You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
On a GB10 (cc 12.1) machine, loading weights using NunchakuFluxTransformer2dModel.from_pretrained(...) takes approximately 37 seconds, while on a 3090 machine it takes approximately 3 seconds.
The profiler display time is almost entirely consumed by numerous cudaMemcpyAsync operations (approximately 2400+ small block copies). Typically, small blocks are moved from the CPU to the GPU memory path, amplifying the fixed overhead of each copy.
Modifications
Solution (How to Locate and Fix)
Location: H2D bandwidth testing confirmed the link was normal (pinned H2D was very high).
The "small copy" experiment verified that the difference between pageable and pinned was ~50×.
The loading logic was found in nunchaku/models/transformers/transformer_flux.py::from_pretrained: load_file(...) first reads the weights into the CPU tensor, then triggers a large number of H2D copies.
Fix: Added a _pin_state_dict(...) helper to perform pin_memory() on the CPU tensors in quantized_part_sd and unquantized_part_sd.
Added a pin_memory switch (default True) to from_pretrained(), enabling it only when device.type == "cuda".
Results (actual test): Loading profiling time decreased from ~37.5s Self CPU time to ~3.9s.
Self CUDA time decreased to the ~154ms range.
Conclusion: This patch is a performance fix, significantly reducing the overhead of "CPU→GPU small block transport during the weight loading phase".
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
On a GB10 (cc 12.1) machine, loading weights using
NunchakuFluxTransformer2dModel.from_pretrained(...)takes approximately 37 seconds, while on a 3090 machine it takes approximately 3 seconds.The profiler display time is almost entirely consumed by numerous
cudaMemcpyAsyncoperations (approximately 2400+ small block copies). Typically, small blocks are moved from the CPU to the GPU memory path, amplifying the fixed overhead of each copy.Modifications
Solution (How to Locate and Fix)
Location: H2D bandwidth testing confirmed the link was normal (pinned H2D was very high).
The "small copy" experiment verified that the difference between pageable and pinned was ~50×.
The loading logic was found in nunchaku/models/transformers/transformer_flux.py::from_pretrained: load_file(...) first reads the weights into the CPU tensor, then triggers a large number of H2D copies.
Fix: Added a _pin_state_dict(...) helper to perform pin_memory() on the CPU tensors in quantized_part_sd and unquantized_part_sd.
Added a pin_memory switch (default True) to from_pretrained(), enabling it only when device.type == "cuda".
Results (actual test): Loading profiling time decreased from ~37.5s Self CPU time to ~3.9s.
Self CUDA time decreased to the ~154ms range.
Conclusion: This patch is a performance fix, significantly reducing the overhead of "CPU→GPU small block transport during the weight loading phase".