From 9d0a0e317e999f79d3758b832ef04be0bb8a1346 Mon Sep 17 00:00:00 2001
From: City <125218114+city96@users.noreply.github.com>
Date: Fri, 7 Mar 2025 22:37:17 +0100
Subject: [PATCH] ltxv bf16 fix

scale shift table loads into nn.parameter, which can't work with BF16
---
 tools/convert.py | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/tools/convert.py b/tools/convert.py
index 6a90960..480de37 100644
--- a/tools/convert.py
+++ b/tools/convert.py
@@ -17,6 +17,7 @@ class ModelTemplate:
     shape_fix = False # whether to reshape tensors
     keys_detect = []  # list of lists to match in state dict
     keys_banned = []  # list of keys that should mark model as invalid for conversion
+    keys_hiprec = []  # list of keys that need to be kept in fp32 for some reason
 
     def handle_nd_tensor(self, key, data):
         raise NotImplementedError(f"Tensor detected that exceeds dims supported by C++ code! ({key} @ {data.shape})")
@@ -82,6 +83,9 @@ class ModelLTXV(ModelTemplate):
             "caption_projection.linear_2.weight",
         )
     ]
+    keys_hiprec = [
+        "scale_shift_table" # nn.parameter, can't load from BF16 base quant
+    ]
 
 class ModelSDXL(ModelTemplate):
     arch = "sdxl"
@@ -227,6 +231,10 @@ def handle_tensors(writer, state_dict, model_arch):
                 # very small tensors
                 data_qtype = gguf.GGMLQuantizationType.F32
 
+            elif any(x in key for x in model_arch.keys_hiprec):
+                # tensors that require max precision
+                data_qtype = gguf.GGMLQuantizationType.F32
+
         if (model_arch.shape_fix                        # NEVER reshape for models such as flux
             and n_dims > 1                              # Skip one-dimensional tensors
             and n_params >= REARRANGE_THRESHOLD         # Only rearrange tensors meeting the size requirement