accelerate

67adb473 - (Part 1) fix: make TP training compatible with new transformers (#3457)

(Part 1) fix: make TP training compatible with new transformers (#3457) * feat: support new tp refactor for training Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com> * fix: @S1ro1 review cmt Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com> * fix: @S1ro1 review cmt - tp_plan flag docstr Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com> * fix: @SunMarc review cmt on un used flag Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com> * fix: pick approach 3 as discussed in the PR see https://github.com/huggingface/accelerate/pull/3457#discussion_r2037909077 for more details Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com> * fix: styling errors Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com> * fix: bump up transformers for tp_size feature Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com> --------- Signed-off-by: Mehant Kammakomati <mehant.kammakomati2@ibm.com>