#unsloth/Mistral-Small-Instruct-2409
#"hf_hub_url": "teknium/trismegistus-project",
#"hf_hub_url": "AIRRC/Eudaimonic",
#"hf_hub_url": "Gryphe/Sonnet3.5-Charcard-Roleplay",
#"hf_hub_url": "anthracite-org/kalo_misc_part2",
#"hf_hub_url": "anthracite-org/kalo_opus_misc_240827",
#"hf_hub_url":"AtlasUnified/atlas-converse",




# Paths
model = '/workspace/model'
output_dir = '/workspace/out'

# Lora configuration
# can use full_fine_tune=true and no quantization to train the whole model instead of a LoRA
#full_fine_tune = true
lora_rank = 1024
lora_alpha = 256
lora_dropout = 0.05

# Train only specific modules. This is passed to the parameter of the same name in the LoraConfig.
# If not set, adapt all linear modules.
# Note, this ALSO affects full fine tuning. In that case, if this is set, only weights containing one
# of these keys as substring will have requires_grad. If not set everything is trained.
#target_modules = ['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj']

# can specify layers to adapt with LoRA if you want
#layers_to_transform = '16:31'

# for Mixtral, set the load balancing coefficient
# load_balancing_loss_coef = 0.02

# Optimization configuration
epochs = 2
lr_scheduler = 'cosine'  # can also be 'constant'
warmup_steps = 50

# might be useful if resuming from a checkpoint and you want to change the LR and force it to something
#force_constant_lr = 5e-5

# hard clamp the magnitude of the LoRA weights
#scale_weight_norms = 1.0

# dynamic batch size, targeting this many tokens per batch, per device
# if set, completely ignores the batch size in the deepspeed JSON config file
# can be thought of as a replacement for sample packing
batch_size_tokens = 10000

# Performance settings
pipeline_stages = 2  # number of pipeline parallel stages, must evenly divide the number of GPUs you launch the script with
logging_steps = 10  # how often to log in Tensorboard
eval_steps = 500
save_steps = 500
checkpoint_every_n_minutes = 60
eval_before_first_step = false  # do an eval before any training happens
# dtype to load the underlying model weights in
model_weight_dtype = 'bfloat16'
# dtype for the LoRA weights
lora_weight_dtype = 'bfloat16'
# Can have the saved weights be different dtype. Don't need to set this. Could be useful for
# training in float32 but saving with float16.
#save_dtype = 'bfloat16'
# Keep this number of stepXXXX (model saves) and global_stepXXX (checkpoint saves) and delete the rest
# (this only applies to the current training session, and resumed training sessions will not touch
# old saves)
keep_states = 5

# sort examples by length before dividing them into batches
# this makes all examples in a batch approximately the same length, to minimize padding
# the batches are still shuffled after that
# you should probably always have this set to true
group_by_length = true

# This can also be 'unsloth' to offload hidden states to CPU, saving potentially a lot of VRAM
# for a minor performance hit.
# Example: 4x4090, PCIE 3.0 16x, pipeline_stages=4, training QLoRA on Llama 3 70B with 4096 sequence length.
# true: 75s step time, 19.7G peak per-GPU VRAM usage.
# 'unsloth': 78s step time, 16.2G peak per-GPU VRAM usage.
activation_checkpointing = 'unsloth'

# Keep MLP weights on system RAM until they are needed. Can save a ton of VRAM with a
# moderate hit to performance. If using an MoE model, this can also be an integer, in
# which case only that many experts are offloaded (tradeoff between VRAM and speed).
offload_mlp_to_cpu = true

# Resume a prior run
# if true, we attempt to resume training from the most recent directory inside output_dir (the directory names are timestamps)
# so, to resume, just run the exact same command but set this to true first
resume_from_checkpoint = false

# Loading the optimizer states seems to cause some kind of unavoidable VRAM memory leak.
# It's very small, only about 0.2 GB in cases I've seen. But if you are very close to the
# limit, it can cause resuming from checkpoint to OOM. As a last resort, you can uncomment
# this to not load the optimizer states and hopefully the resumption won't OOM.
#load_optimizer_states = false


# Dataset configuration

# How to combine multiple datasets if you have more than one.
# Can be 'concatenate' or 'interleave'. Will be 'concatenate' if not set.
dataset_combination_mode = 'concatenate'
# When to stop interleaving datasets when using mode 'interleave'. Either 'first_exhausted' or 'all_exhausted'.
# Default if not set: 'first_exhausted'
# dataset_interleave_stopping_strategy = 'all_exhausted'
# Can set this lower than training, so we don't drop as many examples when trying to make equal-sized batches.
# Default if not set: same as training GAS.
eval_gradient_accumulation_steps = 1

# bitsandbytes 4 bit quantization. The parameters here become arguments to Transformers BitsAndBytesConfig.
#[quantization.bnb]
#load_in_4bit = true
#bnb_4bit_use_double_quant = false
#bnb_4bit_compute_dtype = 'bfloat16'

# HQQ quantization. The parameters here become arguments to CustomHQQConfig.
# [quantization.hqq]
# nbits = 4
# group_size = 64
# compute_dtype = 'bfloat16'

# (Optional) You can override the quant params for certain modules. This does substring matching, e.g. if 'gate_proj'
# is a substring of the full module name, anything specified overwrites the defaults in [quantization.hqq].
# [quantization.hqq.dynamic_config]
# gate_proj = {nbits = 2, group_size = 16, quant_zero = true, quant_scale = true}
# up_proj = {nbits = 2, group_size = 16, quant_zero = true, quant_scale = true}
# down_proj = {nbits = 2, group_size = 16, quant_zero = true, quant_scale = true}

[optimizer]
# options: adamw_kahan, AdamW, AdamW8bit
type = 'adamw_kahan'
lr = 5e-5
beta1 = 0.9
beta2 = 0.99
weight_decay = 0.1

[[datasets]]
# Arbitrary name, used only for separately logging eval metrics. Will be dataset0, dataset1, etc if not set.
name = 'acolyte'
dataset_type = 'axolotl'
dataset_path = './acolyte.yml'
sequence_len = 16384
eval_size = 0.01
# Relative sampling weight, when using combination mode 'interleave'. Will be 1 if not set.
sample_weight = 1