stress test

examples/config_tiny_llama_bench.yaml

@@ -0,0 +1,94 @@
+# /fsx/nouamane/miniconda/envs/2-1-cu121/bin/torchrun --nproc_per_node=8 run_train.py --config-file examples/config_tiny_llama.yaml
+# NANOTRON_BENCHMARK=1 CUDA_DEVICE_MAX_CONNECTIONS=1 /fsx/nouamane/miniconda/envs/2-1-cu121/bin/torchrun --nproc_per_node=8 run_train.py --config-file examples/config_tiny_llama.yaml
+checkpoints:
+  checkpoint_interval: 10000
+  checkpoints_path: checkpoints
+  checkpoints_path_is_shared_file_system: false
+  resume_checkpoint_path: null
+  save_initial_state: false
+data_stages:
+- data:
+    dataset: null # Custom dataloader will be used
+    num_loading_workers: 1
+    seed: 42
+  name: Stable Training Stage
+  start_training_step: 1
+general:
+  benchmark_csv_path: bench.csv
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: debug
+  run: dp2_tp8_seq64k
+  seed: 42
+  step: null
+lighteval: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 25
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    bos_token_id: 0
+    eos_token_id: 0
+    hidden_act: silu
+    hidden_size: 3072
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_llama_config: true
+    max_position_embeddings: 2048
+    num_attention_heads: 24
+    num_hidden_layers: 28
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_scaling: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0003
+    lr_decay_starting_step: null
+    lr_decay_steps: 13
+    lr_decay_style: cosine
+    lr_warmup_steps: 2
+    lr_warmup_style: linear
+    min_decay_lr: 1.0e-05
+  optimizer_factory:
+    adam_beta1: 0.9
+    adam_beta2: 0.95
+    adam_eps: 1.0e-08
+    name: adamW
+    torch_adam_is_fused: true
+  weight_decay: 0.01
+  zero_stage: 0
+parallelism:
+  dp: 1
+  expert_parallel_size: 1
+  pp: 1
+  pp_engine: 1f1b
+  tp: 8
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+profiler: null
+  # profiler_export_path: ./tb_logs
+tokenizer:
+  tokenizer_max_length: null
+  tokenizer_name_or_path: robot-test/dummy-tokenizer-wordlevel
+  tokenizer_revision: null
+tokens:
+  batch_accumulation_per_replica: 1
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 1
+  sequence_length: 2048
+  train_steps: 100
+  val_check_interval: 100

run_multinode.sh

@@ -1,9 +1,10 @@
 #!/bin/bash
 
 #SBATCH --job-name=smolm2-bench    # Job name
-#SBATCH --time=00:15:00
+#SBATCH --time=00:02:00
 #SBATCH --partition=hopper-prod
 #SBATCH --qos=high
+#SBATCH --reservation=huggingface_37
 
 #SBATCH -o /fsx/nouamane/projects/nanotron/logs/%j-%x.out
 
@@ -41,6 +42,7 @@ export CUDA_DEVICE_MAX_CONNECTIONS=1
 # Nanotron specific
 export NANOTRON_BENCHMARK=1
 
+
 # Print some debugging information
 echo "Master node: $MASTER_NODE"
 echo "All nodes: $NODELIST"
@@ -53,5 +55,6 @@ srun torchrun \
     --rdzv_id=$SLURM_JOB_ID \
     --rdzv_backend=c10d \
     --rdzv_endpoint=$MASTER_NODE:$MASTER_PORT \
-    run_train.py \
-    --config-file examples/config_tiny_llama.yaml
+    stress_test.py \
+    # run_train.py \
+    # --config-file examples/config_tiny_llama.yaml