add configs

attention/gqa_16groups.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-gqa-16groups
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 2 #from 4 to 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-gqa-16groups
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

attention/gqa_2groups.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-gqa-2groups
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 15
+    num_key_value_heads: 16 #from 4 to 16
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-gqa-2groups
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

attention/gqa_4groups.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-gqa-4groups
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8 #from 4 to 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-gqa-4groups
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

attention/gqa_4groups_compensate.yaml

@@ -0,0 +1,166 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset: 
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-gqa-4groups-compensate
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 1920
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8 #from 4 to 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    z_loss_coefficient: 1.0e-05
+    z_loss_enabled: false
+    # no_rope_layer: 4
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-gqa-4groups-compensate
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: /fsx/loubna/projects_v2/smollm3/tokenizer/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: A
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/smollm3_eval_decay.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null  # Optional reservation name, can be null

attention/gqa_8groups.yaml

@@ -0,0 +1,166 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-gqa-8groups
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 4
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    z_loss_coefficient: 1.0e-05
+    z_loss_enabled: false
+    # no_rope_layer: 4
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-gqa-8groups
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: /fsx/loubna/projects_v2/smollm3/tokenizer/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 8  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/smollm3_eval_decay.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null  # Optional reservation name, can be null

attention/mha.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-mha-14layers
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 14
+    num_key_value_heads: 32
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-mha-14layers
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

attention/mqa.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-mqa
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 1
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-mqa
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

baseline_config_1B.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

baseline_config_3B.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-3B
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 11008
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 16
+    num_hidden_layers: 36
+    num_key_value_heads: 4
+    pad_token_id: null
+    pretraining_tp: 2
+    rms_norm_eps: 1.0e-06
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0002
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 2.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 4
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 2
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-3B
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 48
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3
+  sequence_length: 4096
+  train_steps: 30000
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 4
+    pp: 1
+    tp: 2
+    tp_linear_async_communication: true
+  batch_size: 2  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/smollm3_eval_decay.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null  

batch_size/gbs_3M.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 500
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-gbs-3M
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 1000
+    lr_decay_steps: 14000
+    lr_decay_style: cosine
+    lr_warmup_steps: 1000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-gbs-3M
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 15000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 500
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

batch_size/gbs_6M.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 500
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-gbs-6M
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 500
+    lr_decay_steps: 7000
+    lr_decay_style: cosine
+    lr_warmup_steps: 500
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 16
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-gbs-6M
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 32
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 7500 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: loubnabnl
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 500
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "low"
+    time: "01:59:00"
+    reservation: null

batch_size/gbs_700M.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 2000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-gbs-750M
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 4000
+    lr_decay_steps: 56000
+    lr_decay_style: cosine
+    lr_warmup_steps: 4000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 16
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-gbs-750M
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 4
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 60000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: loubnabnl
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 2000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

doc_masking/doc_masking.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-doc-masking
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: true
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-doc-masking
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

learning_rate/lr_high.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-lr-5e-3
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-lr-5e-3
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

learning_rate/lr_low.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-lr-1e-4
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0001
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8  
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-lr-1e-4
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16  # changed from 8 to match lr_high
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8  
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation 
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

learning_rate/lr_medium.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 2  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

learning_rate/lr_very_high.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-lr-5e-2
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.05
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-lr-5e-2
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

learning_rate/wsd_10p.yaml

@@ -0,0 +1,166 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-wsd-10p
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    # z_loss_coefficient: 1.0e-05
+    # z_loss_enabled: false
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 27000
+    lr_decay_steps: 3000 # 10% decay
+    lr_decay_style: linear
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 0
+  optimizer_factory:
+    adam_beta1: 0.9
+    adam_beta2: 0.95
+    adam_eps: 1.0e-08
+    name: adamW
+    torch_adam_is_fused: true
+  weight_decay: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-wsd-10p
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

learning_rate/wsd_20p.yaml

@@ -0,0 +1,166 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-wsd-20p
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    # z_loss_coefficient: 1.0e-05
+    # z_loss_enabled: false
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 24000
+    lr_decay_steps: 6000 # 20% decay
+    lr_decay_style: linear
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 0
+  optimizer_factory:
+    adam_beta1: 0.9
+    adam_beta2: 0.95
+    adam_eps: 1.0e-08
+    name: adamW
+    torch_adam_is_fused: true
+  weight_decay: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-wsd-20p
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

nope/nope.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-nope
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: 4
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 16
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-nope
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 8
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: loubnabnl
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

nope/nope_doc_mask.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-nope-doc-mask
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: true
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: 4
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 16
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-nope-doc-mask
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 8
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: loubnabnl
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

requirements.txt

@@ -0,0 +1,2 @@
+git+https://github.com/huggingface/nanotron.git@smollm3
+git+https://github.com/huggingface/datatrove.git@nouamane/avoid-s3

tie_embed/baseline.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

tie_embed/no_tie_embeddings.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 5
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-no-tie-16l-seed5
+  seed: 5
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: false
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 16
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-no-tie-16l-seed5
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 8
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: loubnabnl
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

tie_embed/notie_same_size_as_baseline.yaml

@@ -0,0 +1,164 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-no-tie-12layers
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 12
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: false
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-no-tie-12layers
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: loubnabnl
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

