BennyDaBall commited on
Commit
4da87a3
·
verified ·
1 Parent(s): 277a09f

Attribute Qwen Image 2.1 and update release name and workflows

Browse files
Files changed (49) hide show
  1. .gitattributes +1 -0
  2. NOTICE +14 -14
  3. PROMPTING.md +71 -71
  4. README.md +11 -9
  5. SHA256SUMS +49 -49
  6. VALIDATION.md +125 -125
  7. assets/ab_architecture_20260920.jpg +2 -2
  8. assets/ab_architecture_20260921.jpg +2 -2
  9. assets/ab_astronaut_20260920.jpg +2 -2
  10. assets/ab_astronaut_20260921.jpg +2 -2
  11. assets/ab_botanical_poster_20260920.jpg +2 -2
  12. assets/ab_botanical_poster_20260921.jpg +2 -2
  13. assets/ab_fashion_20260920.jpg +2 -2
  14. assets/ab_fashion_20260921.jpg +2 -2
  15. assets/ab_illustration_20260920.jpg +2 -2
  16. assets/ab_illustration_20260921.jpg +2 -2
  17. assets/ab_landscape_20260920.jpg +2 -2
  18. assets/ab_landscape_20260921.jpg +2 -2
  19. assets/ab_macro_20260920.jpg +2 -2
  20. assets/ab_macro_20260921.jpg +2 -2
  21. assets/ab_night_market_20260920.jpg +2 -2
  22. assets/ab_night_market_20260921.jpg +2 -2
  23. assets/ab_portrait_20260920.jpg +2 -2
  24. assets/ab_portrait_20260921.jpg +2 -2
  25. assets/ab_product_20260920.jpg +2 -2
  26. assets/ab_product_20260921.jpg +2 -2
  27. assets/ab_spatial_20260920.jpg +2 -2
  28. assets/ab_spatial_20260921.jpg +2 -2
  29. assets/ab_typography_20260920.jpg +2 -2
  30. assets/ab_typography_20260921.jpg +2 -2
  31. assets/editing.jpg +2 -2
  32. diffusion_models/qwen_image_2.1_nvfp4.manifest.json +2596 -0
  33. gallery.html +2 -2
  34. input/qwen_image_2.1_edit_reference.png +3 -0
  35. runtime/README.md +36 -36
  36. text_encoders/qwen3vl_8b_nvfp4.manifest.json +0 -0
  37. tools/convert.py +77 -77
  38. tools/run_workflow.py +26 -26
  39. tools/verify.py +36 -36
  40. validation/cases.json +84 -84
  41. validation/roundtrip.json +2 -2
  42. workflows/01_Text_to_Image.json +3 -3
  43. workflows/02_Image_Editing.json +4 -4
  44. workflows/03_Transparent_RGBA.json +3 -3
  45. workflows/04_2K_Typography.json +3 -3
  46. workflows/api/01_Text_to_Image.api.json +3 -3
  47. workflows/api/02_Image_Editing.api.json +4 -4
  48. workflows/api/03_Transparent_RGBA.api.json +3 -3
  49. workflows/api/04_2K_Typography.api.json +3 -3
.gitattributes CHANGED
@@ -119,3 +119,4 @@ assets/typography_20260921_nv-nv.webp filter=lfs diff=lfs merge=lfs -text
119
  assets/typography_2k_bf-bf_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
120
  assets/typography_2k_nv-nv_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
121
  input/prism_edit_reference.png filter=lfs diff=lfs merge=lfs -text
 
 
119
  assets/typography_2k_bf-bf_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
120
  assets/typography_2k_nv-nv_extended_00001_.png filter=lfs diff=lfs merge=lfs -text
121
  input/prism_edit_reference.png filter=lfs diff=lfs merge=lfs -text
122
+ input/qwen_image_2.1_edit_reference.png filter=lfs diff=lfs merge=lfs -text
NOTICE CHANGED
@@ -1,14 +1,14 @@
1
- Built with Qwen
2
-
3
- Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.
4
-
5
- BennyDaBall modified the official BF16 image transformer and Qwen3-VL encoder
6
- checkpoints on 2026-09-20 by converting selected linear weights to native NVFP4.
7
- The per-file manifests identify changed tensors and unchanged protected tensors.
8
- The Qwen Image 2.1 VAE is redistributed unchanged.
9
-
10
- Source: Comfy-Org/Qwen-Image-2.1
11
- Revision: ace0edeb3791a594ddfa36ed5f41a178a394e921
12
-
13
- This is an independent conversion, not an official Qwen or Comfy-Org release.
14
- Model use and redistribution remain subject to the included research license.
 
1
+ Built with Qwen
2
+
3
+ Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.
4
+
5
+ BennyDaBall modified the official BF16 image transformer and Qwen3-VL encoder
6
+ checkpoints on 2026-09-20 by converting selected linear weights to native NVFP4.
7
+ The per-file manifests identify changed tensors and unchanged protected tensors.
8
+ The Qwen Image 2.1 VAE is redistributed unchanged.
9
+
10
+ Source: Comfy-Org/Qwen-Image-2.1
11
+ Revision: ace0edeb3791a594ddfa36ed5f41a178a394e921
12
+
13
+ This is an independent conversion, not an official Qwen or Comfy-Org release.
14
+ Model use and redistribution remain subject to the included research license.
PROMPTING.md CHANGED
@@ -1,71 +1,71 @@
1
- # Prompting Prism / Qwen Image 2.1
2
-
3
- Start with a concrete natural-language paragraph: medium, subject, action,
4
- relationships, composition, lighting, materials, then exact lettering. Keep each
5
- attribute next to the object it describes. The supplied workflows use 40 steps,
6
- Euler/simple, CFG 1, denoise 1, and fixed seeds.
7
-
8
- ## Text to image
9
-
10
- Describe a scene that can be depicted in one image. For a product photo, specify
11
- the object, material, label, supporting objects, surface, background and lighting.
12
- For a portrait, specify the action and framing as well as the person's appearance.
13
- The included product prompt is a complete starting point.
14
-
15
- Set output width and height in Empty Latent Image. The encoder's resolution field
16
- controls reference-image resizing, not the independent text-to-image canvas.
17
- Changing image dimensions may change composition even with the same seed.
18
-
19
- At CFG 1, the negative conditioning branch does not guide sampling. Put essential
20
- requirements in the positive prompt. Higher CFG is not a validated improvement
21
- for this package.
22
-
23
- ## Lettering
24
-
25
- Quote the exact text and specify its position and hierarchy. Example:
26
-
27
- > Large cream headline at the top reads exactly "TAKE THE SCENIC ROUTE".
28
- > A smaller line at the bottom reads exactly "SLOW TRAVEL CLUB".
29
- > Generous margins, flat navy, coral and cream inks, subtle paper grain.
30
-
31
- Proofread the result. Clear headings and short labels worked in the release
32
- examples, but the astronaut test also shows small-label spelling errors in the
33
- BF16 baseline. Neither precision is a guarantee of perfect text.
34
-
35
- ## Image editing
36
-
37
- Upload the reference in Load Image. Connect it and the matching VAE to
38
- TextEncodeQwenImage21. Use that node's latent output for sampling, so the output
39
- matches the processed first reference's size. The included editing workflow
40
- already has those connections.
41
-
42
- Name the changes and the elements to preserve:
43
-
44
- > Change the amber glass bottle to deep translucent cobalt-blue glass and the
45
- > brass cap to brushed silver. Replace the orange with lime and rosemary with
46
- > lavender. Keep the bottle shape, cream EMBER label, marble pedestal, camera
47
- > angle, background and lighting.
48
-
49
- For multiple references, refer explicitly to the first and second images.
50
- Reference preservation is semantic, not pixel-locked: small details, hand poses,
51
- clothing coverage and object placement can change.
52
-
53
- ## Native transparency
54
-
55
- Use an explicit RGBA instruction around the subject description:
56
-
57
- > This is an RGBA image with transparency. A charming hand-painted gouache sticker
58
- > of a small orange fox curled around a glowing blue book, round spectacles,
59
- > moss-green scarf, clean complete silhouette and generous empty margin.
60
- > The image has alpha channel and the background is transparent.
61
-
62
- Save as PNG. Check the alpha channel or composite over a checkerboard. The hidden
63
- RGB under transparent pixels can be colored; an RGB-only viewer is not an alpha
64
- test. The supplied example has alpha values spanning 0 through 255.
65
-
66
- ## Comparisons
67
-
68
- Keep the prompt, seed, dimensions, sampler, steps, backend and encoder mode fixed.
69
- The optional NVFP4 encoder patch changes conditioning arithmetic and may change
70
- the generated image. Compare stock and accelerated modes as distinct settings.
71
- The gallery contains the complete 12-prompt, two-seed comparison set.
 
1
+ # Prompting Prism / Qwen Image 2.1
2
+
3
+ Start with a concrete natural-language paragraph: medium, subject, action,
4
+ relationships, composition, lighting, materials, then exact lettering. Keep each
5
+ attribute next to the object it describes. The supplied workflows use 40 steps,
6
+ Euler/simple, CFG 1, denoise 1, and fixed seeds.
7
+
8
+ ## Text to image
9
+
10
+ Describe a scene that can be depicted in one image. For a product photo, specify
11
+ the object, material, label, supporting objects, surface, background and lighting.
12
+ For a portrait, specify the action and framing as well as the person's appearance.
13
+ The included product prompt is a complete starting point.
14
+
15
+ Set output width and height in Empty Latent Image. The encoder's resolution field
16
+ controls reference-image resizing, not the independent text-to-image canvas.
17
+ Changing image dimensions may change composition even with the same seed.
18
+
19
+ At CFG 1, the negative conditioning branch does not guide sampling. Put essential
20
+ requirements in the positive prompt. Higher CFG is not a validated improvement
21
+ for this package.
22
+
23
+ ## Lettering
24
+
25
+ Quote the exact text and specify its position and hierarchy. Example:
26
+
27
+ > Large cream headline at the top reads exactly "TAKE THE SCENIC ROUTE".
28
+ > A smaller line at the bottom reads exactly "SLOW TRAVEL CLUB".
29
+ > Generous margins, flat navy, coral and cream inks, subtle paper grain.
30
+
31
+ Proofread the result. Clear headings and short labels worked in the release
32
+ examples, but the astronaut test also shows small-label spelling errors in the
33
+ BF16 baseline. Neither precision is a guarantee of perfect text.
34
+
35
+ ## Image editing
36
+
37
+ Upload the reference in Load Image. Connect it and the matching VAE to
38
+ TextEncodeQwenImage21. Use that node's latent output for sampling, so the output
39
+ matches the processed first reference's size. The included editing workflow
40
+ already has those connections.
41
+
42
+ Name the changes and the elements to preserve:
43
+
44
+ > Change the amber glass bottle to deep translucent cobalt-blue glass and the
45
+ > brass cap to brushed silver. Replace the orange with lime and rosemary with
46
+ > lavender. Keep the bottle shape, cream EMBER label, marble pedestal, camera
47
+ > angle, background and lighting.
48
+
49
+ For multiple references, refer explicitly to the first and second images.
50
+ Reference preservation is semantic, not pixel-locked: small details, hand poses,
51
+ clothing coverage and object placement can change.
52
+
53
+ ## Native transparency
54
+
55
+ Use an explicit RGBA instruction around the subject description:
56
+
57
+ > This is an RGBA image with transparency. A charming hand-painted gouache sticker
58
+ > of a small orange fox curled around a glowing blue book, round spectacles,
59
+ > moss-green scarf, clean complete silhouette and generous empty margin.
60
+ > The image has alpha channel and the background is transparent.
61
+
62
+ Save as PNG. Check the alpha channel or composite over a checkerboard. The hidden
63
+ RGB under transparent pixels can be colored; an RGB-only viewer is not an alpha
64
+ test. The supplied example has alpha values spanning 0 through 255.
65
+
66
+ ## Comparisons
67
+
68
+ Keep the prompt, seed, dimensions, sampler, steps, backend and encoder mode fixed.
69
+ The optional NVFP4 encoder patch changes conditioning arithmetic and may change
70
+ the generated image. Compare stock and accelerated modes as distinct settings.
71
+ The gallery contains the complete 12-prompt, two-seed comparison set.
README.md CHANGED
@@ -1,7 +1,7 @@
1
  ---
2
  license: other
3
  license_name: qwen-research-license
4
- license_link: https://huggingface.co/BennyDaBall/Prism-Image-2.1-NVFP4/blob/main/LICENSE
5
  base_model: Comfy-Org/Qwen-Image-2.1
6
  base_model_relation: quantized
7
  pipeline_tag: text-to-image
@@ -16,7 +16,9 @@ tags:
16
  - rtx-5090
17
  ---
18
 
19
- # 🚀 Prism Image 2.1 — NVFP4 for ComfyUI
 
 
20
 
21
  **The image model AND the text encoder. 12.42 GB for the complete package.**
22
 
@@ -43,8 +45,8 @@ NVFP4 support. Older builds without the Qwen Image 2.1 integration will not work
43
 
44
  | File | Where it goes | Size |
45
  |---|---|---:|
46
- | [prism_image_2.1_nvfp4.safetensors](diffusion_models/prism_image_2.1_nvfp4.safetensors) | `ComfyUI/models/diffusion_models/` | 4.20 GB |
47
- | [prism_qwen3vl_8b_nvfp4.safetensors](text_encoders/prism_qwen3vl_8b_nvfp4.safetensors) | `ComfyUI/models/text_encoders/` | 7.55 GB |
48
  | [qwen_image_2.1_vae_bf16.safetensors](vae/qwen_image_2.1_vae_bf16.safetensors) | `ComfyUI/models/vae/` | 0.68 GB |
49
 
50
  The original BF16 set is 32.44 GB. This package is **61.7% smaller on disk**.
@@ -57,7 +59,7 @@ VAE if you already have it.
57
  uses dequantized weights and FP32 activations. Denoising still uses NVFP4.
58
  3. Import one of the workflows below using **Ctrl+O**, then click **Run**.
59
  4. For editing, use **Upload** in Load Image to select the included
60
- [reference image](input/prism_edit_reference.png), or your own image.
61
 
62
  | Working workflow | Purpose |
63
  |---|---|
@@ -111,8 +113,8 @@ revision `ace0edeb3791a594ddfa36ed5f41a178a394e921`.
111
  Use the Python environment of the tested ComfyUI installation:
112
 
113
  ```powershell
114
- python tools/convert.py path/to/qwen_image_2.1_bf16.safetensors diffusion_models/prism_image_2.1_nvfp4.safetensors --component dit --comfy-root path/to/ComfyUI
115
- python tools/convert.py path/to/qwen3vl_8b_bf16.safetensors text_encoders/prism_qwen3vl_8b_nvfp4.safetensors --component encoder --comfy-root path/to/ComfyUI
116
  ```
117
 
118
  The converter uses installed Comfy Kitchen operations, validates matrix alignment,
@@ -126,7 +128,7 @@ Qwen created the base model; Comfy-Org provided the native BF16 checkpoint packa
126
  ComfyUI and Comfy Kitchen provide loading and GPU execution. Independent conversion,
127
  workflow packaging, and RTX 5090 validation by **BennyDaBall_OG**.
128
 
129
- The [Qwen Research License](https://huggingface.co/BennyDaBall/Prism-Image-2.1-NVFP4/blob/main/LICENSE) limits use to non-commercial research and
130
  evaluation; commercial use requires a separate license from Qwen. Read the full
131
- license and [NOTICE](https://huggingface.co/BennyDaBall/Prism-Image-2.1-NVFP4/blob/main/NOTICE). The included ComfyUI patch remains GPL-3.0.
132
  This repo is not affiliated with Qwen or Comfy-Org.
 
1
  ---
2
  license: other
3
  license_name: qwen-research-license
4
+ license_link: https://huggingface.co/BennyDaBall/Qwen-Image-2.1-NVFP4/blob/main/LICENSE
5
  base_model: Comfy-Org/Qwen-Image-2.1
6
  base_model_relation: quantized
7
  pipeline_tag: text-to-image
 
16
  - rtx-5090
17
  ---
18
 
19
+ # 🚀 Qwen Image 2.1 — NVFP4 for ComfyUI
20
+
21
+ **This is an independent NVFP4 quantization of Qwen Image 2.1, created by the Qwen team.** The original model and architecture are Qwen's work. I converted the official [Comfy-Org/Qwen-Image-2.1 BF16 checkpoint package](https://huggingface.co/Comfy-Org/Qwen-Image-2.1) and packaged the quantized weights, workflows, and validation here. This is not a new base model or an official Qwen release.
22
 
23
  **The image model AND the text encoder. 12.42 GB for the complete package.**
24
 
 
45
 
46
  | File | Where it goes | Size |
47
  |---|---|---:|
48
+ | [qwen_image_2.1_nvfp4.safetensors](diffusion_models/qwen_image_2.1_nvfp4.safetensors) | `ComfyUI/models/diffusion_models/` | 4.20 GB |
49
+ | [qwen3vl_8b_nvfp4.safetensors](text_encoders/qwen3vl_8b_nvfp4.safetensors) | `ComfyUI/models/text_encoders/` | 7.55 GB |
50
  | [qwen_image_2.1_vae_bf16.safetensors](vae/qwen_image_2.1_vae_bf16.safetensors) | `ComfyUI/models/vae/` | 0.68 GB |
51
 
52
  The original BF16 set is 32.44 GB. This package is **61.7% smaller on disk**.
 
59
  uses dequantized weights and FP32 activations. Denoising still uses NVFP4.
60
  3. Import one of the workflows below using **Ctrl+O**, then click **Run**.
61
  4. For editing, use **Upload** in Load Image to select the included
62
+ [reference image](input/qwen_image_2.1_edit_reference.png), or your own image.
63
 
64
  | Working workflow | Purpose |
65
  |---|---|
 
113
  Use the Python environment of the tested ComfyUI installation:
114
 
115
  ```powershell
116
+ python tools/convert.py path/to/qwen_image_2.1_bf16.safetensors diffusion_models/qwen_image_2.1_nvfp4.safetensors --component dit --comfy-root path/to/ComfyUI
117
+ python tools/convert.py path/to/qwen3vl_8b_bf16.safetensors text_encoders/qwen3vl_8b_nvfp4.safetensors --component encoder --comfy-root path/to/ComfyUI
118
  ```
119
 
120
  The converter uses installed Comfy Kitchen operations, validates matrix alignment,
 
128
  ComfyUI and Comfy Kitchen provide loading and GPU execution. Independent conversion,
129
  workflow packaging, and RTX 5090 validation by **BennyDaBall_OG**.
130
 
131
+ The [Qwen Research License](https://huggingface.co/BennyDaBall/Qwen-Image-2.1-NVFP4/blob/main/LICENSE) limits use to non-commercial research and
132
  evaluation; commercial use requires a separate license from Qwen. Read the full
133
+ license and [NOTICE](https://huggingface.co/BennyDaBall/Qwen-Image-2.1-NVFP4/blob/main/NOTICE). The included ComfyUI patch remains GPL-3.0.
134
  This repo is not affiliated with Qwen or Comfy-Org.
SHA256SUMS CHANGED
@@ -2,30 +2,30 @@ dbfd3a6bf37d800ab2d5dfbb977fcc6e1b93f4c57ec41ff423598b93a6536ccf assets/01_Text
2
  4e6fa2ad6423a4bf308707d3d892c3616b614f63d634f4202816e92c8c048910 assets/02_Image_Editing.png
3
  14dbba646932489c0bcbe513ec8151fa5d18005832202abe961d03544c8ef6f8 assets/03_Transparent_RGBA.png
4
  95b2ecdf94cf414f51dbb4402176b3deaaad8d984797658ca013825770f6cdf1 assets/04_2K_Typography.png
5
- 22ae98495169b99a864775e7720630cb9efb3de9a99a1b436c209a8508d1eb9e assets/ab_architecture_20260920.jpg
6
- 5dc83f4493aabec649da97c4fccb43778e7d408e65f940b8a36b6f7253cde16f assets/ab_architecture_20260921.jpg
7
- 3657cd495de80e9a383f71b116aacfb3bb3b708f6cb5a84e6087dd9afd1bcb1d assets/ab_astronaut_20260920.jpg
8
- 8651d645ec839a90e97cef397492dd1c6628a0ea0827127a7c44c7ff8994bfcc assets/ab_astronaut_20260921.jpg
9
- 9c8b033d751b02ca57404323e0da59ddc130771246d38b5296491be6c2923d16 assets/ab_botanical_poster_20260920.jpg
10
- f1249b09bed33b1955cc81ed068a165cfe0d7dc33d772e267e7c420085708e17 assets/ab_botanical_poster_20260921.jpg
11
- 47b64564650fbcc7311647f2cdf3b61fad50025babcc8ac22c2f0ffc02060791 assets/ab_fashion_20260920.jpg
12
- 7764788cb7287f1b4bd3f8572add42cb91518e3e0df96044c8536625575cda88 assets/ab_fashion_20260921.jpg
13
- 4ae372493b951d870c62c03eb1f082fa7ec6478846e7364d1154f851362f0f97 assets/ab_illustration_20260920.jpg
14
- fcb1880d6e7affa6d0c51622c937e42fa1618354b6db510608a6295f8d8ef1c2 assets/ab_illustration_20260921.jpg
15
- e55937e67bde8f14b2c7061cab4c18c738fb358b30fd9f04167148c9741de8a5 assets/ab_landscape_20260920.jpg
16
- be6e963c3b087b915ff64a76be4897c7297fde82fe570e706136730709027157 assets/ab_landscape_20260921.jpg
17
- 940640b5ab0535405937dada3d0e88c55c83feb5cd4d4a8e6737b7ec2d022884 assets/ab_macro_20260920.jpg
18
- b5b913b88b4b2d7ec8f37a7c08451a2a53ba6e965817eb7b49f87c46d3e18a0d assets/ab_macro_20260921.jpg
19
- 368181c2642d730adcc1b121041ee88ea5e25f23efe33d593982408ea50b97b5 assets/ab_night_market_20260920.jpg
20
- 5266ca07869dd84f6437a3f65ed3417c98404c5938f1341aeeb2d9766274d116 assets/ab_night_market_20260921.jpg
21
- ccfe23331acbd05c5fcf8b0db46a2d1eeb1872e09a62b5800bac0d983f8f005f assets/ab_portrait_20260920.jpg
22
- 6c4804dac52ce29c2e31a84e00fcc5f7f2e11abcae595f1a7adef3d37a876bdc assets/ab_portrait_20260921.jpg
23
- ab620447fc0e6b17e1bb47df61db047bd4ee15a9f7014c834af9abf595af7847 assets/ab_product_20260920.jpg
24
- 402220da9ae5f5d268fc9caa12f0decbd2d48164ad5453d031ce1b5dbc04b318 assets/ab_product_20260921.jpg
25
- 7498b7a7a3e766c04f27d5d28302f69c10f5c75bc7e7b562172bcd169efd03e3 assets/ab_spatial_20260920.jpg
26
- b3faee6e319a5c675c91167ef0704b63580e9dc8d0340550afeb0c72759b6792 assets/ab_spatial_20260921.jpg
27
- adeb0e4ed51ce0ff6a53793e3c301f4d29713f23c30ba21c4c0f4a2b849bb0d3 assets/ab_typography_20260920.jpg
28
- b1be26c4fa502e0a1fba47227573b932b4fe38d5aa8c61234392b40c94488949 assets/ab_typography_20260921.jpg
29
  81da8fc0e71b2b3373c1cf46734ca449d5f0ba049c77921f0c28e65d19b58a80 assets/architecture_20260920_bf-bf.webp
30
  bd505f621203fa42fc3c0521d1977ec3d2b7899db92c5bad1c8148c9c40a8f2a assets/architecture_20260920_nv-nv.webp
31
  468e043ba915698ccd77c67f5fae8ed53c0cd9219aed2f2465b7d658a8e6c3e8 assets/architecture_20260921_bf-bf.webp
@@ -43,7 +43,7 @@ b756961e69efadf8f66cf13784092777bf00429bbc194606562f4b4850aba697 assets/botanic
43
  b6a97056d829de05da472a06c8c1a7e199a4866ff0db1fed2597431f281a624c assets/edit_portrait_nv-nv_extended_00001_.png
44
  a0117e505673ca323e1c358ab5b6532ccc1fdcf9e08c87ff3a43712f50f0b501 assets/edit_product_bf-bf_extended_00001_.png
45
  4e6fa2ad6423a4bf308707d3d892c3616b614f63d634f4202816e92c8c048910 assets/edit_product_nv-nv_extended_00001_.png
46
- f05044570f14a5c4b75ef5803040b6cf5e8104640da248b64ffecf64c730b2f9 assets/editing.jpg
47
  ef6482fc15134a4ef537883911328b5e209a63d04da2a266972ebf1d435df0e5 assets/fashion_20260920_bf-bf.webp
48
  d7c370c533559d7752cae2593602aa2e6ef615b2ef3f0132686b8cfae6bf034a assets/fashion_20260920_nv-nv.webp
49
  82418e7dcbb5f8c7e48cd74715ce4f5acfa646d51e3de33bb177b945ad377926 assets/fashion_20260921_bf-bf.webp
@@ -92,35 +92,35 @@ ef6723ce3d0814ed25a54ff13fb1bc3efc38c7384d78af66122d78835cc665c0 assets/typogra
92
  6eeab96fd9b89a5c83cc945e7799c9a16b37b503cd4b961dd2364e339a7cacf0 assets/typography_2k_bf-bf_extended_00001_.png
93
  95b2ecdf94cf414f51dbb4402176b3deaaad8d984797658ca013825770f6cdf1 assets/typography_2k_nv-nv_extended_00001_.png
94
  f1352b74e5fa07594113b4fe69a914e8ad36e8dd7ce16a27b1f46df0d3098cfe COMPARISONS.md
95
- 50bfc0c644df3f661447024c380c9c433c8f0f5edcc4878b6695dcab533c7f1b diffusion_models/prism_image_2.1_nvfp4.manifest.json
96
- 4fdbaec94c19f7b8888f1e1ddbae05354c014492f9b2547b1cd8dca0ce923526 diffusion_models/prism_image_2.1_nvfp4.safetensors
97
- a0f54997ba63caa3540823bd5e7e6356f8c3ea16cc796b32477eb45b000066bf gallery.html
98
- 936375ea3fb35c7c33dbe59c7090c35edf019e6119bc6b775f74f3d859b49a54 input/prism_edit_reference.png
99
  06b7d356cd8cb6b79aa2eef3f934d1e1e8e1c2aa1e85dcc23915dd9a68197249 LICENSE
100
- c303de7acfe67db47fceaa8599c156895e462c06482dcd6b1e61d46ce02cadf3 NOTICE
101
- 8a8fafb7836a4d8eb02d176a3beef5d4dd522eff1084661c941494f5feba0f23 PROMPTING.md
102
- e6879db9dc6493b2211823b4891744acc66d3f6a44790a391e44485d3f0da2d9 README.md
103
  230184f60bae2feaf244f10a8bac053c8ff33a183bcc365b4d8b876d2b7f4809 runtime/LICENSE-GPL-3.0
104
  2c13ec9046ffa720484c569acce153f4759da47e9bc008af56d68d583854e2fa runtime/qwen21-nvfp4-conditioning.patch
105
- 524bf47b1f5b061514314d16f6afa6a02d21232545e3e96297c00781fc67733a runtime/README.md
106
- 41ac84507beed8d4bf5f46f7c49fa91c12162f73a354598b9943888411d57381 text_encoders/prism_qwen3vl_8b_nvfp4.manifest.json
107
- cdd9b85bcb60d5cb358ddecfd45a48fa15cb728b70d88bd2028840eaef095d16 text_encoders/prism_qwen3vl_8b_nvfp4.safetensors
108
- c9b72630ba5c3a6bb842afe948a778ae7d742e92d94407310976e8c0f2f7f27d tools/convert.py
109
- b5648e901612c6077ed55b93192bf3a152f842736e08b859082af6635e2339f9 tools/run_workflow.py
110
- 4363e4224c409a8c8a3099a06e75d9881206fbe4f541d44bde05dc872de2c3dc tools/verify.py
111
  bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9 vae/qwen_image_2.1_vae_bf16.safetensors
112
  2a6b6c514cce8041f313e7a68d9e344560b1ebf68feefb576a0c1bb627db0a94 validation/alpha.json
113
  24f0362777e6970f4d71085968478deb5b34e4116c678b11573ad637a66ac3b2 validation/benchmark.json
114
- 135690b380fbe6e4f0362dd5434abcb5736763848a224417bdc556228f39f235 validation/cases.json
115
  a8b8f8ad67db63da99a7dc20701dfdb74235c847d9985c9cd364103f8f6de709 validation/frontend-workflows.json
116
  6892ea5bd1fb4745f677c7e38f6e117b6daa813ad694500b3fd79e32e654e95f validation/kernel_audit.jsonl
117
- da5ab176ad20f4cdfb3994511eea8c8ce8ceb025d4d97c3e3dee6d2041670feb validation/roundtrip.json
118
- 1fd3f3e9d53fbef667628d34057b056d2f6c7f62fdbfe2230bbac1ba77635f38 VALIDATION.md
119
- 9a4ef8c30464cc453a29914904e01c5096051a35ee50cfc63fb9e0d968a07c00 workflows/01_Text_to_Image.json
120
- 42b78026114db814333ecbd7c86117f32f61a736fe630dad67950d9bfd3fec28 workflows/02_Image_Editing.json
121
- df44abe670d355ff885477364f206d70c4086c13df85a1705129ad6247f05524 workflows/03_Transparent_RGBA.json
122
- be45564d77c151fa92fbd3420cc7b62eb428d8e9d65d1755fe72057e1a9c16f7 workflows/04_2K_Typography.json
123
- 13a543282fae430f5a37b70688741300d022eb9a7adea02396898055c5d273de workflows/api/01_Text_to_Image.api.json
124
- 68224d0a803e823f78cf82dc636a6881b75671348b8cdd300f7cf8ac7493c241 workflows/api/02_Image_Editing.api.json
125
- cc742cfda9d55f9ecd5dbd2a6aeb87a8726d8c152f66c0f4b54b629a9ad4e3e2 workflows/api/03_Transparent_RGBA.api.json
126
- ab18f014220866315d52dbb5ff7e5a6f42bd95d6b70219a8c828266b468efff5 workflows/api/04_2K_Typography.api.json
 
2
  4e6fa2ad6423a4bf308707d3d892c3616b614f63d634f4202816e92c8c048910 assets/02_Image_Editing.png
3
  14dbba646932489c0bcbe513ec8151fa5d18005832202abe961d03544c8ef6f8 assets/03_Transparent_RGBA.png
4
  95b2ecdf94cf414f51dbb4402176b3deaaad8d984797658ca013825770f6cdf1 assets/04_2K_Typography.png
5
+ c0f90918906e78b09b81d2bbb290915af495ee8e0b641d32c6d77be8c8f11496 assets/ab_architecture_20260920.jpg
6
+ d7609fd547c9491aad7e67aee3f1343cbe30f3a07905bf5f3ffe3bb426834c0e assets/ab_architecture_20260921.jpg
7
+ cbc371361e53d62a80a0ec9930bd4e2f285cbd915a0354190c59a6b2f2b8f33c assets/ab_astronaut_20260920.jpg
8
+ f1f3636b6447b0fe30c4621964affa4a794d9853a41a600656c93da3da1c9299 assets/ab_astronaut_20260921.jpg
9
+ 30ebe70d150d8bdeb17380bf8f468259fd904b747964062007d3a4d4275b43da assets/ab_botanical_poster_20260920.jpg
10
+ abb1230284bf8ae6139884248fc8336251750cc3f3be2d28747ffe18931186b0 assets/ab_botanical_poster_20260921.jpg
11
+ 7df1054b9494fc0cb11881817369f07bc951496e93df7daa5bb4986aa5c51c15 assets/ab_fashion_20260920.jpg
12
+ faf7099abce238936add7ce9a6f76320606ad995a7f93f9cae4c89c19f3c32e7 assets/ab_fashion_20260921.jpg
13
+ 381f64aba27a935af10b2bf3b2855f58fac9d011463d312cad53b808b321f92e assets/ab_illustration_20260920.jpg
14
+ 434f840c2b773397c5a5f973b5f10852a59205ac6281096d31d3b17eb39dcca8 assets/ab_illustration_20260921.jpg
15
+ 0fcabe92930569f443d1ae23dd386eeb2fe8661d6367ad5a00e5bfc6c377d16a assets/ab_landscape_20260920.jpg
16
+ 1d5c5b8cfd5365fd3359cd45acba7cc5b08a9e656bf6cde1c81197ce4427be9e assets/ab_landscape_20260921.jpg
17
+ 569a63fcde0c992ee6a92165b887c0e10f2254d72c9ac97ea1717d613a6a87ac assets/ab_macro_20260920.jpg
18
+ baa2a5ab9b9ec4117a8805bff79020488c1b8bdd4203c50dbdbe4cda5a7f6591 assets/ab_macro_20260921.jpg
19
+ a1f91a0984720cd3cb805164054da0d9d29445ae2f9e64d12e71843eee9737e1 assets/ab_night_market_20260920.jpg
20
+ 68d5273effcb9481ea99df6c770b3d3ca6a55e90fc3aba2816389bf2a579239d assets/ab_night_market_20260921.jpg
21
+ 2674cdd21e1a3a90882335b13328edf0382fba99537d824cf637f2c6a0682037 assets/ab_portrait_20260920.jpg
22
+ 92dd4e5cb7c4393032eee6b0787bfaf79ef475822a1dbfd8699aee913b807fda assets/ab_portrait_20260921.jpg
23
+ dabc12a4a28d27b15c3206d96c099a5a2345c21929ace89e9ccef7b87230720f assets/ab_product_20260920.jpg
24
+ 92b6e20ea4b4d14249097ae2bf0a67c2485cf3452dc35020fc4817191928f5d3 assets/ab_product_20260921.jpg
25
+ 7bdf59afd58cb6481535e5ff3d540794b31f77dcdc47799fdb3d33a988a8bfb4 assets/ab_spatial_20260920.jpg
26
+ 55cbd086fd0336f23a3b826eb9f92f1a19ccaa04ebcd818eb6d583aa4c51bcc8 assets/ab_spatial_20260921.jpg
27
+ ab25af5746a219b1061d03a5ec63c125bee38c31526eaaee314bb444c01defef assets/ab_typography_20260920.jpg
28
+ 99dbca8e65496f84192c04434ebef768a955a460855cfbdbbf06ff23d00d2814 assets/ab_typography_20260921.jpg
29
  81da8fc0e71b2b3373c1cf46734ca449d5f0ba049c77921f0c28e65d19b58a80 assets/architecture_20260920_bf-bf.webp
30
  bd505f621203fa42fc3c0521d1977ec3d2b7899db92c5bad1c8148c9c40a8f2a assets/architecture_20260920_nv-nv.webp
31
  468e043ba915698ccd77c67f5fae8ed53c0cd9219aed2f2465b7d658a8e6c3e8 assets/architecture_20260921_bf-bf.webp
 
