Commits · 144e6580a4a43d5390769665a5032bb584481ff1 · chenpangpang / ComfyUI

".circleci/vscode:/vscode.git/clone" did not exist on "e808225f1859fbf47e778b73c64e82669d1c1e94"

29 Dec, 2023 1 commit
- Load weights that can't be lowvramed to target device. · e1e322cf
  comfyanonymous authored Dec 28, 2023
  
  e1e322cf
23 Dec, 2023 1 commit
- --disable-smart-memory now unloads everything like it did originally. · a252963f
  comfyanonymous authored Dec 23, 2023
  
  a252963f
22 Dec, 2023 1 commit
- Greatly improve lowvram sampling speed by getting rid of accelerate. · 36a79531
  comfyanonymous authored Dec 22, 2023
```
Let me know if this breaks anything.
```
  36a79531
17 Dec, 2023 1 commit
- Add --deterministic option to make pytorch use deterministic algorithms. · 2f9d6a97
  comfyanonymous authored Dec 17, 2023
  
  2f9d6a97
11 Dec, 2023 3 commits
- Add an option --fp16-unet to force using fp16 for the unet. · b0aab1e4
  comfyanonymous authored Dec 11, 2023
  
  b0aab1e4
- Use faster manual cast for fp8 in unet. · ba07cb74
  comfyanonymous authored Dec 11, 2023
  
  ba07cb74
- Switch text encoder to manual cast. · 57926635
  comfyanonymous authored Dec 10, 2023
```
Use fp16 text encoder weights for CPU inference to lower memory usage.
```
  57926635
10 Dec, 2023 1 commit
- Disable non blocking on mps. · 340177e6
  comfyanonymous authored Dec 10, 2023
  
  340177e6
08 Dec, 2023 1 commit
- Make --gpu-only put intermediate values in GPU memory instead of cpu. · 9ac0b487
  comfyanonymous authored Dec 08, 2023
  
  9ac0b487
06 Dec, 2023 1 commit
- Slightly faster lora applying. · 2db86b46
  comfyanonymous authored Dec 06, 2023
  
  2db86b46
04 Dec, 2023 2 commits
- Use .itemsize to get dtype size for fp8. · ca82ade7
  comfyanonymous authored Dec 04, 2023
  
  ca82ade7
- UNET weights can now be stored in fp8. · 31b0f6f3
  comfyanonymous authored Dec 04, 2023
```
--fp8_e4m3fn-unet and --fp8_e5m2-unet are the two different formats
supported by pytorch.
```
  31b0f6f3
17 Nov, 2023 1 commit

Add some command line arguments to store text encoder weights in fp8. · 0cf4e869

comfyanonymous authored Nov 17, 2023

Pytorch supports two variants of fp8:
--fp8_e4m3fn-text-enc (the one that seems to give better results)
--fp8_e5m2-text-enc

0cf4e869

13 Nov, 2023 1 commit
- Disable xformers when it can't load properly. · 7339479b
  comfyanonymous authored Nov 13, 2023
  
  7339479b
12 Nov, 2023 1 commit
- Allow different models to estimate memory usage differently. · dd4ba68b
  comfyanonymous authored Nov 12, 2023
  
  dd4ba68b
22 Oct, 2023 1 commit
- Empty the cache when torch cache is more than 25% free mem. · 8594c8be
  comfyanonymous authored Oct 22, 2023
  
  8594c8be
16 Oct, 2023 1 commit
- Add some Quadro cards to the list of cards with broken fp16. · c8013f73
  comfyanonymous authored Oct 16, 2023
  
  c8013f73
13 Oct, 2023 2 commits
- Add a --bf16-unet to test running the unet in bf16. · fd4c5f07
  comfyanonymous authored Oct 13, 2023
  
  fd4c5f07
- Refactor code so model can be a dtype other than fp32 or fp16. · 9a55dadb
  comfyanonymous authored Oct 13, 2023
  
  9a55dadb
12 Oct, 2023 2 commits
- pytorch_attention_enabled can now return True when xformers is enabled. · 88733c99
  comfyanonymous authored Oct 11, 2023
  
  88733c99
- Pull some small changes from the other repo. · 20d3852a
  comfyanonymous authored Oct 11, 2023
  
  20d3852a
23 Sep, 2023 1 commit
- Allow Intel GPUs to LoRA cast on GPU since it supports BF16 natively. · eec449ca
  Simon Lui authored Sep 22, 2023
  
  eec449ca
20 Sep, 2023 1 commit
- Only do the cast on the device if the device supports it. · 1cdfb3db
  comfyanonymous authored Sep 20, 2023
  
  1cdfb3db
17 Sep, 2023 1 commit
- Enable pytorch attention by default on xpu. · 321c5fa2
  comfyanonymous authored Sep 17, 2023
  
  321c5fa2
14 Sep, 2023 1 commit
- Don't run text encoders on xpu because there are issues. · 0966d3ce
  comfyanonymous authored Sep 14, 2023
  
  0966d3ce
04 Sep, 2023 1 commit
- Add a force argument to soft_empty_cache to force a cache empty. · 1938f5c5
  comfyanonymous authored Sep 04, 2023
  
  1938f5c5
03 Sep, 2023 1 commit
- Some fixes to generalize CUDA specific functionality to Intel or other GPUs. · 4a0c4ce4
  Simon Lui authored Sep 02, 2023
  
  4a0c4ce4
28 Aug, 2023 1 commit
- Enable bf16-vae by default on ampere and up. · b8c7c770
  comfyanonymous authored Aug 27, 2023
  
  b8c7c770
26 Aug, 2023 1 commit
- Fix lowvram model merging. · a57b0c79
  comfyanonymous authored Aug 26, 2023
  
  a57b0c79
25 Aug, 2023 1 commit
- The new smart memory management makes this unnecessary. · f72780a7
  comfyanonymous authored Aug 25, 2023
  
  f72780a7
24 Aug, 2023 4 commits
- Code cleanups. · 30eb92c3
  comfyanonymous authored Aug 24, 2023
  
  30eb92c3
- Try to free enough vram for control lora inference. · 51dde87e
  comfyanonymous authored Aug 24, 2023
  
  51dde87e
- Always shift text encoder to GPU when the device supports fp16. · cc44ade7
  comfyanonymous authored Aug 23, 2023
  
  cc44ade7
- Even with forced fp16 the cpu device should never use it. · a6ef08a4
  comfyanonymous authored Aug 23, 2023
  
  a6ef08a4
23 Aug, 2023 1 commit
- Save memory by storing text encoder weights in fp16 in most situations. · f081017c
  comfyanonymous authored Aug 23, 2023
```
Do inference in fp32 to make sure quality stays the exact same.
```
  f081017c
20 Aug, 2023 4 commits
- Small cleanups. · 0d7b0a4d
  comfyanonymous authored Aug 20, 2023
  
  0d7b0a4d
- Further tuning and fix mem_free_total. · 92254659
  Simon Lui authored Aug 19, 2023
  
  92254659
- Add ipex optimize and other enhancements for Intel GPUs based on recent memory changes. · 2c096e42
  Simon Lui authored Aug 17, 2023
  
  2c096e42
- --disable-smart-memory now disables loading model directly to vram. · e9469e73
  comfyanonymous authored Aug 20, 2023
  
  e9469e73
17 Aug, 2023 1 commit
- Add --disable-smart-memory for those that want the old behaviour. · 3aee33b5
  comfyanonymous authored Aug 17, 2023
  
  3aee33b5