Implement DeepSpeed Main autotuning for NeoX

configs/autotuning_configs/small_tune.json

@@ -0,0 +1,78 @@
+{
+   "pipe-parallel-size": 1,
+   "model-parallel-size": 1,
+
+   "num-layers": 12,
+   "hidden-size": 768,
+   "num-attention-heads": 12,
+   "seq-length": 2048,
+   "max-position-embeddings": 2048,
+   "norm": "layernorm",
+   "pos-emb": "rotary",
+   "no-weight-tying": true,
+
+   "scaled-upper-triang-masked-softmax-fusion": false,
+   "bias-gelu-fusion": false,
+
+
+   "optimizer": {
+     "type": "Adam",
+     "params": {
+       "lr": 0.0006,
+       "betas": [0.9, 0.999],
+       "eps": 1.0e-8
+     }
+   },
+
+   "train_micro_batch_size_per_gpu": 1,
+   "data-impl": "mmap",
+   "split": "949,50,1",
+
+   "checkpoint-activations": true,
+   "checkpoint-num-layers": 1,
+   "partition-activations": true,
+   "synchronize-each-layer": true,
+
+   "gradient_clipping": 1.0,
+   "weight-decay": 0.0,
+   "hidden-dropout": 0.0,
+   "attention-dropout": 0.0,
+
+   "fp16": {
+     "enabled": true,
+     "loss_scale": 0,
+     "loss_scale_window": 1000,
+     "hysteresis": 2,
+     "min_loss_scale": 1
+   },
+
+   "train-iters": 320000,
+   "lr-decay-iters": 320000,
+   "distributed-backend": "nccl",
+   "lr-decay-style": "cosine",
+   "warmup": 0.01,
+   "save-interval": 10000,
+   "eval-interval": 1000,
+   "eval-iters": 10,
+
+   "log-interval": 100,
+   "steps_per_print": 10,
+   "keep-last-n-checkpoints": 4,
+   "wall_clock_breakdown": true,
+   "launcher": "slurm",
+   "deepspeed_slurm": true,
+   "comment": "neox",
+   "autotuning": {
+       "enabled": true,
+       "arg_mappings": {
+       "train_micro_batch_size_per_gpu": "--train_micro_batch_size_per_gpu",
+       "gradient_accumulation_steps ": "--gradient_accumulation_steps"
+     }
+   },
+   "zero_optimization": {
+      "stage": [0, 1, 2, 3]
+   },
+  "train-data-paths": ["/fsx/pile_deduped/pile_0.87_deduped_text_document"],
+  "valid-data-paths": ["/fsx/pile_deduped/pile_0.87_deduped_text_document"],
+  "test-data-paths": ["/fsx/pile_deduped/pile_0.87_deduped_text_document"]
+}

configs/autotuning_configs/tune.json

@@ -0,0 +1,72 @@
+{
+   "pipe-parallel-size": 1,
+   "model-parallel-size": 1,
+   "num-layers": 12,
+   "hidden-size": 768,
+   "num-attention-heads": 12,
+   "seq-length": 2048,
+   "max-position-embeddings": 2048,
+   "norm": "layernorm",
+   "pos-emb": "rotary",
+   "no-weight-tying": true,
+   "scaled-upper-triang-masked-softmax-fusion": true,
+   "bias-gelu-fusion": true,
+   "optimizer": {
+     "type": "Adam",
+     "params": {
+       "lr": 0.0006,
+       "betas": [0.9, 0.999],
+       "eps": 1.0e-8
+     }
+   },
+   "zero_optimization": {
+    "stage": 0,
+    "allgather_partitions": true,
+    "allgather_bucket_size": 500000000,
+    "overlap_comm": true,
+    "reduce_scatter": true,
+    "reduce_bucket_size": 500000000,
+    "contiguous_gradients": true,
+    "cpu_offload": false
+  },
+   "train_micro_batch_size_per_gpu": 1,
+   "autotuning_config": {
+     "enabled": true,
+     "arg_mappings": {
+       "train_micro_batch_size_per_gpu": "--train_micro_batch_size_per_gpu",
+       "gradient_accumulation_steps ": "--gradient_accumulation_steps"
+     }
+   },
+   "data-impl": "mmap",
+   "split": "949,50,1",
+   "checkpoint-activations": true,
+   "checkpoint-num-layers": 1,
+   "partition-activations": true,
+   "synchronize-each-layer": true,
+   "gradient_clipping": 1.0,
+   "weight-decay": 0.0,
+   "hidden-dropout": 0.0,
+   "attention-dropout": 0.0,
+   "fp16": {
+     "enabled": true,
+     "loss_scale": 0,
+     "loss_scale_window": 1000,
+     "hysteresis": 2,
+     "min_loss_scale": 1
+   },
+   "train-iters": 200,
+   "lr-decay-iters": 320000,
+   "distributed-backend": "nccl",
+   "lr-decay-style": "cosine",
+   "warmup": 0.01,
+   "save-interval": 10000,
+   "eval-interval": 1000,
+   "eval-iters": 10,
+   "log-interval": 100,
+   "steps_per_print": 10,
+   "keep-last-n-checkpoints": 4,
+   "wall_clock_breakdown": true,
+   "launcher": "slurm",
+   "deepspeed_slurm": true,
+   "comment": "neox"
+}

