GLM-DSA: Fix token corruption due to incorrect RoPE type

[usrlocalben]

The DSA lightning indexer hardcoded LLAMA_ROPE_TYPE_NEOX for indexer Q/K positional encoding, but the GLM-5.2 model config.json explicitly sets "indexer_rope_interleave": true, meaning the indexer uses interleaved RoPE (LLAMA_ROPE_TYPE_NORM).

GLM-5.1 config.json also sets "rope_interleave": true.

The mismatch caused incorrect indexer scores, wrong top-k key selection, giving single-character/token errors and other corruption that could grow worse with context length.

Here the setup is changed to use the model's given rope_type instead of hardcoding NEOX. This fixes GLM-5.2, GLM-5.1, and should be less brittle for future GLM variants.

Closes #2083

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[Skelectric]

This fixed it for me, thanks!

[sayap]

This fixed it for me as well, thanks!

The model is now fully coherent until the context limit is reached (I can only fit 100k), and with -dsa the TG speed goes from 12 t/s at the beginning, to 11 t/s at the end. Amazing

[ikawrakow]

Thank you for finding this. The original PR had it hard-coded, so I assumed it needs to be different from the model RoPE type.

[chrislawso]

This fixed it for me, thanks!

This fixed it for me as well, thanks!

The model is now fully coherent until the context limit is reached (I can only fit 100k), and with -dsa the TG speed goes from 12 t/s at the beginning, to 11 t/s at the end. Amazing

I compiled these newest PR bits and ik_llama continues crashing same as before.

What cli settings is everyone running for ik_llama and glm5.2 successfully?

[usrlocalben]

What cli settings is everyone running for ik_llama and glm5.2 successfully?

@chrislawso

Here's my invocation / runscript with the three quants I've been using. gpu is r6kp, cpu is epyc 9b14 x2 nps4.
not shown: drop_caches if the gguf could already be in page cache with different numa distribution.

#!/bin/bash

MEMORY_CONCURRENCY_IN_NUMA_NODES=8
N_BATCH=4096

# COMPUTE_POWER_IN_THREADS_PER_NUMA=17
# M=/model/GLM-5.2/sokann/IQ2_KT/GLM-5.2-GGUF-2.244bpw-q8indexer.gguf

# COMPUTE_POWER_IN_THREADS_PER_NUMA=8
# M=/model/GLM-5.2/SixVolts/GLM-5.2-Q3_K_M-00001-of-00008.gguf

COMPUTE_POWER_IN_THREADS_PER_NUMA=6
M=/model/GLM-5.2/usrlocalben/HQ4_K/GLM-5.2-HQ4_K.gguf

N_THREADS=$[ $MEMORY_CONCURRENCY_IN_NUMA_NODES * $COMPUTE_POWER_IN_THREADS_PER_NUMA ]

GGML_CUDA_NO_PINNED=1 \
exec ./ik_llama.cpp/build/bin/llama-server \
  --host 0.0.0.0 --port 4972 --webui llamacpp \
  --numa distribute \
  -t $N_THREADS \
  --cache-ram 200000 \
  -b $N_BATCH -ub $N_BATCH -amb 1024 \
  -mla 3 -dsa -fidx \
  -ngl 999 \
  -ot exps=CPU \
  --jinja --parallel-tool-calls \
  --chat-template-kwargs '{"enable_thinking": true, "reasoning_effort": "high"}' \
  -ctk q8_0 \
  -c $[ 2**18 ] \
  --alias GLM-5.2 \
  --metrics \
  -m "$M"

  # -ot "blk\.([3-9])\.ffn_.*_exps=CUDA0" \
  # -ot "blk\.(1[0-4])\.ffn_.*_exps=CUDA0" \