| license: other | |
| tags: | |
| - glm | |
| - debug | |
| - synthetic | |
| # GLM-5.2 Debug | |
| Tiny randomly initialized GLM-5.2-shaped `glm_moe_dsa` model for fast trainer/kernel iteration. | |
| Preserves the fused DSA dimensions needed by FlashMLA (`kv_lora_rank=512`, `qk_rope_head_dim=64`, `v_head_dim=512`), plus 64 attention heads, DSA IndexShare, MoE/shared experts, and `model_type=glm_moe_dsa`. | |
| It intentionally uses 8 layers, tiny hidden/MLP/MoE sizes, `vocab_size=2048`, and random weights. Use synthetic token IDs in `[0, 2047]`; this is not intended for natural-language quality or sampling. | |