From 576e01da1400c85af77bfaebc73a414f2bcfb0ba Mon Sep 17 00:00:00 2001
From: Prince Canuma <prince.gdt@gmail.com>
Date: Mon, 9 Mar 2026 18:25:32 +0100
Subject: [PATCH] Implement linking of text encoder and tokenizer directories
 in conversion process. Enhance error handling in LTX2TextEncoder for
 tokenizer loading, providing a fallback model if the specified path is
 unavailable.

---
 mlx_video/models/ltx/convert.py      | 27 +++++++++++++++++++++++++++
 mlx_video/models/ltx/text_encoder.py |  5 ++++-
 2 files changed, 31 insertions(+), 1 deletion(-)

diff --git a/mlx_video/models/ltx/convert.py b/mlx_video/models/ltx/convert.py
index 22dfa2c..b2564a2 100644
--- a/mlx_video/models/ltx/convert.py
+++ b/mlx_video/models/ltx/convert.py
@@ -611,6 +611,33 @@ def convert(source: str, output_path: Path, variant: str = "distilled"):
         else:
             print(f"  {upscaler_file}: not found, skipping")
 
+    # 8. Link text_encoder and tokenizer directories
+    print("\nLinking text encoder & tokenizer...")
+    for subdir in ["text_encoder", "tokenizer"]:
+        dest = output_path / subdir
+        if dest.exists():
+            print(f"  {subdir}/: already exists, skipping")
+            continue
+
+        local_candidate = source_dir / subdir
+        if local_candidate.is_dir():
+            # Resolve through symlinks to get the real directory
+            real_path = local_candidate.resolve()
+            dest.symlink_to(real_path)
+            print(f"  {subdir}/: symlinked to {real_path}")
+        elif is_hf_repo:
+            from huggingface_hub import snapshot_download
+
+            print(f"  {subdir}/: downloading from {source}...")
+            snapshot_download(
+                repo_id=source,
+                allow_patterns=f"{subdir}/*",
+                local_dir=str(output_path),
+            )
+            print(f"  {subdir}/: done")
+        else:
+            print(f"  {subdir}/: not found in source, skipping")
+
     # Summary
     all_converted = (
         len(transformer_weights)
diff --git a/mlx_video/models/ltx/text_encoder.py b/mlx_video/models/ltx/text_encoder.py
index 3fa22bb..f6824f8 100644
--- a/mlx_video/models/ltx/text_encoder.py
+++ b/mlx_video/models/ltx/text_encoder.py
@@ -754,7 +754,10 @@ class LTX2TextEncoder(nn.Module):
         if tokenizer_path.exists():
             self.processor = AutoTokenizer.from_pretrained(str(tokenizer_path), trust_remote_code=True)
         else:
-            self.processor = AutoTokenizer.from_pretrained(text_encoder_path, trust_remote_code=True)
+            try:
+                self.processor = AutoTokenizer.from_pretrained(text_encoder_path, trust_remote_code=True)
+            except Exception:
+                self.processor = AutoTokenizer.from_pretrained("google/gemma-3-12b-it", trust_remote_code=True)
         # Set left padding to match official LTX-2 text encoder
         self.processor.padding_side = "left"