chengzeyi commited on
Commit
140017c
·
verified ·
1 Parent(s): 6ae1bee

Add model card metadata: base_model, license, pipeline_tag, tags

Browse files
Files changed (1) hide show
  1. README.md +56 -0
README.md ADDED
@@ -0,0 +1,56 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: Qwen/Qwen-Image
3
+ library_name: diffusers
4
+ license: apache-2.0
5
+ pipeline_tag: text-to-image
6
+ tags:
7
+ - qwen-image
8
+ - text-to-image
9
+ - quantized
10
+ - fp8
11
+ - e4m3
12
+ - diffusers
13
+ base_model_relation: quantized
14
+ ---
15
+ # Qwen-Image-e4m3
16
+
17
+ FP8 (e4m3) dynamically-quantized [Qwen-Image](https://huggingface.co/Qwen/Qwen-Image),
18
+ saved as a complete `QwenImagePipeline`.
19
+
20
+ ## What was changed
21
+
22
+ All 60 blocks of the `QwenImageTransformer2DModel` are quantized to
23
+ `e4m3_e4m3_dynamic` — `float8_e4m3fn` weights with dynamically scaled
24
+ `float8_e4m3fn` activations. The Qwen2.5-VL text encoder, the VAE and the
25
+ transformer's non-block tensors are untouched and stay in bf16. The transformer
26
+ drops from ~40.9 GB to ~20.5 GB.
27
+
28
+ Quantization was done with WaveSpeed's `xelerate.ao.quantize`. Weights are
29
+ stored as pickled `.bin` shards, so loading requires `use_safetensors=False`.
30
+
31
+ FP8 matmul needs Hopper (H100/H200) or newer to be faster than bf16; on older
32
+ GPUs the weights are dequantized on the fly and you only get the memory saving.
33
+
34
+ One caveat worth repeating from our own testing: on Qwen-Image, running fp8
35
+ *weight-and-activation* matmul under a fully fused fast path produces visible
36
+ quality loss. The configuration published here — dynamic per-tensor activation
37
+ scaling with bf16 accumulation — is the one that holds up.
38
+
39
+ ## Usage
40
+
41
+ ```python
42
+ import torch
43
+ from diffusers import QwenImagePipeline
44
+
45
+ pipe = QwenImagePipeline.from_pretrained(
46
+ "wavespeed/Qwen-Image-e4m3",
47
+ torch_dtype=torch.bfloat16,
48
+ use_safetensors=False,
49
+ ).to("cuda")
50
+
51
+ image = pipe("a chalkboard menu written in neat cursive").images[0]
52
+ ```
53
+
54
+ ## License
55
+
56
+ Apache-2.0, inherited from Qwen-Image.