configs/autotuning_configs/tune_1-3B.json

@@ -0,0 +1,86 @@
+{
+   "pipe-parallel-size": 1,
+   "model-parallel-size": 1,
+
+   "num-layers": 24,
+   "hidden-size": 2048,
+   "num-attention-heads": 16,
+   "seq-length": 2048,
+   "max-position-embeddings": 2048,
+   "norm": "layernorm",
+   "pos-emb": "rotary",
+   "no-weight-tying": true,
+   "gpt_j_residual": false,
+   "output_layer_parallelism": "column",
+   "attention_config": [[["flash"], 24]],
+   "scaled-upper-triang-masked-softmax-fusion": false,
+   "bias-gelu-fusion": false,
+
+   "init_method": "small_init",
+   "output_layer_init_method": "wang_init",
+
+   "optimizer": {
+     "type": "Adam",
+     "params": {
+       "lr": 0.0002,
+       "betas": [0.9, 0.95],
+       "eps":  1.0e-8
+     }
+   },
+   "min_lr": 0.00002,
+
+   "zero_optimization": {
+    "stage": 1,
+    "allgather_partitions": true,
+    "allgather_bucket_size": 500000000,
+    "overlap_comm": true,
+    "reduce_scatter": true,
+    "reduce_bucket_size": 500000000,
+    "contiguous_gradients": true
+  },
+  "train_micro_batch_size_per_gpu": 1,
+   "autotuning": {
+     "enabled": true,
+     "arg_mappings": {
+       "train_micro_batch_size_per_gpu": "--train_micro_batch_size_per_gpu",
+       "gradient_accumulation_steps ": "--gradient_accumulation_steps"
+     }
+   },
+   "data-impl": "mmap",
+
+   "checkpoint-activations": false,
+   "checkpoint-num-layers": 1,
+   "partition-activations": true,
+   "synchronize-each-layer": true,
+
+   "gradient_clipping": 1.0,
+   "weight-decay": 0.1,
+   "hidden-dropout": 0,
+   "attention-dropout": 0,
+
+   "fp16": {
+     "fp16": true,
+     "enabled": true,
+     "loss_scale": 0,
+     "loss_scale_window": 1000,
+     "hysteresis": 2,
+     "min_loss_scale": 1
+   },
+
+   "train-iters": 320000,
+   "lr-decay-iters": 320000,
+   "distributed-backend": "nccl",
+   "lr-decay-style": "cosine",
+   "warmup": 0.01,
+   "checkpoint-factor": 10000,
+   "eval-interval": 1000,
+   "eval-iters": 10,
+   "launcher": "slurm",
+   "deepspeed_slurm": true,
+   "no_ssh_check": true,
+
+   "log-interval": 10,
+   "steps_per_print": 10,
+   "keep-last-n-checkpoints": 1,
+   "wall_clock_breakdown": true
+}

