Wire up CLI for scan and host offloading

Author: tengyifei

.github/workflows/e2e_test.yml

@@ -17,6 +17,7 @@ jobs:
     outputs:
       llama-3-8b-name: ${{ steps.run-llama-3-8b.outputs.name }}
       llama-3_1-8b-sa-name: ${{ steps.run-llama-3_1-8b-SplashAttention.outputs.name }}
+      llama-3_1-8b-scan-offload-name: ${{ steps.run-llama-3_1-8b-scan-offload.outputs.name }}
       llama-3-8b-2d-name: ${{ steps.run-llama-3-8b-2d.outputs.name }}
       llama-3-8b-2-slice-name: ${{ steps.run-llama-3-8b-2-slice.outputs.name }}
       mixtral-8x7b-name: ${{ steps.run-mixtral-8x7b.outputs.name }}
@@ -62,6 +63,7 @@ jobs:
             dataset_config_name=wikitext-2-raw-v1 \
             profile_step=3 \
             max_steps=15
+
       - name: Run Llama 3.1 8B (Splash Attention)
         id: run-llama-3_1-8b-SplashAttention
         env:
@@ -81,6 +83,27 @@ jobs:
             dataset_config_name=wikitext-2-raw-v1 \
             profile_step=3 \
             max_steps=15
+
+      - name: Run Llama 3.1 8B (Scan + Offload)
+        id: run-llama-3_1-8b-scan-offload
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          XLA_IR_DEBUG: 1
+          XLA_HLO_DEBUG: 1
+        run: |
+          name=$(e2e_testing/gen_name.py llama-3dot1-8b-sa)
+          echo "name=$name" >> "$GITHUB_OUTPUT"
+          tp run \
+            --name $name \
+            torchprime/torch_xla_models/train.py \
+            model=llama-3.1-8b \
+            model/remat=llama-scan-offload \
+            global_batch_size=8 \
+            ici_mesh.fsdp=4 \
+            dataset_config_name=wikitext-2-raw-v1 \
+            profile_step=3 \
+            max_steps=15
+
       - name: Run Llama 3.0 8B (2D sharding)
         id: run-llama-3-8b-2d
         env:
@@ -94,7 +117,7 @@ jobs:
             --name $name \
             torchprime/torch_xla_models/train.py \
             model=llama-3-8b \
-            model/scaling=llama-fsdp-tp \
+            model/sharding=llama-fsdp-tp \
             global_batch_size=8 \
             ici_mesh.fsdp=2 \
             ici_mesh.tensor=2 \
@@ -136,7 +159,7 @@ jobs:
             --num-slices 2 \
             torchprime/torch_xla_models/train.py \
             model=llama-3-8b \
-            model/scaling=llama-fsdp \
+            model/sharding=llama-fsdp \
             global_batch_size=16 \
             dcn_mesh.fsdp=2 \
             ici_mesh.fsdp=4 \
@@ -154,14 +177,23 @@ jobs:
     secrets: inherit
 
   llama-3_1-8b-sa:
-    name: Llama 3.1 8B (Spalsh Attention)
+    name: Llama 3.1 8B (Splash Attention)
     needs: tp-run
     uses: ./.github/workflows/reusable_e2e_check.yml
     with:
       jobset_name: ${{ needs.tp-run.outputs.llama-3_1-8b-sa-name }}
       artifact_dir: ${{ needs.tp-run.outputs.artifact-dir }}
     secrets: inherit
 
+  llama-3_1-8b-scan-offload:
+    name: Llama 3.1 8B (Scan + Offload)
+    needs: tp-run
+    uses: ./.github/workflows/reusable_e2e_check.yml
+    with:
+      jobset_name: ${{ needs.tp-run.outputs.llama-3_1-8b-scan-offload-name }}
+      artifact_dir: ${{ needs.tp-run.outputs.artifact-dir }}
+    secrets: inherit
+
   llama-3-8b-2d:
     name: Llama 3.0 8B (2D sharding)
     needs: tp-run

