Summary
Changelog for Release v0.19.3 (notable changes since v0.18.2).
Generality
- Implemented support for
10
additional model variants, including3
new architectures:Phi2
Qwen1.5-0.5B
YOLOX
Performance
Grayskull (e75, e150)
- Enhanced performance observed on e75 include HRNet (41%)
- Notable performance regressions on e150 include OpenPose (-16%)
Wormhole (n150)
- Performance restored on Whisper, Stable Diffusion
Wormhole (n300)
- Generative AI models not supported on n300 (dual-chip)
- Enhanced performance observed on n300 (dual-chip) of 8% average
- Enhanced performance observed on n300 (single-chip) include Falcon-7B (14%)
- Notable performance regressions on n300 (single-chip) include FLAN-T5 (-20%), U-Net (-14%)
- Accuracy regressions on n300 (dual-chip) include YOLOv5, OpenPose, U-Net
Multicard Systems (TT-LoudBox / TT-QuietBox)
- Enabled support for CNN models on 4-chip MMIO and 8-chip configurations
- Generative AI models not supported
General (Improvements/Features/Etc.)
- Introduced functionality for Wormhole n300 (dual-chip), including the TT-LoudBox and TT-QuietBox configurations
- Across-the-board bug fixes and enhancements to compiler stability
- Enhanced customer-facing documentation for improved clarity and accessibility
Known issues
- N/A
Compatibility matrix
OS | Python | PyTorch | Driver | Firmware |
---|---|---|---|---|
Ubuntu 22.04 | 3.10.12 | 2.1.0+cpu.cxx11.abi | ttkmd_1.29 | fw_v80.10.0.0 |
Ubuntu 20.04 | 3.8.10 | 2.1.0+cpu.cxx11.abi | ttkmd_1.29 | fw_v80.10.0.0 |
What's Changed
- Uplift PyBuda changes (week29) by @vmilosevic in #39
Full Changelog: v0.18.2...v0.19.3