Commits · c64098ef768f8367335c21d38d1c38c23f932f2c · OpenDAS / Megatron-LM

11 Mar, 2022 1 commit
- tiny change. · c64098ef
  Lawrence McAfee authored Mar 10, 2022
  
  c64098ef
10 Mar, 2022 2 commits
- small fixes. · 862d70fc
  Lawrence McAfee authored Mar 10, 2022
  
  862d70fc
- updated args for allreduce_embeddings · 08ee8ea2
  Lawrence McAfee authored Mar 10, 2022
  
  08ee8ea2
09 Mar, 2022 14 commits
- updated FP32Optimizer for latest changes. · c88bc979
  Lawrence McAfee authored Mar 09, 2022
  
  c88bc979
- removed old (pre-gbuf-padding) get_model_grad_buffer_dp_views(). · 7ac342b7
  Lawrence McAfee authored Mar 09, 2022
  
  7ac342b7
- more cleanup · 0528bf77
  Lawrence McAfee authored Mar 09, 2022
  
  0528bf77
- removed faulty (already commented out) gather_model_params param iterator. · 94a90215
  Lawrence McAfee authored Mar 09, 2022
  
  94a90215
- renamed reduce_grads/gather_params -> reduce_model_grads/gather_model_params · 64b9d28c
  Lawrence McAfee authored Mar 09, 2022
  
  64b9d28c
- passing args, timers from train_step to optimizer methods. · d58d1762
  Lawrence McAfee authored Mar 09, 2022
  
  d58d1762
- consolidated reduce_grad's sub-methods (i.e., allreduce_embedding_grads) · 45b364b1
  Lawrence McAfee authored Mar 09, 2022
  
  45b364b1
- finished cleaning clip_grads.py · a9b1fc0a
  Lawrence McAfee authored Mar 09, 2022
  
  a9b1fc0a
- passing 'model_parallel_group' to clip_grads, count_zeros · 9546d8f0
  Lawrence McAfee authored Mar 09, 2022
  
  9546d8f0
- moved 'reduce_grads()' to MegatronOptimizer. · e46230dc
  Lawrence McAfee authored Mar 09, 2022
  
  e46230dc
- encapsulated 'get_model_parallel_group()'; renamed BaseFloat16Optimizer -> MixedPrecisionOptimizer · 772a4a2d
  Lawrence McAfee authored Mar 09, 2022
  
  772a4a2d
- removed clip_grad_norm shortcut. · 6e888151
  Lawrence McAfee authored Mar 09, 2022
  
  6e888151
- moved distrib opt to own file. · 28b63ec7
  Lawrence McAfee authored Mar 09, 2022
  
  28b63ec7
- padded ddp's grad_buffer to multiple of data parallel world size · 55695f81
  Lawrence McAfee authored Mar 09, 2022
  
  55695f81
07 Mar, 2022 1 commit
- deleted old distributed_fused_adam · 4f2356dc
  Lawrence McAfee authored Mar 07, 2022
  
  4f2356dc
02 Mar, 2022 3 commits
- working w/ mixed parallelism [ dp, tp, pp ]. · e5bda3c9
  Lawrence McAfee authored Mar 02, 2022
  
  e5bda3c9
- copied t5 embed sync to distrib opt · 2f99b4f4
  Lawrence McAfee authored Mar 02, 2022
  
  2f99b4f4
- fixed gather params; now copying all params, not just local dp subset. · f8fd34e3
  Lawrence McAfee authored Mar 02, 2022
  
  f8fd34e3
01 Mar, 2022 2 commits
- fixed 'count zeros' for distrib opt · cbcd5579
  Lawrence McAfee authored Mar 01, 2022
  
  cbcd5579
- more debugging; some cleanup · 67e23459
  Lawrence McAfee authored Feb 28, 2022
  
  67e23459
28 Feb, 2022 1 commit
- clip grad reduce across world. [ not just model group ] · a58653dc
  Lawrence McAfee authored Feb 28, 2022
  
  a58653dc
23 Feb, 2022 3 commits
- isolated grad discrepency to BaseFloat16Optimizer.step() · 371a8828
  Lawrence McAfee authored Feb 23, 2022
  
  371a8828
- renamed reduce_gradients -> reduce_grads [ matches gather_params ] · e7f0cdee
  Lawrence McAfee authored Feb 23, 2022
  
  e7f0cdee
- fixed param_world_shard bug. · 4b843668
  Lawrence McAfee authored Feb 23, 2022
  
  4b843668
22 Feb, 2022 4 commits
- debugging; localized issue to gather_params() · c13c0a3e
  Lawrence McAfee authored Feb 22, 2022
  
  c13c0a3e
- more cleanup of main params/grads · 9b7854e4
  Lawrence McAfee authored Feb 22, 2022
  
  9b7854e4
- removed zero-size optimizer group shards. · 291592e4
  Lawrence McAfee authored Feb 22, 2022
  
  291592e4
- set main_param_shard==None if local group_size==0 [ e.g., word embedding params · 23f9238d
  Lawrence McAfee authored Feb 22, 2022
  
  23f9238d
18 Feb, 2022 6 commits
- small fixes; training & learning! · ac5ef637
  Lawrence McAfee authored Feb 18, 2022
  
  ac5ef637
- copy model params to main params at very start · a4f41882
  Lawrence McAfee authored Feb 18, 2022
  
  a4f41882
- some cleanup · ae6cd9b5
  Lawrence McAfee authored Feb 18, 2022
  
  ae6cd9b5
- modularized reduce_gradients, gather params; trainin runs, but loss==nan · e5db0fda
  Lawrence McAfee authored Feb 18, 2022
  
  e5db0fda
- fixed clip_grad_norm bug · a7782b21
  Lawrence McAfee authored Feb 18, 2022
  
  a7782b21
- support fp32 training and fix embedding update · b5726555
  Sangkug Lym authored Feb 18, 2022
  
  b5726555
17 Feb, 2022 2 commits
- addressed Jared and Patrick comments. · 90ce932d
  mshoeybi authored Feb 17, 2022
  
  90ce932d
- changed all_gather to _all_gather_base in distributed checkpointing · 37181ef4
  mshoeybi authored Feb 17, 2022
  
  37181ef4
16 Feb, 2022 1 commit

Sangkug Lym authored Feb 13, 2022

remove redundant linear layer class definition

add fuse_gradient_accumulation attribute to weights for simple targetting

reflect feedback and clean up the codes

arg change

83b1e42f