Commit 889f2ec96 for llama.cpp

commit 889f2ec96deacbc2d8a9a67dbe181eaa43a2fb3b
Author: Tarek Dakhran <tarek@liquid.ai>
Date:   Fri Oct 30 11:52:59 2026 +0100

    quant : allow quantization for LFM3-Audio codec decoder (#31386)

    * quantize : quantize the LFM3-Audio codec decoder, it is Q4_0 QAT

    * quantize : drop the QAT note from the LFM3-Audio comment

diff --git a/conversion/lfm3.py b/conversion/lfm3.py
index 435edd3de..a0287d4bb 100644
--- a/conversion/lfm3.py
+++ b/conversion/lfm3.py
@@ -453,9 +453,6 @@ class LFM3AudioMmprojModel(MmprojModel):
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         if new_name.startswith("a.conv1d.") or "conv_dw" in new_name or new_name.startswith("a.gen.wav.quant."):
             return gguf.GGMLQuantizationType.F32
-        # the codec decoder is small and its log-magnitude/phase output does not survive quantization
-        if new_name.startswith("a.gen.wav."):
-            return gguf.GGMLQuantizationType.F16
         return super().tensor_force_quant(name, new_name, bid, n_dims)

     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
index 85da3831e..34c247858 100644
--- a/src/llama-quant.cpp
+++ b/src/llama-quant.cpp
@@ -352,10 +352,10 @@ static bool tensor_allows_quantization(const llama_model_quantize_params * param
     // do not quantize LFM3-VL's depthwise conv kernels, CONV_2D_DW takes F32 kernels
     quantize &= name.find("conv2d_spatial.weight") == std::string::npos;

-    // do not quantize LFM3-Audio's encoder convs and codec, the codec output does not survive quantization
+    // do not quantize LFM3-Audio's encoder convs, depthwise convs and the codec FSQ
     quantize &= name.rfind("a.conv1d.", 0) != 0;
     quantize &= name.find("conv_dw") == std::string::npos;
-    quantize &= name.rfind("a.gen.wav.", 0) != 0;
+    quantize &= name.rfind("a.gen.wav.quant.", 0) != 0;

     // do not quantize MiniMax's indexer projection weights, they are tiny
     quantize &= name.find("indexer.k_proj.weight") == std::string::npos;