8-27B-UD-Q8_K_XL.gguf (arch qwen35, hybrid Gated DeltaNet + full-attention interval layers, native ctx 262144).

Problem description & steps to reproduce

With YaRN rope scaling ×4 to extend Qwen3.8-27B toward 1M context, llama-server dies silently (no assert, no error output, process exits) at ~520K tokens of prefill — suspiciously just under 2× the native training context (2×262144 = 524288).

Command:

llama-server.exe -m Qwen3.8-27B-UD-Q8_K_XL.gguf -ngl 999 -c 1048576 --parallel 1 -fa on ^
  --rope-scaling yarn --rope-scale 4 --yarn-orig-ctx 262144 ^
  --override-kv qwen35.context_length=int:1048576 ^
  --cache-type-k q8_0 --cache-type-v q8_0 --port 8095

Then POST a single ~925K-token chat completion. Prefill proceeds normally (~900 t/s at that depth) until:

slot print_timing: id  0 | task 0 | prompt processing, n_tokens = 518144, progress = 0.56, t = 574.77 s / 901.48 tokens per second
slot print_timing: id  0 | task 0 | prompt processing, n_tokens = 520192, progress = 0.56, t = 578.98 s / 898.46 tokens per second

…then the process is gone. No further output on stdout/stderr. GPU memory is freed. VRAM was not exhausted (~69GB used of 96GB at time of death).

Data points that bracket it:

1. Works: native window, no YaRN, 224K-token prompt → completes, correct retrieval (6/6 planted needles), fa on, q8 KV. (Note: prompts >~90–100K crashed the same way on b10430 and were fixed by b10434 — presumably Initial changes for Recurrent state rollback for nemotron for cuda #26623 ggml: recurrent state rollback for ggml_ssm_scan. This report is about a second, deeper boundary that remains.)
2. Works: YaRN ×2 (--rope-scale 2 --override-kv qwen35.context_length=int:524288, -c 524288), 444K-token prompt → completes, 6/6 retrieval.
3. Crashes: YaRN ×4 as above, dies at n_tokens ≈ 520192 (just under 524288 = 2× yarn-orig-ctx).
4. KV type doesn't matter (q8_0 and f16 both crash at the same place pre-b10434 for the first cliff; the ×4 crash above was with q8_0).

The "just under 2× orig ctx" position plus the silent death (reads like a segfault / device-side fault rather than a caught error) suggests a position/buffer boundary in the hybrid-attention or DeltaNet state path when positions exceed 2× the original training context under YaRN.

Happy to run diagnostic builds/flags if useful.

First Bad Commit

No response

Relevant log output

(see prefill tail above — process exits silently after n_tokens = 520192 with no error output)