Merge pull request #414 from allenai/mitchish65-2

Changes for 65B and 70B runs

CHANGELOG.md

@@ -7,23 +7,25 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 
+### Added
+
+- Added support for Grouped Query Attention.
+- Added commonsense_qa and social_iqa downstream evaluation tasks
+
 ### Changed
 
 - Rename `Olmo` to `OLMo` everywhere in the codebase
 
 ### Removed
 
 - Removed `AMDLayerNorm`, since the original layer norm bug has been fixed and we don't need this workaround anymore.
+- Removed `OLMoParallelBlock`.
 
 ### Fixed
 
 - Don't log garbage on nodes that aren't rank 0
 - Don't crash in the HF code when we are referring to a tokenizer in a local file
 
-### Added
-- Added commonsense_qa and social_iqa downstream evaluation tasks
-
-
 ## [v0.2.5](https://github.com/allenai/OLMo/releases/tag/v0.2.5) - 2024-03-06
 
 ### Fixed

Makefile

@@ -29,7 +29,7 @@ base-image :
 	docker build -f docker/Dockerfile.base -t $(IMAGE_NAME_BASE)-base .
 
 .PHONY : gantry-image
-gantry-image : base-image
+gantry-image :
 	docker build -f docker/Dockerfile.gantry -t $(IMAGE_NAME_BASE)-gantry .
 	beaker image create $(IMAGE_NAME_BASE)-gantry --name $(IMAGE_NAME_BASE)-gantry-tmp --workspace $(BEAKER_WORKSPACE)
 	beaker image delete $(GANTRY_IMAGE) || true

configs/llama7-s3.yaml

@@ -13,7 +13,7 @@ model:
   n_layers: 32
   mlp_hidden_size: 22016
   rope: true
-  flash_attention: true
+  flash_attention: false
   attention_dropout: 0.0
   attention_layer_norm: false
   multi_query_attention: false

configs/llama7.yaml

@@ -13,7 +13,7 @@ model:
   n_layers: 32
   mlp_hidden_size: 22016
   rope: true
-  flash_attention: true
+  flash_attention: false
   attention_dropout: 0.0
   attention_layer_norm: false
   multi_query_attention: false

configs/mcli/ananya-1b-ib.yaml

@@ -31,6 +31,7 @@ model:
   pad_token_id: 1
   init_device: meta
   init_fn: mitchell
+  flash_attention: true
 
 compile: null  # causes instability on AMD GPUs
 

configs/mcli/ananya-1b.yaml

@@ -31,6 +31,7 @@ model:
   pad_token_id: 1
   init_device: meta
   init_fn: normal
+  flash_attention: true
 
 compile: null  # causes instability on AMD GPUs
 

configs/mcli/mitchish-final.yaml

@@ -97,6 +97,7 @@ command: |-
     --save_interval_unsharded=10000 \
     --load_path=/root/checkpoint-unsharded \
     --compile=null \
+    --model.flash_attention=true \
     --activation_checkpointing=fine_grained \
     --fsdp.wrapping_strategy=size_based \
     --remote_save_folder=s3://ai2-llm/checkpoints/7b/${run_name} \

configs/mcli/mitchish-instruct.yml

@@ -98,6 +98,7 @@ command: |-
     --reset_trainer_state \
     --reset_optimizer_state \
     --compile=null \
+    --model.flash_attention=true \
     --activation_checkpointing=whole_layer \
     --fsdp.wrapping_strategy=size_based \
     --max_duration=5ep

configs/mcli/mitchish.yaml

@@ -37,6 +37,7 @@ command: |-
     --save_interval_unsharded=10000 \
     --load_path=${checkpoint} \
     --compile=null \
+    --model.flash_attention=true \
     --activation_checkpointing=fine_grained \
     --fsdp.wrapping_strategy=size_based \
     --remote_save_folder=s3://ai2-llm/checkpoints/7b/${run_name} \

