Eval bug: Qwen3.6 27B segfaults when using -sm tensor and --kv-tail-tokens

[dinerburger]

Name and Version

version: 10855 (176c1a1)
built with GNU 14.2.0 for Linux x86_64

Operating systems

Linux

GGML backends

CPU, CUDA

Hardware

Intel(R) Xeon(R) w5-3435X + 3090 Ti + A4000

Models

Qwen3.6 27B

Problem description & steps to reproduce

llama-server can be invoked with -sm tensor, but adding --kv-tail-tokens causes a segfault. Looks like, unsurprisingly, related to cache context creation.

First Bad Commit

No response

Relevant log output

Logs

0.03.976.118 I llama_context: constructing llama_context
0.03.976.124 I llama_context: n_seq_max     = 4
0.03.976.124 I llama_context: n_ctx         = 65536
0.03.976.125 I llama_context: n_ctx_seq     = 65536
0.03.976.125 I llama_context: n_batch       = 2048
0.03.976.125 I llama_context: n_ubatch      = 512
0.03.976.126 I llama_context: causal_attn   = 1
0.03.976.127 I llama_context: flash_attn    = enabled
0.03.976.127 I llama_context: kv_unified    = true
0.03.976.131 I llama_context: freq_base     = 10000000.0
0.03.976.131 I llama_context: freq_scale    = 1
0.03.976.132 I llama_context: n_rs_seq      = 0
0.03.976.132 I llama_context: n_outputs_max = 4
0.03.976.132 I llama_context: n_ctx_seq (65536) < n_ctx_train (262144) -- the full capacity of the model will not be utilized
0.04.239.980 I llama_context:  CUDA_Host  output buffer size =     3.79 MiB
0.04.241.298 I llama_kv_cache:     Meta() KV buffer size =  1040.00 MiB
[1]    53361 segmentation fault (core dumped)