MiniMaxAI

/

MiniMax-M2.5

like 1.22k

Follow

MiniMax 4.97k

With @ktsaou , @Doctor-Shotgun and @lukealonso , we noticed that the model can be stuck in annoying typos:

• The base FP8 model, @Doctor-Shotgun , when using Transformers v5 instead of Transformers v4
• The NVFP4 quant from @lukealonso , when using Transformers v5 instead of Transformers v4, see
  
  The pattern is a space after a dot . which is really handicapping for code
• My BF16+INT4 AWQ quant, with Transformers v5, see
  
  The pattern is plural being dropped before:
  • an underscore test_layer_fixtures.nim instead of test_layers_fixtures.nim
  • a slash action/checkout instead of actions/checkout
  • a colon tag: instead of tags:
  • an uppercase letter

Investigation is being done in this thread: https://huggingface.co/mratsim/MiniMax-M2.5-BF16-INT4-AWQ/discussions/4