43
  b6a97056d829de05da472a06c8c1a7e199a4866ff0db1fed2597431f281a624c assets/edit_portrait_nv-nv_extended_00001_.png
44
  a0117e505673ca323e1c358ab5b6532ccc1fdcf9e08c87ff3a43712f50f0b501 assets/edit_product_bf-bf_extended_00001_.png
45
  4e6fa2ad6423a4bf308707d3d892c3616b614f63d634f4202816e92c8c048910 assets/edit_product_nv-nv_extended_00001_.png
46
+ 70ac27dd7177e821ea2341d79f09065953f7b9dafcd2ff60f6f89bc32377c61f assets/editing.jpg
47
  ef6482fc15134a4ef537883911328b5e209a63d04da2a266972ebf1d435df0e5 assets/fashion_20260920_bf-bf.webp
48
  d7c370c533559d7752cae2593602aa2e6ef615b2ef3f0132686b8cfae6bf034a assets/fashion_20260920_nv-nv.webp
49
  82418e7dcbb5f8c7e48cd74715ce4f5acfa646d51e3de33bb177b945ad377926 assets/fashion_20260921_bf-bf.webp
 
92
  6eeab96fd9b89a5c83cc945e7799c9a16b37b503cd4b961dd2364e339a7cacf0 assets/typography_2k_bf-bf_extended_00001_.png
93
  95b2ecdf94cf414f51dbb4402176b3deaaad8d984797658ca013825770f6cdf1 assets/typography_2k_nv-nv_extended_00001_.png
94
  f1352b74e5fa07594113b4fe69a914e8ad36e8dd7ce16a27b1f46df0d3098cfe COMPARISONS.md
95
+ 0f60367e0d01f795f2afd62466d44119653258592d586bc9d64c131208b08b8b diffusion_models/qwen_image_2.1_nvfp4.manifest.json
96
+ 4fdbaec94c19f7b8888f1e1ddbae05354c014492f9b2547b1cd8dca0ce923526 diffusion_models/qwen_image_2.1_nvfp4.safetensors
97
+ fa6f566d4cf29b171bc996ad81857cf0dbdbb9cc5b8524297c1f1ca1738db99f gallery.html
98
+ 936375ea3fb35c7c33dbe59c7090c35edf019e6119bc6b775f74f3d859b49a54 input/qwen_image_2.1_edit_reference.png
99
  06b7d356cd8cb6b79aa2eef3f934d1e1e8e1c2aa1e85dcc23915dd9a68197249 LICENSE
100
+ 0215c805cf5ee96af84731daffb90ba39942c33ef8159b897c754156ffbc95f6 NOTICE
101
+ 09a924579e9efea1ca3c3bccec2bf04ab6bd00dda6e1719a171f32a115aebcad PROMPTING.md
102
+ 075a0e74648bc1bbd0dae2649244fec67801ba7407dc604057f11f5d915feb2c README.md
103
  230184f60bae2feaf244f10a8bac053c8ff33a183bcc365b4d8b876d2b7f4809 runtime/LICENSE-GPL-3.0
104
  2c13ec9046ffa720484c569acce153f4759da47e9bc008af56d68d583854e2fa runtime/qwen21-nvfp4-conditioning.patch
105
+ 787b656910e289283b0671311b479949b2496f9a233413fa71ce0f478866d9dc runtime/README.md
106
+ 640b0140cb6cd3b93cfd4b7899f6bac48b0c596fc707e03c926fdc15aa57352e text_encoders/qwen3vl_8b_nvfp4.manifest.json
107
+ cdd9b85bcb60d5cb358ddecfd45a48fa15cb728b70d88bd2028840eaef095d16 text_encoders/qwen3vl_8b_nvfp4.safetensors
108
+ f0a09696345c6fb96149f230ea095a5b77d31ebfeda24fcbab95ff3b6dc7e0c0 tools/convert.py
109
+ 22f7cf73b46589cea8f8c78169f0f7c6a3a4c8ce98957ef9ab7cfdd6929e37ad tools/run_workflow.py
110
+ b99db3f33715e108ba69180b6afd61d237b4f588e5cb2af34c1a8cbe5e1285e5 tools/verify.py
111
  bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9 vae/qwen_image_2.1_vae_bf16.safetensors
112
  2a6b6c514cce8041f313e7a68d9e344560b1ebf68feefb576a0c1bb627db0a94 validation/alpha.json
113
  24f0362777e6970f4d71085968478deb5b34e4116c678b11573ad637a66ac3b2 validation/benchmark.json
114
+ 127eebed48bd8f385296f307b8640f6d1f1525839559e61f8a3fbb27dfd624c7 validation/cases.json
115
  a8b8f8ad67db63da99a7dc20701dfdb74235c847d9985c9cd364103f8f6de709 validation/frontend-workflows.json
116
  6892ea5bd1fb4745f677c7e38f6e117b6daa813ad694500b3fd79e32e654e95f validation/kernel_audit.jsonl
117
+ d43d76af370457d5c9e4517ad34edc100f3309a9706e446e3949b56bf1e0f5dc validation/roundtrip.json
118
+ e76d0ce239c0abb38eb191e46a2403b3a37e3d450d46ec8ef01ad7be348788c9 VALIDATION.md
119
+ d7061cab3abe77078c72b8ef55f4068e90958223d2b29dd9f94251d862e92a30 workflows/01_Text_to_Image.json
120
+ 23a68d924c07a6aa9b879f0738fcf297cf0336d59e86a54cd8ea3eaa31735a29 workflows/02_Image_Editing.json
121
+ d7000d73f6247c1152449f92fd62d49cdb3a44f39b147306e85c48b49311b694 workflows/03_Transparent_RGBA.json
122
+ 4976a78af3b1211fe6a81804a7f136739a02e38109d7327dd824126d230a5de8 workflows/04_2K_Typography.json
123
+ b8d796cc3fb392cd18d90958c75fd939302eda71ca00c81e4c74ab598a7fa9b1 workflows/api/01_Text_to_Image.api.json
124
+ a246089819e2cf1305d2a89e44d6f547ec7f38cfc8c93ee16c52bfdd433750b0 workflows/api/02_Image_Editing.api.json
125
+ d30e4fbbc6c3869bf12db07b8dfb8f152cfaa9886d8d5ca0b5fd6af4c7da95cc workflows/api/03_Transparent_RGBA.api.json
126
+ a4661936b491d4990397861a9d1dfff4c08d52633e1b06caef7474c29ef59911 workflows/api/04_2K_Typography.api.json
VALIDATION.md CHANGED
@@ -1,125 +1,125 @@
1
- # Validation — RTX 5090, 20 September 2026
2
-
3
- ## What passed
4
-
5
- - Both native safetensors checkpoints loaded through standard ComfyUI loaders.
6
- - 48 main renders: 12 prompts × two seeds × BF16/NVFP4, all at 1024×1024.
7
- - Eight extended renders: product edit, portrait edit, transparent RGBA and 2K
8
- typography, each with BF16 and NVFP4.
9
- - Six held-out renders: Chinese lettering, 1536×864 landscape and two-reference
10
- image editing, each with BF16 and NVFP4. These prompts were not used to choose
11
- the conversion recipe.
12
- - Four shipped workflows imported and executed through the actual frontend.
13
- - Three subjects tested across all four combinations of BF16/NVFP4 transformer
14
- and encoder weights, using the original stock conditioning path.
15
- - Real SM120 block-scaled FP4 CUDA kernels observed in both language conditioning
16
- and image denoising with the accelerated encoder patch.
17
- - All 571 protected tensors checked for exact equality against the source files.
18
- All 444 quantized matrices checked for packed shape, marker, scale dtype and
19
- finite scales. Source and release file hashes recorded.
20
- - Hub-uploaded weights downloaded into a separate directory and matched against
21
- the release SHA-256 hashes. The VAE hash matches the official original.
22
- - All four API workflows also passed using the Hub-downloaded files on the
23
- unpatched base runtime, with custom nodes disabled. Thus stock compatibility
24
- and optional accelerated conditioning were tested separately.
25
-
26
- No calibration, optimization against the test prompts, or fine-tuning was used.
27
- The quantization recipe remained unchanged after its first conversion.
28
-
29
- ## Measured performance
30
-
31
- 1024×1024 EMBER product prompt from the shipped generation workflow, seed
32
- 20260922, 40 steps, Euler/simple, CFG 1, denoise 1. One first-after-unload run,
33
- then three warm runs per variant. **Every node executed on every repeat** using
34
- `--cache-none`; the histories were checked for an empty cached-node list.
35
- Elapsed time comes from server execution-start/success timestamps, not polling
36
- latency. Includes conditioning, denoising, VAE decode and PNG save.
37
-
38
- | Weights / encoder mode | Warm median | Warm range | First after unload | Largest sampled GPU use, warm |
39
- |---|---:|---:|---:|---:|
40
- | Official BF16 / stock FP32 conditioning | 15.629 s | 15.553–15.650 s | 18.932 s | 31,677 MiB |
41
- | Official INT8 ConvRot / stock conditioning | 7.553 s | 7.496–7.591 s | 9.973 s | 18,433 MiB |
42
- | Prism NVFP4 / patched BF16 + NVFP4 conditioning | **6.746 s** | **6.706–6.749 s** | **8.083 s** | **13,249 MiB** |
43
-
44
- This NVFP4 workflow delivered **2.32× the BF16 throughput** and took **10.7% less
45
- time than INT8** in this one controlled test. These are local measurements, not
46
- general speed guarantees. First-after-unload is not a cold filesystem-cache or
47
- fresh-boot test. The quality-suite runs used ordinary node caching and are not
48
- the basis of this performance table.
49
-
50
- GPU memory was sampled through `nvidia-smi` every 100 ms. Values are **whole-GPU
51
- usage**, including Windows and other idle desktop services, not PyTorch peak
52
- allocated memory or a minimum VRAM requirement. The largest sampled NVFP4 warm
53
- value is about 12.94 GiB. No other model inference job was running during this
54
- benchmark. Hardware other than the RTX 5090 was not benchmarked.
55
-
56
- Raw measurements: [benchmark.json](validation/benchmark.json).
57
-
58
- ## Quality observations and limits
59
-
60
- The main set covers portraits, landscapes, glass products, typography, object
61
- counts and spatial relationships, illustration, fashion, night scenes,
62
- architecture, macro, astronaut portraits and botanical posters. All 24 matched
63
- pairs are published in [COMPARISONS.md](COMPARISONS.md).
64
-
65
- Visual review found coherent images and preserved useful text, material and
66
- editing capabilities. It also found real differences: changed portrait details,
67
- object placement, typography layout and small material/color details. In one
68
- macro pair the quantized spider is less green. Small astronaut-label spelling
69
- was imperfect in the BF16 baseline too. The Chinese probe retained the requested
70
- wording, but NVFP4 moved the subtitle above the cup rather than below it.
71
-
72
- The two-reference test combined the reference woman with the EMBER bottle and
73
- removed the violin, while changing hand pose. The product edit preserved the
74
- label and scene while changing glass, metal and props. Portrait recoloring can
75
- also change sleeve coverage. This is semantic reference preservation, not a
76
- pixel-locked edit.
77
-
78
- The transparent example has genuine RGBA output: alpha minimum 0, maximum 255,
79
- 49.3% of pixels below alpha 16 and 49.7% above alpha 240. See the original
80
- [RGBA PNG](assets/03_Transparent_RGBA.png) and the
81
- [checkerboard composite](assets/transparency-checker.jpg). RGB-only viewers may
82
- show the purple RGB values stored underneath transparent pixels.
83
-
84
- This is a small local evaluation, not a blinded preference study or a benchmark
85
- proving no quality loss. There was no LoRA, ControlNet, training or video test.
86
- Long text, exhaustive multilingual coverage and every reference-image count were
87
- not tested. Stock encoder arithmetic and accelerated encoder arithmetic can
88
- produce different images even with identical weights and seeds.
89
-
90
- ## Runtime and kernel evidence
91
-
92
- - Windows, Python 3.13.12, RTX 5090 32 GB.
93
- - ComfyUI 0.36.0, base commit `99073836d45f66053c45ba8564984e6def9cebba`.
94
- - Included Qwen-specific encoder patch; local validation commit `2d324220`.
95
- - PyTorch `2.14.0+cu130`; comfy-kitchen `0.2.35`; comfy-aimdo `0.5.5`.
96
- - Frontend package `1.53.6`; safetensors `0.8.0`.
97
- - Comfy Kitchen attention, normal VRAM mode, async offload, batch size one.
98
- - All tests ran with `--disable-all-custom-nodes`.
99
-
100
- The profiler captured kernels beginning
101
- `cutlass3x_sm120_bstensorop_s16864gemm_block_scaled_ue4m3xe2m1_ue4m3xe2m1`
102
- for short conditioning sequences and image-denoising sequences. This confirms
103
- hardware FP4 execution, rather than relying on checkpoint names or successful
104
- loading. Raw observations: [kernel_audit.jsonl](validation/kernel_audit.jsonl).
105
- No NVFP4 matmul-fallback warning occurred in the successful audited runs.
106
-
107
- The first acceleration probe exposed FP32 activations entering a BF16/FP16-only
108
- quantizer. The included patch fixes that at the Qwen Image 2.1 encoding boundary:
109
- vision preprocessing stays unchanged, the resulting multimodal embeddings become
110
- BF16 only for NVFP4 conditioning on supported hardware, and existing ComfyUI
111
- matmul dispatch is used. BF16 and INT8 checkpoints retain their original path.
112
-
113
- ## Precision and provenance
114
-
115
- The DiT converts 192 matrices and retains its other 73 tensors. The encoder
116
- converts 252 matrices and retains its other 498 tensors, including the complete
117
- vision tower, embeddings and output head. Per-matrix relative reconstruction
118
- RMSE is recorded in the manifests; it is not an image-quality score.
119
-
120
- Source revision: `Comfy-Org/Qwen-Image-2.1@ace0edeb3791a594ddfa36ed5f41a178a394e921`.
121
- The image transformer source SHA-256 is
122
- `89f4158d066cc33906a199fca85634f766892dd78f49b6698dabf187ac86c4bc`.
123
- The encoder source SHA-256 is
124
- `68bdc82bc1b66851162ae656225e7e2068166b603db19bd5d5a3b90eb12669a9`.
125
- Full tensor manifests sit beside each checkpoint; `SHA256SUMS` covers the release.
 
1
+ # Validation — RTX 5090, 20 September 2026
2
+
3
+ ## What passed
4
+
5
+ - Both native safetensors checkpoints loaded through standard ComfyUI loaders.
6
+ - 48 main renders: 12 prompts × two seeds × BF16/NVFP4, all at 1024×1024.
7
+ - Eight extended renders: product edit, portrait edit, transparent RGBA and 2K
8
+ typography, each with BF16 and NVFP4.
9
+ - Six held-out renders: Chinese lettering, 1536×864 landscape and two-reference
10
+ image editing, each with BF16 and NVFP4. These prompts were not used to choose
11
+ the conversion recipe.
12
+ - Four shipped workflows imported and executed through the actual frontend.
13
+ - Three subjects tested across all four combinations of BF16/NVFP4 transformer
14
+ and encoder weights, using the original stock conditioning path.
15
+ - Real SM120 block-scaled FP4 CUDA kernels observed in both language conditioning
16
+ and image denoising with the accelerated encoder patch.
17
+ - All 571 protected tensors checked for exact equality against the source files.
18
+ All 444 quantized matrices checked for packed shape, marker, scale dtype and
19
+ finite scales. Source and release file hashes recorded.
20
+ - Hub-uploaded weights downloaded into a separate directory and matched against
21
+ the release SHA-256 hashes. The VAE hash matches the official original.
22
+ - All four API workflows also passed using the Hub-downloaded files on the
23
+ unpatched base runtime, with custom nodes disabled. Thus stock compatibility
24
+ and optional accelerated conditioning were tested separately.
25
+
26
+ No calibration, optimization against the test prompts, or fine-tuning was used.
27
+ The quantization recipe remained unchanged after its first conversion.
28
+
29
+ ## Measured performance
30
+
31
+ 1024×1024 EMBER product prompt from the shipped generation workflow, seed
32
+ 20260922, 40 steps, Euler/simple, CFG 1, denoise 1. One first-after-unload run,
33
+ then three warm runs per variant. **Every node executed on every repeat** using
34
+ `--cache-none`; the histories were checked for an empty cached-node list.
35
+ Elapsed time comes from server execution-start/success timestamps, not polling
36
+ latency. Includes conditioning, denoising, VAE decode and PNG save.
37
+
38
+ | Weights / encoder mode | Warm median | Warm range | First after unload | Largest sampled GPU use, warm |
39
+ |---|---:|---:|---:|---:|
40
+ | Official BF16 / stock FP32 conditioning | 15.629 s | 15.553–15.650 s | 18.932 s | 31,677 MiB |
41
+ | Official INT8 ConvRot / stock conditioning | 7.553 s | 7.496–7.591 s | 9.973 s | 18,433 MiB |
42
+ | Qwen Image 2.1 NVFP4 / patched BF16 + NVFP4 conditioning | **6.746 s** | **6.706–6.749 s** | **8.083 s** | **13,249 MiB** |
43
+
44
+ This NVFP4 workflow delivered **2.32× the BF16 throughput** and took **10.7% less
45
+ time than INT8** in this one controlled test. These are local measurements, not
46
+ general speed guarantees. First-after-unload is not a cold filesystem-cache or
47
+ fresh-boot test. The quality-suite runs used ordinary node caching and are not
48
+ the basis of this performance table.
49
+
50
+ GPU memory was sampled through `nvidia-smi` every 100 ms. Values are **whole-GPU
51
+ usage**, including Windows and other idle desktop services, not PyTorch peak
52
+ allocated memory or a minimum VRAM requirement. The largest sampled NVFP4 warm
53
+ value is about 12.94 GiB. No other model inference job was running during this
54
+ benchmark. Hardware other than the RTX 5090 was not benchmarked.
55
+
56
+ Raw measurements: [benchmark.json](validation/benchmark.json).
57
+
58
+ ## Quality observations and limits
59
+
60
+ The main set covers portraits, landscapes, glass products, typography, object
61
+ counts and spatial relationships, illustration, fashion, night scenes,
62
+ architecture, macro, astronaut portraits and botanical posters. All 24 matched
63
+ pairs are published in [COMPARISONS.md](COMPARISONS.md).
64
+
65
+ Visual review found coherent images and preserved useful text, material and
66
+ editing capabilities. It also found real differences: changed portrait details,
67
+ object placement, typography layout and small material/color details. In one
68
+ macro pair the quantized spider is less green. Small astronaut-label spelling
69
+ was imperfect in the BF16 baseline too. The Chinese probe retained the requested
70
+ wording, but NVFP4 moved the subtitle above the cup rather than below it.
71
+
72
+ The two-reference test combined the reference woman with the EMBER bottle and
73
+ removed the violin, while changing hand pose. The product edit preserved the
74
+ label and scene while changing glass, metal and props. Portrait recoloring can
75
+ also change sleeve coverage. This is semantic reference preservation, not a
76
+ pixel-locked edit.
77
+
78
+ The transparent example has genuine RGBA output: alpha minimum 0, maximum 255,
79
+ 49.3% of pixels below alpha 16 and 49.7% above alpha 240. See the original
80
+ [RGBA PNG](assets/03_Transparent_RGBA.png) and the
81
+ [checkerboard composite](assets/transparency-checker.jpg). RGB-only viewers may
82
+ show the purple RGB values stored underneath transparent pixels.
83
+
84
+ This is a small local evaluation, not a blinded preference study or a benchmark
85
+ proving no quality loss. There was no LoRA, ControlNet, training or video test.
86
+ Long text, exhaustive multilingual coverage and every reference-image count were
87
+ not tested. Stock encoder arithmetic and accelerated encoder arithmetic can
88
+ produce different images even with identical weights and seeds.
89
+
90
+ ## Runtime and kernel evidence
91
+
92
+ - Windows, Python 3.13.12, RTX 5090 32 GB.
93
+ - ComfyUI 0.36.0, base commit `99073836d45f66053c45ba8564984e6def9cebba`.
94
+ - Included Qwen-specific encoder patch; local validation commit `2d324220`.
95
+ - PyTorch `2.14.0+cu130`; comfy-kitchen `0.2.35`; comfy-aimdo `0.5.5`.
96
+ - Frontend package `1.53.6`; safetensors `0.8.0`.
97
+ - Comfy Kitchen attention, normal VRAM mode, async offload, batch size one.
98
+ - All tests ran with `--disable-all-custom-nodes`.
99
+
100
+ The profiler captured kernels beginning
101
+ `cutlass3x_sm120_bstensorop_s16864gemm_block_scaled_ue4m3xe2m1_ue4m3xe2m1`
102
+ for short conditioning sequences and image-denoising sequences. This confirms
103
+ hardware FP4 execution, rather than relying on checkpoint names or successful
104
+ loading. Raw observations: [kernel_audit.jsonl](validation/kernel_audit.jsonl).
105
+ No NVFP4 matmul-fallback warning occurred in the successful audited runs.
106
+
107
+ The first acceleration probe exposed FP32 activations entering a BF16/FP16-only
108
+ quantizer. The included patch fixes that at the Qwen Image 2.1 encoding boundary:
109
+ vision preprocessing stays unchanged, the resulting multimodal embeddings become
110
+ BF16 only for NVFP4 conditioning on supported hardware, and existing ComfyUI
111
+ matmul dispatch is used. BF16 and INT8 checkpoints retain their original path.
112
+
113
+ ## Precision and provenance
114
+
115
+ The DiT converts 192 matrices and retains its other 73 tensors. The encoder
116
+ converts 252 matrices and retains its other 498 tensors, including the complete
117
+ vision tower, embeddings and output head. Per-matrix relative reconstruction
118
+ RMSE is recorded in the manifests; it is not an image-quality score.
119
+
120
+ Source revision: `Comfy-Org/Qwen-Image-2.1@ace0edeb3791a594ddfa36ed5f41a178a394e921`.
121
+ The image transformer source SHA-256 is
122
+ `89f4158d066cc33906a199fca85634f766892dd78f49b6698dabf187ac86c4bc`.
123
+ The encoder source SHA-256 is
124
+ `68bdc82bc1b66851162ae656225e7e2068166b603db19bd5d5a3b90eb12669a9`.
125
+ Full tensor manifests sit beside each checkpoint; `SHA256SUMS` covers the release.
assets/ab_architecture_20260920.jpg CHANGED

Git LFS Details

  • SHA256: 22ae98495169b99a864775e7720630cb9efb3de9a99a1b436c209a8508d1eb9e
  • Pointer size: 131 Bytes
  • Size of remote file: 676 kB

Git LFS Details

  • SHA256: c0f90918906e78b09b81d2bbb290915af495ee8e0b641d32c6d77be8c8f11496
  • Pointer size: 131 Bytes
  • Size of remote file: 677 kB
assets/ab_architecture_20260921.jpg CHANGED

Git LFS Details

  • SHA256: 5dc83f4493aabec649da97c4fccb43778e7d408e65f940b8a36b6f7253cde16f
  • Pointer size: 131 Bytes
  • Size of remote file: 744 kB

Git LFS Details

  • SHA256: d7609fd547c9491aad7e67aee3f1343cbe30f3a07905bf5f3ffe3bb426834c0e
  • Pointer size: 131 Bytes
  • Size of remote file: 746 kB
assets/ab_astronaut_20260920.jpg CHANGED

Git LFS Details

  • SHA256: 3657cd495de80e9a383f71b116aacfb3bb3b708f6cb5a84e6087dd9afd1bcb1d
  • Pointer size: 131 Bytes
  • Size of remote file: 710 kB

Git LFS Details

  • SHA256: cbc371361e53d62a80a0ec9930bd4e2f285cbd915a0354190c59a6b2f2b8f33c
  • Pointer size: 131 Bytes
  • Size of remote file: 712 kB
assets/ab_astronaut_20260921.jpg CHANGED

Git LFS Details

  • SHA256: 8651d645ec839a90e97cef397492dd1c6628a0ea0827127a7c44c7ff8994bfcc
  • Pointer size: 131 Bytes
  • Size of remote file: 744 kB

Git LFS Details

  • SHA256: f1f3636b6447b0fe30c4621964affa4a794d9853a41a600656c93da3da1c9299
  • Pointer size: 131 Bytes
  • Size of remote file: 746 kB
assets/ab_botanical_poster_20260920.jpg CHANGED

Git LFS Details

  • SHA256: 9c8b033d751b02ca57404323e0da59ddc130771246d38b5296491be6c2923d16
  • Pointer size: 131 Bytes
  • Size of remote file: 340 kB

Git LFS Details

  • SHA256: 30ebe70d150d8bdeb17380bf8f468259fd904b747964062007d3a4d4275b43da
  • Pointer size: 131 Bytes
  • Size of remote file: 342 kB
assets/ab_botanical_poster_20260921.jpg CHANGED

Git LFS Details

  • SHA256: f1249b09bed33b1955cc81ed068a165cfe0d7dc33d772e267e7c420085708e17
  • Pointer size: 131 Bytes
  • Size of remote file: 333 kB

Git LFS Details

  • SHA256: abb1230284bf8ae6139884248fc8336251750cc3f3be2d28747ffe18931186b0
  • Pointer size: 131 Bytes
  • Size of remote file: 335 kB
assets/ab_fashion_20260920.jpg CHANGED

Git LFS Details

  • SHA256: 47b64564650fbcc7311647f2cdf3b61fad50025babcc8ac22c2f0ffc02060791
  • Pointer size: 131 Bytes
  • Size of remote file: 565 kB

Git LFS Details

  • SHA256: 7df1054b9494fc0cb11881817369f07bc951496e93df7daa5bb4986aa5c51c15
  • Pointer size: 131 Bytes
  • Size of remote file: 567 kB
assets/ab_fashion_20260921.jpg CHANGED

Git LFS Details

  • SHA256: 7764788cb7287f1b4bd3f8572add42cb91518e3e0df96044c8536625575cda88
  • Pointer size: 131 Bytes
  • Size of remote file: 554 kB

Git LFS Details

  • SHA256: faf7099abce238936add7ce9a6f76320606ad995a7f93f9cae4c89c19f3c32e7
  • Pointer size: 131 Bytes
  • Size of remote file: 555 kB
assets/ab_illustration_20260920.jpg CHANGED

Git LFS Details

  • SHA256: 4ae372493b951d870c62c03eb1f082fa7ec6478846e7364d1154f851362f0f97
  • Pointer size: 131 Bytes
  • Size of remote file: 930 kB

Git LFS Details

  • SHA256: 381f64aba27a935af10b2bf3b2855f58fac9d011463d312cad53b808b321f92e
  • Pointer size: 131 Bytes
  • Size of remote file: 932 kB
assets/ab_illustration_20260921.jpg CHANGED

Git LFS Details

  • SHA256: fcb1880d6e7affa6d0c51622c937e42fa1618354b6db510608a6295f8d8ef1c2
  • Pointer size: 131 Bytes
  • Size of remote file: 955 kB

Git LFS Details

  • SHA256: 434f840c2b773397c5a5f973b5f10852a59205ac6281096d31d3b17eb39dcca8
  • Pointer size: 131 Bytes
  • Size of remote file: 957 kB
assets/ab_landscape_20260920.jpg CHANGED

Git LFS Details

  • SHA256: e55937e67bde8f14b2c7061cab4c18c738fb358b30fd9f04167148c9741de8a5
  • Pointer size: 131 Bytes
  • Size of remote file: 688 kB

Git LFS Details

  • SHA256: 0fcabe92930569f443d1ae23dd386eeb2fe8661d6367ad5a00e5bfc6c377d16a
  • Pointer size: 131 Bytes
  • Size of remote file: 690 kB
assets/ab_landscape_20260921.jpg CHANGED

Git LFS Details

  • SHA256: be6e963c3b087b915ff64a76be4897c7297fde82fe570e706136730709027157
  • Pointer size: 131 Bytes
  • Size of remote file: 614 kB

Git LFS Details

  • SHA256: 1d5c5b8cfd5365fd3359cd45acba7cc5b08a9e656bf6cde1c81197ce4427be9e
  • Pointer size: 131 Bytes
  • Size of remote file: 616 kB
assets/ab_macro_20260920.jpg CHANGED

Git LFS Details

  • SHA256: 940640b5ab0535405937dada3d0e88c55c83feb5cd4d4a8e6737b7ec2d022884
  • Pointer size: 131 Bytes
  • Size of remote file: 440 kB

Git LFS Details

  • SHA256: 569a63fcde0c992ee6a92165b887c0e10f2254d72c9ac97ea1717d613a6a87ac
  • Pointer size: 131 Bytes
  • Size of remote file: 442 kB
assets/ab_macro_20260921.jpg CHANGED

Git LFS Details

  • SHA256: b5b913b88b4b2d7ec8f37a7c08451a2a53ba6e965817eb7b49f87c46d3e18a0d
  • Pointer size: 131 Bytes
  • Size of remote file: 452 kB

Git LFS Details

  • SHA256: baa2a5ab9b9ec4117a8805bff79020488c1b8bdd4203c50dbdbe4cda5a7f6591
  • Pointer size: 131 Bytes
  • Size of remote file: 454 kB
assets/ab_night_market_20260920.jpg CHANGED

Git LFS Details

  • SHA256: 368181c2642d730adcc1b121041ee88ea5e25f23efe33d593982408ea50b97b5
  • Pointer size: 131 Bytes
  • Size of remote file: 709 kB

Git LFS Details

  • SHA256: a1f91a0984720cd3cb805164054da0d9d29445ae2f9e64d12e71843eee9737e1
  • Pointer size: 131 Bytes
  • Size of remote file: 711 kB
assets/ab_night_market_20260921.jpg CHANGED

Git LFS Details

  • SHA256: 5266ca07869dd84f6437a3f65ed3417c98404c5938f1341aeeb2d9766274d116
  • Pointer size: 131 Bytes
  • Size of remote file: 759 kB

Git LFS Details

  • SHA256: 68d5273effcb9481ea99df6c770b3d3ca6a55e90fc3aba2816389bf2a579239d
  • Pointer size: 131 Bytes
  • Size of remote file: 761 kB
assets/ab_portrait_20260920.jpg CHANGED

Git LFS Details

  • SHA256: ccfe23331acbd05c5fcf8b0db46a2d1eeb1872e09a62b5800bac0d983f8f005f
  • Pointer size: 131 Bytes
  • Size of remote file: 637 kB

Git LFS Details

  • SHA256: 2674cdd21e1a3a90882335b13328edf0382fba99537d824cf637f2c6a0682037
  • Pointer size: 131 Bytes
  • Size of remote file: 639 kB
assets/ab_portrait_20260921.jpg CHANGED

Git LFS Details

  • SHA256: 6c4804dac52ce29c2e31a84e00fcc5f7f2e11abcae595f1a7adef3d37a876bdc
  • Pointer size: 131 Bytes
  • Size of remote file: 628 kB

Git LFS Details

  • SHA256: 92dd4e5cb7c4393032eee6b0787bfaf79ef475822a1dbfd8699aee913b807fda
  • Pointer size: 131 Bytes
  • Size of remote file: 630 kB
assets/ab_product_20260920.jpg CHANGED

Git LFS Details

  • SHA256: ab620447fc0e6b17e1bb47df61db047bd4ee15a9f7014c834af9abf595af7847
  • Pointer size: 131 Bytes
  • Size of remote file: 402 kB

Git LFS Details

  • SHA256: dabc12a4a28d27b15c3206d96c099a5a2345c21929ace89e9ccef7b87230720f
  • Pointer size: 131 Bytes
  • Size of remote file: 403 kB
assets/ab_product_20260921.jpg CHANGED

Git LFS Details

  • SHA256: 402220da9ae5f5d268fc9caa12f0decbd2d48164ad5453d031ce1b5dbc04b318
  • Pointer size: 131 Bytes
  • Size of remote file: 360 kB

Git LFS Details

  • SHA256: 92b6e20ea4b4d14249097ae2bf0a67c2485cf3452dc35020fc4817191928f5d3
  • Pointer size: 131 Bytes
  • Size of remote file: 362 kB
assets/ab_spatial_20260920.jpg CHANGED

Git LFS Details

  • SHA256: 7498b7a7a3e766c04f27d5d28302f69c10f5c75bc7e7b562172bcd169efd03e3
  • Pointer size: 131 Bytes
  • Size of remote file: 254 kB

Git LFS Details

  • SHA256: 7bdf59afd58cb6481535e5ff3d540794b31f77dcdc47799fdb3d33a988a8bfb4
  • Pointer size: 131 Bytes
  • Size of remote file: 255 kB
assets/ab_spatial_20260921.jpg CHANGED

Git LFS Details

  • SHA256: b3faee6e319a5c675c91167ef0704b63580e9dc8d0340550afeb0c72759b6792
  • Pointer size: 131 Bytes
  • Size of remote file: 253 kB

Git LFS Details

  • SHA256: 55cbd086fd0336f23a3b826eb9f92f1a19ccaa04ebcd818eb6d583aa4c51bcc8
  • Pointer size: 131 Bytes
  • Size of remote file: 254 kB
assets/ab_typography_20260920.jpg CHANGED

Git LFS Details

  • SHA256: adeb0e4ed51ce0ff6a53793e3c301f4d29713f23c30ba21c4c0f4a2b849bb0d3
  • Pointer size: 131 Bytes
  • Size of remote file: 789 kB

Git LFS Details

  • SHA256: ab25af5746a219b1061d03a5ec63c125bee38c31526eaaee314bb444c01defef
  • Pointer size: 131 Bytes
  • Size of remote file: 790 kB
assets/ab_typography_20260921.jpg CHANGED

Git LFS Details

  • SHA256: b1be26c4fa502e0a1fba47227573b932b4fe38d5aa8c61234392b40c94488949
  • Pointer size: 131 Bytes
  • Size of remote file: 811 kB

Git LFS Details

  • SHA256: 99dbca8e65496f84192c04434ebef768a955a460855cfbdbbf06ff23d00d2814
  • Pointer size: 131 Bytes
  • Size of remote file: 812 kB
assets/editing.jpg CHANGED

Git LFS Details

  • SHA256: f05044570f14a5c4b75ef5803040b6cf5e8104640da248b64ffecf64c730b2f9
  • Pointer size: 131 Bytes
  • Size of remote file: 376 kB

Git LFS Details

  • SHA256: 70ac27dd7177e821ea2341d79f09065953f7b9dafcd2ff60f6f89bc32377c61f
  • Pointer size: 131 Bytes
  • Size of remote file: 377 kB
diffusion_models/qwen_image_2.1_nvfp4.manifest.json ADDED
@@ -0,0 +1,2596 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "source": "qwen_image_2.1_bf16.safetensors",
3
+ "source_sha256": "89f4158d066cc33906a199fca85634f766892dd78f49b6698dabf187ac86c4bc",
4
+ "output": "qwen_image_2.1_nvfp4.safetensors",
5
+ "output_sha256": "4fdbaec94c19f7b8888f1e1ddbae05354c014492f9b2547b1cd8dca0ce923526",
6
+ "output_bytes": 4197574664,
7
+ "nvfp4_matrices": 192,
8
+ "elapsed_seconds": 22.873350399997435,
9
+ "tensors": [
10
+ {
11
+ "name": "img_in.weight",
12
+ "shape": [
13
+ 4096,
14
+ 64
15
+ ],
16
+ "source_dtype": "torch.bfloat16",
17
+ "storage": "torch.bfloat16",
18
+ "sha256": "34310ebf0edb66ef00944eb6fd46e7ce70e7df7928d03283df0a28dd8fab2c68"
19
+ },
20
+ {
21
+ "name": "modulation.1.weight",
22
+ "shape": [
23
+ 16384,
24
+ 4096
25
+ ],
26
+ "source_dtype": "torch.bfloat16",
27
+ "storage": "torch.bfloat16",
28
+ "sha256": "4360117d34f9659f36d154af517f54dc7f521ea914a1894a837d660eb8048e0f"
29
+ },
30
+ {
31
+ "name": "norm_out.linear.weight",
32
+ "shape": [
33
+ 4096,
34
+ 4096
35
+ ],
36
+ "source_dtype": "torch.bfloat16",
37
+ "storage": "torch.bfloat16",
38
+ "sha256": "e9bcf4457d9deb7a0fd874fdfa7c950efcf36f7442b4e34878e3933524514ad3"
39
+ },
40
+ {
41
+ "name": "proj_out.weight",
42
+ "shape": [
43
+ 64,
44
+ 4096
45
+ ],
46
+ "source_dtype": "torch.bfloat16",
47
+ "storage": "torch.bfloat16",
48
+ "sha256": "36cc3c201fbe549b207588243e4ed9995582fa506db8313b3c6b30f0e26d7e9d"
49
+ },
50
+ {
51
+ "name": "time_text_embed.timestep_embedder.linear_1.weight",
52
+ "shape": [
53
+ 4096,
54
+ 256
55
+ ],
56
+ "source_dtype": "torch.bfloat16",
57
+ "storage": "torch.bfloat16",
58
+ "sha256": "c15dc973e402662d4dc7d2b8a5d256080f10d39e5aeec563723c327d22a3a712"
59
+ },
60
+ {
61
+ "name": "time_text_embed.timestep_embedder.linear_2.weight",
62
+ "shape": [
63
+ 4096,
64
+ 4096
65
+ ],
66
+ "source_dtype": "torch.bfloat16",
67
+ "storage": "torch.bfloat16",
68
+ "sha256": "c96b67bbcb7c19d2a31a517959c3090adfa420bc16735810abf0367f20a71615"
69
+ },
70
+ {
71
+ "name": "transformer_blocks.0.attn.norm_k.weight",
72
+ "shape": [
73
+ 128
74
+ ],
75
+ "source_dtype": "torch.bfloat16",
76
+ "storage": "torch.bfloat16",
77
+ "sha256": "572f201a7f1f0eb25c33d65be6a3afc6e237af9753c94cb0c924c54622d33277"
78
+ },
79
+ {
80
+ "name": "transformer_blocks.0.attn.norm_q.weight",
81
+ "shape": [
82
+ 128
83
+ ],
84
+ "source_dtype": "torch.bfloat16",
85
+ "storage": "torch.bfloat16",
86
+ "sha256": "3faf374413f8b2585700421c2f9bb39c1e846227f105e3ff10e1a8220ffde022"
87
+ },
88
+ {
89
+ "name": "transformer_blocks.0.attn.to_k.weight",
90
+ "shape": [
91
+ 4096,
92
+ 4096
93
+ ],
94
+ "source_dtype": "torch.bfloat16",
95
+ "storage": "nvfp4",
96
+ "relative_rmse": 0.09442301839590073
97
+ },
98
+ {
99
+ "name": "transformer_blocks.0.attn.to_out.0.weight",
100
+ "shape": [
101
+ 4096,
102
+ 4096
103
+ ],
104
+ "source_dtype": "torch.bfloat16",
105
+ "storage": "nvfp4",
106
+ "relative_rmse": 0.09725861996412277
107
+ },
108
+ {
109
+ "name": "transformer_blocks.0.attn.to_q.weight",
110
+ "shape": [
111
+ 4096,
112
+ 4096
113
+ ],
114
+ "source_dtype": "torch.bfloat16",
115
+ "storage": "nvfp4",
116
+ "relative_rmse": 0.0944296196103096
117
+ },
118
+ {
119
+ "name": "transformer_blocks.0.attn.to_v.weight",
120
+ "shape": [
121
+ 4096,
122
+ 4096
123
+ ],
124
+ "source_dtype": "torch.bfloat16",
125
+ "storage": "nvfp4",
126
+ "relative_rmse": 0.09454155713319778
127
+ },
128
+ {
129
+ "name": "transformer_blocks.0.img_mlp.gate_up.weight",
130
+ "shape": [
131
+ 24576,
132
+ 4096
133
+ ],
134
+ "source_dtype": "torch.bfloat16",
135
+ "storage": "nvfp4",
136
+ "relative_rmse": 0.0946829617023468
137
+ },
138
+ {
139
+ "name": "transformer_blocks.0.img_mlp.out.weight",
140
+ "shape": [
141
+ 4096,
142
+ 12288
143
+ ],
144
+ "source_dtype": "torch.bfloat16",
145
+ "storage": "nvfp4",
146
+ "relative_rmse": 0.09461568295955658
147
+ },
148
+ {
149
+ "name": "transformer_blocks.1.attn.norm_k.weight",
150
+ "shape": [
151
+ 128
152
+ ],
153
+ "source_dtype": "torch.bfloat16",
154
+ "storage": "torch.bfloat16",
155
+ "sha256": "92458b847b05a1da9cd8087668002bf9b31e9bdf4319b41082ac12a55eb556e0"
156
+ },
157
+ {
158
+ "name": "transformer_blocks.1.attn.norm_q.weight",
159
+ "shape": [
160
+ 128
161
+ ],
162
+ "source_dtype": "torch.bfloat16",
163
+ "storage": "torch.bfloat16",
164
+ "sha256": "d5e3edf26d8478f4bada9caae2b2b6b86b22338c2c7a6c6abfad38f337f8f246"
165
+ },
166
+ {
167
+ "name": "transformer_blocks.1.attn.to_k.weight",
168
+ "shape": [
169
+ 4096,
170
+ 4096
171
+ ],
172
+ "source_dtype": "torch.bfloat16",
173
+ "storage": "nvfp4",
174
+ "relative_rmse": 0.09432397782802582
175
+ },
176
+ {
177
+ "name": "transformer_blocks.1.attn.to_out.0.weight",
178
+ "shape": [
179
+ 4096,
180
+ 4096
181
+ ],
182
+ "source_dtype": "torch.bfloat16",
183
+ "storage": "nvfp4",
184
+ "relative_rmse": 0.09523904323577881
185
+ },
186
+ {
187
+ "name": "transformer_blocks.1.attn.to_q.weight",
188
+ "shape": [
189
+ 4096,
190
+ 4096
191
+ ],
192
+ "source_dtype": "torch.bfloat16",
193
+ "storage": "nvfp4",
194
+ "relative_rmse": 0.09456050395965576
195
+ },
196
+ {
197
+ "name": "transformer_blocks.1.attn.to_v.weight",
198
+ "shape": [
199
+ 4096,
200
+ 4096
201
+ ],
202
+ "source_dtype": "torch.bfloat16",
203
+ "storage": "nvfp4",
204
+ "relative_rmse": 0.09415525943040848
205
+ },
206
+ {
207
+ "name": "transformer_blocks.1.img_mlp.gate_up.weight",
208
+ "shape": [
209
+ 24576,
210
+ 4096
211
+ ],
212
+ "source_dtype": "torch.bfloat16",
213
+ "storage": "nvfp4",
214
+ "relative_rmse": 0.0945473164319992
215
+ },
216
+ {
217
+ "name": "transformer_blocks.1.img_mlp.out.weight",
218
+ "shape": [
219
+ 4096,
220
+ 12288
221
+ ],
222
+ "source_dtype": "torch.bfloat16",
223
+ "storage": "nvfp4",
224
+ "relative_rmse": 0.0944310873746872
225
+ },
226
+ {
227
+ "name": "transformer_blocks.10.attn.norm_k.weight",
228
+ "shape": [
229
+ 128
230
+ ],
231
+ "source_dtype": "torch.bfloat16",
232
+ "storage": "torch.bfloat16",
233
+ "sha256": "89dd760dbc3d89539c68b4c79e1b1c13feb28a4b334cc518717d6f25a5330c03"
234
+ },
235
+ {
236
+ "name": "transformer_blocks.10.attn.norm_q.weight",
237
+ "shape": [
238
+ 128
239
+ ],
240
+ "source_dtype": "torch.bfloat16",
241
+ "storage": "torch.bfloat16",
242
+ "sha256": "f3ee1eaebb22009e5f372b32269f48683f4c79759c0a941be40b9f7778a39b3d"
243
+ },
244
+ {
245
+ "name": "transformer_blocks.10.attn.to_k.weight",
246
+ "shape": [
247
+ 4096,
248
+ 4096
249
+ ],
250
+ "source_dtype": "torch.bfloat16",
251
+ "storage": "nvfp4",
252
+ "relative_rmse": 0.09438031911849976
253
+ },
254
+ {
255
+ "name": "transformer_blocks.10.attn.to_out.0.weight",
256
+ "shape": [
257
+ 4096,
258
+ 4096
259
+ ],
260
+ "source_dtype": "torch.bfloat16",
261
+ "storage": "nvfp4",
262
+ "relative_rmse": 0.09511259198188782
263
+ },
264
+ {
265
+ "name": "transformer_blocks.10.attn.to_q.weight",
266
+ "shape": [
267
+ 4096,
268
+ 4096
269
+ ],
270
+ "source_dtype": "torch.bfloat16",
271
+ "storage": "nvfp4",
272
+ "relative_rmse": 0.09445391595363617
273
+ },
274
+ {
275
+ "name": "transformer_blocks.10.attn.to_v.weight",
276
+ "shape": [
277
+ 4096,
278
+ 4096
279
+ ],
280
+ "source_dtype": "torch.bfloat16",
281
+ "storage": "nvfp4",
282
+ "relative_rmse": 0.09419218450784683
283
+ },
284
+ {
285
+ "name": "transformer_blocks.10.img_mlp.gate_up.weight",
286
+ "shape": [
287
+ 24576,
288
+ 4096
289
+ ],
290
+ "source_dtype": "torch.bfloat16",
291
+ "storage": "nvfp4",
292
+ "relative_rmse": 0.09455808252096176
293
+ },
294
+ {
295
+ "name": "transformer_blocks.10.img_mlp.out.weight",
296
+ "shape": [
297
+ 4096,
298
+ 12288
299
+ ],
300
+ "source_dtype": "torch.bfloat16",
301
+ "storage": "nvfp4",
302
+ "relative_rmse": 0.09482181072235107
303
+ },
304
+ {
305
+ "name": "transformer_blocks.11.attn.norm_k.weight",
306
+ "shape": [
307
+ 128
308
+ ],
309
+ "source_dtype": "torch.bfloat16",
310
+ "storage": "torch.bfloat16",
311
+ "sha256": "c738b8925971f5f3cce78ef067876764565a358d82605a9d59479712d421cb2d"
312
+ },
313
+ {
314
+ "name": "transformer_blocks.11.attn.norm_q.weight",
315
+ "shape": [
316
+ 128
317
+ ],
318
+ "source_dtype": "torch.bfloat16",
319
+ "storage": "torch.bfloat16",
320
+ "sha256": "59cdf03f52fc4dced1699703c6303847638b1c0b5ec4996abb5e49d323c5b400"
321
+ },
322
+ {
323
+ "name": "transformer_blocks.11.attn.to_k.weight",
324
+ "shape": [
325
+ 4096,
326
+ 4096
327
+ ],
328
+ "source_dtype": "torch.bfloat16",
329
+ "storage": "nvfp4",
330
+ "relative_rmse": 0.09461991488933563
331
+ },
332
+ {
333
+ "name": "transformer_blocks.11.attn.to_out.0.weight",
334
+ "shape": [
335
+ 4096,
336
+ 4096
337
+ ],
338
+ "source_dtype": "torch.bfloat16",
339
+ "storage": "nvfp4",
340
+ "relative_rmse": 0.09524192661046982
341
+ },
342
+ {
343
+ "name": "transformer_blocks.11.attn.to_q.weight",
344
+ "shape": [
345
+ 4096,
346
+ 4096
347
+ ],
348
+ "source_dtype": "torch.bfloat16",
349
+ "storage": "nvfp4",
350
+ "relative_rmse": 0.09446550160646439
351
+ },
352
+ {
353
+ "name": "transformer_blocks.11.attn.to_v.weight",
354
+ "shape": [
355
+ 4096,
356
+ 4096
357
+ ],
358
+ "source_dtype": "torch.bfloat16",
359
+ "storage": "nvfp4",
360
+ "relative_rmse": 0.09410617500543594
361
+ },
362
+ {
363
+ "name": "transformer_blocks.11.img_mlp.gate_up.weight",
364
+ "shape": [
365
+ 24576,
366
+ 4096
367
+ ],
368
+ "source_dtype": "torch.bfloat16",
369
+ "storage": "nvfp4",
370
+ "relative_rmse": 0.09447803348302841
371
+ },
372
+ {
373
+ "name": "transformer_blocks.11.img_mlp.out.weight",
374
+ "shape": [
375
+ 4096,
376
+ 12288
377
+ ],
378
+ "source_dtype": "torch.bfloat16",
379
+ "storage": "nvfp4",
380
+ "relative_rmse": 0.09461723268032074
381
+ },
382
+ {
383
+ "name": "transformer_blocks.12.attn.norm_k.weight",
384
+ "shape": [
385
+ 128
386
+ ],
387
+ "source_dtype": "torch.bfloat16",
388
+ "storage": "torch.bfloat16",
389
+ "sha256": "515145e5eb450bd0615815dc8728ab1e4922bc7f038853eaf61d8e4b110876dd"
390
+ },
391
+ {
392
+ "name": "transformer_blocks.12.attn.norm_q.weight",
393
+ "shape": [
394
+ 128
395
+ ],
396
+ "source_dtype": "torch.bfloat16",
397
+ "storage": "torch.bfloat16",
398
+ "sha256": "ab59bae2c2af0dafa123601c0411f2d24307f533691ce36ccf8bc694da54f72a"
399
+ },
400
+ {
401
+ "name": "transformer_blocks.12.attn.to_k.weight",
402
+ "shape": [
403
+ 4096,
404
+ 4096
405
+ ],
406
+ "source_dtype": "torch.bfloat16",
407
+ "storage": "nvfp4",
408
+ "relative_rmse": 0.09440026432275772
409
+ },
410
+ {
411
+ "name": "transformer_blocks.12.attn.to_out.0.weight",
412
+ "shape": [
413
+ 4096,
414
+ 4096
415
+ ],
416
+ "source_dtype": "torch.bfloat16",
417
+ "storage": "nvfp4",
418
+ "relative_rmse": 0.09543906152248383
419
+ },
420
+ {
421
+ "name": "transformer_blocks.12.attn.to_q.weight",
422
+ "shape": [
423
+ 4096,
424
+ 4096
425
+ ],
426
+ "source_dtype": "torch.bfloat16",
427
+ "storage": "nvfp4",
428
+ "relative_rmse": 0.09447003901004791
429
+ },
430
+ {
431
+ "name": "transformer_blocks.12.attn.to_v.weight",
432
+ "shape": [
433
+ 4096,
434
+ 4096
435
+ ],
436
+ "source_dtype": "torch.bfloat16",
437
+ "storage": "nvfp4",
438
+ "relative_rmse": 0.09415226429700851
439
+ },
440
+ {
441
+ "name": "transformer_blocks.12.img_mlp.gate_up.weight",
442
+ "shape": [
443
+ 24576,
444
+ 4096
445
+ ],
446
+ "source_dtype": "torch.bfloat16",
447
+ "storage": "nvfp4",
448
+ "relative_rmse": 0.09448692947626114
449
+ },
450
+ {
451
+ "name": "transformer_blocks.12.img_mlp.out.weight",
452
+ "shape": [
453
+ 4096,
454
+ 12288
455
+ ],
456
+ "source_dtype": "torch.bfloat16",
457
+ "storage": "nvfp4",
458
+ "relative_rmse": 0.09464255720376968
459
+ },
460
+ {
461
+ "name": "transformer_blocks.13.attn.norm_k.weight",
462
+ "shape": [
463
+ 128
464
+ ],
465
+ "source_dtype": "torch.bfloat16",
466
+ "storage": "torch.bfloat16",
467
+ "sha256": "bf992f464eba69ccaacc832235583c4a617724483c15c0b808b5bd29d39037a0"
468
+ },
469
+ {
470
+ "name": "transformer_blocks.13.attn.norm_q.weight",
471
+ "shape": [
472
+ 128
473
+ ],
474
+ "source_dtype": "torch.bfloat16",
475
+ "storage": "torch.bfloat16",
476
+ "sha256": "d69e4d8abe599ea9730fb81f74d527a7ee7c1eced6dd530f33eaa178d1fa3ec7"
477
+ },
478
+ {
479
+ "name": "transformer_blocks.13.attn.to_k.weight",
480
+ "shape": [
481
+ 4096,
482
+ 4096
483
+ ],
484
+ "source_dtype": "torch.bfloat16",
485
+ "storage": "nvfp4",
486
+ "relative_rmse": 0.09427323937416077
487
+ },
488
+ {
489
+ "name": "transformer_blocks.13.attn.to_out.0.weight",
490
+ "shape": [
491
+ 4096,
492
+ 4096
493
+ ],
494
+ "source_dtype": "torch.bfloat16",
495
+ "storage": "nvfp4",
496
+ "relative_rmse": 0.09524256736040115
497
+ },
498
+ {
499
+ "name": "transformer_blocks.13.attn.to_q.weight",
500
+ "shape": [
501
+ 4096,
502
+ 4096
503
+ ],
504
+ "source_dtype": "torch.bfloat16",
505
+ "storage": "nvfp4",
506
+ "relative_rmse": 0.09445545077323914
507
+ },
508
+ {
509
+ "name": "transformer_blocks.13.attn.to_v.weight",
510
+ "shape": [
511
+ 4096,
512
+ 4096
513
+ ],
514
+ "source_dtype": "torch.bfloat16",
515
+ "storage": "nvfp4",
516
+ "relative_rmse": 0.09404991567134857
517
+ },
518
+ {
519
+ "name": "transformer_blocks.13.img_mlp.gate_up.weight",
520
+ "shape": [
521
+ 24576,
522
+ 4096
523
+ ],
524
+ "source_dtype": "torch.bfloat16",
525
+ "storage": "nvfp4",
526
+ "relative_rmse": 0.09441465139389038
527
+ },
528
+ {
529
+ "name": "transformer_blocks.13.img_mlp.out.weight",
530
+ "shape": [
531
+ 4096,
532
+ 12288
533
+ ],
534
+ "source_dtype": "torch.bfloat16",
535
+ "storage": "nvfp4",
536
+ "relative_rmse": 0.09465336799621582
537
+ },
538
+ {
539
+ "name": "transformer_blocks.14.attn.norm_k.weight",
540
+ "shape": [
541
+ 128
542
+ ],
543
+ "source_dtype": "torch.bfloat16",
544
+ "storage": "torch.bfloat16",
545
+ "sha256": "dddc04478d062739db7d47f38fc403054d7ebb6689a41b47eb357b6aefd2280f"
546
+ },
547
+ {
548
+ "name": "transformer_blocks.14.attn.norm_q.weight",
549
+ "shape": [
550
+ 128
551
+ ],
552
+ "source_dtype": "torch.bfloat16",
553
+ "storage": "torch.bfloat16",
554
+ "sha256": "5957f7c5eb010c214df3cf6214d87c916b331926a726c43d7bc312b211a3a577"
555
+ },
556
+ {
557
+ "name": "transformer_blocks.14.attn.to_k.weight",
558
+ "shape": [
559
+ 4096,
560
+ 4096
561
+ ],
562
+ "source_dtype": "torch.bfloat16",
563
+ "storage": "nvfp4",
564
+ "relative_rmse": 0.09433333575725555
565
+ },
566
+ {
567
+ "name": "transformer_blocks.14.attn.to_out.0.weight",
568
+ "shape": [
569
+ 4096,
570
+ 4096
571
+ ],
572
+ "source_dtype": "torch.bfloat16",
573
+ "storage": "nvfp4",
574
+ "relative_rmse": 0.09543435275554657
575
+ },
576
+ {
577
+ "name": "transformer_blocks.14.attn.to_q.weight",
578
+ "shape": [
579
+ 4096,
580
+ 4096
581
+ ],
582
+ "source_dtype": "torch.bfloat16",
583
+ "storage": "nvfp4",
584
+ "relative_rmse": 0.0944102481007576
585
+ },
586
+ {
587
+ "name": "transformer_blocks.14.attn.to_v.weight",
588
+ "shape": [
589
+ 4096,
590
+ 4096
591
+ ],
592
+ "source_dtype": "torch.bfloat16",
593
+ "storage": "nvfp4",
594
+ "relative_rmse": 0.09410250931978226
595
+ },
596
+ {
597
+ "name": "transformer_blocks.14.img_mlp.gate_up.weight",
598
+ "shape": [
599
+ 24576,
600
+ 4096
601
+ ],
602
+ "source_dtype": "torch.bfloat16",
603
+ "storage": "nvfp4",
604
+ "relative_rmse": 0.0944635346531868
605
+ },
606
+ {
607
+ "name": "transformer_blocks.14.img_mlp.out.weight",
608
+ "shape": [
609
+ 4096,
610
+ 12288
611
+ ],
612
+ "source_dtype": "torch.bfloat16",
613
+ "storage": "nvfp4",
614
+ "relative_rmse": 0.09459885954856873
615
+ },
616
+ {
617
+ "name": "transformer_blocks.15.attn.norm_k.weight",
618
+ "shape": [
619
+ 128
620
+ ],
621
+ "source_dtype": "torch.bfloat16",
622
+ "storage": "torch.bfloat16",
623
+ "sha256": "65470fddf382b681da0afef7172ec8b30499e36935c921f058498e7a3f827dc7"
624
+ },
625
+ {
626
+ "name": "transformer_blocks.15.attn.norm_q.weight",
627
+ "shape": [
628
+ 128
629
+ ],
630
+ "source_dtype": "torch.bfloat16",
631
+ "storage": "torch.bfloat16",
632
+ "sha256": "3f7664105a6c56532b60a8e1893b907c2f5473f735e290ae00ee33594756a7c1"
633
+ },
634
+ {
635
+ "name": "transformer_blocks.15.attn.to_k.weight",
636
+ "shape": [
637
+ 4096,
638
+ 4096
639
+ ],
640
+ "source_dtype": "torch.bfloat16",
641
+ "storage": "nvfp4",
642
+ "relative_rmse": 0.09426062554121017
643
+ },
644
+ {
645
+ "name": "transformer_blocks.15.attn.to_out.0.weight",
646
+ "shape": [
647
+ 4096,
648
+ 4096
649
+ ],
650
+ "source_dtype": "torch.bfloat16",
651
+ "storage": "nvfp4",
652
+ "relative_rmse": 0.09513749182224274
653
+ },
654
+ {
655
+ "name": "transformer_blocks.15.attn.to_q.weight",
656
+ "shape": [
657
+ 4096,
658
+ 4096
659
+ ],
660
+ "source_dtype": "torch.bfloat16",
661
+ "storage": "nvfp4",
662
+ "relative_rmse": 0.09444401413202286
663
+ },
664
+ {
665
+ "name": "transformer_blocks.15.attn.to_v.weight",
666
+ "shape": [
667
+ 4096,
668
+ 4096
669
+ ],
670
+ "source_dtype": "torch.bfloat16",
671
+ "storage": "nvfp4",
672
+ "relative_rmse": 0.09422986209392548
673
+ },
674
+ {
675
+ "name": "transformer_blocks.15.img_mlp.gate_up.weight",
676
+ "shape": [
677
+ 24576,
678
+ 4096
679
+ ],
680
+ "source_dtype": "torch.bfloat16",
681
+ "storage": "nvfp4",
682
+ "relative_rmse": 0.09437613189220428
683
+ },
684
+ {
685
+ "name": "transformer_blocks.15.img_mlp.out.weight",
686
+ "shape": [
687
+ 4096,
688
+ 12288
689
+ ],
690
+ "source_dtype": "torch.bfloat16",
691
+ "storage": "nvfp4",
692
+ "relative_rmse": 0.09464667737483978
693
+ },
694
+ {
695
+ "name": "transformer_blocks.16.attn.norm_k.weight",
696
+ "shape": [
697
+ 128
698
+ ],
699
+ "source_dtype": "torch.bfloat16",
700
+ "storage": "torch.bfloat16",
701
+ "sha256": "476dc2fe95fb3c82698af316f7ae5bf1d79e3dc3504a5a67a80375cd62de21ca"
702
+ },
703
+ {
704
+ "name": "transformer_blocks.16.attn.norm_q.weight",
705
+ "shape": [
706
+ 128
707
+ ],
708
+ "source_dtype": "torch.bfloat16",
709
+ "storage": "torch.bfloat16",
710
+ "sha256": "19e6ab3408978ac437e3a1caac230feda2cbf50587b72532b934221ca87f8a09"
711
+ },
712
+ {
713
+ "name": "transformer_blocks.16.attn.to_k.weight",
714
+ "shape": [
715
+ 4096,
716
+ 4096
717
+ ],
718
+ "source_dtype": "torch.bfloat16",
719
+ "storage": "nvfp4",
720
+ "relative_rmse": 0.0942886620759964
721
+ },
722
+ {
723
+ "name": "transformer_blocks.16.attn.to_out.0.weight",
724
+ "shape": [
725
+ 4096,
726
+ 4096
727
+ ],
728
+ "source_dtype": "torch.bfloat16",
729
+ "storage": "nvfp4",
730
+ "relative_rmse": 0.09513603895902634
731
+ },
732
+ {
733
+ "name": "transformer_blocks.16.attn.to_q.weight",
734
+ "shape": [
735
+ 4096,
736
+ 4096
737
+ ],
738
+ "source_dtype": "torch.bfloat16",
739
+ "storage": "nvfp4",
740
+ "relative_rmse": 0.09439306706190109
741
+ },
742
+ {
743
+ "name": "transformer_blocks.16.attn.to_v.weight",
744
+ "shape": [
745
+ 4096,
746
+ 4096
747
+ ],
748
+ "source_dtype": "torch.bfloat16",
749
+ "storage": "nvfp4",
750
+ "relative_rmse": 0.09432097524404526
751
+ },
752
+ {
753
+ "name": "transformer_blocks.16.img_mlp.gate_up.weight",
754
+ "shape": [
755
+ 24576,
756
+ 4096
757
+ ],
758
+ "source_dtype": "torch.bfloat16",
759
+ "storage": "nvfp4",
760
+ "relative_rmse": 0.09429603815078735
761
+ },
762
+ {
763
+ "name": "transformer_blocks.16.img_mlp.out.weight",
764
+ "shape": [
765
+ 4096,
766
+ 12288
767
+ ],
768
+ "source_dtype": "torch.bfloat16",
769
+ "storage": "nvfp4",
770
+ "relative_rmse": 0.09470091760158539
771
+ },
772
+ {
773
+ "name": "transformer_blocks.17.attn.norm_k.weight",
774
+ "shape": [
775
+ 128
776
+ ],
777
+ "source_dtype": "torch.bfloat16",
778
+ "storage": "torch.bfloat16",
779
+ "sha256": "eb990a79a552d7b4b4077e5433dbdf08416716e7bfe44dd329d0212de411784a"
780
+ },
781
+ {
782
+ "name": "transformer_blocks.17.attn.norm_q.weight",
783
+ "shape": [
784
+ 128
785
+ ],
786
+ "source_dtype": "torch.bfloat16",
787
+ "storage": "torch.bfloat16",
788
+ "sha256": "12d5339fca65163052cab415ffc38d8006527aa66d60e3e1e6eb3d2ecd9b57d1"
789
+ },
790
+ {
791
+ "name": "transformer_blocks.17.attn.to_k.weight",
792
+ "shape": [
793
+ 4096,
794
+ 4096
795
+ ],
796
+ "source_dtype": "torch.bfloat16",
797
+ "storage": "nvfp4",
798
+ "relative_rmse": 0.0939578264951706
799
+ },
800
+ {
801
+ "name": "transformer_blocks.17.attn.to_out.0.weight",
802
+ "shape": [
803
+ 4096,
804
+ 4096
805
+ ],
806
+ "source_dtype": "torch.bfloat16",
807
+ "storage": "nvfp4",
808
+ "relative_rmse": 0.09532124549150467
809
+ },
810
+ {
811
+ "name": "transformer_blocks.17.attn.to_q.weight",
812
+ "shape": [
813
+ 4096,
814
+ 4096
815
+ ],
816
+ "source_dtype": "torch.bfloat16",
817
+ "storage": "nvfp4",
818
+ "relative_rmse": 0.09449957311153412
819
+ },
820
+ {
821
+ "name": "transformer_blocks.17.attn.to_v.weight",
822
+ "shape": [
823
+ 4096,
824
+ 4096
825
+ ],
826
+ "source_dtype": "torch.bfloat16",
827
+ "storage": "nvfp4",
828
+ "relative_rmse": 0.09404770284891129
829
+ },
830
+ {
831
+ "name": "transformer_blocks.17.img_mlp.gate_up.weight",
832
+ "shape": [
833
+ 24576,
834
+ 4096
835
+ ],
836
+ "source_dtype": "torch.bfloat16",
837
+ "storage": "nvfp4",
838
+ "relative_rmse": 0.09432825446128845
839
+ },
840
+ {
841
+ "name": "transformer_blocks.17.img_mlp.out.weight",
842
+ "shape": [
843
+ 4096,
844
+ 12288
845
+ ],
846
+ "source_dtype": "torch.bfloat16",
847
+ "storage": "nvfp4",
848
+ "relative_rmse": 0.09457230567932129
849
+ },
850
+ {
851
+ "name": "transformer_blocks.18.attn.norm_k.weight",
852
+ "shape": [
853
+ 128
854
+ ],
855
+ "source_dtype": "torch.bfloat16",
856
+ "storage": "torch.bfloat16",
857
+ "sha256": "de98af927d0b67feaf772e235f8570f0e8af43562bff458173f71cb633e60cd9"
858
+ },
859
+ {
860
+ "name": "transformer_blocks.18.attn.norm_q.weight",
861
+ "shape": [
862
+ 128
863
+ ],
864
+ "source_dtype": "torch.bfloat16",
865
+ "storage": "torch.bfloat16",
866
+ "sha256": "d83c826d686ec9066b1f11bc1f082d9dee32e8cf7a3238d9af7b7acbe88ad4f5"
867
+ },
868
+ {
869
+ "name": "transformer_blocks.18.attn.to_k.weight",
870
+ "shape": [
871
+ 4096,
872
+ 4096
873
+ ],
874
+ "source_dtype": "torch.bfloat16",
875
+ "storage": "nvfp4",
876
+ "relative_rmse": 0.09411294013261795
877
+ },
878
+ {
879
+ "name": "transformer_blocks.18.attn.to_out.0.weight",
880
+ "shape": [
881
+ 4096,
882
+ 4096
883
+ ],
884
+ "source_dtype": "torch.bfloat16",
885
+ "storage": "nvfp4",
886
+ "relative_rmse": 0.09533113241195679
887
+ },
888
+ {
889
+ "name": "transformer_blocks.18.attn.to_q.weight",
890
+ "shape": [
891
+ 4096,
892
+ 4096
893
+ ],
894
+ "source_dtype": "torch.bfloat16",
895
+ "storage": "nvfp4",
896
+ "relative_rmse": 0.0944882333278656
897
+ },
898
+ {
899
+ "name": "transformer_blocks.18.attn.to_v.weight",
900
+ "shape": [
901
+ 4096,
902
+ 4096
903
+ ],
904
+ "source_dtype": "torch.bfloat16",
905
+ "storage": "nvfp4",
906
+ "relative_rmse": 0.09433288872241974
907
+ },
908
+ {
909
+ "name": "transformer_blocks.18.img_mlp.gate_up.weight",
910
+ "shape": [
911
+ 24576,
912
+ 4096
913
+ ],
914
+ "source_dtype": "torch.bfloat16",
915
+ "storage": "nvfp4",
916
+ "relative_rmse": 0.09435984492301941
917
+ },
918
+ {
919
+ "name": "transformer_blocks.18.img_mlp.out.weight",
920
+ "shape": [
921
+ 4096,
922
+ 12288
923
+ ],
924
+ "source_dtype": "torch.bfloat16",
925
+ "storage": "nvfp4",
926
+ "relative_rmse": 0.09458048641681671
927
+ },
928
+ {
929
+ "name": "transformer_blocks.19.attn.norm_k.weight",
930
+ "shape": [
931
+ 128
932
+ ],
933
+ "source_dtype": "torch.bfloat16",
934
+ "storage": "torch.bfloat16",
935
+ "sha256": "0ac76478bf22c4c63dc51dd0652d36c746c938a52875b4a0832c2f1610be40e9"
936
+ },
937
+ {
938
+ "name": "transformer_blocks.19.attn.norm_q.weight",
939
+ "shape": [
940
+ 128
941
+ ],
942
+ "source_dtype": "torch.bfloat16",
943
+ "storage": "torch.bfloat16",
944
+ "sha256": "b08bac0e42f6da3a7273dc53c0a92614b39ab445e7a35676fc8bb6d760e05744"
945
+ },
946
+ {
947
+ "name": "transformer_blocks.19.attn.to_k.weight",
948
+ "shape": [
949
+ 4096,
950
+ 4096
951
+ ],
952
+ "source_dtype": "torch.bfloat16",
953
+ "storage": "nvfp4",
954
+ "relative_rmse": 0.09398675709962845
955
+ },
956
+ {
957
+ "name": "transformer_blocks.19.attn.to_out.0.weight",
958
+ "shape": [
959
+ 4096,
960
+ 4096
961
+ ],
962
+ "source_dtype": "torch.bfloat16",
963
+ "storage": "nvfp4",
964
+ "relative_rmse": 0.09518875926733017
965
+ },
966
+ {
967
+ "name": "transformer_blocks.19.attn.to_q.weight",
968
+ "shape": [
969
+ 4096,
970
+ 4096
971
+ ],
972
+ "source_dtype": "torch.bfloat16",
973
+ "storage": "nvfp4",
974
+ "relative_rmse": 0.09432109445333481
975
+ },
976
+ {
977
+ "name": "transformer_blocks.19.attn.to_v.weight",
978
+ "shape": [
979
+ 4096,
980
+ 4096
981
+ ],
982
+ "source_dtype": "torch.bfloat16",
983
+ "storage": "nvfp4",
984
+ "relative_rmse": 0.09421292692422867
985
+ },
986
+ {
987
+ "name": "transformer_blocks.19.img_mlp.gate_up.weight",
988
+ "shape": [
989
+ 24576,
990
+ 4096
991
+ ],
992
+ "source_dtype": "torch.bfloat16",
993
+ "storage": "nvfp4",
994
+ "relative_rmse": 0.09432484209537506
995
+ },
996
+ {
997
+ "name": "transformer_blocks.19.img_mlp.out.weight",
998
+ "shape": [
999
+ 4096,
1000
+ 12288
1001
+ ],
1002
+ "source_dtype": "torch.bfloat16",
1003
+ "storage": "nvfp4",
1004
+ "relative_rmse": 0.0945013165473938
1005
+ },
1006
+ {
1007
+ "name": "transformer_blocks.2.attn.norm_k.weight",
1008
+ "shape": [
1009
+ 128
1010
+ ],
1011
+ "source_dtype": "torch.bfloat16",
1012
+ "storage": "torch.bfloat16",
1013
+ "sha256": "49896138bb58de842c31f43422b134f3100886d6449d1cc682d534d4a84643be"
1014
+ },
1015
+ {
1016
+ "name": "transformer_blocks.2.attn.norm_q.weight",
1017
+ "shape": [
1018
+ 128
1019
+ ],
1020
+ "source_dtype": "torch.bfloat16",
1021
+ "storage": "torch.bfloat16",
1022
+ "sha256": "ba0386bcdc5803f9500afbb627a10221af1dc7f8caa0ba174441da14be01777e"
1023
+ },
1024
+ {
1025
+ "name": "transformer_blocks.2.attn.to_k.weight",
1026
+ "shape": [
1027
+ 4096,
1028
+ 4096
1029
+ ],
1030
+ "source_dtype": "torch.bfloat16",
1031
+ "storage": "nvfp4",
1032
+ "relative_rmse": 0.09443528950214386
1033
+ },
1034
+ {
1035
+ "name": "transformer_blocks.2.attn.to_out.0.weight",
1036
+ "shape": [
1037
+ 4096,
1038
+ 4096
1039
+ ],
1040
+ "source_dtype": "torch.bfloat16",
1041
+ "storage": "nvfp4",
1042
+ "relative_rmse": 0.0953395888209343
1043
+ },
1044
+ {
1045
+ "name": "transformer_blocks.2.attn.to_q.weight",
1046
+ "shape": [
1047
+ 4096,
1048
+ 4096
1049
+ ],
1050
+ "source_dtype": "torch.bfloat16",
1051
+ "storage": "nvfp4",
1052
+ "relative_rmse": 0.09463594853878021
1053
+ },
1054
+ {
1055
+ "name": "transformer_blocks.2.attn.to_v.weight",
1056
+ "shape": [
1057
+ 4096,
1058
+ 4096
1059
+ ],
1060
+ "source_dtype": "torch.bfloat16",
1061
+ "storage": "nvfp4",
1062
+ "relative_rmse": 0.094529889523983
1063
+ },
1064
+ {
1065
+ "name": "transformer_blocks.2.img_mlp.gate_up.weight",
1066
+ "shape": [
1067
+ 24576,
1068
+ 4096
1069
+ ],
1070
+ "source_dtype": "torch.bfloat16",
1071
+ "storage": "nvfp4",
1072
+ "relative_rmse": 0.09461885690689087
1073
+ },
1074
+ {
1075
+ "name": "transformer_blocks.2.img_mlp.out.weight",
1076
+ "shape": [
1077
+ 4096,
1078
+ 12288
1079
+ ],
1080
+ "source_dtype": "torch.bfloat16",
1081
+ "storage": "nvfp4",
1082
+ "relative_rmse": 0.09440790116786957
1083
+ },
1084
+ {
1085
+ "name": "transformer_blocks.20.attn.norm_k.weight",
1086
+ "shape": [
1087
+ 128
1088
+ ],
1089
+ "source_dtype": "torch.bfloat16",
1090
+ "storage": "torch.bfloat16",
1091
+ "sha256": "f7d6a3d1f15d81d6ec47bd81d80f001865bb35c7c1abaa6ab799e4a1eb146e61"
1092
+ },
1093
+ {
1094
+ "name": "transformer_blocks.20.attn.norm_q.weight",
1095
+ "shape": [
1096
+ 128
1097
+ ],
1098
+ "source_dtype": "torch.bfloat16",
1099
+ "storage": "torch.bfloat16",
1100
+ "sha256": "425e716e33509e8d528985eccaba3850451546fda76a2368944170d5ba5b9062"
1101
+ },
1102
+ {
1103
+ "name": "transformer_blocks.20.attn.to_k.weight",
1104
+ "shape": [
1105
+ 4096,
1106
+ 4096
1107
+ ],
1108
+ "source_dtype": "torch.bfloat16",
1109
+ "storage": "nvfp4",
1110
+ "relative_rmse": 0.09418923407793045
1111
+ },
1112
+ {
1113
+ "name": "transformer_blocks.20.attn.to_out.0.weight",
1114
+ "shape": [
1115
+ 4096,
1116
+ 4096
1117
+ ],
1118
+ "source_dtype": "torch.bfloat16",
1119
+ "storage": "nvfp4",
1120
+ "relative_rmse": 0.0951521173119545
1121
+ },
1122
+ {
1123
+ "name": "transformer_blocks.20.attn.to_q.weight",
1124
+ "shape": [
1125
+ 4096,
1126
+ 4096
1127
+ ],
1128
+ "source_dtype": "torch.bfloat16",
1129
+ "storage": "nvfp4",
1130
+ "relative_rmse": 0.09440641105175018
1131
+ },
1132
+ {
1133
+ "name": "transformer_blocks.20.attn.to_v.weight",
1134
+ "shape": [
1135
+ 4096,
1136
+ 4096
1137
+ ],
1138
+ "source_dtype": "torch.bfloat16",
1139
+ "storage": "nvfp4",
1140
+ "relative_rmse": 0.094474196434021
1141
+ },
1142
+ {
1143
+ "name": "transformer_blocks.20.img_mlp.gate_up.weight",
1144
+ "shape": [
1145
+ 24576,
1146
+ 4096
1147
+ ],
1148
+ "source_dtype": "torch.bfloat16",
1149
+ "storage": "nvfp4",
1150
+ "relative_rmse": 0.09444796293973923
1151
+ },
1152
+ {
1153
+ "name": "transformer_blocks.20.img_mlp.out.weight",
1154
+ "shape": [
1155
+ 4096,
1156
+ 12288
1157
+ ],
1158
+ "source_dtype": "torch.bfloat16",
1159
+ "storage": "nvfp4",
1160
+ "relative_rmse": 0.09421083331108093
1161
+ },
1162
+ {
1163
+ "name": "transformer_blocks.21.attn.norm_k.weight",
1164
+ "shape": [
1165
+ 128
1166
+ ],
1167
+ "source_dtype": "torch.bfloat16",
1168
+ "storage": "torch.bfloat16",
1169
+ "sha256": "d78604ccfcd6f4f8f8834cbb8ee6d18d16c7e6b11431623377cd84876de83d0e"
1170
+ },
1171
+ {
1172
+ "name": "transformer_blocks.21.attn.norm_q.weight",
1173
+ "shape": [
1174
+ 128
1175
+ ],
1176
+ "source_dtype": "torch.bfloat16",
1177
+ "storage": "torch.bfloat16",
1178
+ "sha256": "bc53eb71732e3a2758d210a5a1ea161419ad556135a8125bf1d7b775d2103b64"
1179
+ },
1180
+ {
1181
+ "name": "transformer_blocks.21.attn.to_k.weight",
1182
+ "shape": [
1183
+ 4096,
1184
+ 4096
1185
+ ],
1186
+ "source_dtype": "torch.bfloat16",
1187
+ "storage": "nvfp4",
1188
+ "relative_rmse": 0.09433793276548386
1189
+ },
1190
+ {
1191
+ "name": "transformer_blocks.21.attn.to_out.0.weight",
1192
+ "shape": [
1193
+ 4096,
1194
+ 4096
1195
+ ],
1196
+ "source_dtype": "torch.bfloat16",
1197
+ "storage": "nvfp4",
1198
+ "relative_rmse": 0.09503240883350372
1199
+ },
1200
+ {
1201
+ "name": "transformer_blocks.21.attn.to_q.weight",
1202
+ "shape": [
1203
+ 4096,
1204
+ 4096
1205
+ ],
1206
+ "source_dtype": "torch.bfloat16",
1207
+ "storage": "nvfp4",
1208
+ "relative_rmse": 0.09457851201295853
1209
+ },
1210
+ {
1211
+ "name": "transformer_blocks.21.attn.to_v.weight",
1212
+ "shape": [
1213
+ 4096,
1214
+ 4096
1215
+ ],
1216
+ "source_dtype": "torch.bfloat16",
1217
+ "storage": "nvfp4",
1218
+ "relative_rmse": 0.09436649829149246
1219
+ },
1220
+ {
1221
+ "name": "transformer_blocks.21.img_mlp.gate_up.weight",
1222
+ "shape": [
1223
+ 24576,
1224
+ 4096
1225
+ ],
1226
+ "source_dtype": "torch.bfloat16",
1227
+ "storage": "nvfp4",
1228
+ "relative_rmse": 0.09445636719465256
1229
+ },
1230
+ {
1231
+ "name": "transformer_blocks.21.img_mlp.out.weight",
1232
+ "shape": [
1233
+ 4096,
1234
+ 12288
1235
+ ],
1236
+ "source_dtype": "torch.bfloat16",
1237
+ "storage": "nvfp4",
1238
+ "relative_rmse": 0.09428934752941132
1239
+ },
1240
+ {
1241
+ "name": "transformer_blocks.22.attn.norm_k.weight",
1242
+ "shape": [
1243
+ 128
1244
+ ],
1245
+ "source_dtype": "torch.bfloat16",
1246
+ "storage": "torch.bfloat16",
1247
+ "sha256": "ffb7961b6099752067ad7dbb3de075bfb14d65816448a84f99c95e886436345e"
1248
+ },
1249
+ {
1250
+ "name": "transformer_blocks.22.attn.norm_q.weight",
1251
+ "shape": [
1252
+ 128
1253
+ ],
1254
+ "source_dtype": "torch.bfloat16",
1255
+ "storage": "torch.bfloat16",
1256
+ "sha256": "5f4e26623063a53c95ee7f6ace3e072377b0f2edbefe3bd5af1f1af1b02b8ab9"
1257
+ },
1258
+ {
1259
+ "name": "transformer_blocks.22.attn.to_k.weight",
1260
+ "shape": [
1261
+ 4096,
1262
+ 4096
1263
+ ],
1264
+ "source_dtype": "torch.bfloat16",
1265
+ "storage": "nvfp4",
1266
+ "relative_rmse": 0.09428413957357407
1267
+ },
1268
+ {
1269
+ "name": "transformer_blocks.22.attn.to_out.0.weight",
1270
+ "shape": [
1271
+ 4096,
1272
+ 4096
1273
+ ],
1274
+ "source_dtype": "torch.bfloat16",
1275
+ "storage": "nvfp4",
1276
+ "relative_rmse": 0.09517910331487656
1277
+ },
1278
+ {
1279
+ "name": "transformer_blocks.22.attn.to_q.weight",
1280
+ "shape": [
1281
+ 4096,
1282
+ 4096
1283
+ ],
1284
+ "source_dtype": "torch.bfloat16",
1285
+ "storage": "nvfp4",
1286
+ "relative_rmse": 0.09460055083036423
1287
+ },
1288
+ {
1289
+ "name": "transformer_blocks.22.attn.to_v.weight",
1290
+ "shape": [
1291
+ 4096,
1292
+ 4096
1293
+ ],
1294
+ "source_dtype": "torch.bfloat16",
1295
+ "storage": "nvfp4",
1296
+ "relative_rmse": 0.09407515823841095
1297
+ },
1298
+ {
1299
+ "name": "transformer_blocks.22.img_mlp.gate_up.weight",
1300
+ "shape": [
1301
+ 24576,
1302
+ 4096
1303
+ ],
1304
+ "source_dtype": "torch.bfloat16",
1305
+ "storage": "nvfp4",
1306
+ "relative_rmse": 0.09422037750482559
1307
+ },
1308
+ {
1309
+ "name": "transformer_blocks.22.img_mlp.out.weight",
1310
+ "shape": [
1311
+ 4096,
1312
+ 12288
1313
+ ],
1314
+ "source_dtype": "torch.bfloat16",
1315
+ "storage": "nvfp4",
1316
+ "relative_rmse": 0.09452139586210251
1317
+ },
1318
+ {
1319
+ "name": "transformer_blocks.23.attn.norm_k.weight",
1320
+ "shape": [
1321
+ 128
1322
+ ],
1323
+ "source_dtype": "torch.bfloat16",
1324
+ "storage": "torch.bfloat16",
1325
+ "sha256": "b2af62fa9d8f64e07dc6e05c378ddc174579dc4f5c9978f45ecc8a1c3b1bc61f"
1326
+ },
1327
+ {
1328
+ "name": "transformer_blocks.23.attn.norm_q.weight",
1329
+ "shape": [
1330
+ 128
1331
+ ],
1332
+ "source_dtype": "torch.bfloat16",
1333
+ "storage": "torch.bfloat16",
1334
+ "sha256": "1d607ccc2b59a46d1bcf8890c712e2a96f2b8cbb961f3900ab259b3094d0e9d7"
1335
+ },
1336
+ {
1337
+ "name": "transformer_blocks.23.attn.to_k.weight",
1338
+ "shape": [
1339
+ 4096,
1340
+ 4096
1341
+ ],
1342
+ "source_dtype": "torch.bfloat16",
1343
+ "storage": "nvfp4",
1344
+ "relative_rmse": 0.09408468753099442
1345
+ },
1346
+ {
1347
+ "name": "transformer_blocks.23.attn.to_out.0.weight",
1348
+ "shape": [
1349
+ 4096,
1350
+ 4096
1351
+ ],
1352
+ "source_dtype": "torch.bfloat16",
1353
+ "storage": "nvfp4",
1354
+ "relative_rmse": 0.09505080431699753
1355
+ },
1356
+ {
1357
+ "name": "transformer_blocks.23.attn.to_q.weight",
1358
+ "shape": [
1359
+ 4096,
1360
+ 4096
1361
+ ],
1362
+ "source_dtype": "torch.bfloat16",
1363
+ "storage": "nvfp4",
1364
+ "relative_rmse": 0.09434711188077927
1365
+ },
1366
+ {
1367
+ "name": "transformer_blocks.23.attn.to_v.weight",
1368
+ "shape": [
1369
+ 4096,
1370
+ 4096
1371
+ ],
1372
+ "source_dtype": "torch.bfloat16",
1373
+ "storage": "nvfp4",
1374
+ "relative_rmse": 0.09392070770263672
1375
+ },
1376
+ {
1377
+ "name": "transformer_blocks.23.img_mlp.gate_up.weight",
1378
+ "shape": [
1379
+ 24576,
1380
+ 4096
1381
+ ],
1382
+ "source_dtype": "torch.bfloat16",
1383
+ "storage": "nvfp4",
1384
+ "relative_rmse": 0.09421507269144058
1385
+ },
1386
+ {
1387
+ "name": "transformer_blocks.23.img_mlp.out.weight",
1388
+ "shape": [
1389
+ 4096,
1390
+ 12288
1391
+ ],
1392
+ "source_dtype": "torch.bfloat16",
1393
+ "storage": "nvfp4",
1394
+ "relative_rmse": 0.09436953812837601
1395
+ },
1396
+ {
1397
+ "name": "transformer_blocks.24.attn.norm_k.weight",
1398
+ "shape": [
1399
+ 128
1400
+ ],
1401
+ "source_dtype": "torch.bfloat16",
1402
+ "storage": "torch.bfloat16",
1403
+ "sha256": "7cf55d07732b4460b4f4032d756677186dfe451fd580adc2e335794d3efc24e5"
1404
+ },
1405
+ {
1406
+ "name": "transformer_blocks.24.attn.norm_q.weight",
1407
+ "shape": [
1408
+ 128
1409
+ ],
1410
+ "source_dtype": "torch.bfloat16",
1411
+ "storage": "torch.bfloat16",
1412
+ "sha256": "5c0019afbf7531a43cab1c18b944ea42e5358338b22afdb7c3b63ccd96c0cc9a"
1413
+ },
1414
+ {
1415
+ "name": "transformer_blocks.24.attn.to_k.weight",
1416
+ "shape": [
1417
+ 4096,
1418
+ 4096
1419
+ ],
1420
+ "source_dtype": "torch.bfloat16",
1421
+ "storage": "nvfp4",
1422
+ "relative_rmse": 0.09408159554004669
1423
+ },
1424
+ {
1425
+ "name": "transformer_blocks.24.attn.to_out.0.weight",
1426
+ "shape": [
1427
+ 4096,
1428
+ 4096
1429
+ ],
1430
+ "source_dtype": "torch.bfloat16",
1431
+ "storage": "nvfp4",
1432
+ "relative_rmse": 0.09515871852636337
1433
+ },
1434
+ {
1435
+ "name": "transformer_blocks.24.attn.to_q.weight",
1436
+ "shape": [
1437
+ 4096,
1438
+ 4096
1439
+ ],
1440
+ "source_dtype": "torch.bfloat16",
1441
+ "storage": "nvfp4",
1442
+ "relative_rmse": 0.09425819665193558
1443
+ },
1444
+ {
1445
+ "name": "transformer_blocks.24.attn.to_v.weight",
1446
+ "shape": [
1447
+ 4096,
1448
+ 4096
1449
+ ],
1450
+ "source_dtype": "torch.bfloat16",
1451
+ "storage": "nvfp4",
1452
+ "relative_rmse": 0.09295009821653366
1453
+ },
1454
+ {
1455
+ "name": "transformer_blocks.24.img_mlp.gate_up.weight",
1456
+ "shape": [
1457
+ 24576,
1458
+ 4096
1459
+ ],
1460
+ "source_dtype": "torch.bfloat16",
1461
+ "storage": "nvfp4",
1462
+ "relative_rmse": 0.0942286029458046
1463
+ },
1464
+ {
1465
+ "name": "transformer_blocks.24.img_mlp.out.weight",
1466
+ "shape": [
1467
+ 4096,
1468
+ 12288
1469
+ ],
1470
+ "source_dtype": "torch.bfloat16",
1471
+ "storage": "nvfp4",
1472
+ "relative_rmse": 0.09437375515699387
1473
+ },
1474
+ {
1475
+ "name": "transformer_blocks.25.attn.norm_k.weight",
1476
+ "shape": [
1477
+ 128
1478
+ ],
1479
+ "source_dtype": "torch.bfloat16",
1480
+ "storage": "torch.bfloat16",
1481
+ "sha256": "9912d37ce907c913ebe90ce90f8b74d951401f3fc6d081d4268cf6aa0fd530fb"
1482
+ },
1483
+ {
1484
+ "name": "transformer_blocks.25.attn.norm_q.weight",
1485
+ "shape": [
1486
+ 128
1487
+ ],
1488
+ "source_dtype": "torch.bfloat16",
1489
+ "storage": "torch.bfloat16",
1490
+ "sha256": "41b0f2d2d46510be1bca296cc1981fd6bfbf1d897b9ad95cbfeb6efb23b46504"
1491
+ },
1492
+ {
1493
+ "name": "transformer_blocks.25.attn.to_k.weight",
1494
+ "shape": [
1495
+ 4096,
1496
+ 4096
1497
+ ],
1498
+ "source_dtype": "torch.bfloat16",
1499
+ "storage": "nvfp4",
1500
+ "relative_rmse": 0.09419506043195724
1501
+ },
1502
+ {
1503
+ "name": "transformer_blocks.25.attn.to_out.0.weight",
1504
+ "shape": [
1505
+ 4096,
1506
+ 4096
1507
+ ],
1508
+ "source_dtype": "torch.bfloat16",
1509
+ "storage": "nvfp4",
1510
+ "relative_rmse": 0.09516096860170364
1511
+ },
1512
+ {
1513
+ "name": "transformer_blocks.25.attn.to_q.weight",
1514
+ "shape": [
1515
+ 4096,
1516
+ 4096
1517
+ ],
1518
+ "source_dtype": "torch.bfloat16",
1519
+ "storage": "nvfp4",
1520
+ "relative_rmse": 0.09440808743238449
1521
+ },
1522
+ {
1523
+ "name": "transformer_blocks.25.attn.to_v.weight",
1524
+ "shape": [
1525
+ 4096,
1526
+ 4096
1527
+ ],
1528
+ "source_dtype": "torch.bfloat16",
1529
+ "storage": "nvfp4",
1530
+ "relative_rmse": 0.09340789914131165
1531
+ },
1532
+ {
1533
+ "name": "transformer_blocks.25.img_mlp.gate_up.weight",
1534
+ "shape": [
1535
+ 24576,
1536
+ 4096
1537
+ ],
1538
+ "source_dtype": "torch.bfloat16",
1539
+ "storage": "nvfp4",
1540
+ "relative_rmse": 0.09431573748588562
1541
+ },
1542
+ {
1543
+ "name": "transformer_blocks.25.img_mlp.out.weight",
1544
+ "shape": [
1545
+ 4096,
1546
+ 12288
1547
+ ],
1548
+ "source_dtype": "torch.bfloat16",
1549
+ "storage": "nvfp4",
1550
+ "relative_rmse": 0.09465906769037247
1551
+ },
1552
+ {
1553
+ "name": "transformer_blocks.26.attn.norm_k.weight",
1554
+ "shape": [
1555
+ 128
1556
+ ],
1557
+ "source_dtype": "torch.bfloat16",
1558
+ "storage": "torch.bfloat16",
1559
+ "sha256": "1770794279275ff0e355fc4c967f6fd15725898c41be4247ecf3208369e6db90"
1560
+ },
1561
+ {
1562
+ "name": "transformer_blocks.26.attn.norm_q.weight",
1563
+ "shape": [
1564
+ 128
1565
+ ],
1566
+ "source_dtype": "torch.bfloat16",
1567
+ "storage": "torch.bfloat16",
1568
+ "sha256": "7d75f4f6ba1e2ffc36f448fc91e122ce73ed7a3ddfd578b90f6af614b0b05931"
1569
+ },
1570
+ {
1571
+ "name": "transformer_blocks.26.attn.to_k.weight",
1572
+ "shape": [
1573
+ 4096,
1574
+ 4096
1575
+ ],
1576
+ "source_dtype": "torch.bfloat16",
1577
+ "storage": "nvfp4",
1578
+ "relative_rmse": 0.09402094036340714
1579
+ },
1580
+ {
1581
+ "name": "transformer_blocks.26.attn.to_out.0.weight",
1582
+ "shape": [
1583
+ 4096,
1584
+ 4096
1585
+ ],
1586
+ "source_dtype": "torch.bfloat16",
1587
+ "storage": "nvfp4",
1588
+ "relative_rmse": 0.09514297544956207
1589
+ },
1590
+ {
1591
+ "name": "transformer_blocks.26.attn.to_q.weight",
1592
+ "shape": [
1593
+ 4096,
1594
+ 4096
1595
+ ],
1596
+ "source_dtype": "torch.bfloat16",
1597
+ "storage": "nvfp4",
1598
+ "relative_rmse": 0.09432660043239594
1599
+ },
1600
+ {
1601
+ "name": "transformer_blocks.26.attn.to_v.weight",
1602
+ "shape": [
1603
+ 4096,
1604
+ 4096
1605
+ ],
1606
+ "source_dtype": "torch.bfloat16",
1607
+ "storage": "nvfp4",
1608
+ "relative_rmse": 0.09392236918210983
1609
+ },
1610
+ {
1611
+ "name": "transformer_blocks.26.img_mlp.gate_up.weight",
1612
+ "shape": [
1613
+ 24576,
1614
+ 4096
1615
+ ],
1616
+ "source_dtype": "torch.bfloat16",
1617
+ "storage": "nvfp4",
1618
+ "relative_rmse": 0.09436123818159103
1619
+ },
1620
+ {
1621
+ "name": "transformer_blocks.26.img_mlp.out.weight",
1622
+ "shape": [
1623
+ 4096,
1624
+ 12288
1625
+ ],
1626
+ "source_dtype": "torch.bfloat16",
1627
+ "storage": "nvfp4",
1628
+ "relative_rmse": 0.09454623609781265
1629
+ },
1630
+ {
1631
+ "name": "transformer_blocks.27.attn.norm_k.weight",
1632
+ "shape": [
1633
+ 128
1634
+ ],
1635
+ "source_dtype": "torch.bfloat16",
1636
+ "storage": "torch.bfloat16",
1637
+ "sha256": "e617551c3aea1ec17fbfd87d17d70439108df684584c801a33ae3b20134f2b93"
1638
+ },
1639
+ {
1640
+ "name": "transformer_blocks.27.attn.norm_q.weight",
1641
+ "shape": [
1642
+ 128
1643
+ ],
1644
+ "source_dtype": "torch.bfloat16",
1645
+ "storage": "torch.bfloat16",
1646
+ "sha256": "8b4cb1eb2bde149b81d1a09d3de880668ba8f6a62406546df9ee9b0a07c95e78"
1647
+ },
1648
+ {
1649
+ "name": "transformer_blocks.27.attn.to_k.weight",
1650
+ "shape": [
1651
+ 4096,
1652
+ 4096
1653
+ ],
1654
+ "source_dtype": "torch.bfloat16",
1655
+ "storage": "nvfp4",
1656
+ "relative_rmse": 0.093889020383358
1657
+ },
1658
+ {
1659
+ "name": "transformer_blocks.27.attn.to_out.0.weight",
1660
+ "shape": [
1661
+ 4096,
1662
+ 4096
1663
+ ],
1664
+ "source_dtype": "torch.bfloat16",
1665
+ "storage": "nvfp4",
1666
+ "relative_rmse": 0.09523267298936844
1667
+ },
1668
+ {
1669
+ "name": "transformer_blocks.27.attn.to_q.weight",
1670
+ "shape": [
1671
+ 4096,
1672
+ 4096
1673
+ ],
1674
+ "source_dtype": "torch.bfloat16",
1675
+ "storage": "nvfp4",
1676
+ "relative_rmse": 0.09420020878314972
1677
+ },
1678
+ {
1679
+ "name": "transformer_blocks.27.attn.to_v.weight",
1680
+ "shape": [
1681
+ 4096,
1682
+ 4096
1683
+ ],
1684
+ "source_dtype": "torch.bfloat16",
1685
+ "storage": "nvfp4",
1686
+ "relative_rmse": 0.09274192154407501
1687
+ },
1688
+ {
1689
+ "name": "transformer_blocks.27.img_mlp.gate_up.weight",
1690
+ "shape": [
1691
+ 24576,
1692
+ 4096
1693
+ ],
1694
+ "source_dtype": "torch.bfloat16",
1695
+ "storage": "nvfp4",
1696
+ "relative_rmse": 0.09421180933713913
1697
+ },
1698
+ {
1699
+ "name": "transformer_blocks.27.img_mlp.out.weight",
1700
+ "shape": [
1701
+ 4096,
1702
+ 12288
1703
+ ],
1704
+ "source_dtype": "torch.bfloat16",
1705
+ "storage": "nvfp4",
1706
+ "relative_rmse": 0.09465836733579636
1707
+ },
1708
+ {
1709
+ "name": "transformer_blocks.28.attn.norm_k.weight",
1710
+ "shape": [
1711
+ 128
1712
+ ],
1713
+ "source_dtype": "torch.bfloat16",
1714
+ "storage": "torch.bfloat16",
1715
+ "sha256": "e61f36f30099a0029417ef24a5ce24237e9e8c57e72199ce9380b31cad296537"
1716
+ },
1717
+ {
1718
+ "name": "transformer_blocks.28.attn.norm_q.weight",
1719
+ "shape": [
1720
+ 128
1721
+ ],
1722
+ "source_dtype": "torch.bfloat16",
1723
+ "storage": "torch.bfloat16",
1724
+ "sha256": "d2420a2ed03503525c0309d1a5758c05db515fdee1bc61e1413704e58426433c"
1725
+ },
1726
+ {
1727
+ "name": "transformer_blocks.28.attn.to_k.weight",
1728
+ "shape": [
1729
+ 4096,
1730
+ 4096
1731
+ ],
1732
+ "source_dtype": "torch.bfloat16",
1733
+ "storage": "nvfp4",
1734
+ "relative_rmse": 0.09418927133083344
1735
+ },
1736
+ {
1737
+ "name": "transformer_blocks.28.attn.to_out.0.weight",
1738
+ "shape": [
1739
+ 4096,
1740
+ 4096
1741
+ ],
1742
+ "source_dtype": "torch.bfloat16",
1743
+ "storage": "nvfp4",
1744
+ "relative_rmse": 0.09510087966918945
1745
+ },
1746
+ {
1747
+ "name": "transformer_blocks.28.attn.to_q.weight",
1748
+ "shape": [
1749
+ 4096,
1750
+ 4096
1751
+ ],
1752
+ "source_dtype": "torch.bfloat16",
1753
+ "storage": "nvfp4",
1754
+ "relative_rmse": 0.09439761936664581
1755
+ },
1756
+ {
1757
+ "name": "transformer_blocks.28.attn.to_v.weight",
1758
+ "shape": [
1759
+ 4096,
1760
+ 4096
1761
+ ],
1762
+ "source_dtype": "torch.bfloat16",
1763
+ "storage": "nvfp4",
1764
+ "relative_rmse": 0.09371969848871231
1765
+ },
1766
+ {
1767
+ "name": "transformer_blocks.28.img_mlp.gate_up.weight",
1768
+ "shape": [
1769
+ 24576,
1770
+ 4096
1771
+ ],
1772
+ "source_dtype": "torch.bfloat16",
1773
+ "storage": "nvfp4",
1774
+ "relative_rmse": 0.09451578557491302
1775
+ },
1776
+ {
1777
+ "name": "transformer_blocks.28.img_mlp.out.weight",
1778
+ "shape": [
1779
+ 4096,
1780
+ 12288
1781
+ ],
1782
+ "source_dtype": "torch.bfloat16",
1783
+ "storage": "nvfp4",
1784
+ "relative_rmse": 0.09452327340841293
1785
+ },
1786
+ {
1787
+ "name": "transformer_blocks.29.attn.norm_k.weight",
1788
+ "shape": [
1789
+ 128
1790
+ ],
1791
+ "source_dtype": "torch.bfloat16",
1792
+ "storage": "torch.bfloat16",
1793
+ "sha256": "0e952868b18d5a499239a5122a51c4e64f72896a8eac1e2782073dbafae627b8"
1794
+ },
1795
+ {
1796
+ "name": "transformer_blocks.29.attn.norm_q.weight",
1797
+ "shape": [
1798
+ 128
1799
+ ],
1800
+ "source_dtype": "torch.bfloat16",
1801
+ "storage": "torch.bfloat16",
1802
+ "sha256": "19cc1be26f7ae2d010a01ec1ee1e8a4d685578ef470750c77afc2d1d59d291df"
1803
+ },
1804
+ {
1805
+ "name": "transformer_blocks.29.attn.to_k.weight",
1806
+ "shape": [
1807
+ 4096,
1808
+ 4096
1809
+ ],
1810
+ "source_dtype": "torch.bfloat16",
1811
+ "storage": "nvfp4",
1812
+ "relative_rmse": 0.09389638900756836
1813
+ },
1814
+ {
1815
+ "name": "transformer_blocks.29.attn.to_out.0.weight",
1816
+ "shape": [
1817
+ 4096,
1818
+ 4096
1819
+ ],
1820
+ "source_dtype": "torch.bfloat16",
1821
+ "storage": "nvfp4",
1822
+ "relative_rmse": 0.09516238421201706
1823
+ },
1824
+ {
1825
+ "name": "transformer_blocks.29.attn.to_q.weight",
1826
+ "shape": [
1827
+ 4096,
1828
+ 4096
1829
+ ],
1830
+ "source_dtype": "torch.bfloat16",
1831
+ "storage": "nvfp4",
1832
+ "relative_rmse": 0.09444045275449753
1833
+ },
1834
+ {
1835
+ "name": "transformer_blocks.29.attn.to_v.weight",
1836
+ "shape": [
1837
+ 4096,
1838
+ 4096
1839
+ ],
1840
+ "source_dtype": "torch.bfloat16",
1841
+ "storage": "nvfp4",
1842
+ "relative_rmse": 0.09318964183330536
1843
+ },
1844
+ {
1845
+ "name": "transformer_blocks.29.img_mlp.gate_up.weight",
1846
+ "shape": [
1847
+ 24576,
1848
+ 4096
1849
+ ],
1850
+ "source_dtype": "torch.bfloat16",
1851
+ "storage": "nvfp4",
1852
+ "relative_rmse": 0.09439058601856232
1853
+ },
1854
+ {
1855
+ "name": "transformer_blocks.29.img_mlp.out.weight",
1856
+ "shape": [
1857
+ 4096,
1858
+ 12288
1859
+ ],
1860
+ "source_dtype": "torch.bfloat16",
1861
+ "storage": "nvfp4",
1862
+ "relative_rmse": 0.09467429667711258
1863
+ },
1864
+ {
1865
+ "name": "transformer_blocks.3.attn.norm_k.weight",
1866
+ "shape": [
1867
+ 128
1868
+ ],
1869
+ "source_dtype": "torch.bfloat16",
1870
+ "storage": "torch.bfloat16",
1871
+ "sha256": "9adafcf36fbcf1a1fc90864119107f5aaf1034c0af2e842a2e7e76abf67249dc"
1872
+ },
1873
+ {
1874
+ "name": "transformer_blocks.3.attn.norm_q.weight",
1875
+ "shape": [
1876
+ 128
1877
+ ],
1878
+ "source_dtype": "torch.bfloat16",
1879
+ "storage": "torch.bfloat16",
1880
+ "sha256": "11fe2b7ce8993bc09ff3d12afaef320f363fc5b9bdd84e7bef02d223ab09a990"
1881
+ },
1882
+ {
1883
+ "name": "transformer_blocks.3.attn.to_k.weight",
1884
+ "shape": [
1885
+ 4096,
1886
+ 4096
1887
+ ],
1888
+ "source_dtype": "torch.bfloat16",
1889
+ "storage": "nvfp4",
1890
+ "relative_rmse": 0.09428946673870087
1891
+ },
1892
+ {
1893
+ "name": "transformer_blocks.3.attn.to_out.0.weight",
1894
+ "shape": [
1895
+ 4096,
1896
+ 4096
1897
+ ],
1898
+ "source_dtype": "torch.bfloat16",
1899
+ "storage": "nvfp4",
1900
+ "relative_rmse": 0.09520148485898972
1901
+ },
1902
+ {
1903
+ "name": "transformer_blocks.3.attn.to_q.weight",
1904
+ "shape": [
1905
+ 4096,
1906
+ 4096
1907
+ ],
1908
+ "source_dtype": "torch.bfloat16",
1909
+ "storage": "nvfp4",
1910
+ "relative_rmse": 0.09453629702329636
1911
+ },
1912
+ {
1913
+ "name": "transformer_blocks.3.attn.to_v.weight",
1914
+ "shape": [
1915
+ 4096,
1916
+ 4096
1917
+ ],
1918
+ "source_dtype": "torch.bfloat16",
1919
+ "storage": "nvfp4",
1920
+ "relative_rmse": 0.09431635588407516
1921
+ },
1922
+ {
1923
+ "name": "transformer_blocks.3.img_mlp.gate_up.weight",
1924
+ "shape": [
1925
+ 24576,
1926
+ 4096
1927
+ ],
1928
+ "source_dtype": "torch.bfloat16",
1929
+ "storage": "nvfp4",
1930
+ "relative_rmse": 0.09454379230737686
1931
+ },
1932
+ {
1933
+ "name": "transformer_blocks.3.img_mlp.out.weight",
1934
+ "shape": [
1935
+ 4096,
1936
+ 12288
1937
+ ],
1938
+ "source_dtype": "torch.bfloat16",
1939
+ "storage": "nvfp4",
1940
+ "relative_rmse": 0.09459960460662842
1941
+ },
1942
+ {
1943
+ "name": "transformer_blocks.30.attn.norm_k.weight",
1944
+ "shape": [
1945
+ 128
1946
+ ],
1947
+ "source_dtype": "torch.bfloat16",
1948
+ "storage": "torch.bfloat16",
1949
+ "sha256": "69689ec6f8424448347567aaf0a7f577c21cd02dd9d329c977f485671c8c39a6"
1950
+ },
1951
+ {
1952
+ "name": "transformer_blocks.30.attn.norm_q.weight",
1953
+ "shape": [
1954
+ 128
1955
+ ],
1956
+ "source_dtype": "torch.bfloat16",
1957
+ "storage": "torch.bfloat16",
1958
+ "sha256": "49f667f017d3164db4b840b087903801a85b350a103d1173c8f5fc5c5a771165"
1959
+ },
1960
+ {
1961
+ "name": "transformer_blocks.30.attn.to_k.weight",
1962
+ "shape": [
1963
+ 4096,
1964
+ 4096
1965
+ ],
1966
+ "source_dtype": "torch.bfloat16",
1967
+ "storage": "nvfp4",
1968
+ "relative_rmse": 0.09390967339277267
1969
+ },
1970
+ {
1971
+ "name": "transformer_blocks.30.attn.to_out.0.weight",
1972
+ "shape": [
1973
+ 4096,
1974
+ 4096
1975
+ ],
1976
+ "source_dtype": "torch.bfloat16",
1977
+ "storage": "nvfp4",
1978
+ "relative_rmse": 0.095180943608284
1979
+ },
1980
+ {
1981
+ "name": "transformer_blocks.30.attn.to_q.weight",
1982
+ "shape": [
1983
+ 4096,
1984
+ 4096
1985
+ ],
1986
+ "source_dtype": "torch.bfloat16",
1987
+ "storage": "nvfp4",
1988
+ "relative_rmse": 0.09430109709501266
1989
+ },
1990
+ {
1991
+ "name": "transformer_blocks.30.attn.to_v.weight",
1992
+ "shape": [
1993
+ 4096,
1994
+ 4096
1995
+ ],
1996
+ "source_dtype": "torch.bfloat16",
1997
+ "storage": "nvfp4",
1998
+ "relative_rmse": 0.09337132424116135
1999
+ },
2000
+ {
2001
+ "name": "transformer_blocks.30.img_mlp.gate_up.weight",
2002
+ "shape": [
2003
+ 24576,
2004
+ 4096
2005
+ ],
2006
+ "source_dtype": "torch.bfloat16",
2007
+ "storage": "nvfp4",
2008
+ "relative_rmse": 0.09421910345554352
2009
+ },
2010
+ {
2011
+ "name": "transformer_blocks.30.img_mlp.out.weight",
2012
+ "shape": [
2013
+ 4096,
2014
+ 12288
2015
+ ],
2016
+ "source_dtype": "torch.bfloat16",
2017
+ "storage": "nvfp4",
2018
+ "relative_rmse": 0.09415393322706223
2019
+ },
2020
+ {
2021
+ "name": "transformer_blocks.31.attn.norm_k.weight",
2022
+ "shape": [
2023
+ 128
2024
+ ],
2025
+ "source_dtype": "torch.bfloat16",
2026
+ "storage": "torch.bfloat16",
2027
+ "sha256": "c35261d78c7d5177ab9ec1e8412a59853981c038e6ba41066dc216d9a394f1e9"
2028
+ },
2029
+ {
2030
+ "name": "transformer_blocks.31.attn.norm_q.weight",
2031
+ "shape": [
2032
+ 128
2033
+ ],
2034
+ "source_dtype": "torch.bfloat16",
2035
+ "storage": "torch.bfloat16",
2036
+ "sha256": "e0f1c31a771f5948f3f2d2bb96d4e0ed2a897b5d275251486e07c7d7ebb1651a"
2037
+ },
2038
+ {
2039
+ "name": "transformer_blocks.31.attn.to_k.weight",
2040
+ "shape": [
2041
+ 4096,
2042
+ 4096
2043
+ ],
2044
+ "source_dtype": "torch.bfloat16",
2045
+ "storage": "nvfp4",
2046
+ "relative_rmse": 0.09357605874538422
2047
+ },
2048
+ {
2049
+ "name": "transformer_blocks.31.attn.to_out.0.weight",
2050
+ "shape": [
2051
+ 4096,
2052
+ 4096
2053
+ ],
2054
+ "source_dtype": "torch.bfloat16",
2055
+ "storage": "nvfp4",
2056
+ "relative_rmse": 0.09499607235193253
2057
+ },
2058
+ {
2059
+ "name": "transformer_blocks.31.attn.to_q.weight",
2060
+ "shape": [
2061
+ 4096,
2062
+ 4096
2063
+ ],
2064
+ "source_dtype": "torch.bfloat16",
2065
+ "storage": "nvfp4",
2066
+ "relative_rmse": 0.0939827635884285
2067
+ },
2068
+ {
2069
+ "name": "transformer_blocks.31.attn.to_v.weight",
2070
+ "shape": [
2071
+ 4096,
2072
+ 4096
2073
+ ],
2074
+ "source_dtype": "torch.bfloat16",
2075
+ "storage": "nvfp4",
2076
+ "relative_rmse": 0.09221894294023514
2077
+ },
2078
+ {
2079
+ "name": "transformer_blocks.31.img_mlp.gate_up.weight",
2080
+ "shape": [
2081
+ 24576,
2082
+ 4096
2083
+ ],
2084
+ "source_dtype": "torch.bfloat16",
2085
+ "storage": "nvfp4",
2086
+ "relative_rmse": 0.09380938112735748
2087
+ },
2088
+ {
2089
+ "name": "transformer_blocks.31.img_mlp.out.weight",
2090
+ "shape": [
2091
+ 4096,
2092
+ 12288
2093
+ ],
2094
+ "source_dtype": "torch.bfloat16",
2095
+ "storage": "nvfp4",
2096
+ "relative_rmse": 0.09404856711626053
2097
+ },
2098
+ {
2099
+ "name": "transformer_blocks.4.attn.norm_k.weight",
2100
+ "shape": [
2101
+ 128
2102
+ ],
2103
+ "source_dtype": "torch.bfloat16",
2104
+ "storage": "torch.bfloat16",
2105
+ "sha256": "bbee8dc5c62f81c8286c2485ace11ab9eda0f9107596ce5c587c54dae1174406"
2106
+ },
2107
+ {
2108
+ "name": "transformer_blocks.4.attn.norm_q.weight",
2109
+ "shape": [
2110
+ 128
2111
+ ],
2112
+ "source_dtype": "torch.bfloat16",
2113
+ "storage": "torch.bfloat16",
2114
+ "sha256": "dad41e5972a227e07a6e2885f463546b7b171b1a120b9de5199527148ca942aa"
2115
+ },
2116
+ {
2117
+ "name": "transformer_blocks.4.attn.to_k.weight",
2118
+ "shape": [
2119
+ 4096,
2120
+ 4096
2121
+ ],
2122
+ "source_dtype": "torch.bfloat16",
2123
+ "storage": "nvfp4",
2124
+ "relative_rmse": 0.09441725164651871
2125
+ },
2126
+ {
2127
+ "name": "transformer_blocks.4.attn.to_out.0.weight",
2128
+ "shape": [
2129
+ 4096,
2130
+ 4096
2131
+ ],
2132
+ "source_dtype": "torch.bfloat16",
2133
+ "storage": "nvfp4",
2134
+ "relative_rmse": 0.09523670375347137
2135
+ },
2136
+ {
2137
+ "name": "transformer_blocks.4.attn.to_q.weight",
2138
+ "shape": [
2139
+ 4096,
2140
+ 4096
2141
+ ],
2142
+ "source_dtype": "torch.bfloat16",
2143
+ "storage": "nvfp4",
2144
+ "relative_rmse": 0.09450657665729523
2145
+ },
2146
+ {
2147
+ "name": "transformer_blocks.4.attn.to_v.weight",
2148
+ "shape": [
2149
+ 4096,
2150
+ 4096
2151
+ ],
2152
+ "source_dtype": "torch.bfloat16",
2153
+ "storage": "nvfp4",
2154
+ "relative_rmse": 0.09446003288030624
2155
+ },
2156
+ {
2157
+ "name": "transformer_blocks.4.img_mlp.gate_up.weight",
2158
+ "shape": [
2159
+ 24576,
2160
+ 4096
2161
+ ],
2162
+ "source_dtype": "torch.bfloat16",
2163
+ "storage": "nvfp4",
2164
+ "relative_rmse": 0.09445067495107651
2165
+ },
2166
+ {
2167
+ "name": "transformer_blocks.4.img_mlp.out.weight",
2168
+ "shape": [
2169
+ 4096,
2170
+ 12288
2171
+ ],
2172
+ "source_dtype": "torch.bfloat16",
2173
+ "storage": "nvfp4",
2174
+ "relative_rmse": 0.09469588100910187
2175
+ },
2176
+ {
2177
+ "name": "transformer_blocks.5.attn.norm_k.weight",
2178
+ "shape": [
2179
+ 128
2180
+ ],
2181
+ "source_dtype": "torch.bfloat16",
2182
+ "storage": "torch.bfloat16",
2183
+ "sha256": "1d8bde5394b8c899d54bdb9bdbe83db069204b16286c4aeaf05c7c9f166090ad"
2184
+ },
2185
+ {
2186
+ "name": "transformer_blocks.5.attn.norm_q.weight",
2187
+ "shape": [
2188
+ 128
2189
+ ],
2190
+ "source_dtype": "torch.bfloat16",
2191
+ "storage": "torch.bfloat16",
2192
+ "sha256": "d48c5fe07fcd71a357af1295c6e00109d4e2f23fb5b213ace6dc88fc7fd33de2"
2193
+ },
2194
+ {
2195
+ "name": "transformer_blocks.5.attn.to_k.weight",
2196
+ "shape": [
2197
+ 4096,
2198
+ 4096
2199
+ ],
2200
+ "source_dtype": "torch.bfloat16",
2201
+ "storage": "nvfp4",
2202
+ "relative_rmse": 0.09460146725177765
2203
+ },
2204
+ {
2205
+ "name": "transformer_blocks.5.attn.to_out.0.weight",
2206
+ "shape": [
2207
+ 4096,
2208
+ 4096
2209
+ ],
2210
+ "source_dtype": "torch.bfloat16",
2211
+ "storage": "nvfp4",
2212
+ "relative_rmse": 0.09520139545202255
2213
+ },
2214
+ {
2215
+ "name": "transformer_blocks.5.attn.to_q.weight",
2216
+ "shape": [
2217
+ 4096,
2218
+ 4096
2219
+ ],
2220
+ "source_dtype": "torch.bfloat16",
2221
+ "storage": "nvfp4",
2222
+ "relative_rmse": 0.09447811543941498
2223
+ },
2224
+ {
2225
+ "name": "transformer_blocks.5.attn.to_v.weight",
2226
+ "shape": [
2227
+ 4096,
2228
+ 4096
2229
+ ],
2230
+ "source_dtype": "torch.bfloat16",
2231
+ "storage": "nvfp4",
2232
+ "relative_rmse": 0.09432633221149445
2233
+ },
2234
+ {
2235
+ "name": "transformer_blocks.5.img_mlp.gate_up.weight",
2236
+ "shape": [
2237
+ 24576,
2238
+ 4096
2239
+ ],
2240
+ "source_dtype": "torch.bfloat16",
2241
+ "storage": "nvfp4",
2242
+ "relative_rmse": 0.0945073738694191
2243
+ },
2244
+ {
2245
+ "name": "transformer_blocks.5.img_mlp.out.weight",
2246
+ "shape": [
2247
+ 4096,
2248
+ 12288
2249
+ ],
2250
+ "source_dtype": "torch.bfloat16",
2251
+ "storage": "nvfp4",
2252
+ "relative_rmse": 0.09461148828268051
2253
+ },
2254
+ {
2255
+ "name": "transformer_blocks.6.attn.norm_k.weight",
2256
+ "shape": [
2257
+ 128
2258
+ ],
2259
+ "source_dtype": "torch.bfloat16",
2260
+ "storage": "torch.bfloat16",
2261
+ "sha256": "fbe5113744aff19a2be1ad19a9e3c2c97bdd83cd8d468168d1c5ada895f6b0e1"
2262
+ },
2263
+ {
2264
+ "name": "transformer_blocks.6.attn.norm_q.weight",
2265
+ "shape": [
2266
+ 128
2267
+ ],
2268
+ "source_dtype": "torch.bfloat16",
2269
+ "storage": "torch.bfloat16",
2270
+ "sha256": "2d100b9bb3855e307372e7f60875268e9f760ad441936bc0bc15072a997143cc"
2271
+ },
2272
+ {
2273
+ "name": "transformer_blocks.6.attn.to_k.weight",
2274
+ "shape": [
2275
+ 4096,
2276
+ 4096
2277
+ ],
2278
+ "source_dtype": "torch.bfloat16",
2279
+ "storage": "nvfp4",
2280
+ "relative_rmse": 0.09451349079608917
2281
+ },
2282
+ {
2283
+ "name": "transformer_blocks.6.attn.to_out.0.weight",
2284
+ "shape": [
2285
+ 4096,
2286
+ 4096
2287
+ ],
2288
+ "source_dtype": "torch.bfloat16",
2289
+ "storage": "nvfp4",
2290
+ "relative_rmse": 0.09513034671545029
2291
+ },
2292
+ {
2293
+ "name": "transformer_blocks.6.attn.to_q.weight",
2294
+ "shape": [
2295
+ 4096,
2296
+ 4096
2297
+ ],
2298
+ "source_dtype": "torch.bfloat16",
2299
+ "storage": "nvfp4",
2300
+ "relative_rmse": 0.09451977163553238
2301
+ },
2302
+ {
2303
+ "name": "transformer_blocks.6.attn.to_v.weight",
2304
+ "shape": [
2305
+ 4096,
2306
+ 4096
2307
+ ],
2308
+ "source_dtype": "torch.bfloat16",
2309
+ "storage": "nvfp4",
2310
+ "relative_rmse": 0.09427080303430557
2311
+ },
2312
+ {
2313
+ "name": "transformer_blocks.6.img_mlp.gate_up.weight",
2314
+ "shape": [
2315
+ 24576,
2316
+ 4096
2317
+ ],
2318
+ "source_dtype": "torch.bfloat16",
2319
+ "storage": "nvfp4",
2320
+ "relative_rmse": 0.09461503475904465
2321
+ },
2322
+ {
2323
+ "name": "transformer_blocks.6.img_mlp.out.weight",
2324
+ "shape": [
2325
+ 4096,
2326
+ 12288
2327
+ ],
2328
+ "source_dtype": "torch.bfloat16",
2329
+ "storage": "nvfp4",
2330
+ "relative_rmse": 0.0945393294095993
2331
+ },
2332
+ {
2333
+ "name": "transformer_blocks.7.attn.norm_k.weight",
2334
+ "shape": [
2335
+ 128
2336
+ ],
2337
+ "source_dtype": "torch.bfloat16",
2338
+ "storage": "torch.bfloat16",
2339
+ "sha256": "1702076f2303eb64dc660d2934811de2ac7ef90e816f60fa319d422fc34867d2"
2340
+ },
2341
+ {
2342
+ "name": "transformer_blocks.7.attn.norm_q.weight",
2343
+ "shape": [
2344
+ 128
2345
+ ],
2346
+ "source_dtype": "torch.bfloat16",
2347
+ "storage": "torch.bfloat16",
2348
+ "sha256": "b52ac50f9154d388aa3178a1cc21b1400616120ee3095ad984658f75b5ca0b1c"
2349
+ },
2350
+ {
2351
+ "name": "transformer_blocks.7.attn.to_k.weight",
2352
+ "shape": [
2353
+ 4096,
2354
+ 4096
2355
+ ],
2356
+ "source_dtype": "torch.bfloat16",
2357
+ "storage": "nvfp4",
2358
+ "relative_rmse": 0.09451562166213989
2359
+ },
2360
+ {
2361
+ "name": "transformer_blocks.7.attn.to_out.0.weight",
2362
+ "shape": [
2363
+ 4096,
2364
+ 4096
2365
+ ],
2366
+ "source_dtype": "torch.bfloat16",
2367
+ "storage": "nvfp4",
2368
+ "relative_rmse": 0.0953247919678688
2369
+ },
2370
+ {
2371
+ "name": "transformer_blocks.7.attn.to_q.weight",
2372
+ "shape": [
2373
+ 4096,
2374
+ 4096
2375
+ ],
2376
+ "source_dtype": "torch.bfloat16",
2377
+ "storage": "nvfp4",
2378
+ "relative_rmse": 0.09449370205402374
2379
+ },
2380
+ {
2381
+ "name": "transformer_blocks.7.attn.to_v.weight",
2382
+ "shape": [
2383
+ 4096,
2384
+ 4096
2385
+ ],
2386
+ "source_dtype": "torch.bfloat16",
2387
+ "storage": "nvfp4",
2388
+ "relative_rmse": 0.09425489604473114
2389
+ },
2390
+ {
2391
+ "name": "transformer_blocks.7.img_mlp.gate_up.weight",
2392
+ "shape": [
2393
+ 24576,
2394
+ 4096
2395
+ ],
2396
+ "source_dtype": "torch.bfloat16",
2397
+ "storage": "nvfp4",
2398
+ "relative_rmse": 0.09459647536277771
2399
+ },
2400
+ {
2401
+ "name": "transformer_blocks.7.img_mlp.out.weight",
2402
+ "shape": [
2403
+ 4096,
2404
+ 12288
2405
+ ],
2406
+ "source_dtype": "torch.bfloat16",
2407
+ "storage": "nvfp4",
2408
+ "relative_rmse": 0.09462863206863403
2409
+ },
2410
+ {
2411
+ "name": "transformer_blocks.8.attn.norm_k.weight",
2412
+ "shape": [
2413
+ 128
2414
+ ],
2415
+ "source_dtype": "torch.bfloat16",
2416
+ "storage": "torch.bfloat16",
2417
+ "sha256": "9a56df6095d2a425d0f66275b41bd6c356056735cd3bfa023724e0cd90e8dfff"
2418
+ },
2419
+ {
2420
+ "name": "transformer_blocks.8.attn.norm_q.weight",
2421
+ "shape": [
2422
+ 128
2423
+ ],
2424
+ "source_dtype": "torch.bfloat16",
2425
+ "storage": "torch.bfloat16",
2426
+ "sha256": "ff2c3f870cad641f784f36f95efddfbc098752999384ee69463b39bb1e7f5533"
2427
+ },
2428
+ {
2429
+ "name": "transformer_blocks.8.attn.to_k.weight",
2430
+ "shape": [
2431
+ 4096,
2432
+ 4096
2433
+ ],
2434
+ "source_dtype": "torch.bfloat16",
2435
+ "storage": "nvfp4",
2436
+ "relative_rmse": 0.09443946182727814
2437
+ },
2438
+ {
2439
+ "name": "transformer_blocks.8.attn.to_out.0.weight",
2440
+ "shape": [
2441
+ 4096,
2442
+ 4096
2443
+ ],
2444
+ "source_dtype": "torch.bfloat16",
2445
+ "storage": "nvfp4",
2446
+ "relative_rmse": 0.09550652652978897
2447
+ },
2448
+ {
2449
+ "name": "transformer_blocks.8.attn.to_q.weight",
2450
+ "shape": [
2451
+ 4096,
2452
+ 4096
2453
+ ],
2454
+ "source_dtype": "torch.bfloat16",
2455
+ "storage": "nvfp4",
2456
+ "relative_rmse": 0.0944790169596672
2457
+ },
2458
+ {
2459
+ "name": "transformer_blocks.8.attn.to_v.weight",
2460
+ "shape": [
2461
+ 4096,
2462
+ 4096
2463
+ ],
2464
+ "source_dtype": "torch.bfloat16",
2465
+ "storage": "nvfp4",
2466
+ "relative_rmse": 0.09414148330688477
2467
+ },
2468
+ {
2469
+ "name": "transformer_blocks.8.img_mlp.gate_up.weight",
2470
+ "shape": [
2471
+ 24576,
2472
+ 4096
2473
+ ],
2474
+ "source_dtype": "torch.bfloat16",
2475
+ "storage": "nvfp4",
2476
+ "relative_rmse": 0.09464308619499207
2477
+ },
2478
+ {
2479
+ "name": "transformer_blocks.8.img_mlp.out.weight",
2480
+ "shape": [
2481
+ 4096,
2482
+ 12288
2483
+ ],
2484
+ "source_dtype": "torch.bfloat16",
2485
+ "storage": "nvfp4",
2486
+ "relative_rmse": 0.09467563033103943
2487
+ },
2488
+ {
2489
+ "name": "transformer_blocks.9.attn.norm_k.weight",
2490
+ "shape": [
2491
+ 128
2492
+ ],
2493
+ "source_dtype": "torch.bfloat16",
2494
+ "storage": "torch.bfloat16",
2495
+ "sha256": "7e20746b285a1d254a56aed717f9e22293ea36bb699d5f599b7f43384bd1441a"
2496
+ },
2497
+ {
2498
+ "name": "transformer_blocks.9.attn.norm_q.weight",
2499
+ "shape": [
2500
+ 128
2501
+ ],
2502
+ "source_dtype": "torch.bfloat16",
2503
+ "storage": "torch.bfloat16",
2504
+ "sha256": "c0bb2e40385389a5f7312b3a6c37883d70306a20f0affdd2b8303109431b8c3d"
2505
+ },
2506
+ {
2507
+ "name": "transformer_blocks.9.attn.to_k.weight",
2508
+ "shape": [
2509
+ 4096,
2510
+ 4096
2511
+ ],
2512
+ "source_dtype": "torch.bfloat16",
2513
+ "storage": "nvfp4",
2514
+ "relative_rmse": 0.09447592496871948
2515
+ },
2516
+ {
2517
+ "name": "transformer_blocks.9.attn.to_out.0.weight",
2518
+ "shape": [
2519
+ 4096,
2520
+ 4096
2521
+ ],
2522
+ "source_dtype": "torch.bfloat16",
2523
+ "storage": "nvfp4",
2524
+ "relative_rmse": 0.09517411887645721
2525
+ },
2526
+ {
2527
+ "name": "transformer_blocks.9.attn.to_q.weight",
2528
+ "shape": [
2529
+ 4096,
2530
+ 4096
2531
+ ],
2532
+ "source_dtype": "torch.bfloat16",
2533
+ "storage": "nvfp4",
2534
+ "relative_rmse": 0.09470388293266296
2535
+ },
2536
+ {
2537
+ "name": "transformer_blocks.9.attn.to_v.weight",
2538
+ "shape": [
2539
+ 4096,
2540
+ 4096
2541
+ ],
2542
+ "source_dtype": "torch.bfloat16",
2543
+ "storage": "nvfp4",
2544
+ "relative_rmse": 0.09404205530881882
2545
+ },
2546
+ {
2547
+ "name": "transformer_blocks.9.img_mlp.gate_up.weight",
2548
+ "shape": [
2549
+ 24576,
2550
+ 4096
2551
+ ],
2552
+ "source_dtype": "torch.bfloat16",
2553
+ "storage": "nvfp4",
2554
+ "relative_rmse": 0.09456857293844223
2555
+ },
2556
+ {
2557
+ "name": "transformer_blocks.9.img_mlp.out.weight",
2558
+ "shape": [
2559
+ 4096,
2560
+ 12288
2561
+ ],
2562
+ "source_dtype": "torch.bfloat16",
2563
+ "storage": "nvfp4",
2564
+ "relative_rmse": 0.09464238584041595
2565
+ },
2566
+ {
2567
+ "name": "txt_in.in_layer.weight",
2568
+ "shape": [
2569
+ 4096,
2570
+ 4096
2571
+ ],
2572
+ "source_dtype": "torch.bfloat16",
2573
+ "storage": "torch.bfloat16",
2574
+ "sha256": "2afeaaac182536117eb984cd878e94b3d5447fd1e0989f1c925b55f5b2aba828"
2575
+ },
2576
+ {
2577
+ "name": "txt_in.out_layer.weight",
2578
+ "shape": [
2579
+ 4096,
2580
+ 4096
2581
+ ],
2582
+ "source_dtype": "torch.bfloat16",
2583
+ "storage": "torch.bfloat16",
2584
+ "sha256": "e4c37943834c699fc4e101b3141bee53839b78c096e00cfbd836b8b061162672"
2585
+ },
2586
+ {
2587
+ "name": "txt_in.text_norm.weight",
2588
+ "shape": [
2589
+ 4096
2590
+ ],
2591
+ "source_dtype": "torch.bfloat16",
2592
+ "storage": "torch.bfloat16",
2593
+ "sha256": "b496795d142460d2f72d13b560d7cf54fba3081cc895c2d41b8e00b2470f9efe"
2594
+ }
2595
+ ]
2596
+ }
gallery.html CHANGED
@@ -1,2 +1,2 @@
1
- <!doctype html><html lang="en"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Prism Image 2.1 | NVFP4 comparisons</title><style>
2
- body{margin:0;background:#101621;color:#edf4ff;font:16px/1.6 system-ui}main{max-width:1250px;margin:auto;padding:48px 24px}h1{font-size:clamp(32px,5vw,64px);line-height:1.1;margin:12px 0}p{color:#b7c5db;max-width:850px}.tag{color:#8be8c9;letter-spacing:.12em;text-transform:uppercase;font-size:13px}select,input{font:inherit;background:#202d40;color:white;border:1px solid #526278;padding:10px;border-radius:8px}header{margin-bottom:32px}.controls{display:flex;gap:15px;flex-wrap:wrap;align-items:center;margin:20px 0}.compare{position:relative;aspect-ratio:1;background:#182030;max-width:1024px;border-radius:14px;overflow:hidden}.compare img{width:100%;height:100%;object-fit:contain;position:absolute;inset:0}.compare .over{clip-path:inset(0 50% 0 0)}.labels{position:absolute;top:16px;left:16px;right:16px;display:flex;justify-content:space-between}.labels span{padding:6px 12px;background:#101621dd;border-radius:8px}.line{position:absolute;top:0;bottom:0;left:50%;width:2px;background:#fff}.prompt{white-space:pre-wrap;max-width:1024px}a{color:#8be8c9}footer{margin-top:40px;border-top:1px solid #344158;padding-top:22px}</style><main><header><div class="tag">BennyDaBall / RTX 5090 / Built with Qwen</div><h1>Two weights.<br>One prompt.</h1><p>BF16 source weights versus Prism's mixed NVFP4 image transformer and accelerated language encoder. Same prompt, seed, size, 40-step Euler/simple sampler, CFG 1. Quantization can change composition and detail. Every pair is included, not just the best-looking ones.</p></header><div class="controls"><select id="case"></select><label>Comparison split <input id="split" type="range" min="0" max="100" value="50"></label></div><div class="compare"><img id="nv"><img id="bf" class="over"><div class="line" id="line"></div><div class="labels"><span>BF16 weights</span><span>Prism NVFP4</span></div></div><p id="prompt" class="prompt"></p><footer><a href="https://huggingface.co/BennyDaBall/Prism-Image-2.1-NVFP4">Models, tested workflows, conversion scripts and validation</a><p>Research/evaluation license. Native ComfyUI nodes. The accelerated encoder requires the included runtime patch. Vision, VAE and protected tensors remain BF16.</p></footer></main><script>const data=[{"name": "portrait", "seed": 20260920, "prompt": "Editorial environmental photograph of an elderly female violin maker in her sunlit wooden workshop. She holds a small carving tool in her right hand and steadies an unfinished violin with her left hand. Natural skin texture, thoughtful expression, curled wood shavings on the workbench, soft window light from the left, warm subdued colors, shallow depth of field, 50mm lens. Waist-up composition, candid and believable.", "bf": "portrait_20260920_bf-bf.webp", "nv": "portrait_20260920_nv-nv.webp"}, {"name": "portrait", "seed": 20260921, "prompt": "Editorial environmental photograph of an elderly female violin maker in her sunlit wooden workshop. She holds a small carving tool in her right hand and steadies an unfinished violin with her left hand. Natural skin texture, thoughtful expression, curled wood shavings on the workbench, soft window light from the left, warm subdued colors, shallow depth of field, 50mm lens. Waist-up composition, candid and believable.", "bf": "portrait_20260921_bf-bf.webp", "nv": "portrait_20260921_nv-nv.webp"}, {"name": "landscape", "seed": 20260920, "prompt": "Cinematic landscape photograph of a tiny red mountain refuge on a rocky ridge above a sea of clouds at sunrise. A winding stone footpath leads from the bottom left toward the refuge. Jagged distant peaks, warm sunlight grazing the roof, cool blue shadows, thin luminous mist, realistic geology, restrained natural colors. Wide composition with a strong sense of scale, no people, no lettering.", "bf": "landscape_20260920_bf-bf.webp", "nv": "landscape_20260920_nv-nv.webp"}, {"name": "landscape", "seed": 20260921, "prompt": "Cinematic landscape photograph of a tiny red mountain refuge on a rocky ridge above a sea of clouds at sunrise. A winding stone footpath leads from the bottom left toward the refuge. Jagged distant peaks, warm sunlight grazing the roof, cool blue shadows, thin luminous mist, realistic geology, restrained natural colors. Wide composition with a strong sense of scale, no people, no lettering.", "bf": "landscape_20260921_bf-bf.webp", "nv": "landscape_20260921_nv-nv.webp"}, {"name": "product", "seed": 20260920, "prompt": "Premium studio product photograph of one amber glass perfume bottle with a brushed brass cap, standing on a dark green marble block. A small cream paper label on the bottle reads \"EMBER\" in elegant black capital letters. A sliced orange and a sprig of rosemary rest beside the block. Warm rim lighting, soft front fill, rich realistic materials, controlled glass reflections, clean charcoal background, centered luxury advertising composition.", "bf": "product_20260920_bf-bf.webp", "nv": "product_20260920_nv-nv.webp"}, {"name": "product", "seed": 20260921, "prompt": "Premium studio product photograph of one amber glass perfume bottle with a brushed brass cap, standing on a dark green marble block. A small cream paper label on the bottle reads \"EMBER\" in elegant black capital letters. A sliced orange and a sprig of rosemary rest beside the block. Warm rim lighting, soft front fill, rich realistic materials, controlled glass reflections, clean charcoal background, centered luxury advertising composition.", "bf": "product_20260921_bf-bf.webp", "nv": "product_20260921_nv-nv.webp"}, {"name": "typography", "seed": 20260920, "prompt": "Design a polished square travel poster with a restrained mid-century screen-print illustration of a red tram crossing a stone bridge over a blue river at sunset. Large cream headline at the top reads exactly \"TAKE THE SCENIC ROUTE\". A smaller line at the bottom reads exactly \"SLOW TRAVEL CLUB\". Strong typographic hierarchy, generous margins, flat navy, coral and cream inks, subtle paper grain. All lettering sharp and clearly legible.", "bf": "typography_20260920_bf-bf.webp", "nv": "typography_20260920_nv-nv.webp"}, {"name": "typography", "seed": 20260921, "prompt": "Design a polished square travel poster with a restrained mid-century screen-print illustration of a red tram crossing a stone bridge over a blue river at sunset. Large cream headline at the top reads exactly \"TAKE THE SCENIC ROUTE\". A smaller line at the bottom reads exactly \"SLOW TRAVEL CLUB\". Strong typographic hierarchy, generous margins, flat navy, coral and cream inks, subtle paper grain. All lettering sharp and clearly legible.", "bf": "typography_20260921_bf-bf.webp", "nv": "typography_20260921_nv-nv.webp"}, {"name": "spatial", "seed": 20260920, "prompt": "A clean studio photograph on a light gray tabletop showing exactly three objects arranged in a straight horizontal row: a red ceramic cube on the left, a blue glass sphere in the center, and a yellow wooden cone on the right. The sphere is slightly larger than the cube, and the cone is tallest. A single soft light from the upper left casts shadows toward the lower right. Neutral pale gray background, eye-level camera, every object fully visible and separated by equal gaps, no other objects, no text.", "bf": "spatial_20260920_bf-bf.webp", "nv": "spatial_20260920_nv-nv.webp"}, {"name": "spatial", "seed": 20260921, "prompt": "A clean studio photograph on a light gray tabletop showing exactly three objects arranged in a straight horizontal row: a red ceramic cube on the left, a blue glass sphere in the center, and a yellow wooden cone on the right. The sphere is slightly larger than the cube, and the cone is tallest. A single soft light from the upper left casts shadows toward the lower right. Neutral pale gray background, eye-level camera, every object fully visible and separated by equal gaps, no other objects, no text.", "bf": "spatial_20260921_bf-bf.webp", "nv": "spatial_20260921_nv-nv.webp"}, {"name": "illustration", "seed": 20260920, "prompt": "A richly detailed storybook gouache illustration of a small fox librarian wearing round spectacles and a moss-green waistcoat, standing on a rolling wooden ladder in a towering circular library inside an ancient tree. The fox reaches toward a glowing blue book with one paw. Curving shelves, warm amber lamps, tiny floating dust motes, a rainy moonlit window, hand-painted brush texture, charming expressive face, coherent architecture, dramatic but cozy composition, no lettering.", "bf": "illustration_20260920_bf-bf.webp", "nv": "illustration_20260920_nv-nv.webp"}, {"name": "illustration", "seed": 20260921, "prompt": "A richly detailed storybook gouache illustration of a small fox librarian wearing round spectacles and a moss-green waistcoat, standing on a rolling wooden ladder in a towering circular library inside an ancient tree. The fox reaches toward a glowing blue book with one paw. Curving shelves, warm amber lamps, tiny floating dust motes, a rainy moonlit window, hand-painted brush texture, charming expressive face, coherent architecture, dramatic but cozy composition, no lettering.", "bf": "illustration_20260921_bf-bf.webp", "nv": "illustration_20260921_nv-nv.webp"}, {"name": "fashion", "seed": 20260920, "prompt": "A candid editorial photograph of an adult woman with short dark curly hair and freckles, wearing a cobalt-blue tailored wool coat over a cream turtleneck, standing outside a small Parisian cafe on a rainy evening. She glances back over her shoulder with a slight natural smile, holding a folded red umbrella by its wooden handle. Fine raindrops on the coat, realistic skin pores, wet pavement reflections, warm cafe windows against cool blue dusk, waist-up framing, 85mm lens, subtle film grain, believable unretouched photography.", "bf": "fashion_20260920_bf-bf.webp", "nv": "fashion_20260920_nv-nv.webp"}, {"name": "fashion", "seed": 20260921, "prompt": "A candid editorial photograph of an adult woman with short dark curly hair and freckles, wearing a cobalt-blue tailored wool coat over a cream turtleneck, standing outside a small Parisian cafe on a rainy evening. She glances back over her shoulder with a slight natural smile, holding a folded red umbrella by its wooden handle. Fine raindrops on the coat, realistic skin pores, wet pavement reflections, warm cafe windows against cool blue dusk, waist-up framing, 85mm lens, subtle film grain, believable unretouched photography.", "bf": "fashion_20260921_bf-bf.webp", "nv": "fashion_20260921_nv-nv.webp"}, {"name": "night_market", "seed": 20260920, "prompt": "A cinematic street photograph of a crowded night market in Taipei during gentle rain. In the foreground an elderly cook in a white apron lifts a steaming bamboo basket from a street-food stall, with hands clearly visible. A small illuminated sign above him reads exactly \"NIGHT BITES\". Rich red lanterns and teal shop lights reflect in wet asphalt; layered pedestrians with umbrellas recede into the scene. Realistic steam, appetizing food texture, natural human proportions, 35mm lens, documentary photography, atmospheric but restrained color.", "bf": "night_market_20260920_bf-bf.webp", "nv": "night_market_20260920_nv-nv.webp"}, {"name": "night_market", "seed": 20260921, "prompt": "A cinematic street photograph of a crowded night market in Taipei during gentle rain. In the foreground an elderly cook in a white apron lifts a steaming bamboo basket from a street-food stall, with hands clearly visible. A small illuminated sign above him reads exactly \"NIGHT BITES\". Rich red lanterns and teal shop lights reflect in wet asphalt; layered pedestrians with umbrellas recede into the scene. Realistic steam, appetizing food texture, natural human proportions, 35mm lens, documentary photography, atmospheric but restrained color.", "bf": "night_market_20260921_bf-bf.webp", "nv": "night_market_20260921_nv-nv.webp"}, {"name": "architecture", "seed": 20260920, "prompt": "Architectural interior photograph of a sunken conversation pit in a restored 1970s coastal house. Rust-orange corduroy sofas surround a low circular walnut table holding a clear glass vase with three white tulips. Floor-to-ceiling windows reveal a stormy ocean. Rain beads on the windows; a suspended black fireplace glows softly at the left. Pale travertine floor, detailed walnut ceiling, soft overcast daylight, balanced wide-angle composition, realistic materials and straight architectural lines, no people or lettering.", "bf": "architecture_20260920_bf-bf.webp", "nv": "architecture_20260920_nv-nv.webp"}, {"name": "architecture", "seed": 20260921, "prompt": "Architectural interior photograph of a sunken conversation pit in a restored 1970s coastal house. Rust-orange corduroy sofas surround a low circular walnut table holding a clear glass vase with three white tulips. Floor-to-ceiling windows reveal a stormy ocean. Rain beads on the windows; a suspended black fireplace glows softly at the left. Pale travertine floor, detailed walnut ceiling, soft overcast daylight, balanced wide-angle composition, realistic materials and straight architectural lines, no people or lettering.", "bf": "architecture_20260921_bf-bf.webp", "nv": "architecture_20260921_nv-nv.webp"}, {"name": "macro", "seed": 20260920, "prompt": "Extreme macro nature photograph of a tiny emerald jumping spider perched on the curled edge of a copper-colored autumn leaf. A single spherical dew drop beside the spider reflects a miniature garden. Fine individual hairs on the spider, realistic eight-legged anatomy, sharp jewel-like eyes, translucent dew, intricate leaf veins. Soft warm backlight, muted olive background with circular bokeh, narrow but carefully placed focus plane, natural colors, scientifically plausible wildlife photography, no lettering.", "bf": "macro_20260920_bf-bf.webp", "nv": "macro_20260920_nv-nv.webp"}, {"name": "macro", "seed": 20260921, "prompt": "Extreme macro nature photograph of a tiny emerald jumping spider perched on the curled edge of a copper-colored autumn leaf. A single spherical dew drop beside the spider reflects a miniature garden. Fine individual hairs on the spider, realistic eight-legged anatomy, sharp jewel-like eyes, translucent dew, intricate leaf veins. Soft warm backlight, muted olive background with circular bokeh, narrow but carefully placed focus plane, natural colors, scientifically plausible wildlife photography, no lettering.", "bf": "macro_20260921_bf-bf.webp", "nv": "macro_20260921_nv-nv.webp"}, {"name": "astronaut", "seed": 20260920, "prompt": "A cinematic close-up photograph of an adult female astronaut inside a weathered lunar habitat, looking directly into the camera through a clear helmet visor. Her brown eyes, small freckles and loose strands of hair remain clearly visible behind subtle reflections. White spacesuit with stitched fabric, orange fittings and a small rectangular chest patch reading \"LUNA 09\". Warm practical light on one cheek and cool Earthlight on the other, Earth visible through a round window in the background, restrained analog science-fiction realism, finely detailed materials, no illustration style.", "bf": "astronaut_20260920_bf-bf.webp", "nv": "astronaut_20260920_nv-nv.webp"}, {"name": "astronaut", "seed": 20260921, "prompt": "A cinematic close-up photograph of an adult female astronaut inside a weathered lunar habitat, looking directly into the camera through a clear helmet visor. Her brown eyes, small freckles and loose strands of hair remain clearly visible behind subtle reflections. White spacesuit with stitched fabric, orange fittings and a small rectangular chest patch reading \"LUNA 09\". Warm practical light on one cheek and cool Earthlight on the other, Earth visible through a round window in the background, restrained analog science-fiction realism, finely detailed materials, no illustration style.", "bf": "astronaut_20260921_bf-bf.webp", "nv": "astronaut_20260921_nv-nv.webp"}, {"name": "botanical_poster", "seed": 20260920, "prompt": "An elegant contemporary botanical exhibition poster on textured ivory paper. A lifelike ink-and-watercolor illustration of one flowering magnolia branch curves diagonally across the center, pale pink petals with delicate translucent edges and deep green leaves. At the top, a large refined serif headline reads exactly \"THE QUIET GARDEN\". Beneath it, smaller text reads exactly \"BOTANICAL STUDIES\". At the bottom, a small centered line reads exactly \"APRIL 12 - MAY 30\". Generous negative space, exceptional typography, restrained burgundy and sage palette, museum-quality printed design.", "bf": "botanical_poster_20260920_bf-bf.webp", "nv": "botanical_poster_20260920_nv-nv.webp"}, {"name": "botanical_poster", "seed": 20260921, "prompt": "An elegant contemporary botanical exhibition poster on textured ivory paper. A lifelike ink-and-watercolor illustration of one flowering magnolia branch curves diagonally across the center, pale pink petals with delicate translucent edges and deep green leaves. At the top, a large refined serif headline reads exactly \"THE QUIET GARDEN\". Beneath it, smaller text reads exactly \"BOTANICAL STUDIES\". At the bottom, a small centered line reads exactly \"APRIL 12 - MAY 30\". Generous negative space, exceptional typography, restrained burgundy and sage palette, museum-quality printed design.", "bf": "botanical_poster_20260921_bf-bf.webp", "nv": "botanical_poster_20260921_nv-nv.webp"}];const select=document.getElementById('case');data.forEach((r,i)=>{let o=document.createElement('option');o.value=i;o.textContent=r.name.replaceAll('_',' ')+' / seed '+r.seed;select.append(o)});function show(){let r=data[select.value];document.getElementById('bf').src='assets/'+r.bf;document.getElementById('nv').src='assets/'+r.nv;document.getElementById('prompt').textContent=r.prompt}select.onchange=show;document.getElementById('split').oninput=e=>{let v=e.target.value;document.getElementById('bf').style.clipPath='inset(0 '+(100-v)+'% 0 0)';document.getElementById('line').style.left=v+'%'};show();</script></html>
 