wd_on_embeddings/no_wd.yaml

@@ -0,0 +1,165 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-no-wd
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params:
+  - .*token_embedding.*
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-no-wd
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

wd_on_embeddings/no_wd_nope_docmask.yaml

@@ -0,0 +1,165 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-no-wd-nope-docmask
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: true
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    no_rope_layer: 4
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params:
+  - .*token_embedding.*
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 8
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-no-wd-nope-docmask
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 16
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 3 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: huggingface
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "normal"
+    time: "01:59:00"
+    reservation: null

wd_on_embeddings/zloss.yaml

@@ -0,0 +1,166 @@
+checkpoints:
+  checkpoint_interval: 1000
+  checkpoints_path: /scratch/loubna/checkpoints
+  checkpoints_path_is_shared_file_system: false
+  load_lr_scheduler: true
+  load_optimizer: true
+  resume_checkpoint_path: null
+  save_final_state: true
+  save_initial_state: false
+data_stages:
+- data:
+    dataset:
+      dataset_read_path:
+      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
+      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
+      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
+      dataset_folder:
+      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
+      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
+      dataset_weights:
+      - 0.7
+      - 0.1
+      - 0.2
+      pad_samples_to_global_batch_size: false
+      return_positions: true
+      token_size_in_bytes: 4
+      tokenizer_name: HuggingFaceTB/SmolLM3-3B
+      use_old_brrr_dataloader: false
+      vocab_size: 128256
+    num_loading_workers: 0
+    seed: 6
+  name: stable
+  start_training_step: 1
+general:
+  benchmark_csv_path: null
+  consumed_train_samples: null
+  ignore_sanity_checks: true
+  project: smollm3-blog
+  run: baseline-zloss
+  seed: 6
+  step: null
+logging:
+  iteration_step_info_interval: 1
+  log_level: info
+  log_level_replica: info
+model:
+  ddp_bucket_cap_mb: 50
+  dtype: bfloat16
+  init_method:
+    std: 0.02
+  make_vocab_size_divisible_by: 1
+  model_config:
+    _attn_implementation: flash_attention_2
+    _fused_rms_norm: true
+    _fused_rotary_emb: true
+    _use_doc_masking: false
+    _use_qkv_packed: true
+    attention_bias: false
+    bos_token_id: 128000
+    eos_token_id: 128001
+    flex_attention_mask: null
+    hidden_act: silu
+    hidden_size: 2048
+    initializer_range: 0.02
+    intermediate_size: 8192
+    is_qwen2_config: true
+    max_position_embeddings: 4096
+    moe_config: null
+    num_attention_heads: 32
+    num_hidden_layers: 16
+    num_key_value_heads: 8
+    pad_token_id: null
+    pretraining_tp: 1
+    rms_norm_eps: 1.0e-05
+    rope_interleaved: false
+    rope_scaling: null
+    rope_theta: 50000.0
+    sliding_window_size: null
+    tie_word_embeddings: true
+    use_cache: true
+    vocab_size: 128256
+    z_loss_coefficient: 1.0e-05
+    z_loss_enabled: true
+    no_rope_layer: null
+optimizer:
+  accumulate_grad_in_fp32: true
+  clip_grad: 1.0
+  learning_rate_scheduler:
+    learning_rate: 0.0005
+    lr_decay_starting_step: 2000
+    lr_decay_steps: 28000
+    lr_decay_style: cosine
+    lr_warmup_steps: 2000
+    lr_warmup_style: linear
+    min_decay_lr: 5.0e-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: 0.1
+  weight_decay_exclude_named_params: null
+  zero_stage: 0
+parallelism:
+  context_parallel_size: 1
+  dp: 16
+  expert_parallel_size: 1
+  moe_layer_recompute: false
+  pp: 1
+  pp_engine: 1f1b
+  recompute_layer: false
+  tp: 1
+  tp_linear_async_communication: true
+  tp_mode: REDUCE_SCATTER
+  tp_recompute_allgather: true
+profiler: null
+s3_upload:
+  remove_after_upload: true
+  s5cmd_concurrency: 5
+  s5cmd_numworkers: 16
+  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
+  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-zloss
+tokenizer:
+  tokenizer_max_length: 4096
+  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
+  tokenizer_revision: null
+metrics_logging:
+  log_level: 1
+  log_detail_interval: 100
+tokens:
+  batch_accumulation_per_replica: 12
+  limit_test_batches: 0
+  limit_val_batches: 0
+  micro_batch_size: 2 # 1.5M tokens
+  sequence_length: 4096
+  train_steps: 30000 # 30B tokens
+  val_check_interval: 100
+lighteval:
+  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
+  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
+  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
+  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
+  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
+  s3_save_path: "s3://smollm3/blogpost-evals"
+  upload_to_wandb: true
+  wandb_project: smollm3-blog-evals
+  wandb_entity: loubnabnl
+  parallelism:
+    dp: 8
+    pp: 1
+    tp: 1
+    tp_linear_async_communication: true
+  batch_size: 1  # Optional batch size for evaluation
+  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
+  eval_interval: 1000
+  eval_interval_file: null
+  slurm:
+    gpus_per_node: 8
+    partition: "hopper-prod"
+    hf_cache: "/fsx/loubna/.cache/huggingface"
+    cpus_per_task: 88
+    qos: "low"
+    time: "01:59:00"
+    reservation: null