Add GLM-5.2 IndexShare DSA, MTP, and GGUF export support - #404

Open

justinchuby wants to merge 25 commits into

mainfrom

glm5.2-moe-export

Commits on Jul 21, 2026

Add GLM-5.2 MoE export support

justinchuby
and
Copilot
committed
Implement GLM-5.2 IndexShare DSA and MTP export

justinchuby
and
Copilot
committed
fix(glm_moe_dsa): indexer RoPE must rotate full index_head_dim

justinchuby
and
Copilot
committed
test(export): add tiny DeepSeek-V2 ONNX helper

justinchuby
and
Copilot
committed
tools: add quantized tiny glm_moe_dsa export helper for onnx-genai E2E

justinchuby
and
Copilot
committed
feat(moe): fused com.microsoft::QMoE emitter for GLM/DeepSeek routed experts

justinchuby
and
Copilot
committed
fix(deepseek): restore group top-k selection

justinchuby
and
Copilot
committed
style: fix GLM export lint

justinchuby
and
Copilot
committed
test: update onnx-genai runtime writer mock

justinchuby
and
Copilot
committed
test: cover streamed GLM GGUF quantization

justinchuby
and
Copilot
committed
ci: keep GPU validation on CUDA 12 ORT

justinchuby
and
Copilot
committed
test: handle keyed recurrent cache states

justinchuby
and
Copilot
committed
test: close ONNX temp files before Windows saves

justinchuby
and
Copilot
committed

Commits on Jul 24, 2026

Lower GLM DSA sparse-attention to pkg.nxrt::IndexShare

justinchuby
and
Copilot
committed
test(glm-dsa): assert IndexShare lowering instead of dense attention

justinchuby
and
GitHub Copilot
committed
fix(export): address PR 404 review feedback

justinchuby
and
Copilot
committed

Commits on Jul 29, 2026

Commits on Jul 30, 2026