This quant needs fixing

#1
by trohrbaugh - opened

config.json's ignore is missing three entries present in the model card's creation snippet, and 86 weight_scale tensors exist for modules the recipe excludes, which trips assert key == "w2_weight" in vLLM's sink_experts loader.

@trohrbaugh Can you share a snippet of config.json?

Solution 2: Edit config.json (Permanent Fix)

  • Locate the "ignore" array inside "quantization_config".
  • Change "re:.gate." to a stricter string so it only targets your model's actual unquantized text/vision gates, leaving the MoE layers alone. Change it to:
    json
    "re:.*\.gate$"

Sign up or log in to comment