From a18e2c0db226b7118ed7ebbaecd8edb57dc59335 Mon Sep 17 00:00:00 2001
From: jacobfulano <62222220+jacobfulano@users.noreply.github.com>
Date: Thu, 3 Aug 2023 14:14:45 -0400
Subject: [PATCH] Update BERT yamls for compatibility with main branch (#427)

* typo

* update bf16 --> amp_bf16 in bert yamls

* update mcloud yamls to run on mosaicml/examples main branch with correct file structure

* updated git repo branch in 40gb yaml
---
 .../benchmarks/bert/yamls/main/mcloud_run_a100_40gb.yaml | 9 +++++----
 .../benchmarks/bert/yamls/main/mcloud_run_a100_80gb.yaml | 9 +++++----
 2 files changed, 10 insertions(+), 8 deletions(-)

diff --git a/examples/benchmarks/bert/yamls/main/mcloud_run_a100_40gb.yaml b/examples/benchmarks/bert/yamls/main/mcloud_run_a100_40gb.yaml
index d945eacc4..02880d0d7 100644
--- a/examples/benchmarks/bert/yamls/main/mcloud_run_a100_40gb.yaml
+++ b/examples/benchmarks/bert/yamls/main/mcloud_run_a100_40gb.yaml
@@ -21,7 +21,7 @@ compute:
 integrations:
 - integration_type: git_repo
   git_repo: mosaicml/examples
-  git_branch: v0.0.4 # use your branch
+  git_branch: main # use your branch
   # git_commit: # OR use your commit hash
   pip_install: -e .[bert]
   ssh_clone: false # Should be true if using a private repo
@@ -31,8 +31,9 @@ integrations:
 # For real training runs, follow the instructions in `examples/bert/README.md`
 # to convert and host the full 'train' dataset.
 command: |
-  cd examples/examples/bert
-  python ../common/convert_dataset.py --dataset c4 --data_subset en --out_root ./my-copy-c4 --splits train_small val
+  cd examples/examples/benchmarks/bert
+  pip install -r requirements.txt
+  python src/convert_dataset.py --dataset c4 --data_subset en --out_root ./my-copy-c4 --splits train_small val
   composer main.py /mnt/config/parameters.yaml \
     train_loader.dataset.split=train_small
 
@@ -117,7 +118,7 @@ parameters:
   seed: 17
   device_eval_batch_size: 128
   device_train_microbatch_size: 128
-  precision: bf16
+  precision: amp_bf16 # update for Composer release/v0.12.1
 
   # Logging
   progress_bar: false
diff --git a/examples/benchmarks/bert/yamls/main/mcloud_run_a100_80gb.yaml b/examples/benchmarks/bert/yamls/main/mcloud_run_a100_80gb.yaml
index 324caae3f..48f2b0563 100644
--- a/examples/benchmarks/bert/yamls/main/mcloud_run_a100_80gb.yaml
+++ b/examples/benchmarks/bert/yamls/main/mcloud_run_a100_80gb.yaml
@@ -22,7 +22,7 @@ compute:
 integrations:
 - integration_type: git_repo
   git_repo: mosaicml/examples
-  git_branch: v0.0.4 # use your branch
+  git_branch: main # use your branch
   # git_commit: # OR use your commit hash
   pip_install: -e .[bert]
   ssh_clone: false # Should be true if using a private repo
@@ -32,8 +32,9 @@ integrations:
 # For real training runs, follow the instructions in `examples/bert/README.md`
 # to convert and host the full 'train' dataset.
 command: |
-  cd examples/examples/bert
-  python ../common/convert_dataset.py --dataset c4 --data_subset en --out_root ./my-copy-c4 --splits train_small val
+  cd examples/examples/benchmarks/bert
+  pip install -r requirements.txt
+  python src/convert_dataset.py --dataset c4 --data_subset en --out_root ./my-copy-c4 --splits train_small val
   composer main.py /mnt/config/parameters.yaml \
     train_loader.dataset.split=train_small
 
@@ -117,7 +118,7 @@ parameters:
   seed: 17
   device_eval_batch_size: 512
   device_train_microbatch_size: 512
-  precision: bf16
+  precision: amp_bf16 # update for Composer release/v0.12.1
 
   # Logging
   progress_bar: false