From 58b9064cc343b5236ecba7805772fc0a77060486 Mon Sep 17 00:00:00 2001
From: Francis Couture-Harpin <git@compilade.net>
Date: Wed, 26 Jun 2024 16:24:40 -0400
Subject: [PATCH] convert-hf : simplify BitNet pre-quantization

This still results in the exact same tensor weights and scales,
but it reveals some weirdness in the current algorithm.
---
 convert-hf-to-gguf.py | 20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)

diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py
index 635f649a071964..ee8350e43219e9 100755
--- a/convert-hf-to-gguf.py
+++ b/convert-hf-to-gguf.py
@@ -263,7 +263,10 @@ def write_tensors(self):
                     break
 
             for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
-                data: np.ndarray = data  # type hint
+                data: np.ndarray  # type hint
+                if len(data.shape) == 0:
+                    # otherwise single-value tensors get squeezed
+                    data = data.reshape((1,))
                 n_dims = len(data.shape)
                 data_dtype = data.dtype
                 data_qtype: gguf.GGMLQuantizationType | None = None
@@ -334,7 +337,7 @@ def write_tensors(self):
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 
                 # reverse shape to make it similar to the internal ggml dimension order
-                shape_str = f"{{{', '.join(str(n) for n in reversed(shape)) or '1'}}}"
+                shape_str = f"{{{', '.join(str(n) for n in reversed(shape))}}}"
 
                 # n_dims is implicit in the shape
                 logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
@@ -1442,12 +1445,13 @@ def set_gguf_parameters(self):
     def weight_quant(self, weight):
         dtype = weight.dtype
         weight = weight.float()
-        s = 1 / weight.abs().mean().clamp(min=1e-5)
-        weight = (weight * s).round().clamp(-1, 1) / s
-        scale = weight.abs().max().unsqueeze(0)
-        weight = torch.where(weight.abs().less(1e-6), 0, weight).type(dtype)
-        weight = torch.sign(weight).type(dtype)
-        return weight.type(dtype), scale.type(torch.float32)
+        scale = weight.abs().mean().clamp(min=1e-5)
+        iscale = 1 / scale
+        weight = (weight * iscale).round().clamp(-1, 1)
+        # TODO: use the scale directly instead of inverting it twice
+        # (this is also unnecessarily doubly inverted upstream)
+        # ref: https://huggingface.co/1bitLLM/bitnet_b1_58-3B/blob/af89e318d78a70802061246bf037199d2fb97020/utils_quant.py#L10
+        return weight.type(dtype), (1 / iscale).type(torch.float32)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         new_name = self.map_tensor_name(name)