1
+ <!doctype html><html lang="en"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>Qwen Image 2.1 | NVFP4 comparisons</title><style>
2
+ body{margin:0;background:#101621;color:#edf4ff;font:16px/1.6 system-ui}main{max-width:1250px;margin:auto;padding:48px 24px}h1{font-size:clamp(32px,5vw,64px);line-height:1.1;margin:12px 0}p{color:#b7c5db;max-width:850px}.tag{color:#8be8c9;letter-spacing:.12em;text-transform:uppercase;font-size:13px}select,input{font:inherit;background:#202d40;color:white;border:1px solid #526278;padding:10px;border-radius:8px}header{margin-bottom:32px}.controls{display:flex;gap:15px;flex-wrap:wrap;align-items:center;margin:20px 0}.compare{position:relative;aspect-ratio:1;background:#182030;max-width:1024px;border-radius:14px;overflow:hidden}.compare img{width:100%;height:100%;object-fit:contain;position:absolute;inset:0}.compare .over{clip-path:inset(0 50% 0 0)}.labels{position:absolute;top:16px;left:16px;right:16px;display:flex;justify-content:space-between}.labels span{padding:6px 12px;background:#101621dd;border-radius:8px}.line{position:absolute;top:0;bottom:0;left:50%;width:2px;background:#fff}.prompt{white-space:pre-wrap;max-width:1024px}a{color:#8be8c9}footer{margin-top:40px;border-top:1px solid #344158;padding-top:22px}</style><main><header><div class="tag">BennyDaBall / RTX 5090 / Built with Qwen</div><h1>Two weights.<br>One prompt.</h1><p>BF16 source weights versus the Qwen Image 2.1 mixed NVFP4 image transformer and accelerated language encoder. Same prompt, seed, size, 40-step Euler/simple sampler, CFG 1. Quantization can change composition and detail. Every pair is included, not just the best-looking ones.</p></header><div class="controls"><select id="case"></select><label>Comparison split <input id="split" type="range" min="0" max="100" value="50"></label></div><div class="compare"><img id="nv"><img id="bf" class="over"><div class="line" id="line"></div><div class="labels"><span>BF16 weights</span><span>Qwen Image 2.1 NVFP4</span></div></div><p id="prompt" class="prompt"></p><footer><a href="https://huggingface.co/BennyDaBall/Qwen-Image-2.1-NVFP4">Models, tested workflows, conversion scripts and validation</a><p>Research/evaluation license. Native ComfyUI nodes. The accelerated encoder requires the included runtime patch. Vision, VAE and protected tensors remain BF16.</p></footer></main><script>const data=[{"name": "portrait", "seed": 20260920, "prompt": "Editorial environmental photograph of an elderly female violin maker in her sunlit wooden workshop. She holds a small carving tool in her right hand and steadies an unfinished violin with her left hand. Natural skin texture, thoughtful expression, curled wood shavings on the workbench, soft window light from the left, warm subdued colors, shallow depth of field, 50mm lens. Waist-up composition, candid and believable.", "bf": "portrait_20260920_bf-bf.webp", "nv": "portrait_20260920_nv-nv.webp"}, {"name": "portrait", "seed": 20260921, "prompt": "Editorial environmental photograph of an elderly female violin maker in her sunlit wooden workshop. She holds a small carving tool in her right hand and steadies an unfinished violin with her left hand. Natural skin texture, thoughtful expression, curled wood shavings on the workbench, soft window light from the left, warm subdued colors, shallow depth of field, 50mm lens. Waist-up composition, candid and believable.", "bf": "portrait_20260921_bf-bf.webp", "nv": "portrait_20260921_nv-nv.webp"}, {"name": "landscape", "seed": 20260920, "prompt": "Cinematic landscape photograph of a tiny red mountain refuge on a rocky ridge above a sea of clouds at sunrise. A winding stone footpath leads from the bottom left toward the refuge. Jagged distant peaks, warm sunlight grazing the roof, cool blue shadows, thin luminous mist, realistic geology, restrained natural colors. Wide composition with a strong sense of scale, no people, no lettering.", "bf": "landscape_20260920_bf-bf.webp", "nv": "landscape_20260920_nv-nv.webp"}, {"name": "landscape", "seed": 20260921, "prompt": "Cinematic landscape photograph of a tiny red mountain refuge on a rocky ridge above a sea of clouds at sunrise. A winding stone footpath leads from the bottom left toward the refuge. Jagged distant peaks, warm sunlight grazing the roof, cool blue shadows, thin luminous mist, realistic geology, restrained natural colors. Wide composition with a strong sense of scale, no people, no lettering.", "bf": "landscape_20260921_bf-bf.webp", "nv": "landscape_20260921_nv-nv.webp"}, {"name": "product", "seed": 20260920, "prompt": "Premium studio product photograph of one amber glass perfume bottle with a brushed brass cap, standing on a dark green marble block. A small cream paper label on the bottle reads \"EMBER\" in elegant black capital letters. A sliced orange and a sprig of rosemary rest beside the block. Warm rim lighting, soft front fill, rich realistic materials, controlled glass reflections, clean charcoal background, centered luxury advertising composition.", "bf": "product_20260920_bf-bf.webp", "nv": "product_20260920_nv-nv.webp"}, {"name": "product", "seed": 20260921, "prompt": "Premium studio product photograph of one amber glass perfume bottle with a brushed brass cap, standing on a dark green marble block. A small cream paper label on the bottle reads \"EMBER\" in elegant black capital letters. A sliced orange and a sprig of rosemary rest beside the block. Warm rim lighting, soft front fill, rich realistic materials, controlled glass reflections, clean charcoal background, centered luxury advertising composition.", "bf": "product_20260921_bf-bf.webp", "nv": "product_20260921_nv-nv.webp"}, {"name": "typography", "seed": 20260920, "prompt": "Design a polished square travel poster with a restrained mid-century screen-print illustration of a red tram crossing a stone bridge over a blue river at sunset. Large cream headline at the top reads exactly \"TAKE THE SCENIC ROUTE\". A smaller line at the bottom reads exactly \"SLOW TRAVEL CLUB\". Strong typographic hierarchy, generous margins, flat navy, coral and cream inks, subtle paper grain. All lettering sharp and clearly legible.", "bf": "typography_20260920_bf-bf.webp", "nv": "typography_20260920_nv-nv.webp"}, {"name": "typography", "seed": 20260921, "prompt": "Design a polished square travel poster with a restrained mid-century screen-print illustration of a red tram crossing a stone bridge over a blue river at sunset. Large cream headline at the top reads exactly \"TAKE THE SCENIC ROUTE\". A smaller line at the bottom reads exactly \"SLOW TRAVEL CLUB\". Strong typographic hierarchy, generous margins, flat navy, coral and cream inks, subtle paper grain. All lettering sharp and clearly legible.", "bf": "typography_20260921_bf-bf.webp", "nv": "typography_20260921_nv-nv.webp"}, {"name": "spatial", "seed": 20260920, "prompt": "A clean studio photograph on a light gray tabletop showing exactly three objects arranged in a straight horizontal row: a red ceramic cube on the left, a blue glass sphere in the center, and a yellow wooden cone on the right. The sphere is slightly larger than the cube, and the cone is tallest. A single soft light from the upper left casts shadows toward the lower right. Neutral pale gray background, eye-level camera, every object fully visible and separated by equal gaps, no other objects, no text.", "bf": "spatial_20260920_bf-bf.webp", "nv": "spatial_20260920_nv-nv.webp"}, {"name": "spatial", "seed": 20260921, "prompt": "A clean studio photograph on a light gray tabletop showing exactly three objects arranged in a straight horizontal row: a red ceramic cube on the left, a blue glass sphere in the center, and a yellow wooden cone on the right. The sphere is slightly larger than the cube, and the cone is tallest. A single soft light from the upper left casts shadows toward the lower right. Neutral pale gray background, eye-level camera, every object fully visible and separated by equal gaps, no other objects, no text.", "bf": "spatial_20260921_bf-bf.webp", "nv": "spatial_20260921_nv-nv.webp"}, {"name": "illustration", "seed": 20260920, "prompt": "A richly detailed storybook gouache illustration of a small fox librarian wearing round spectacles and a moss-green waistcoat, standing on a rolling wooden ladder in a towering circular library inside an ancient tree. The fox reaches toward a glowing blue book with one paw. Curving shelves, warm amber lamps, tiny floating dust motes, a rainy moonlit window, hand-painted brush texture, charming expressive face, coherent architecture, dramatic but cozy composition, no lettering.", "bf": "illustration_20260920_bf-bf.webp", "nv": "illustration_20260920_nv-nv.webp"}, {"name": "illustration", "seed": 20260921, "prompt": "A richly detailed storybook gouache illustration of a small fox librarian wearing round spectacles and a moss-green waistcoat, standing on a rolling wooden ladder in a towering circular library inside an ancient tree. The fox reaches toward a glowing blue book with one paw. Curving shelves, warm amber lamps, tiny floating dust motes, a rainy moonlit window, hand-painted brush texture, charming expressive face, coherent architecture, dramatic but cozy composition, no lettering.", "bf": "illustration_20260921_bf-bf.webp", "nv": "illustration_20260921_nv-nv.webp"}, {"name": "fashion", "seed": 20260920, "prompt": "A candid editorial photograph of an adult woman with short dark curly hair and freckles, wearing a cobalt-blue tailored wool coat over a cream turtleneck, standing outside a small Parisian cafe on a rainy evening. She glances back over her shoulder with a slight natural smile, holding a folded red umbrella by its wooden handle. Fine raindrops on the coat, realistic skin pores, wet pavement reflections, warm cafe windows against cool blue dusk, waist-up framing, 85mm lens, subtle film grain, believable unretouched photography.", "bf": "fashion_20260920_bf-bf.webp", "nv": "fashion_20260920_nv-nv.webp"}, {"name": "fashion", "seed": 20260921, "prompt": "A candid editorial photograph of an adult woman with short dark curly hair and freckles, wearing a cobalt-blue tailored wool coat over a cream turtleneck, standing outside a small Parisian cafe on a rainy evening. She glances back over her shoulder with a slight natural smile, holding a folded red umbrella by its wooden handle. Fine raindrops on the coat, realistic skin pores, wet pavement reflections, warm cafe windows against cool blue dusk, waist-up framing, 85mm lens, subtle film grain, believable unretouched photography.", "bf": "fashion_20260921_bf-bf.webp", "nv": "fashion_20260921_nv-nv.webp"}, {"name": "night_market", "seed": 20260920, "prompt": "A cinematic street photograph of a crowded night market in Taipei during gentle rain. In the foreground an elderly cook in a white apron lifts a steaming bamboo basket from a street-food stall, with hands clearly visible. A small illuminated sign above him reads exactly \"NIGHT BITES\". Rich red lanterns and teal shop lights reflect in wet asphalt; layered pedestrians with umbrellas recede into the scene. Realistic steam, appetizing food texture, natural human proportions, 35mm lens, documentary photography, atmospheric but restrained color.", "bf": "night_market_20260920_bf-bf.webp", "nv": "night_market_20260920_nv-nv.webp"}, {"name": "night_market", "seed": 20260921, "prompt": "A cinematic street photograph of a crowded night market in Taipei during gentle rain. In the foreground an elderly cook in a white apron lifts a steaming bamboo basket from a street-food stall, with hands clearly visible. A small illuminated sign above him reads exactly \"NIGHT BITES\". Rich red lanterns and teal shop lights reflect in wet asphalt; layered pedestrians with umbrellas recede into the scene. Realistic steam, appetizing food texture, natural human proportions, 35mm lens, documentary photography, atmospheric but restrained color.", "bf": "night_market_20260921_bf-bf.webp", "nv": "night_market_20260921_nv-nv.webp"}, {"name": "architecture", "seed": 20260920, "prompt": "Architectural interior photograph of a sunken conversation pit in a restored 1970s coastal house. Rust-orange corduroy sofas surround a low circular walnut table holding a clear glass vase with three white tulips. Floor-to-ceiling windows reveal a stormy ocean. Rain beads on the windows; a suspended black fireplace glows softly at the left. Pale travertine floor, detailed walnut ceiling, soft overcast daylight, balanced wide-angle composition, realistic materials and straight architectural lines, no people or lettering.", "bf": "architecture_20260920_bf-bf.webp", "nv": "architecture_20260920_nv-nv.webp"}, {"name": "architecture", "seed": 20260921, "prompt": "Architectural interior photograph of a sunken conversation pit in a restored 1970s coastal house. Rust-orange corduroy sofas surround a low circular walnut table holding a clear glass vase with three white tulips. Floor-to-ceiling windows reveal a stormy ocean. Rain beads on the windows; a suspended black fireplace glows softly at the left. Pale travertine floor, detailed walnut ceiling, soft overcast daylight, balanced wide-angle composition, realistic materials and straight architectural lines, no people or lettering.", "bf": "architecture_20260921_bf-bf.webp", "nv": "architecture_20260921_nv-nv.webp"}, {"name": "macro", "seed": 20260920, "prompt": "Extreme macro nature photograph of a tiny emerald jumping spider perched on the curled edge of a copper-colored autumn leaf. A single spherical dew drop beside the spider reflects a miniature garden. Fine individual hairs on the spider, realistic eight-legged anatomy, sharp jewel-like eyes, translucent dew, intricate leaf veins. Soft warm backlight, muted olive background with circular bokeh, narrow but carefully placed focus plane, natural colors, scientifically plausible wildlife photography, no lettering.", "bf": "macro_20260920_bf-bf.webp", "nv": "macro_20260920_nv-nv.webp"}, {"name": "macro", "seed": 20260921, "prompt": "Extreme macro nature photograph of a tiny emerald jumping spider perched on the curled edge of a copper-colored autumn leaf. A single spherical dew drop beside the spider reflects a miniature garden. Fine individual hairs on the spider, realistic eight-legged anatomy, sharp jewel-like eyes, translucent dew, intricate leaf veins. Soft warm backlight, muted olive background with circular bokeh, narrow but carefully placed focus plane, natural colors, scientifically plausible wildlife photography, no lettering.", "bf": "macro_20260921_bf-bf.webp", "nv": "macro_20260921_nv-nv.webp"}, {"name": "astronaut", "seed": 20260920, "prompt": "A cinematic close-up photograph of an adult female astronaut inside a weathered lunar habitat, looking directly into the camera through a clear helmet visor. Her brown eyes, small freckles and loose strands of hair remain clearly visible behind subtle reflections. White spacesuit with stitched fabric, orange fittings and a small rectangular chest patch reading \"LUNA 09\". Warm practical light on one cheek and cool Earthlight on the other, Earth visible through a round window in the background, restrained analog science-fiction realism, finely detailed materials, no illustration style.", "bf": "astronaut_20260920_bf-bf.webp", "nv": "astronaut_20260920_nv-nv.webp"}, {"name": "astronaut", "seed": 20260921, "prompt": "A cinematic close-up photograph of an adult female astronaut inside a weathered lunar habitat, looking directly into the camera through a clear helmet visor. Her brown eyes, small freckles and loose strands of hair remain clearly visible behind subtle reflections. White spacesuit with stitched fabric, orange fittings and a small rectangular chest patch reading \"LUNA 09\". Warm practical light on one cheek and cool Earthlight on the other, Earth visible through a round window in the background, restrained analog science-fiction realism, finely detailed materials, no illustration style.", "bf": "astronaut_20260921_bf-bf.webp", "nv": "astronaut_20260921_nv-nv.webp"}, {"name": "botanical_poster", "seed": 20260920, "prompt": "An elegant contemporary botanical exhibition poster on textured ivory paper. A lifelike ink-and-watercolor illustration of one flowering magnolia branch curves diagonally across the center, pale pink petals with delicate translucent edges and deep green leaves. At the top, a large refined serif headline reads exactly \"THE QUIET GARDEN\". Beneath it, smaller text reads exactly \"BOTANICAL STUDIES\". At the bottom, a small centered line reads exactly \"APRIL 12 - MAY 30\". Generous negative space, exceptional typography, restrained burgundy and sage palette, museum-quality printed design.", "bf": "botanical_poster_20260920_bf-bf.webp", "nv": "botanical_poster_20260920_nv-nv.webp"}, {"name": "botanical_poster", "seed": 20260921, "prompt": "An elegant contemporary botanical exhibition poster on textured ivory paper. A lifelike ink-and-watercolor illustration of one flowering magnolia branch curves diagonally across the center, pale pink petals with delicate translucent edges and deep green leaves. At the top, a large refined serif headline reads exactly \"THE QUIET GARDEN\". Beneath it, smaller text reads exactly \"BOTANICAL STUDIES\". At the bottom, a small centered line reads exactly \"APRIL 12 - MAY 30\". Generous negative space, exceptional typography, restrained burgundy and sage palette, museum-quality printed design.", "bf": "botanical_poster_20260921_bf-bf.webp", "nv": "botanical_poster_20260921_nv-nv.webp"}];const select=document.getElementById('case');data.forEach((r,i)=>{let o=document.createElement('option');o.value=i;o.textContent=r.name.replaceAll('_',' ')+' / seed '+r.seed;select.append(o)});function show(){let r=data[select.value];document.getElementById('bf').src='assets/'+r.bf;document.getElementById('nv').src='assets/'+r.nv;document.getElementById('prompt').textContent=r.prompt}select.onchange=show;document.getElementById('split').oninput=e=>{let v=e.target.value;document.getElementById('bf').style.clipPath='inset(0 '+(100-v)+'% 0 0)';document.getElementById('line').style.left=v+'%'};show();</script></html>
input/qwen_image_2.1_edit_reference.png ADDED

