Add DASH-Q remote-code inference (Triton decode kernel)

dashq_kernel.py

@@ -165,29 +165,6 @@ class TritonQuantLinear(nn.Module):
             in_features // self.group_size,
         )
 
-    @classmethod
-    def from_packed(cls, module: nn.Module, **kwargs) -> "TritonQuantLinear":
-        """Build from a dashq.quantization.PackedQuantizedLinear instance."""
-        from dashq.quantization import _unpack_int_values
-
-        K = int(getattr(module, "quant_in_features", module.in_features))
-        N = int(module.out_features)
-        W_int = _unpack_int_values(module.W_q_packed, module.nbits, module.numel).view(N, K)
-        num_groups = K // int(module.group_size)
-        scale = module.scale.view(N, num_groups)
-        zero = module.zero.view(N, num_groups)
-        bias = module.bias if getattr(module, "bias", None) is not None else None
-        return cls(
-            W_int,
-            scale,
-            zero,
-            int(module.nbits),
-            int(module.group_size),
-            bias=bias,
-            out_dtype=getattr(module, "linear_dtype", torch.float16),
-            **kwargs,
-        )
-
     def dequantize_weight(self, dtype: torch.dtype) -> torch.Tensor:
         """Returns W^T as (in_features, out_features), matching the K-major layout."""
         if self.nbits == 3: