turturtur250 commited on
Commit
c7b7735
·
verified ·
1 Parent(s): 3381c43

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +26 -0
README.md CHANGED
@@ -63,6 +63,32 @@ sha256sum model.safetensors
63
  | `tokenizer.json` | Tokenizer |
64
  | `tokenizer_config.json` | Tokenizer configuration |
65
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
66
  ## Evaluation
67
 
68
  Internal evaluation of this checkpoint was run with
 
63
  | `tokenizer.json` | Tokenizer |
64
  | `tokenizer_config.json` | Tokenizer configuration |
65
 
66
+ ## Usage
67
+
68
+ Verified with `transformers` 5.3.0. All 760 tensors in `model.safetensors` map
69
+ cleanly onto the model (no missing or unexpected keys); total parameter count is
70
+ 9.41B, of which 0.46B is the vision tower.
71
+
72
+ ```python
73
+ import torch
74
+ from transformers import AutoModelForImageTextToText, AutoProcessor
75
+
76
+ path = "InnovatorLab/InnovatorVL-v2"
77
+ processor = AutoProcessor.from_pretrained(path) # Qwen3VLProcessor
78
+ model = AutoModelForImageTextToText.from_pretrained( # Qwen3_5ForConditionalGeneration
79
+ path, dtype=torch.bfloat16, device_map="auto"
80
+ )
81
+ ```
82
+
83
+ ### Inference dependencies
84
+
85
+ The text stack uses hybrid attention with `linear_attention` layers. Without
86
+ [`flash-linear-attention`](https://github.com/fla-org/flash-linear-attention) and
87
+ [`causal-conv1d`](https://github.com/Dao-AILab/causal-conv1d) installed,
88
+ `transformers` logs *"The fast path is not available"* and falls back to a slower
89
+ pure-PyTorch implementation. Results are unaffected, but throughput is lower —
90
+ install both for production inference.
91
+
92
  ## Evaluation
93
 
94
  Internal evaluation of this checkpoint was run with