Git LFS Details

  • SHA256: 936375ea3fb35c7c33dbe59c7090c35edf019e6119bc6b775f74f3d859b49a54
  • Pointer size: 132 Bytes
  • Size of remote file: 1.22 MB
runtime/README.md CHANGED
@@ -1,36 +1,36 @@
1
- # Native NVFP4 text conditioning
2
-
3
- The checkpoints load with existing ComfyUI nodes. On the tested ComfyUI revision,
4
- normal text conditioning runs in FP32 and dequantizes the encoder weights. The
5
- image transformer already uses native NVFP4 compute.
6
-
7
- `qwen21-nvfp4-conditioning.patch` enables native NVFP4 matrix multiplication for
8
- Qwen Image 2.1 conditioning. It detects NVFP4 checkpoint metadata, converts the
9
- completed multimodal embeddings to BF16 on supported GPUs, and uses ComfyUI's
10
- existing quantized-matmul context. The context restores the previous operation
11
- settings after encoding. Vision preprocessing and the unnormalized final hidden
12
- state contract are preserved. BF16 and INT8 checkpoints keep the original path.
13
-
14
- Tested base: `99073836d45f66053c45ba8564984e6def9cebba`.
15
- No custom nodes, monkey-patching node pack, or additional runtime dependencies.
16
-
17
- Stop ComfyUI. From its Git checkout, apply the supplied patch:
18
-
19
- ```powershell
20
- git apply --check C:/path/to/Prism-Image-2.1-NVFP4/runtime/qwen21-nvfp4-conditioning.patch
21
- git apply C:/path/to/Prism-Image-2.1-NVFP4/runtime/qwen21-nvfp4-conditioning.patch
22
- ```
23
-
24
- Restart ComfyUI. If the check fails, inspect the installed version and local
25
- changes; do not force the patch onto a different implementation. An upstream
26
- version may already provide equivalent support. To undo this exact patch, stop
27
- ComfyUI and use `git apply -R` with the same patch file.
28
-
29
- The same workflows also run without the patch, using native NVFP4 denoising and
30
- weight-only encoder quantization with FP32 conditioning. That mode is slower for
31
- uncached encoding and can produce a different image. All advertised accelerated
32
- results use the patch.
33
-
34
- Validation used an RTX 5090, CUDA 13.0 PyTorch, and comfy-kitchen 0.2.35.
35
- Other GPUs and CPU execution were not benchmarked. This is an independent patch,
36
- not an upstream ComfyUI release. The patch is subject to ComfyUI's GPL-3.0 license.
 