configs/mcli/mitchish70.yaml

@@ -0,0 +1,47 @@
+name: olmo-70b
+image: mosaicml/pytorch:2.2.1_cu121-python3.11-ubuntu20.04
+compute:
+  cluster: r9z3
+  gpus: 256
+  gpu_type: h100_80gb
+integrations:
+  - integration_type: git_repo
+    git_repo: allenai/OLMo
+    git_branch: mitchish65-2
+    pip_install: -e .[train]
+    ssh_clone: true
+env_variables:
+  PIP_DISABLE_PIP_VERSION_CHECK: "1"
+  OMP_NUM_THREADS: "8"
+  LOG_FILTER_TYPE: local_rank0_only
+command: |-
+  # Install AWS CLI (for download unsharded checkpoints).
+  #apt-get update
+  #apt-get install zip unzip
+  #curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"
+  #unzip awscliv2.zip
+  #sudo ./aws/install
+
+  # Make sure we have a recent flash-attn.
+  # NOTE: only pinning flash-attn here to future proof it.
+  pip install flash-attn==2.5.3 --no-build-isolation
+
+  # Show packages for debugging.
+  pip freeze
+
+  # Prepare environment.
+  cd OLMo
+  mkdir -p /root/.cache/torch
+
+  torchrun \
+  --master_addr "$MASTER_ADDR" \
+  --master_port "$MASTER_PORT" \
+  --nnodes "$NUM_NODES" \
+  --node_rank "$NODE_RANK" \
+  --nproc_per_node 8 \
+  scripts/train.py configs/mitchish70-s3.yaml \
+    --run_name=mitchish70-001 \
+    --wandb.group=mitchish70 \
+    --global_train_batch_size=768 \
+    --device_train_microbatch_size=3 \
+    --save_overwrite

configs/mcli/v1-mix-medium-mitch-ish.yaml

@@ -28,4 +28,5 @@ command: |-
   --nproc_per_node 8 \
   scripts/train.py configs/v1-mix-medium-mitch-ish-s3.yaml \
     --run_name=v1-mix-mitch-ish \
+    --model.flash_attention=true \
     --global_train_batch_size=2160

configs/mcli/v1-mix-medium.yaml

@@ -28,5 +28,6 @@ command: |-
   --nproc_per_node 8 \
   scripts/train.py configs/v1-mix-medium-s3.yaml \
     --run_name=v1-mix-medium \
+    --model.flash_attention=true \
     --scheduler.name=linear_with_warmup \
     --global_train_batch_size=2160

configs/mcli/v1_5-mix-medium-mitch-ish.yaml

@@ -31,6 +31,7 @@ command: |-
     --run_name=v1_5-mix-mitch-ish \
     --wandb.name=v1_5-mix-mitch-ish-mcli-final \
     --global_train_batch_size=2160 \
+    --model.flash_attention=true \
     --time_limit=169200
 
 # We added these flags in order to get a final checkpoint where we decayed the LR down to 0.

configs/mcli/v1_5-mix-medium.yaml

@@ -29,4 +29,5 @@ command: |-
   scripts/train.py configs/v1_5-mix-medium-s3.yaml \
     --run_name=v1_5-mix-mcli \
     --scheduler.name=linear_with_warmup \
+    --model.flash_attention=true \
     --global_train_batch_size=2160

configs/mitchish35.yaml

@@ -16,7 +16,7 @@ model:
   weight_tying: false
   alibi: false
   rope: true
-  flash_attention: true
+  flash_attention: false
   attention_dropout: 0.0
   attention_layer_norm: false
   multi_query_attention: false

configs/mitchish50.yaml

@@ -16,7 +16,7 @@ model:
   weight_tying: false
   alibi: false
   rope: true
-  flash_attention: true
+  flash_attention: false
   attention_dropout: 0.0
   attention_layer_norm: false
   multi_query_attention: false