Cannot load w/ sglang 0.5.6

#1
by huggyfaceenjoyer - opened

AttributeError: 'Glm4vMoeConfig' object has no attribute 'rope_scaling'

4.6v

    "rope_parameters": {
      "mrope_section": [
        8,
        12,
        12
      ],
      "partial_rotary_factor": 0.5,
      "rope_theta": 500000,
      "rope_type": "default"
    },

4.5 Air

  "rope_scaling": null,
  "rope_theta": 1000000,

use sglang 0.5.6.post1

use sglang 0.5.6.post1

That got me to be able to get further along, but I run out of vram with same settings as Air FP8 with 4.6V FP8, even drastically reducing context window to minimum. Dual RTX 6000 Pro 96G.
Also having a lot of problems with 0.5.6.post1 and/or prerelease transformers where even GLM 4.5 Air is no longer stable and had to revert.

Sign up or log in to comment