feat: ia3 variant, real bnb 4bit/8bit smoke, dev guide split, user-only readme

src/lora_lite/variants/__init__.py

@@ -1 +1 @@
-from . import lora, pissa, delora  # noqa: F401  side-effect: register
+from . import lora, pissa, delora, ia3  # noqa: F401  side-effect: register

src/lora_lite/variants/ia3.py

@@ -0,0 +1,22 @@
+"""IA3-style output gating. y_new = y * g, with g initialized to ones."""
+import torch
+from torch import nn
+
+from ..variant import register, ParamSpec
+
+
+@register
+class IA3:
+    name = "ia3"
+
+    @staticmethod
+    def param_specs(d_in, d_out, cfg):
+        return {"lora_g": ParamSpec((d_out,), init="ones", trainable=True)}
+
+    @staticmethod
+    def init(layer: nn.Linear, cfg) -> None:
+        return
+
+    @staticmethod
+    def forward(layer: nn.Linear, x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+        return y * layer.lora_g