1
+ # Native NVFP4 text conditioning
2
+
3
+ The checkpoints load with existing ComfyUI nodes. On the tested ComfyUI revision,
4
+ normal text conditioning runs in FP32 and dequantizes the encoder weights. The
5
+ image transformer already uses native NVFP4 compute.
6
+
7
+ `qwen21-nvfp4-conditioning.patch` enables native NVFP4 matrix multiplication for
8
+ Qwen Image 2.1 conditioning. It detects NVFP4 checkpoint metadata, converts the
9
+ completed multimodal embeddings to BF16 on supported GPUs, and uses ComfyUI's
10
+ existing quantized-matmul context. The context restores the previous operation
11
+ settings after encoding. Vision preprocessing and the unnormalized final hidden
12
+ state contract are preserved. BF16 and INT8 checkpoints keep the original path.
13
+
14
+ Tested base: `99073836d45f66053c45ba8564984e6def9cebba`.
15
+ No custom nodes, monkey-patching node pack, or additional runtime dependencies.
16
+
17
+ Stop ComfyUI. From its Git checkout, apply the supplied patch:
18
+
19
+ ```powershell
20
+ git apply --check C:/path/to/Qwen-Image-2.1-NVFP4/runtime/qwen21-nvfp4-conditioning.patch
21
+ git apply C:/path/to/Qwen-Image-2.1-NVFP4/runtime/qwen21-nvfp4-conditioning.patch
22
+ ```
23
+
24
+ Restart ComfyUI. If the check fails, inspect the installed version and local
25
+ changes; do not force the patch onto a different implementation. An upstream
26
+ version may already provide equivalent support. To undo this exact patch, stop
27
+ ComfyUI and use `git apply -R` with the same patch file.
28
+
29
+ The same workflows also run without the patch, using native NVFP4 denoising and
30
+ weight-only encoder quantization with FP32 conditioning. That mode is slower for
31
+ uncached encoding and can produce a different image. All advertised accelerated
32
+ results use the patch.
33
+
34
+ Validation used an RTX 5090, CUDA 13.0 PyTorch, and comfy-kitchen 0.2.35.
35
+ Other GPUs and CPU execution were not benchmarked. This is an independent patch,
36
+ not an upstream ComfyUI release. The patch is subject to ComfyUI's GPL-3.0 license.
text_encoders/qwen3vl_8b_nvfp4.manifest.json ADDED
The diff for this file is too large to render. See raw diff
 
tools/convert.py CHANGED
@@ -1,77 +1,77 @@
1
- """Convert official BF16 checkpoints to ComfyUI's native mixed NVFP4 format."""
2
- import argparse
3
- import hashlib
4
- import json
5
- from pathlib import Path
6
- import re
7
- import sys
8
- import time
9
-
10
- import torch
11
- from safetensors import safe_open
12
- from safetensors.torch import save_file
13
-
14
-
15
- def digest(path):
16
- with open(path, "rb") as stream:
17
- return hashlib.file_digest(stream, "sha256").hexdigest()
18
-
19
-
20
- def main():
21
- parser = argparse.ArgumentParser()
22
- parser.add_argument("source", type=Path)
23
- parser.add_argument("output", type=Path)
24
- parser.add_argument("--component", choices=["dit", "encoder"], required=True)
25
- parser.add_argument("--comfy-root", type=Path, required=True)
26
- args = parser.parse_args()
27
- sys.path.insert(0, str(args.comfy_root.resolve()))
28
- # ComfyUI parses process arguments when its operation registry is imported.
29
- sys.argv = [sys.argv[0]]
30
- from comfy.quant_ops import TensorCoreNVFP4Layout
31
-
32
- patterns = {
33
- "dit": r"transformer_blocks\.\d+\.(attn\.(to_q|to_k|to_v|to_out\.0)|img_mlp\.(gate_up|out))\.weight",
34
- "encoder": r"model\.layers\.\d+\.(self_attn\.(q_proj|k_proj|v_proj|o_proj)|mlp\.(gate_proj|up_proj|down_proj))\.weight",
35
- }
36
- output, records = {}, []
37
- start = time.perf_counter()
38
- with safe_open(args.source, framework="pt", device="cpu") as source:
39
- metadata = dict(source.metadata() or {})
40
- for key in source.keys():
41
- tensor = source.get_tensor(key)
42
- record = {"name": key, "shape": list(tensor.shape), "source_dtype": str(tensor.dtype)}
43
- if re.fullmatch(patterns[args.component], key):
44
- if tensor.shape[1] % 32 or tensor.shape[0] % 8:
45
- raise ValueError(f"Unsupported NVFP4 GEMM alignment: {key} {tensor.shape}")
46
- original = tensor.to("cuda")
47
- scale = original.float().abs().amax().clamp_min(1e-12) / (448 * 6)
48
- packed, params = TensorCoreNVFP4Layout.quantize(original, scale=scale)
49
- reconstructed = TensorCoreNVFP4Layout.dequantize(packed, params)
50
- error = (original.float() - reconstructed.float()).square().mean()
51
- record.update(storage="nvfp4", relative_rmse=float((error / original.float().square().mean().clamp_min(1e-30)).sqrt()))
52
- for suffix, value in TensorCoreNVFP4Layout.state_dict_tensors(packed, params).items():
53
- output[key + suffix] = value.cpu().contiguous()
54
- marker = json.dumps({"format": "nvfp4"}, separators=(",", ":")).encode()
55
- output[key.removesuffix("weight") + "comfy_quant"] = torch.tensor(list(marker), dtype=torch.uint8)
56
- del original, reconstructed, packed, params
57
- print(f"NVFP4 {key} relative_rmse={record['relative_rmse']:.5f}", flush=True)
58
- else:
59
- output[key] = tensor.clone()
60
- record.update(storage=str(tensor.dtype), sha256=hashlib.sha256(tensor.view(torch.uint8).numpy().tobytes()).hexdigest())
61
- records.append(record)
62
- count = sum(r["storage"] == "nvfp4" for r in records)
63
- expected = 192 if args.component == "dit" else 252
64
- if count != expected:
65
- raise ValueError(f"Recipe matched {count} matrices, expected {expected}")
66
- metadata.update(format="pt", conversion="Prism native NVFP4; protected tensors unchanged; Built with Qwen")
67
- args.output.parent.mkdir(parents=True, exist_ok=True)
68
- save_file(output, str(args.output), metadata=metadata)
69
- report = {"source": args.source.name, "source_sha256": digest(args.source), "output": args.output.name,
70
- "output_sha256": digest(args.output), "output_bytes": args.output.stat().st_size,
71
- "nvfp4_matrices": count, "elapsed_seconds": time.perf_counter() - start, "tensors": records}
72
- args.output.with_suffix(".manifest.json").write_text(json.dumps(report, indent=2), encoding="utf-8")
73
- print(json.dumps({k: v for k, v in report.items() if k != "tensors"}), flush=True)
74
-
75
-
76
- if __name__ == "__main__":
77
- main()
 
1
+ """Convert official BF16 checkpoints to ComfyUI's native mixed NVFP4 format."""
2
+ import argparse
3
+ import hashlib
4
+ import json
5
+ from pathlib import Path
6
+ import re
7
+ import sys
8
+ import time
9
+
10
+ import torch
11
+ from safetensors import safe_open
12
+ from safetensors.torch import save_file
13
+
14
+
15
+ def digest(path):
16
+ with open(path, "rb") as stream:
17
+ return hashlib.file_digest(stream, "sha256").hexdigest()
18
+
19
+
20
+ def main():
21
+ parser = argparse.ArgumentParser()
22
+ parser.add_argument("source", type=Path)
23
+ parser.add_argument("output", type=Path)
24
+ parser.add_argument("--component", choices=["dit", "encoder"], required=True)
25
+ parser.add_argument("--comfy-root", type=Path, required=True)
26
+ args = parser.parse_args()
27
+ sys.path.insert(0, str(args.comfy_root.resolve()))
28
+ # ComfyUI parses process arguments when its operation registry is imported.
29
+ sys.argv = [sys.argv[0]]
30
+ from comfy.quant_ops import TensorCoreNVFP4Layout
31
+
32
+ patterns = {
33
+ "dit": r"transformer_blocks\.\d+\.(attn\.(to_q|to_k|to_v|to_out\.0)|img_mlp\.(gate_up|out))\.weight",
34
+ "encoder": r"model\.layers\.\d+\.(self_attn\.(q_proj|k_proj|v_proj|o_proj)|mlp\.(gate_proj|up_proj|down_proj))\.weight",
35
+ }
36
+ output, records = {}, []
37
+ start = time.perf_counter()
38
+ with safe_open(args.source, framework="pt", device="cpu") as source:
39
+ metadata = dict(source.metadata() or {})
40
+ for key in source.keys():
41
+ tensor = source.get_tensor(key)
42
+ record = {"name": key, "shape": list(tensor.shape), "source_dtype": str(tensor.dtype)}
43
+ if re.fullmatch(patterns[args.component], key):
44
+ if tensor.shape[1] % 32 or tensor.shape[0] % 8:
45
+ raise ValueError(f"Unsupported NVFP4 GEMM alignment: {key} {tensor.shape}")
46
+ original = tensor.to("cuda")
47
+ scale = original.float().abs().amax().clamp_min(1e-12) / (448 * 6)
48
+ packed, params = TensorCoreNVFP4Layout.quantize(original, scale=scale)
49
+ reconstructed = TensorCoreNVFP4Layout.dequantize(packed, params)
50
+ error = (original.float() - reconstructed.float()).square().mean()
51
+ record.update(storage="nvfp4", relative_rmse=float((error / original.float().square().mean().clamp_min(1e-30)).sqrt()))
52
+ for suffix, value in TensorCoreNVFP4Layout.state_dict_tensors(packed, params).items():
53
+ output[key + suffix] = value.cpu().contiguous()
54
+ marker = json.dumps({"format": "nvfp4"}, separators=(",", ":")).encode()
55
+ output[key.removesuffix("weight") + "comfy_quant"] = torch.tensor(list(marker), dtype=torch.uint8)
56
+ del original, reconstructed, packed, params
57
+ print(f"NVFP4 {key} relative_rmse={record['relative_rmse']:.5f}", flush=True)
58
+ else:
59
+ output[key] = tensor.clone()
60
+ record.update(storage=str(tensor.dtype), sha256=hashlib.sha256(tensor.view(torch.uint8).numpy().tobytes()).hexdigest())
61
+ records.append(record)
62
+ count = sum(r["storage"] == "nvfp4" for r in records)
63
+ expected = 192 if args.component == "dit" else 252
64
+ if count != expected:
65
+ raise ValueError(f"Recipe matched {count} matrices, expected {expected}")
66
+ metadata.update(format="pt", conversion="Prism native NVFP4; protected tensors unchanged; Built with Qwen")
67
+ args.output.parent.mkdir(parents=True, exist_ok=True)
68
+ save_file(output, str(args.output), metadata=metadata)
69
+ report = {"source": args.source.name, "source_sha256": digest(args.source), "output": args.output.name,
70
+ "output_sha256": digest(args.output), "output_bytes": args.output.stat().st_size,
71
+ "nvfp4_matrices": count, "elapsed_seconds": time.perf_counter() - start, "tensors": records}
72
+ args.output.with_suffix(".manifest.json").write_text(json.dumps(report, indent=2), encoding="utf-8")
73
+ print(json.dumps({k: v for k, v in report.items() if k != "tensors"}), flush=True)
74
+
75
+
76
+ if __name__ == "__main__":
77
+ main()
tools/run_workflow.py CHANGED
@@ -1,26 +1,26 @@
1
- """Run an included API workflow against a local ComfyUI server."""
2
- import argparse
3
- import json
4
- from pathlib import Path
5
- import time
6
- from urllib.request import Request, urlopen
7
-
8
- parser = argparse.ArgumentParser()
9
- parser.add_argument('workflow', type=Path)
10
- parser.add_argument('--server', default='http://127.0.0.1:8188')
11
- args = parser.parse_args()
12
- graph = json.loads(args.workflow.read_text(encoding='utf-8'))
13
- request = Request(args.server + '/prompt', data=json.dumps({'prompt':graph}).encode(), headers={'Content-Type':'application/json'})
14
- with urlopen(request, timeout=30) as response:
15
- prompt_id = json.load(response)['prompt_id']
16
- print('Queued:',prompt_id,flush=True)
17
- while True:
18
- with urlopen(args.server + '/history/' + prompt_id, timeout=30) as response:
19
- history = json.load(response).get(prompt_id)
20
- if history:
21
- print(json.dumps(history['status'],indent=2))
22
- print(json.dumps(history['outputs'],indent=2))
23
- if history['status']['status_str'] != 'success':
24
- raise SystemExit(1)
25
- break
26
- time.sleep(0.5)
 
1
+ """Run an included API workflow against a local ComfyUI server."""
2
+ import argparse
3
+ import json
4
+ from pathlib import Path
5
+ import time
6
+ from urllib.request import Request, urlopen
7
+
8
+ parser = argparse.ArgumentParser()
9
+ parser.add_argument('workflow', type=Path)
10
+ parser.add_argument('--server', default='http://127.0.0.1:8188')
11
+ args = parser.parse_args()
12
+ graph = json.loads(args.workflow.read_text(encoding='utf-8'))
13
+ request = Request(args.server + '/prompt', data=json.dumps({'prompt':graph}).encode(), headers={'Content-Type':'application/json'})
14
+ with urlopen(request, timeout=30) as response:
15
+ prompt_id = json.load(response)['prompt_id']
16
+ print('Queued:',prompt_id,flush=True)
17
+ while True:
18
+ with urlopen(args.server + '/history/' + prompt_id, timeout=30) as response:
19
+ history = json.load(response).get(prompt_id)
20
+ if history:
21
+ print(json.dumps(history['status'],indent=2))
22
+ print(json.dumps(history['outputs'],indent=2))
23
+ if history['status']['status_str'] != 'success':
24
+ raise SystemExit(1)
25
+ break
26
+ time.sleep(0.5)
tools/verify.py CHANGED
@@ -1,36 +1,36 @@
1
- """Verify saved native NVFP4 shapes, metadata, scales, and protected tensors."""
2
- import argparse
3
- import hashlib
4
- import json
5
- from pathlib import Path
6
- import torch
7
- from safetensors import safe_open
8
-
9
- parser = argparse.ArgumentParser()
10
- parser.add_argument('source', type=Path)
11
- parser.add_argument('converted', type=Path)
12
- args = parser.parse_args()
13
- manifest = json.loads(args.converted.with_suffix('.manifest.json').read_text())
14
- with args.converted.open('rb') as stream:
15
- assert hashlib.file_digest(stream, 'sha256').hexdigest() == manifest['output_sha256']
16
- with safe_open(args.source, framework='pt') as original, safe_open(args.converted, framework='pt') as result:
17
- expected_keys = set(original.keys())
18
- for record in manifest['tensors']:
19
- key = record['name']
20
- value = result.get_tensor(key)
21
- if record['storage'] == 'nvfp4':
22
- marker_key = key.removesuffix('weight') + 'comfy_quant'
23
- assert json.loads(result.get_tensor(marker_key).numpy().tobytes()) == {'format':'nvfp4'}
24
- assert value.dtype == torch.uint8
25
- assert list(value.shape) == [record['shape'][0], record['shape'][1] // 2]
26
- block = result.get_tensor(key + '_scale')
27
- scale = result.get_tensor(key + '_scale_2')
28
- assert block.dtype == torch.float8_e4m3fn and scale.dtype == torch.float32
29
- assert bool(torch.isfinite(block.float()).all()) and bool(torch.isfinite(scale).all())
30
- assert bool((scale > 0).all())
31
- expected_keys.update([key + '_scale', key + '_scale_2', marker_key])
32
- else:
33
- source = original.get_tensor(key)
34
- assert source.dtype == value.dtype and torch.equal(source, value), key
35
- assert set(result.keys()) == expected_keys
36
- print(json.dumps({'file':args.converted.name, 'verified_tensors':len(manifest['tensors']), 'nvfp4_matrices':manifest['nvfp4_matrices'], 'protected_exact':True, 'sha256':manifest['output_sha256']}))
 
1
+ """Verify saved native NVFP4 shapes, metadata, scales, and protected tensors."""
2
+ import argparse
3
+ import hashlib
4
+ import json
5
+ from pathlib import Path
6
+ import torch
7
+ from safetensors import safe_open
8
+
9
+ parser = argparse.ArgumentParser()
10
+ parser.add_argument('source', type=Path)
11
+ parser.add_argument('converted', type=Path)
12
+ args = parser.parse_args()
13
+ manifest = json.loads(args.converted.with_suffix('.manifest.json').read_text())
14
+ with args.converted.open('rb') as stream:
15
+ assert hashlib.file_digest(stream, 'sha256').hexdigest() == manifest['output_sha256']
16
+ with safe_open(args.source, framework='pt') as original, safe_open(args.converted, framework='pt') as result:
17
+ expected_keys = set(original.keys())
18
+ for record in manifest['tensors']:
19
+ key = record['name']
20
+ value = result.get_tensor(key)
21
+ if record['storage'] == 'nvfp4':
22
+ marker_key = key.removesuffix('weight') + 'comfy_quant'
23
+ assert json.loads(result.get_tensor(marker_key).numpy().tobytes()) == {'format':'nvfp4'}
24
+ assert value.dtype == torch.uint8
25
+ assert list(value.shape) == [record['shape'][0], record['shape'][1] // 2]
26
+ block = result.get_tensor(key + '_scale')
27
+ scale = result.get_tensor(key + '_scale_2')
28
+ assert block.dtype == torch.float8_e4m3fn and scale.dtype == torch.float32
29
+ assert bool(torch.isfinite(block.float()).all()) and bool(torch.isfinite(scale).all())
30
+ assert bool((scale > 0).all())
31
+ expected_keys.update([key + '_scale', key + '_scale_2', marker_key])
32
+ else:
33
+ source = original.get_tensor(key)
34
+ assert source.dtype == value.dtype and torch.equal(source, value), key
35
+ assert set(result.keys()) == expected_keys
36
+ print(json.dumps({'file':args.converted.name, 'verified_tensors':len(manifest['tensors']), 'nvfp4_matrices':manifest['nvfp4_matrices'], 'protected_exact':True, 'sha256':manifest['output_sha256']}))
validation/cases.json CHANGED
@@ -227,14 +227,14 @@
227
  "1": {
228
  "class_type": "UNETLoader",
229
  "inputs": {
230
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
231
  "weight_dtype": "default"
232
  }
233
  },
234
  "2": {
235
  "class_type": "CLIPLoader",
236
  "inputs": {
237
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
238
  "type": "qwen_image",
239
  "device": "default"
240
  }
@@ -337,14 +337,14 @@
337
  "1": {
338
  "class_type": "UNETLoader",
339
  "inputs": {
340
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
341
  "weight_dtype": "default"
342
  }
343
  },
344
  "2": {
345
  "class_type": "CLIPLoader",
346
  "inputs": {
347
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
348
  "type": "qwen_image",
349
  "device": "default"
350
  }
@@ -667,14 +667,14 @@
667
  "1": {
668
  "class_type": "UNETLoader",
669
  "inputs": {
670
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
671
  "weight_dtype": "default"
672
  }
673
  },
674
  "2": {
675
  "class_type": "CLIPLoader",
676
  "inputs": {
677
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
678
  "type": "qwen_image",
679
  "device": "default"
680
  }
@@ -777,14 +777,14 @@
777
  "1": {
778
  "class_type": "UNETLoader",
779
  "inputs": {
780
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
781
  "weight_dtype": "default"
782
  }
783
  },
784
  "2": {
785
  "class_type": "CLIPLoader",
786
  "inputs": {
787
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
788
  "type": "qwen_image",
789
  "device": "default"
790
  }
@@ -1107,14 +1107,14 @@
1107
  "1": {
1108
  "class_type": "UNETLoader",
1109
  "inputs": {
1110
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
1111
  "weight_dtype": "default"
1112
  }
1113
  },
1114
  "2": {
1115
  "class_type": "CLIPLoader",
1116
  "inputs": {
1117
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
1118
  "type": "qwen_image",
1119
  "device": "default"
1120
  }
@@ -1217,14 +1217,14 @@
1217
  "1": {
1218
  "class_type": "UNETLoader",
1219
  "inputs": {
1220
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
1221
  "weight_dtype": "default"
1222
  }
1223
  },
1224
  "2": {
1225
  "class_type": "CLIPLoader",
1226
  "inputs": {
1227
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
1228
  "type": "qwen_image",
1229
  "device": "default"
1230
  }
@@ -1443,14 +1443,14 @@
1443
  "1": {
1444
  "class_type": "UNETLoader",
1445
  "inputs": {
1446
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
1447
  "weight_dtype": "default"
1448
  }
1449
  },
1450
  "2": {
1451
  "class_type": "CLIPLoader",
1452
  "inputs": {
1453
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
1454
  "type": "qwen_image",
1455
  "device": "default"
1456
  }
@@ -1675,14 +1675,14 @@
1675
  "1": {
1676
  "class_type": "UNETLoader",
1677
  "inputs": {
1678
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
1679
  "weight_dtype": "default"
1680
  }
1681
  },
1682
  "2": {
1683
  "class_type": "CLIPLoader",
1684
  "inputs": {
1685
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
1686
  "type": "qwen_image",
1687
  "device": "default"
1688
  }
@@ -1791,14 +1791,14 @@
1791
  "1": {
1792
  "class_type": "UNETLoader",
1793
  "inputs": {
1794
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
1795
  "weight_dtype": "default"
1796
  }
1797
  },
1798
  "2": {
1799
  "class_type": "CLIPLoader",
1800
  "inputs": {
1801
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
1802
  "type": "qwen_image",
1803
  "device": "default"
1804
  }
@@ -2127,14 +2127,14 @@
2127
  "1": {
2128
  "class_type": "UNETLoader",
2129
  "inputs": {
2130
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
2131
  "weight_dtype": "default"
2132
  }
2133
  },
2134
  "2": {
2135
  "class_type": "CLIPLoader",
2136
  "inputs": {
2137
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
2138
  "type": "qwen_image",
2139
  "device": "default"
2140
  }
@@ -2237,14 +2237,14 @@
2237
  "1": {
2238
  "class_type": "UNETLoader",
2239
  "inputs": {
2240
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
2241
  "weight_dtype": "default"
2242
  }
2243
  },
2244
  "2": {
2245
  "class_type": "CLIPLoader",
2246
  "inputs": {
2247
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
2248
  "type": "qwen_image",
2249
  "device": "default"
2250
  }
@@ -2457,14 +2457,14 @@
2457
  "1": {
2458
  "class_type": "UNETLoader",
2459
  "inputs": {
2460
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
2461
  "weight_dtype": "default"
2462
  }
2463
  },
2464
  "2": {
2465
  "class_type": "CLIPLoader",
2466
  "inputs": {
2467
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
2468
  "type": "qwen_image",
2469
  "device": "default"
2470
  }
@@ -2693,14 +2693,14 @@
2693
  "1": {
2694
  "class_type": "UNETLoader",
2695
  "inputs": {
2696
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
2697
  "weight_dtype": "default"
2698
  }
2699
  },
2700
  "2": {
2701
  "class_type": "CLIPLoader",
2702
  "inputs": {
2703
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
2704
  "type": "qwen_image",
2705
  "device": "default"
2706
  }
@@ -2929,14 +2929,14 @@
2929
  "1": {
2930
  "class_type": "UNETLoader",
2931
  "inputs": {
2932
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
2933
  "weight_dtype": "default"
2934
  }
2935
  },
2936
  "2": {
2937
  "class_type": "CLIPLoader",
2938
  "inputs": {
2939
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
2940
  "type": "qwen_image",
2941
  "device": "default"
2942
  }
@@ -3259,14 +3259,14 @@
3259
  "1": {
3260
  "class_type": "UNETLoader",
3261
  "inputs": {
3262
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
3263
  "weight_dtype": "default"
3264
  }
3265
  },
3266
  "2": {
3267
  "class_type": "CLIPLoader",
3268
  "inputs": {
3269
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
3270
  "type": "qwen_image",
3271
  "device": "default"
3272
  }
@@ -3369,14 +3369,14 @@
3369
  "1": {
3370
  "class_type": "UNETLoader",
3371
  "inputs": {
3372
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
3373
  "weight_dtype": "default"
3374
  }
3375
  },
3376
  "2": {
3377
  "class_type": "CLIPLoader",
3378
  "inputs": {
3379
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
3380
  "type": "qwen_image",
3381
  "device": "default"
3382
  }
@@ -3699,14 +3699,14 @@
3699
  "1": {
3700
  "class_type": "UNETLoader",
3701
  "inputs": {
3702
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
3703
  "weight_dtype": "default"
3704
  }
3705
  },
3706
  "2": {
3707
  "class_type": "CLIPLoader",
3708
  "inputs": {
3709
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
3710
  "type": "qwen_image",
3711
  "device": "default"
3712
  }
@@ -3809,14 +3809,14 @@
3809
  "1": {
3810
  "class_type": "UNETLoader",
3811
  "inputs": {
3812
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
3813
  "weight_dtype": "default"
3814
  }
3815
  },
3816
  "2": {
3817
  "class_type": "CLIPLoader",
3818
  "inputs": {
3819
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
3820
  "type": "qwen_image",
3821
  "device": "default"
3822
  }
@@ -4139,14 +4139,14 @@
4139
  "1": {
4140
  "class_type": "UNETLoader",
4141
  "inputs": {
4142
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
4143
  "weight_dtype": "default"
4144
  }
4145
  },
4146
  "2": {
4147
  "class_type": "CLIPLoader",
4148
  "inputs": {
4149
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
4150
  "type": "qwen_image",
4151
  "device": "default"
4152
  }
@@ -4249,14 +4249,14 @@
4249
  "1": {
4250
  "class_type": "UNETLoader",
4251
  "inputs": {
4252
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
4253
  "weight_dtype": "default"
4254
  }
4255
  },
4256
  "2": {
4257
  "class_type": "CLIPLoader",
4258
  "inputs": {
4259
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
4260
  "type": "qwen_image",
4261
  "device": "default"
4262
  }
@@ -4579,14 +4579,14 @@
4579
  "1": {
4580
  "class_type": "UNETLoader",
4581
  "inputs": {
4582
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
4583
  "weight_dtype": "default"
4584
  }
4585
  },
4586
  "2": {
4587
  "class_type": "CLIPLoader",
4588
  "inputs": {
4589
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
4590
  "type": "qwen_image",
4591
  "device": "default"
4592
  }
@@ -4689,14 +4689,14 @@
4689
  "1": {
4690
  "class_type": "UNETLoader",
4691
  "inputs": {
4692
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
4693
  "weight_dtype": "default"
4694
  }
4695
  },
4696
  "2": {
4697
  "class_type": "CLIPLoader",
4698
  "inputs": {
4699
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
4700
  "type": "qwen_image",
4701
  "device": "default"
4702
  }
@@ -5136,7 +5136,7 @@
5136
  "2": {
5137
  "class_type": "CLIPLoader",
5138
  "inputs": {
5139
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
5140
  "type": "qwen_image",
5141
  "device": "default"
5142
  }
@@ -5239,7 +5239,7 @@
5239
  "1": {
5240
  "class_type": "UNETLoader",
5241
  "inputs": {
5242
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
5243
  "weight_dtype": "default"
5244
  }
5245
  },
@@ -5349,14 +5349,14 @@
5349
  "1": {
5350
  "class_type": "UNETLoader",
5351
  "inputs": {
5352
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
5353
  "weight_dtype": "default"
5354
  }
5355
  },
5356
  "2": {
5357
  "class_type": "CLIPLoader",
5358
  "inputs": {
5359
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
5360
  "type": "qwen_image",
5361
  "device": "default"
5362
  }
@@ -5459,14 +5459,14 @@
5459
  "1": {
5460
  "class_type": "UNETLoader",
5461
  "inputs": {
5462
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
5463
  "weight_dtype": "default"
5464
  }
5465
  },
5466
  "2": {
5467
  "class_type": "CLIPLoader",
5468
  "inputs": {
5469
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
5470
  "type": "qwen_image",
5471
  "device": "default"
5472
  }
@@ -5569,14 +5569,14 @@
5569
  "1": {
5570
  "class_type": "UNETLoader",
5571
  "inputs": {
5572
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
5573
  "weight_dtype": "default"
5574
  }
5575
  },
5576
  "2": {
5577
  "class_type": "CLIPLoader",
5578
  "inputs": {
5579
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
5580
  "type": "qwen_image",
5581
  "device": "default"
5582
  }
@@ -5679,14 +5679,14 @@
5679
  "1": {
5680
  "class_type": "UNETLoader",
5681
  "inputs": {
5682
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
5683
  "weight_dtype": "default"
5684
  }
5685
  },
5686
  "2": {
5687
  "class_type": "CLIPLoader",
5688
  "inputs": {
5689
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
5690
  "type": "qwen_image",
5691
  "device": "default"
5692
  }
@@ -6126,7 +6126,7 @@
6126
  "2": {
6127
  "class_type": "CLIPLoader",
6128
  "inputs": {
6129
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
6130
  "type": "qwen_image",
6131
  "device": "default"
6132
  }
@@ -6229,7 +6229,7 @@
6229
  "1": {
6230
  "class_type": "UNETLoader",
6231
  "inputs": {
6232
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6233
  "weight_dtype": "default"
6234
  }
6235
  },
@@ -6339,14 +6339,14 @@
6339
  "1": {
6340
  "class_type": "UNETLoader",
6341
  "inputs": {
6342
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6343
  "weight_dtype": "default"
6344
  }
6345
  },
6346
  "2": {
6347
  "class_type": "CLIPLoader",
6348
  "inputs": {
6349
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
6350
  "type": "qwen_image",
6351
  "device": "default"
6352
  }
@@ -6449,14 +6449,14 @@
6449
  "1": {
6450
  "class_type": "UNETLoader",
6451
  "inputs": {
6452
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6453
  "weight_dtype": "default"
6454
  }
6455
  },
6456
  "2": {
6457
  "class_type": "CLIPLoader",
6458
  "inputs": {
6459
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
6460
  "type": "qwen_image",
6461
  "device": "default"
6462
  }
@@ -6559,14 +6559,14 @@
6559
  "1": {
6560
  "class_type": "UNETLoader",
6561
  "inputs": {
6562
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6563
  "weight_dtype": "default"
6564
  }
6565
  },
6566
  "2": {
6567
  "class_type": "CLIPLoader",
6568
  "inputs": {
6569
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
6570
  "type": "qwen_image",
6571
  "device": "default"
6572
  }
@@ -6669,14 +6669,14 @@
6669
  "1": {
6670
  "class_type": "UNETLoader",
6671
  "inputs": {
6672
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6673
  "weight_dtype": "default"
6674
  }
6675
  },
6676
  "2": {
6677
  "class_type": "CLIPLoader",
6678
  "inputs": {
6679
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
6680
  "type": "qwen_image",
6681
  "device": "default"
6682
  }
@@ -6999,14 +6999,14 @@
6999
  "1": {
7000
  "class_type": "UNETLoader",
7001
  "inputs": {
7002
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
7003
  "weight_dtype": "default"
7004
  }
7005
  },
7006
  "2": {
7007
  "class_type": "CLIPLoader",
7008
  "inputs": {
7009
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
7010
  "type": "qwen_image",
7011
  "device": "default"
7012
  }
@@ -7109,14 +7109,14 @@
7109
  "1": {
7110
  "class_type": "UNETLoader",
7111
  "inputs": {
7112
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
7113
  "weight_dtype": "default"
7114
  }
7115
  },
7116
  "2": {
7117
  "class_type": "CLIPLoader",
7118
  "inputs": {
7119
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
7120
  "type": "qwen_image",
7121
  "device": "default"
7122
  }
@@ -7329,14 +7329,14 @@
7329
  "1": {
7330
  "class_type": "UNETLoader",
7331
  "inputs": {
7332
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
7333
  "weight_dtype": "default"
7334
  }
7335
  },
7336
  "2": {
7337
  "class_type": "CLIPLoader",
7338
  "inputs": {
7339
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
7340
  "type": "qwen_image",
7341
  "device": "default"
7342
  }
@@ -7439,14 +7439,14 @@
7439
  "1": {
7440
  "class_type": "UNETLoader",
7441
  "inputs": {
7442
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
7443
  "weight_dtype": "default"
7444
  }
7445
  },
7446
  "2": {
7447
  "class_type": "CLIPLoader",
7448
  "inputs": {
7449
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
7450
  "type": "qwen_image",
7451
  "device": "default"
7452
  }
@@ -7659,14 +7659,14 @@
7659
  "1": {
7660
  "class_type": "UNETLoader",
7661
  "inputs": {
7662
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
7663
  "weight_dtype": "default"
7664
  }
7665
  },
7666
  "2": {
7667
  "class_type": "CLIPLoader",
7668
  "inputs": {
7669
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
7670
  "type": "qwen_image",
7671
  "device": "default"
7672
  }
@@ -8106,7 +8106,7 @@
8106
  "2": {
8107
  "class_type": "CLIPLoader",
8108
  "inputs": {
8109
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
8110
  "type": "qwen_image",
8111
  "device": "default"
8112
  }
@@ -8209,7 +8209,7 @@
8209
  "1": {
8210
  "class_type": "UNETLoader",
8211
  "inputs": {
8212
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
8213
  "weight_dtype": "default"
8214
  }
8215
  },
@@ -8319,14 +8319,14 @@
8319
  "1": {
8320
  "class_type": "UNETLoader",
8321
  "inputs": {
8322
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
8323
  "weight_dtype": "default"
8324
  }
8325
  },
8326
  "2": {
8327
  "class_type": "CLIPLoader",
8328
  "inputs": {
8329
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
8330
  "type": "qwen_image",
8331
  "device": "default"
8332
  }
@@ -8429,14 +8429,14 @@
8429
  "1": {
8430
  "class_type": "UNETLoader",
8431
  "inputs": {
8432
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
8433
  "weight_dtype": "default"
8434
  }
8435
  },
8436
  "2": {
8437
  "class_type": "CLIPLoader",
8438
  "inputs": {
8439
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
8440
  "type": "qwen_image",
8441
  "device": "default"
8442
  }
@@ -8539,14 +8539,14 @@
8539
  "1": {
8540
  "class_type": "UNETLoader",
8541
  "inputs": {
8542
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
8543
  "weight_dtype": "default"
8544
  }
8545
  },
8546
  "2": {
8547
  "class_type": "CLIPLoader",
8548
  "inputs": {
8549
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
8550
  "type": "qwen_image",
8551
  "device": "default"
8552
  }
@@ -8649,14 +8649,14 @@
8649
  "1": {
8650
  "class_type": "UNETLoader",
8651
  "inputs": {
8652
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
8653
  "weight_dtype": "default"
8654
  }
8655
  },
8656
  "2": {
8657
  "class_type": "CLIPLoader",
8658
  "inputs": {
8659
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
8660
  "type": "qwen_image",
8661
  "device": "default"
8662
  }
 
