imvladikon/GLM-5.2-9B-LoRA-Surgery-Dummy-FP8

GLM-5.2 10L-16E Surgery Dummy (fp8-rollout)

Test-only checkpoint for GLM-5.2 SFT, RL, LoRA, sharding and live-adapter integration. It is not a usable chat or benchmark model.

This is the mixed E4M3/BF16 rollout base with the released 128x128 scale grids half of pair glm52-9b-2db222dcbd5d236a. It keeps the released hidden width, MLA/DSA dimensions, top-8 routing and three dense layers, while reducing 78 decoder layers to 10 and 256 routed experts to 16. Source layers are [0, 1, 2, 3, 15, 27, 38, 51, 63, 77]. It contains 8,763,269,120 model parameters. The exact source revision, byte ranges, BF16-anchor layer map and selected expert IDs are in surgery_plan.json; output hashes are in surgery_manifest.json.

Selected source experts: {'3': [8, 22, 40, 87, 94, 95, 101, 131, 141, 205, 222, 229, 234, 245, 246, 247], '4': [21, 29, 48, 77, 103, 135, 143, 150, 154, 185, 189, 207, 214, 231, 234, 245], '5': [7, 69, 82, 115, 119, 124, 127, 190, 198, 200, 212, 220, 231, 236, 250, 254], '6': [0, 3, 14, 33, 42, 60, 81, 83, 88, 96, 112, 181, 187, 194, 236, 240], '7': [34, 44, 54, 55, 78, 80, 102, 145, 187, 198, 199, 207, 216, 217, 246, 251], '8': [34, 40, 66, 68, 84, 110, 115, 121, 126, 163, 185, 206, 223, 225, 249, 254], '9': [11, 32, 38, 39, 48, 58, 61, 83, 114, 144, 158, 165, 173, 193, 237, 242]}.