Official Release: AMD Strix Point Native Optimization and Director St… - #796
Open
Cyb3rLab5 wants to merge 23 commits into
Open
Official Release: AMD Strix Point Native Optimization and Director St…#796Cyb3rLab5 wants to merge 23 commits into
Cyb3rLab5 wants to merge 23 commits into
Conversation
Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
Cached the weight and bias tensors in `vae_decode_fake` based on device and dtype. This prevents re-allocation and conversion on every call, significantly reducing overhead in hot loops. Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
Refactored `HunyuanVideoRotaryPosEmbed` to calculate Rotary Positional Embeddings (RoPE) iteratively over 1D components (T, H, W). PyTorch's zero-copy `.expand()` and views are then used to map these vectors into 3D. This eliminates massive redundant 3D grid creations (via `torch.meshgrid`) and sequential looping over batches. Performance benchmark shows ~4.3x forward pass speedup for batch coordinate generations. Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
…6079084948 ⚡ Bolt: [performance improvement] 1D RoPE Expand
…858351843 ⚡ Bolt: [performance improvement] Cache tensors in vae_decode_fake
Batch the torch.cuda.memory_stats queries to check only every 25 modules, significantly reducing CPU-GPU synchronization stalls. Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
…head-55594135257129418 ⚡ Bolt: Optimize VRAM polling overhead
…27942520 ⚡ Bolt: [performance improvement]
Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
…09869604236974047 ⚡ Bolt: Replace np.poly1d with unrolled native Python evaluation
Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
…792659738826133 ⚡ Bolt: Vectorize cu_seqlens loop in framepack_transformer
- Replaced `torch.tensor([val] * bs).to(device)` with `torch.full((bs,), val, device=device)` in framepack pipeline. - Replaced `torch.tensor([0.5], device=device)` with `torch.full((1,), 0.5, device=device)` in UniPC solver. - Creating a tensor from a Python list forces a CPU allocation and a blocking CPU-GPU synchronization, which introduces a ~1.5x-2x overhead for small tensors compared to natively creating them on the device via `torch.full`. Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
…-15565401181004026508 ⚡ Bolt: Optimize implicit CPU-GPU synchronization during tensor initialization
- Shifted sum operation for `attention_mask` to happen *before* the tensor is transferred to the GPU in `framepack_helpers.py`. - Added learning to `.jules/bolt.md` reflecting that using `.sum()` on a GPU tensor in PyTorch blocks host execution by syncing data over the PCIe bus, identical to `.item()`. Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
…10562887828396246823 ⚡ Bolt: [Remove CPU-GPU Sync on Device Tensors]
…C by using torch.stack Co-authored-by: Cyb3rLab5 <224908985+Cyb3rLab5@users.noreply.github.com>
…198123106505 ⚡ Bolt: [performance improvement] Avoid CPU-GPU sync in FlowMatchUniPC by using torch.stack
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
…udio