Tags: fasterinnerlooper/DeepSpeed
Tags
fix fused_qkv model accuracy issue (deepspeedai#5217) Fused_qkv model can not correctly choose the fused_qkv type. Need to update the module_name_matches. Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com>
Add script to check for `--extra-index-url` (deepspeedai#5184) Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com>
Switch cpu-inference workflow from --extra-index-url to --index-url (d… …eepspeedai#5182) This switch should cause no impact to the workflow, but ensures that we only download this package from the correct feed rather than defaulting/falling back to the index-url [default](https://pip.pypa.io/en/stable/cli/pip_install/#cmdoption-i) (PyPI) for if a package existed with a higher version there that would be chosen instead.
Remove optimizer step on initialization (deepspeedai#5104) All ZeRO 1/2/3 stages call the optimizer's `step()` on its initialization. This increments a counter in the optimizer and produces a different result in parameter update with the normal usage of PyTorch. This PR eliminates `step()` in the initialization and lazily configures some internal states (linking *hp_params*) after the first `step()` call. --------- Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com>
Refactor the Qwen positional emebdding config code (deepspeedai#4955) follow PR deepspeedai#4920 on Qwen inference code Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com>
Mixtral FastGen Support (deepspeedai#4828) Adds support for Mixtral with FastGen. Key features implemented: 1. Top-2 MoE support 2. Better support for RoPE thetas 3. The mistral model implementation --------- Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com>
Fix 4649 (deepspeedai#4650) Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com>
Add safetensors support (deepspeedai#4659) Co-authored-by: Logan Adams <114770087+loadams@users.noreply.github.com> Co-authored-by: Jeff Rasley <jerasley@microsoft.com> Co-authored-by: Michael Wyatt <michaelwyatt@microsoft.com>
PreviousNext