torchprime/torch_xla_models/README.md

@@ -137,8 +137,10 @@ tp run torchprime/torch_xla_models/train.py \
 - `train.py`: Main training script that sets up the model, data, and training loop
 - `configs/base.yaml`: Configuration file for the training script
 - `configs/model`: Configuration files for models
-- `configs/model/scaling`: Configuration files for scaling the training of a model, e.g.
-  rematerialization, sharding tensors.
+- `configs/model/sharding`: Configuration files for distributing the training
+  over many chips.
+- `configs/model/remat`: Configuration files for rematerialization strategy e.g.
+  activation checkpointing, host offloading
 - `llama/model.py`: Implementation of the Llama model family
 - `mixtral/model.py`: Implementation of the Mixtral model family
 

torchprime/torch_xla_models/configs/default.yaml

@@ -47,3 +47,16 @@ dcn_mesh:
   fsdp: 1
   tensor: 1
   expert: 1
+
+# These are default values for model activation rematerialization configuration.
+# They can be overridden on the command line or by importing one of the presets
+# in the `model/remat` directory.
+model:
+  remat:
+    # If specified, compile a module of type `HomogeneousSequential` located at the
+    # given path in the module tree using `torch_xla.experimental.scan_layers`.
+    scan_layers: null
+
+    # If specified, offload these tensors to host RAM during the forward pass and
+    # move them back during the backward pass.
+    offload_tensors: []

torchprime/torch_xla_models/configs/model/llama-3-8b.yaml

@@ -1,6 +1,7 @@
 defaults:
   - _self_  # refers to this config file
-  - scaling: llama-fsdp  # refers to scaling/llama-fsdp.yaml
+  - sharding: llama-fsdp  # refers to sharding/llama-fsdp.yaml
+  - remat: llama  # refers to remat/llama.yaml
 
 model_class: llama.LlamaForCausalLM  # Used to import the model from this class
 vocab_size: 128256

torchprime/torch_xla_models/configs/model/llama-3.1-405b.yaml

@@ -1,6 +1,7 @@
 defaults:
   - _self_
-  - scaling: llama-fsdp-tp
+  - sharding: llama-fsdp-tp
+  - remat: llama
 
 model_class: llama.LlamaForCausalLM
 vocab_size: 128256

torchprime/torch_xla_models/configs/model/llama-3.1-8b.yaml

@@ -1,6 +1,7 @@
 defaults:
   - _self_  # refers to this config file
-  - scaling: llama-fsdp  # refers to scaling/llama-fsdp.yaml
+  - sharding: llama-fsdp  # refers to sharding/llama-fsdp.yaml
+  - remat: llama  # refers to remat/llama.yaml
 
 model_class: llama.LlamaForCausalLM  # Used to import the model from this class
 vocab_size: 128256
@@ -25,4 +26,4 @@ rope_scaling:
   factor: 8.0
   low_freq_factor: 1.0
   high_freq_factor: 4.0
-  original_context_len: 8192
+  original_context_len: 8192

torchprime/torch_xla_models/configs/model/mixtral-8x7b.yaml

@@ -1,6 +1,7 @@
 defaults:
   - _self_  # refers to this config file
-  - scaling: mixtral-fsdp  # refers to scaling/mixtral-fsdp.yaml
+  - sharding: mixtral-fsdp  # refers to sharding/mixtral-fsdp.yaml
+  - remat: mixtral  # refers to remat/mixtral.yaml
 
 model_class: mixtral.MixtralForCausalLM  # Used to import the model from this class
 bos_token_id: 1

torchprime/torch_xla_models/configs/model/remat/llama-scan-offload.yaml

@@ -0,0 +1,6 @@
+defaults:
+ - llama-scan
+ - _self_
+
+offload_tensors:
+ - decoder_input

torchprime/torch_xla_models/configs/model/remat/llama-scan.yaml

@@ -0,0 +1,6 @@
+defaults:
+ - llama
+ - _self_
+
+# Compile the sequence of decoder layers at `model.layers` using scan.
+scan_layers: 'model.layers'

torchprime/torch_xla_models/configs/model/remat/llama.yaml

@@ -0,0 +1,6 @@
+activation_checkpoint_layers:
+ - LlamaDecoderLayer
+
+# Refer to https://github.com/pytorch/xla/issues/6379 for backward optimization barrier info.
+optimization_barrier_layers:
+ - LlamaDecoderLayer

torchprime/torch_xla_models/configs/model/remat/mixtral.yaml

@@ -0,0 +1,6 @@
+activation_checkpoint_layers:
+ - MixtralDecoderLayer
+
+# Refer to https://github.com/pytorch/xla/issues/6379 for backward optimization barrier info.
+optimization_barrier_layers:
+ - MixtralDecoderLayer