227
  "1": {
228
  "class_type": "UNETLoader",
229
  "inputs": {
230
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
231
  "weight_dtype": "default"
232
  }
233
  },
234
  "2": {
235
  "class_type": "CLIPLoader",
236
  "inputs": {
237
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
238
  "type": "qwen_image",
239
  "device": "default"
240
  }
 
337
  "1": {
338
  "class_type": "UNETLoader",
339
  "inputs": {
340
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
341
  "weight_dtype": "default"
342
  }
343
  },
344
  "2": {
345
  "class_type": "CLIPLoader",
346
  "inputs": {
347
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
348
  "type": "qwen_image",
349
  "device": "default"
350
  }
 
667
  "1": {
668
  "class_type": "UNETLoader",
669
  "inputs": {
670
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
671
  "weight_dtype": "default"
672
  }
673
  },
674
  "2": {
675
  "class_type": "CLIPLoader",
676
  "inputs": {
677
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
678
  "type": "qwen_image",
679
  "device": "default"
680
  }
 
777
  "1": {
778
  "class_type": "UNETLoader",
779
  "inputs": {
780
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
781
  "weight_dtype": "default"
782
  }
783
  },
784
  "2": {
785
  "class_type": "CLIPLoader",
786
  "inputs": {
787
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
788
  "type": "qwen_image",
789
  "device": "default"
790
  }
 
1107
  "1": {
1108
  "class_type": "UNETLoader",
1109
  "inputs": {
1110
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
1111
  "weight_dtype": "default"
1112
  }
1113
  },
1114
  "2": {
1115
  "class_type": "CLIPLoader",
1116
  "inputs": {
1117
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
1118
  "type": "qwen_image",
1119
  "device": "default"
1120
  }
 
1217
  "1": {
1218
  "class_type": "UNETLoader",
1219
  "inputs": {
1220
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
1221
  "weight_dtype": "default"
1222
  }
1223
  },
1224
  "2": {
1225
  "class_type": "CLIPLoader",
1226
  "inputs": {
1227
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
1228
  "type": "qwen_image",
1229
  "device": "default"
1230
  }
 
1443
  "1": {
1444
  "class_type": "UNETLoader",
1445
  "inputs": {
1446
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
1447
  "weight_dtype": "default"
1448
  }
1449
  },
1450
  "2": {
1451
  "class_type": "CLIPLoader",
1452
  "inputs": {
1453
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
1454
  "type": "qwen_image",
1455
  "device": "default"
1456
  }
 
1675
  "1": {
1676
  "class_type": "UNETLoader",
1677
  "inputs": {
1678
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
1679
  "weight_dtype": "default"
1680
  }
1681
  },
1682
  "2": {
1683
  "class_type": "CLIPLoader",
1684
  "inputs": {
1685
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
1686
  "type": "qwen_image",
1687
  "device": "default"
1688
  }
 
1791
  "1": {
1792
  "class_type": "UNETLoader",
1793
  "inputs": {
1794
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
1795
  "weight_dtype": "default"
1796
  }
1797
  },
1798
  "2": {
1799
  "class_type": "CLIPLoader",
1800
  "inputs": {
1801
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
1802
  "type": "qwen_image",
1803
  "device": "default"
1804
  }
 
2127
  "1": {
2128
  "class_type": "UNETLoader",
2129
  "inputs": {
2130
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
2131
  "weight_dtype": "default"
2132
  }
2133
  },
2134
  "2": {
2135
  "class_type": "CLIPLoader",
2136
  "inputs": {
2137
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
2138
  "type": "qwen_image",
2139
  "device": "default"
2140
  }
 
2237
  "1": {
2238
  "class_type": "UNETLoader",
2239
  "inputs": {
2240
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
2241
  "weight_dtype": "default"
2242
  }
2243
  },
2244
  "2": {
2245
  "class_type": "CLIPLoader",
2246
  "inputs": {
2247
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
2248
  "type": "qwen_image",
2249
  "device": "default"
2250
  }
 
2457
  "1": {
2458
  "class_type": "UNETLoader",
2459
  "inputs": {
2460
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
2461
  "weight_dtype": "default"
2462
  }
2463
  },
2464
  "2": {
2465
  "class_type": "CLIPLoader",
2466
  "inputs": {
2467
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
2468
  "type": "qwen_image",
2469
  "device": "default"
2470
  }
 
2693
  "1": {
2694
  "class_type": "UNETLoader",
2695
  "inputs": {
2696
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
2697
  "weight_dtype": "default"
2698
  }
2699
  },
2700
  "2": {
2701
  "class_type": "CLIPLoader",
2702
  "inputs": {
2703
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
2704
  "type": "qwen_image",
2705
  "device": "default"
2706
  }
 
2929
  "1": {
2930
  "class_type": "UNETLoader",
2931
  "inputs": {
2932
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
2933
  "weight_dtype": "default"
2934
  }
2935
  },
2936
  "2": {
2937
  "class_type": "CLIPLoader",
2938
  "inputs": {
2939
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
2940
  "type": "qwen_image",
2941
  "device": "default"
2942
  }
 
3259
  "1": {
3260
  "class_type": "UNETLoader",
3261
  "inputs": {
3262
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
3263
  "weight_dtype": "default"
3264
  }
3265
  },
3266
  "2": {
3267
  "class_type": "CLIPLoader",
3268
  "inputs": {
3269
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
3270
  "type": "qwen_image",
3271
  "device": "default"
3272
  }
 
3369
  "1": {
3370
  "class_type": "UNETLoader",
3371
  "inputs": {
3372
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
3373
  "weight_dtype": "default"
3374
  }
3375
  },
3376
  "2": {
3377
  "class_type": "CLIPLoader",
3378
  "inputs": {
3379
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
3380
  "type": "qwen_image",
3381
  "device": "default"
3382
  }
 
3699
  "1": {
3700
  "class_type": "UNETLoader",
3701
  "inputs": {
3702
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
3703
  "weight_dtype": "default"
3704
  }
3705
  },
3706
  "2": {
3707
  "class_type": "CLIPLoader",
3708
  "inputs": {
3709
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
3710
  "type": "qwen_image",
3711
  "device": "default"
3712
  }
 
3809
  "1": {
3810
  "class_type": "UNETLoader",
3811
  "inputs": {
3812
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
3813
  "weight_dtype": "default"
3814
  }
3815
  },
3816
  "2": {
3817
  "class_type": "CLIPLoader",
3818
  "inputs": {
3819
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
3820
  "type": "qwen_image",
3821
  "device": "default"
3822
  }
 
4139
  "1": {
4140
  "class_type": "UNETLoader",
4141
  "inputs": {
4142
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
4143
  "weight_dtype": "default"
4144
  }
4145
  },
4146
  "2": {
4147
  "class_type": "CLIPLoader",
4148
  "inputs": {
4149
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
4150
  "type": "qwen_image",
4151
  "device": "default"
4152
  }
 
4249
  "1": {
4250
  "class_type": "UNETLoader",
4251
  "inputs": {
4252
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
4253
  "weight_dtype": "default"
4254
  }
4255
  },
4256
  "2": {
4257
  "class_type": "CLIPLoader",
4258
  "inputs": {
4259
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
4260
  "type": "qwen_image",
4261
  "device": "default"
4262
  }
 
4579
  "1": {
4580
  "class_type": "UNETLoader",
4581
  "inputs": {
4582
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
4583
  "weight_dtype": "default"
4584
  }
4585
  },
4586
  "2": {
4587
  "class_type": "CLIPLoader",
4588
  "inputs": {
4589
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
4590
  "type": "qwen_image",
4591
  "device": "default"
4592
  }
 
4689
  "1": {
4690
  "class_type": "UNETLoader",
4691
  "inputs": {
4692
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
4693
  "weight_dtype": "default"
4694
  }
4695
  },
4696
  "2": {
4697
  "class_type": "CLIPLoader",
4698
  "inputs": {
4699
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
4700
  "type": "qwen_image",
4701
  "device": "default"
4702
  }
 
5136
  "2": {
5137
  "class_type": "CLIPLoader",
5138
  "inputs": {
5139
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
5140
  "type": "qwen_image",
5141
  "device": "default"
5142
  }
 
5239
  "1": {
5240
  "class_type": "UNETLoader",
5241
  "inputs": {
5242
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
5243
  "weight_dtype": "default"
5244
  }
5245
  },
 
5349
  "1": {
5350
  "class_type": "UNETLoader",
5351
  "inputs": {
5352
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
5353
  "weight_dtype": "default"
5354
  }
5355
  },
5356
  "2": {
5357
  "class_type": "CLIPLoader",
5358
  "inputs": {
5359
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
5360
  "type": "qwen_image",
5361
  "device": "default"
5362
  }
 
5459
  "1": {
5460
  "class_type": "UNETLoader",
5461
  "inputs": {
5462
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
5463
  "weight_dtype": "default"
5464
  }
5465
  },
5466
  "2": {
5467
  "class_type": "CLIPLoader",
5468
  "inputs": {
5469
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
5470
  "type": "qwen_image",
5471
  "device": "default"
5472
  }
 
5569
  "1": {
5570
  "class_type": "UNETLoader",
5571
  "inputs": {
5572
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
5573
  "weight_dtype": "default"
5574
  }
5575
  },
5576
  "2": {
5577
  "class_type": "CLIPLoader",
5578
  "inputs": {
5579
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
5580
  "type": "qwen_image",
5581
  "device": "default"
5582
  }
 
5679
  "1": {
5680
  "class_type": "UNETLoader",
5681
  "inputs": {
5682
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
5683
  "weight_dtype": "default"
5684
  }
5685
  },
5686
  "2": {
5687
  "class_type": "CLIPLoader",
5688
  "inputs": {
5689
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
5690
  "type": "qwen_image",
5691
  "device": "default"
5692
  }
 
6126
  "2": {
6127
  "class_type": "CLIPLoader",
6128
  "inputs": {
6129
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
6130
  "type": "qwen_image",
6131
  "device": "default"
6132
  }
 
6229
  "1": {
6230
  "class_type": "UNETLoader",
6231
  "inputs": {
6232
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6233
  "weight_dtype": "default"
6234
  }
6235
  },
 
6339
  "1": {
6340
  "class_type": "UNETLoader",
6341
  "inputs": {
6342
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6343
  "weight_dtype": "default"
6344
  }
6345
  },
6346
  "2": {
6347
  "class_type": "CLIPLoader",
6348
  "inputs": {
6349
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
6350
  "type": "qwen_image",
6351
  "device": "default"
6352
  }
 
6449
  "1": {
6450
  "class_type": "UNETLoader",
6451
  "inputs": {
6452
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6453
  "weight_dtype": "default"
6454
  }
6455
  },
6456
  "2": {
6457
  "class_type": "CLIPLoader",
6458
  "inputs": {
6459
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
6460
  "type": "qwen_image",
6461
  "device": "default"
6462
  }
 
6559
  "1": {
6560
  "class_type": "UNETLoader",
6561
  "inputs": {
6562
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6563
  "weight_dtype": "default"
6564
  }
6565
  },
6566
  "2": {
6567
  "class_type": "CLIPLoader",
6568
  "inputs": {
6569
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
6570
  "type": "qwen_image",
6571
  "device": "default"
6572
  }
 
6669
  "1": {
6670
  "class_type": "UNETLoader",
6671
  "inputs": {
6672
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6673
  "weight_dtype": "default"
6674
  }
6675
  },
6676
  "2": {
6677
  "class_type": "CLIPLoader",
6678
  "inputs": {
6679
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
6680
  "type": "qwen_image",
6681
  "device": "default"
6682
  }
 
6999
  "1": {
7000
  "class_type": "UNETLoader",
7001
  "inputs": {
7002
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
7003
  "weight_dtype": "default"
7004
  }
7005
  },
7006
  "2": {
7007
  "class_type": "CLIPLoader",
7008
  "inputs": {
7009
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
7010
  "type": "qwen_image",
7011
  "device": "default"
7012
  }
 
7109
  "1": {
7110
  "class_type": "UNETLoader",
7111
  "inputs": {
7112
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
7113
  "weight_dtype": "default"
7114
  }
7115
  },
7116
  "2": {
7117
  "class_type": "CLIPLoader",
7118
  "inputs": {
7119
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
7120
  "type": "qwen_image",
7121
  "device": "default"
7122
  }
 
7329
  "1": {
7330
  "class_type": "UNETLoader",
7331
  "inputs": {
7332
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
7333
  "weight_dtype": "default"
7334
  }
7335
  },
7336
  "2": {
7337
  "class_type": "CLIPLoader",
7338
  "inputs": {
7339
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
7340
  "type": "qwen_image",
7341
  "device": "default"
7342
  }
 
7439
  "1": {
7440
  "class_type": "UNETLoader",
7441
  "inputs": {
7442
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
7443
  "weight_dtype": "default"
7444
  }
7445
  },
7446
  "2": {
7447
  "class_type": "CLIPLoader",
7448
  "inputs": {
7449
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
7450
  "type": "qwen_image",
7451
  "device": "default"
7452
  }
 
7659
  "1": {
7660
  "class_type": "UNETLoader",
7661
  "inputs": {
7662
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
7663
  "weight_dtype": "default"
7664
  }
7665
  },
7666
  "2": {
7667
  "class_type": "CLIPLoader",
7668
  "inputs": {
7669
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
7670
  "type": "qwen_image",
7671
  "device": "default"
7672
  }
 
8106
  "2": {
8107
  "class_type": "CLIPLoader",
8108
  "inputs": {
8109
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
8110
  "type": "qwen_image",
8111
  "device": "default"
8112
  }
 
8209
  "1": {
8210
  "class_type": "UNETLoader",
8211
  "inputs": {
8212
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
8213
  "weight_dtype": "default"
8214
  }
8215
  },
 
8319
  "1": {
8320
  "class_type": "UNETLoader",
8321
  "inputs": {
8322
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
8323
  "weight_dtype": "default"
8324
  }
8325
  },
8326
  "2": {
8327
  "class_type": "CLIPLoader",
8328
  "inputs": {
8329
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
8330
  "type": "qwen_image",
8331
  "device": "default"
8332
  }
 
8429
  "1": {
8430
  "class_type": "UNETLoader",
8431
  "inputs": {
8432
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
8433
  "weight_dtype": "default"
8434
  }
8435
  },
8436
  "2": {
8437
  "class_type": "CLIPLoader",
8438
  "inputs": {
8439
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
8440
  "type": "qwen_image",
8441
  "device": "default"
8442
  }
 
8539
  "1": {
8540
  "class_type": "UNETLoader",
8541
  "inputs": {
8542
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
8543
  "weight_dtype": "default"
8544
  }
8545
  },
8546
  "2": {
8547
  "class_type": "CLIPLoader",
8548
  "inputs": {
8549
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
8550
  "type": "qwen_image",
8551
  "device": "default"
8552
  }
 
8649
  "1": {
8650
  "class_type": "UNETLoader",
8651
  "inputs": {
8652
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
8653
  "weight_dtype": "default"
8654
  }
8655
  },
8656
  "2": {
8657
  "class_type": "CLIPLoader",
8658
  "inputs": {
8659
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
8660
  "type": "qwen_image",
8661
  "device": "default"
8662
  }
validation/roundtrip.json CHANGED
@@ -1,11 +1,11 @@
1
  [
2
  {
3
- "file": "prism_image_2.1_nvfp4.safetensors",
4
  "sha256": "4fdbaec94c19f7b8888f1e1ddbae05354c014492f9b2547b1cd8dca0ce923526",
5
  "match": true
6
  },
7
  {
8
- "file": "prism_qwen3vl_8b_nvfp4.safetensors",
9
  "sha256": "cdd9b85bcb60d5cb358ddecfd45a48fa15cb728b70d88bd2028840eaef095d16",
10
  "match": true
11
  },
 
1
  [
2
  {
3
+ "file": "qwen_image_2.1_nvfp4.safetensors",
4
  "sha256": "4fdbaec94c19f7b8888f1e1ddbae05354c014492f9b2547b1cd8dca0ce923526",
5
  "match": true
6
  },
7
  {
8
+ "file": "qwen3vl_8b_nvfp4.safetensors",
9
  "sha256": "cdd9b85bcb60d5cb358ddecfd45a48fa15cb728b70d88bd2028840eaef095d16",
10
  "match": true
11
  },
workflows/01_Text_to_Image.json CHANGED
@@ -30,7 +30,7 @@
30
  "Node name for S&R": "UNETLoader"
31
  },
32
  "widgets_values": [
33
- "prism_image_2.1_nvfp4.safetensors",
34
  "default"
35
  ]
36
  },
@@ -62,7 +62,7 @@
62
  "Node name for S&R": "CLIPLoader"
63
  },
64
  "widgets_values": [
65
- "prism_qwen3vl_8b_nvfp4.safetensors",
66
  "qwen_image",
67
  "default"
68
  ]
@@ -306,7 +306,7 @@
306
  "Node name for S&R": "SaveImage"
307
  },
308
  "widgets_values": [
309
- "Prism/01_Text_to_Image"
310
  ]
311
  }
312
  ],
 
30
  "Node name for S&R": "UNETLoader"
31
  },
32
  "widgets_values": [
33
+ "qwen_image_2.1_nvfp4.safetensors",
34
  "default"
35
  ]
36
  },
 
62
  "Node name for S&R": "CLIPLoader"
63
  },
64
  "widgets_values": [
65
+ "qwen3vl_8b_nvfp4.safetensors",
66
  "qwen_image",
67
  "default"
68
  ]
 
306
  "Node name for S&R": "SaveImage"
307
  },
308
  "widgets_values": [
309
+ "Qwen21_NVFP4/01_Text_to_Image"
310
  ]
311
  }
312
  ],
workflows/02_Image_Editing.json CHANGED
@@ -30,7 +30,7 @@
30
  "Node name for S&R": "UNETLoader"
31
  },
32
  "widgets_values": [
33
- "prism_image_2.1_nvfp4.safetensors",
34
  "default"
35
  ]
36
  },
@@ -62,7 +62,7 @@
62
  "Node name for S&R": "CLIPLoader"
63
  },
64
  "widgets_values": [
65
- "prism_qwen3vl_8b_nvfp4.safetensors",
66
  "qwen_image",
67
  "default"
68
  ]
@@ -286,7 +286,7 @@
286
  "Node name for S&R": "SaveImage"
287
  },
288
  "widgets_values": [
289
- "Prism/02_Image_Editing"
290
  ]
291
  },
292
  {
@@ -322,7 +322,7 @@
322
  "Node name for S&R": "LoadImage"
323
  },
324
  "widgets_values": [
325
- "prism_edit_reference.png"
326
  ]
327
  }
328
  ],
 
30
  "Node name for S&R": "UNETLoader"
31
  },
32
  "widgets_values": [
33
+ "qwen_image_2.1_nvfp4.safetensors",
34
  "default"
35
  ]
36
  },
 
62
  "Node name for S&R": "CLIPLoader"
63
  },
64
  "widgets_values": [
65
+ "qwen3vl_8b_nvfp4.safetensors",
66
  "qwen_image",
67
  "default"
68
  ]
 
286
  "Node name for S&R": "SaveImage"
287
  },
288
  "widgets_values": [
289
+ "Qwen21_NVFP4/02_Image_Editing"
290
  ]
291
  },
292
  {
 
322
  "Node name for S&R": "LoadImage"
323
  },
324
  "widgets_values": [
325
+ "qwen_image_2.1_edit_reference.png"
326
  ]
327
  }
328
  ],
workflows/03_Transparent_RGBA.json CHANGED
@@ -30,7 +30,7 @@
30
  "Node name for S&R": "UNETLoader"
31
  },
32
  "widgets_values": [
33
- "prism_image_2.1_nvfp4.safetensors",
34
  "default"
35
  ]
36
  },
@@ -62,7 +62,7 @@
62
  "Node name for S&R": "CLIPLoader"
63
  },
64
  "widgets_values": [
65
- "prism_qwen3vl_8b_nvfp4.safetensors",
66
  "qwen_image",
67
  "default"
68
  ]
@@ -306,7 +306,7 @@
306
  "Node name for S&R": "SaveImage"
307
  },
308
  "widgets_values": [
309
- "Prism/03_Transparent_RGBA"
310
  ]
311
  }
312
  ],
 
30
  "Node name for S&R": "UNETLoader"
31
  },
32
  "widgets_values": [
33
+ "qwen_image_2.1_nvfp4.safetensors",
34
  "default"
35
  ]
36
  },
 
62
  "Node name for S&R": "CLIPLoader"
63
  },
64
  "widgets_values": [
65
+ "qwen3vl_8b_nvfp4.safetensors",
66
  "qwen_image",
67
  "default"
68
  ]
 
306
  "Node name for S&R": "SaveImage"
307
  },
308
  "widgets_values": [
309
+ "Qwen21_NVFP4/03_Transparent_RGBA"
310
  ]
311
  }
312
  ],
workflows/04_2K_Typography.json CHANGED
@@ -30,7 +30,7 @@
30
  "Node name for S&R": "UNETLoader"
31
  },
32
  "widgets_values": [
33
- "prism_image_2.1_nvfp4.safetensors",
34
  "default"
35
  ]
36
  },
@@ -62,7 +62,7 @@
62
  "Node name for S&R": "CLIPLoader"
63
  },
64
  "widgets_values": [
65
- "prism_qwen3vl_8b_nvfp4.safetensors",
66
  "qwen_image",
67
  "default"
68
  ]
@@ -306,7 +306,7 @@
306
  "Node name for S&R": "SaveImage"
307
  },
308
  "widgets_values": [
309
- "Prism/04_2K_Typography"
310
  ]
311
  }
312
  ],
 
30
  "Node name for S&R": "UNETLoader"
31
  },
32
  "widgets_values": [
33
+ "qwen_image_2.1_nvfp4.safetensors",
34
  "default"
35
  ]
36
  },
 
62
  "Node name for S&R": "CLIPLoader"
63
  },
64
  "widgets_values": [
65
+ "qwen3vl_8b_nvfp4.safetensors",
66
  "qwen_image",
67
  "default"
68
  ]
 
306
  "Node name for S&R": "SaveImage"
307
  },
308
  "widgets_values": [
309
+ "Qwen21_NVFP4/04_2K_Typography"
310
  ]
311
  }
312
  ],
workflows/api/01_Text_to_Image.api.json CHANGED
@@ -2,14 +2,14 @@
2
  "1": {
3
  "class_type": "UNETLoader",
4
  "inputs": {
5
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6
  "weight_dtype": "default"
7
  }
8
  },
9
  "2": {
10
  "class_type": "CLIPLoader",
11
  "inputs": {
12
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
13
  "type": "qwen_image",
14
  "device": "default"
15
  }
@@ -87,7 +87,7 @@
87
  "7",
88
  0
89
  ],
90
- "filename_prefix": "Prism/01_Text_to_Image"
91
  }
92
  }
93
  }
 
2
  "1": {
3
  "class_type": "UNETLoader",
4
  "inputs": {
5
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6
  "weight_dtype": "default"
7
  }
8
  },
9
  "2": {
10
  "class_type": "CLIPLoader",
11
  "inputs": {
12
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
13
  "type": "qwen_image",
14
  "device": "default"
15
  }
 
87
  "7",
88
  0
89
  ],
90
+ "filename_prefix": "Qwen21_NVFP4/01_Text_to_Image"
91
  }
92
  }
93
  }
workflows/api/02_Image_Editing.api.json CHANGED
@@ -2,14 +2,14 @@
2
  "1": {
3
  "class_type": "UNETLoader",
4
  "inputs": {
5
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6
  "weight_dtype": "default"
7
  }
8
  },
9
  "2": {
10
  "class_type": "CLIPLoader",
11
  "inputs": {
12
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
13
  "type": "qwen_image",
14
  "device": "default"
15
  }
@@ -87,13 +87,13 @@
87
  "7",
88
  0
89
  ],
90
- "filename_prefix": "Prism/02_Image_Editing"
91
  }
92
  },
93
  "10": {
94
  "class_type": "LoadImage",
95
  "inputs": {
96
- "image": "prism_edit_reference.png"
97
  }
98
  }
99
  }
 
2
  "1": {
3
  "class_type": "UNETLoader",
4
  "inputs": {
5
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6
  "weight_dtype": "default"
7
  }
8
  },
9
  "2": {
10
  "class_type": "CLIPLoader",
11
  "inputs": {
12
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
13
  "type": "qwen_image",
14
  "device": "default"
15
  }
 
87
  "7",
88
  0
89
  ],
90
+ "filename_prefix": "Qwen21_NVFP4/02_Image_Editing"
91
  }
92
  },
93
  "10": {
94
  "class_type": "LoadImage",
95
  "inputs": {
96
+ "image": "qwen_image_2.1_edit_reference.png"
97
  }
98
  }
99
  }
workflows/api/03_Transparent_RGBA.api.json CHANGED
@@ -2,14 +2,14 @@
2
  "1": {
3
  "class_type": "UNETLoader",
4
  "inputs": {
5
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6
  "weight_dtype": "default"
7
  }
8
  },
9
  "2": {
10
  "class_type": "CLIPLoader",
11
  "inputs": {
12
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
13
  "type": "qwen_image",
14
  "device": "default"
15
  }
@@ -87,7 +87,7 @@
87
  "7",
88
  0
89
  ],
90
- "filename_prefix": "Prism/03_Transparent_RGBA"
91
  }
92
  }
93
  }
 
2
  "1": {
3
  "class_type": "UNETLoader",
4
  "inputs": {
5
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6
  "weight_dtype": "default"
7
  }
8
  },
9
  "2": {
10
  "class_type": "CLIPLoader",
11
  "inputs": {
12
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
13
  "type": "qwen_image",
14
  "device": "default"
15
  }
 
87
  "7",
88
  0
89
  ],
90
+ "filename_prefix": "Qwen21_NVFP4/03_Transparent_RGBA"
91
  }
92
  }
93
  }
workflows/api/04_2K_Typography.api.json CHANGED
@@ -2,14 +2,14 @@
2
  "1": {
3
  "class_type": "UNETLoader",
4
  "inputs": {
5
- "unet_name": "prism_image_2.1_nvfp4.safetensors",
6
  "weight_dtype": "default"
7
  }
8
  },
9
  "2": {
10
  "class_type": "CLIPLoader",
11
  "inputs": {
12
- "clip_name": "prism_qwen3vl_8b_nvfp4.safetensors",
13
  "type": "qwen_image",
14
  "device": "default"
15
  }
@@ -87,7 +87,7 @@
87
  "7",
88
  0
89
  ],
90
- "filename_prefix": "Prism/04_2K_Typography"
91
  }
92
  }
93
  }
 
2
  "1": {
3
  "class_type": "UNETLoader",
4
  "inputs": {
5
+ "unet_name": "qwen_image_2.1_nvfp4.safetensors",
6
  "weight_dtype": "default"
7
  }
8
  },
9
  "2": {
10
  "class_type": "CLIPLoader",
11
  "inputs": {
12
+ "clip_name": "qwen3vl_8b_nvfp4.safetensors",
13
  "type": "qwen_image",
14
  "device": "default"
15
  }
 
87
  "7",
88
  0
89
  ],
90
+ "filename_prefix": "Qwen21_NVFP4/04_2K_Typography"
91
  }
92
  }
93
  }