Commits on Mar 7, 2024

1. SymIntify torchrec variable batch size path (pytorch#2394) …

   Summary:
   
   Variable Batch parameters are SymInt in dynamo tracing:
   
   1/ SymInt does not support bit shifts => Skipping adjust_info_B_num_bits logic for dynamo case (when SymInt arrive into kernel) defaulting the values.
   The main idea of `adjust_info_B_num_bits` is to fit BatchSize (B) and number of Tables (T) into 32 bit.
   Default value for B_bits == 26, T_bits == 6. `adjust_info_B_num_bits` will at runtime change those bit numbers to fit specified T and B using control flow and bit shifts to check that T and B fits in bits.
   
   We have flows with T ~ 600, so picking default values for dynamo T_bits == 10 and B_bits == 22, guarding to fail in runtime if those bit numbers does not fit specified T and B.
   
   For Dynamo we will use constant values for now.
   
   2/ fbcode/deeplearning/fbgemm/fbgemm_gpu/fbgemm_gpu/split_table_batched_embeddings_ops_training.py - changing List comprehension with python sum to torch.sum. 
   
   3/ Removing int() convertions for SymInt.
   
   4/ Adding torch._check() for VB parameters
   
   Reviewed By: ezyang
   
   Differential Revision: D54554735

   

   Ivan Kobzarev authored and facebook-github-bot committed Mar 7, 2024
   Configuration menu

   • View commit details
   • Copy full SHA for 6dd5241
   • Browse repository at this point

   Copy the full SHA

   6dd5241 View commit details

   Browse the repository at this point in the history