From 294fc70ca0998990cc575be80cb0c8cbafdeee8c Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 24 Sep 2026 08:29:07 +0300 Subject: [PATCH] tmp --- conversion/mimo.py | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) diff --git a/conversion/mimo.py b/conversion/mimo.py index 8a2689b965..7a76723db3 100644 --- a/conversion/mimo.py +++ b/conversion/mimo.py @@ -101,7 +101,7 @@ class MimoV2Model(TextModel): qkv_overrides: dict[str, tuple[Callable, Callable, int]] = {} qc = self.hparams.get("quantization_config") if isinstance(qc, dict) and qc.get("quant_method") == "fp8": - pat = re.compile(r"^model\.layers\.(\d+)\.self_attn\.qkv_proj\.weight_scale_inv$") + pat = re.compile(r"^model\.(mtp\.)?layers\.(\d+)\.self_attn\.qkv_proj\.weight_scale_inv$") for name in list(self.model_tensors.keys()): m = pat.match(name) if not m: @@ -109,10 +109,13 @@ class MimoV2Model(TextModel): weight_name = name.removesuffix("_scale_inv") if weight_name not in self.model_tensors: continue + bid = int(m.group(2)) + if m.group(1) is not None: + bid += self.hparams["num_hidden_layers"] qkv_overrides[weight_name] = ( self.model_tensors[weight_name], self.model_tensors[name], - int(m.group(1)), + bid, ) super().dequant_model()