torchprime/torch_xla_models/configs/model/scaling/llama-fsdp-tp.yaml

@@ -0,0 +1,29 @@
+# 2D (FSDP + TP) sharding configuration for Llama models.
+
+# Weights
+
+# TODO(https://github.com/AI-Hypercomputer/torchprime/issues/114): This
+# cannot be `[tensor, fsdp]`, or the gradients will sometimes become NaN.
+model.embed_tokens.weight: [fsdp, tensor]
+
+model.layers.*.self_attn.q_proj.weight: [tensor, fsdp]
+model.layers.*.self_attn.k_proj.weight: [tensor, fsdp]
+model.layers.*.self_attn.v_proj.weight: [tensor, fsdp]
+model.layers.*.self_attn.o_proj.weight: [fsdp, tensor]
+model.layers.*.mlp.gate_proj.weight: [tensor, fsdp]
+model.layers.*.mlp.up_proj.weight: [tensor, fsdp]
+model.layers.*.mlp.down_proj.weight: [fsdp, tensor]
+model.layers.*.input_layernorm.weight: [fsdp]
+model.layers.*.post_attention_layernorm.weight: [fsdp]
+model.norm.weight: [fsdp]
+lm_head.weight: [tensor, fsdp]
+
+# Activations
+model.layers.*.self_attn.q_proj: [fsdp, null, tensor]
+model.layers.*.self_attn.k_proj: [fsdp, null, tensor]
+model.layers.*.self_attn.v_proj: [fsdp, null, tensor]
+model.layers.*.self_attn.o_proj: [fsdp, null, tensor]
+model.layers.*.input_layernorm: [fsdp, null, tensor]
+model.layers.*.post_attention_layernorm: [fsdp, null, tensor]
+model.layers.*.mlp: [fsdp, null, tensor]
+model.layers.*: [fsdp, null, tensor]

torchprime/torch_xla_models/configs/model/scaling/llama-fsdp.yaml

@@ -0,0 +1,17 @@
+# Weights
+model.embed_tokens.weight: [fsdp, null]
+model.layers.*.self_attn.q_proj.weight: [fsdp, null]
+model.layers.*.self_attn.k_proj.weight: [null, fsdp]
+model.layers.*.self_attn.v_proj.weight: [null, fsdp]
+model.layers.*.self_attn.o_proj.weight: [fsdp, null]
+model.layers.*.mlp.gate_proj.weight: [fsdp, null]
+model.layers.*.mlp.up_proj.weight: [fsdp, null]
+model.layers.*.mlp.down_proj.weight: [null, fsdp]
+model.layers.*.input_layernorm.weight: [fsdp]
+model.layers.*.post_attention_layernorm.weight: [fsdp]
+model.norm.weight: [fsdp]
+lm_head.weight: [fsdp, null]
+
+# Activations
+model.layers.*: [fsdp, null, null]
+lm_head: [fsdp, null, null]

torchprime/torch_xla_models/configs/model/scaling/mixtral-fsdp.yaml

@@ -0,0 +1,18 @@
+# Weights
+model.embed_tokens.weight: ['fsdp', null]
+model.layers.*.self_attn.q_proj.weight: ['fsdp', null]
+model.layers.*.self_attn.k_proj.weight: [null, 'fsdp']
+model.layers.*.self_attn.v_proj.weight: [null, 'fsdp']
+model.layers.*.self_attn.o_proj.weight: ['fsdp', null]
+model.layers.*.block_sparse_moe.gate.weight: [null, 'fsdp']
+model.layers.*.block_sparse_moe.experts.w1: [null, null, 'fsdp']
+model.layers.*.block_sparse_moe.experts.w2: [null, 'fsdp', null]
+model.layers.*.block_sparse_moe.experts.w3: [null, null, 'fsdp']
+model.layers.*.input_layernorm.weight: ['fsdp']
+model.layers.*.post_attention_layernorm.weight: ['fsdp']
+model.norm.weight: ['fsdp']
+lm_head.weight: ['fsdp', null]
+
+# Activations
+model.layers.*[0]: [fsdp, null, null]  # Shard the first output of the decoder layer
+lm_head: [fsdp, null, null]