quant: include MTP/NextN block when counting FFN layers (GLM-5.2) - #24832

Closed

Deviad wants to merge 9 commits into

ggml-org:masterfrom

Deviad:feature/patch_used_to_create_mixed_quantization_of_glm5.2

Commits on Jun 20, 2026

quant: include MTP/NextN block when counting FFN layers (GLM-5.2)
Deviad
committed
88c3c51
Update src/llama-quant.cpp

Deviad
and
CISC
authored
f30cf36

Commits on Jun 21, 2026

trace-moe: new example for MoE expert-routing tracing
Deviad
committed
95ceec5
trace-moe: batched multi-prompt mode (load model once, trace N prompts)
Deviad
committed
fff701b
trace-moe: real reproducibility provenance in meta sidecar
Deviad
committed
37dd8f0
trace-moe: populate n_expert_total + per-prompt speed metrics
Deviad
committed
969787c
trace-moe: bounded activation summaries (Phase 4 / Story 6)
Deviad
committed
0d0e48a
trace-moe: emit Block Influence (BI) records for ShortGPT
Deviad
committed
6f67b8a

Commits on Jun 24, 2026

feat(glm-dsa): sparse-gather decode attention + forward-path override + remediation
Deviad
committed
7651134