Jedrzej Kosinski
|
5329180fce
|
Made flipflop consider partial_unload, partial_offload, and add flip+flop to mem counters
|
2025-10-03 16:21:01 -07:00 |
|
Jedrzej Kosinski
|
ee01002e63
|
Add flipflop support to (base) WAN, fix issue with applying loras to flipflop weights being done on CPU instead of GPU, left some timing functions as the lora application time could use some reduction
|
2025-10-02 22:02:50 -07:00 |
|
Jedrzej Kosinski
|
8d7b22b720
|
Fixed FlipFlipModule.execute_blocks having hardcoded strings from Qwen
|
2025-10-02 17:49:43 -07:00 |
|
Jedrzej Kosinski
|
6d3ec9fcf3
|
Simplified flipflop setup by adding FlipFlopModule.execute_blocks helper
|
2025-10-02 16:46:37 -07:00 |
|
Jedrzej Kosinski
|
0df61b5032
|
Fix improper index slicing for flipflop get blocks, add extra log message
|
2025-10-01 21:21:36 -07:00 |
|
Jedrzej Kosinski
|
7c896c5567
|
Initial automatic support for flipflop within ModelPatcher - only Qwen Image diffusion_model uses FlipFlopModule currently
|
2025-10-01 20:13:50 -07:00 |
|
Jedrzej Kosinski
|
01f4512bf8
|
In-progress commit on making flipflop async weight streaming native, made loaded partially/loaded completely log messages have labels because having to memorize their meaning for dev work is annoying
|
2025-09-30 23:08:08 -07:00 |
|
Jedrzej Kosinski
|
8a8162e8da
|
Fix percentage logic, begin adding elements to ModelPatcher to track flip flop compatibility
|
2025-09-29 22:49:12 -07:00 |
|
Jedrzej Kosinski
|
6b240b0bce
|
Refactored old flip flop into a new implementation that allows for controlling the percentage of blocks getting flip flopped, converted nodes to v3 schema
|
2025-09-25 22:41:41 -07:00 |
|
Jedrzej Kosinski
|
f9fbf902d5
|
Added missing Qwen block params, further subdivided blocks function
|
2025-09-25 17:49:39 -07:00 |
|
Jedrzej Kosinski
|
f083720eb4
|
Refactored FlipFlopTransformer.__call__ to fully separate out actions between flip and flop
|
2025-09-25 16:16:51 -07:00 |
|
Jedrzej Kosinski
|
84e73f2aa5
|
Brought over flip flop prototype from contentis' fork, limiting it to only Qwen to ease the process of adapting it to be a native feature
|
2025-09-25 16:15:46 -07:00 |
|