configs/autotuning_configs/tune_6-7B.json

@@ -0,0 +1,77 @@
+{
+   "pipe-parallel-size": 1,
+   "model-parallel-size": 8,
+
+   "num-layers": 32,
+   "hidden-size": 4096,
+   "num-attention-heads": 32,
+   "seq-length": 2048,
+   "max-position-embeddings": 2048,
+   "norm": "layernorm",
+   "pos-emb": "rotary",
+   "no-weight-tying": true,
+
+   "scaled-upper-triang-masked-softmax-fusion": false,
+   "bias-gelu-fusion": false,
+
+
+   "optimizer": {
+     "type": "Adam",
+     "params": {
+       "lr": 0.00012,
+       "betas": [0.9, 0.999],
+       "eps": 1.0e-8
+     }
+   },
+
+   "train_micro_batch_size_per_gpu": 1,
+   "zero_optimization": {
+      "stage": [0, 1, 2, 3]
+   },
+   "data-impl": "mmap",
+   "split": "949,50,1",
+
+   "checkpoint-activations": true,
+   "checkpoint-num-layers": 1,
+   "partition-activations": true,
+   "synchronize-each-layer": true,
+
+   "gradient_clipping": 1.0,
+   "weight-decay": 0,
+   "hidden-dropout": 0,
+   "attention-dropout": 0,
+
+   "fp16": {
+     "fp16": true,
+     "enabled": true,
+     "loss_scale": 0,
+     "loss_scale_window": 1000,
+     "hysteresis": 2,
+     "min_loss_scale": 1
+   },
+
+   "train-iters": 100,
+   "lr-decay-iters": 320000,
+   "distributed-backend": "nccl",
+   "lr-decay-style": "cosine",
+   "warmup": 0.01,
+   "checkpoint-factor": 10000,
+   "eval-interval": 1000,
+   "eval-iters": 10,
+   "log-interval": 100,
+   "steps_per_print": 10,
+   "keep-last-n-checkpoints": 4,
+   "wall_clock_breakdown": true,
+   "launcher": "slurm",
+   "deepspeed_slurm": true,
+   "no_ssh_check": true,
+   "comment": "neox",
+   "autotuning": {
+       "enabled": true,
+       "mp_size": 8,
+       "arg_mappings": {
+       "train_micro_batch_size_per_gpu": "--train_micro_batch_size_per_gpu",
+       "gradient_accumulation_steps ": "--gradient_accumulation_steps"
+     }
+   }
+}

configs/neox_arguments.md

@@ -592,7 +592,7 @@ Optimizer Arguments
 
 
 
-- **zero_stage**: int
+- **zero_stage**: typing.Union[int, typing.List[int], typing.Literal['all']]
 
     Default = None
 
@@ -1732,6 +1732,14 @@ Args for deepspeed config
 
 
 
+- **autotuning**: dict
+
+    Default = None
+
+    Dictionary as described in DeepSpeed autotuning documentation: https://github.com/microsoft/DeepSpeed/tree/master/deepspeed/autotuning
+
+
+
 ## NeoXArgsDeepspeedRunner
 
 Args for deepspeed runner (deepspeed.launcher.runner).
@@ -1801,7 +1809,7 @@ Args for deepspeed runner (deepspeed.launcher.runner).
 
 
 
-- **launcher**: str
+- **launcher**: typing.Literal['pdsh', 'openmpi', 'mvapich', 'slurm']
 
     Default = pdsh
 
@@ -1817,6 +1825,12 @@ Args for deepspeed runner (deepspeed.launcher.runner).
 
 
 
+- **autotuning_run**: str
+
+    Default = None
+
+    Either "tune", "run", or `None`.
+
 - **no_ssh_check**: bool
 
     Default = False
@@ -1831,3 +1845,11 @@ Args for deepspeed runner (deepspeed.launcher.runner).
 
     Adds a `--comment` to the DeepSpeed launch command. In DeeperSpeed this is passed on to the SlurmLauncher as well. Sometime necessary for cluster rules, or so I've heard.
 
+
+
+- **no_ssh_check**: bool
+
+    Default = False
+
+    If `True` and running with multiple nodes, then DeepSpeedd doesn't conduct a check to ensure the head node is reachable with ssh.
+

configs/slurm_local.json

@@ -0,0 +1,11 @@
+{
+  "vocab-file": "data/gpt2-vocab.json",
+  "merge-file": "data/gpt2-merges.txt",
+  "save": "checkpoints",
+  "checkpoint_validation_with_forward_pass": false,
+  "tensorboard-dir": "tensorboard",
+  "log-dir": "logs",
+  "use_wandb": true,
+  "wandb_host": "https://api.wandb.ai",
+  "wandb_project": "neox"
+}