Feat: Add example for Mistral (#644)
Browse files* Feat: Add example for Mistral
* chore: turn off flash
* chore: add is_mistral_derived_model
* chore: update following PR
- README.md +2 -1
- examples/mistral/config.yml +62 -0
- src/axolotl/utils/config.py +15 -2
README.md
CHANGED
|
@@ -413,9 +413,10 @@ tokenizer_legacy:
|
|
| 413 |
# this is reported to improve training speed on some models
|
| 414 |
resize_token_embeddings_to_32x:
|
| 415 |
|
| 416 |
-
# used to identify
|
| 417 |
is_falcon_derived_model:
|
| 418 |
is_llama_derived_model:
|
|
|
|
| 419 |
|
| 420 |
# whether you are training a 4-bit GPTQ quantized model
|
| 421 |
gptq: true
|
|
|
|
| 413 |
# this is reported to improve training speed on some models
|
| 414 |
resize_token_embeddings_to_32x:
|
| 415 |
|
| 416 |
+
# used to identify which the model is based on
|
| 417 |
is_falcon_derived_model:
|
| 418 |
is_llama_derived_model:
|
| 419 |
+
is_mistral_derived_model:
|
| 420 |
|
| 421 |
# whether you are training a 4-bit GPTQ quantized model
|
| 422 |
gptq: true
|
examples/mistral/config.yml
ADDED
|
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
base_model: mistralai/Mistral-7B-v0.1
|
| 2 |
+
base_model_config: mistralai/Mistral-7B-v0.1
|
| 3 |
+
model_type: MistralForCausalLM
|
| 4 |
+
tokenizer_type: LlamaTokenizer
|
| 5 |
+
is_mistral_derived_model: true
|
| 6 |
+
|
| 7 |
+
load_in_8bit: false
|
| 8 |
+
load_in_4bit: false
|
| 9 |
+
strict: false
|
| 10 |
+
|
| 11 |
+
datasets:
|
| 12 |
+
- path: mhenrichsen/alpaca_2k_test
|
| 13 |
+
type: alpaca
|
| 14 |
+
dataset_prepared_path: last_run_prepared
|
| 15 |
+
val_set_size: 0.01
|
| 16 |
+
output_dir: ./out
|
| 17 |
+
|
| 18 |
+
sequence_len: 8192
|
| 19 |
+
sample_packing:
|
| 20 |
+
pad_to_sequence_len:
|
| 21 |
+
|
| 22 |
+
wandb_project:
|
| 23 |
+
wandb_entity:
|
| 24 |
+
wandb_watch:
|
| 25 |
+
wandb_run_id:
|
| 26 |
+
wandb_log_model:
|
| 27 |
+
|
| 28 |
+
gradient_accumulation_steps: 4
|
| 29 |
+
micro_batch_size: 2
|
| 30 |
+
num_epochs: 3
|
| 31 |
+
optimizer: adamw_bnb_8bit
|
| 32 |
+
lr_scheduler: cosine
|
| 33 |
+
learning_rate: 0.0002
|
| 34 |
+
|
| 35 |
+
train_on_inputs: false
|
| 36 |
+
group_by_length: false
|
| 37 |
+
bf16: true
|
| 38 |
+
fp16: false
|
| 39 |
+
tf32: false
|
| 40 |
+
|
| 41 |
+
gradient_checkpointing: true
|
| 42 |
+
early_stopping_patience:
|
| 43 |
+
resume_from_checkpoint:
|
| 44 |
+
local_rank:
|
| 45 |
+
logging_steps: 1
|
| 46 |
+
xformers_attention:
|
| 47 |
+
flash_attention: true
|
| 48 |
+
|
| 49 |
+
warmup_steps: 10
|
| 50 |
+
eval_steps: 20
|
| 51 |
+
eval_table_size: 5
|
| 52 |
+
eval_table_max_new_tokens: 128
|
| 53 |
+
save_steps:
|
| 54 |
+
debug:
|
| 55 |
+
deepspeed:
|
| 56 |
+
weight_decay: 0.0
|
| 57 |
+
fsdp:
|
| 58 |
+
fsdp_config:
|
| 59 |
+
special_tokens:
|
| 60 |
+
bos_token: "<s>"
|
| 61 |
+
eos_token: "</s>"
|
| 62 |
+
unk_token: "<unk>"
|
src/axolotl/utils/config.py
CHANGED
|
@@ -82,7 +82,7 @@ def normalize_config(cfg):
|
|
| 82 |
cfg.is_llama_derived_model = (
|
| 83 |
(hasattr(model_config, "model_type") and model_config.model_type == "llama")
|
| 84 |
or cfg.is_llama_derived_model
|
| 85 |
-
or "llama" in cfg.base_model
|
| 86 |
or (cfg.model_type and "llama" in cfg.model_type.lower())
|
| 87 |
)
|
| 88 |
|
|
@@ -98,10 +98,23 @@ def normalize_config(cfg):
|
|
| 98 |
]
|
| 99 |
)
|
| 100 |
or cfg.is_falcon_derived_model
|
| 101 |
-
or "falcon" in cfg.base_model
|
| 102 |
or (cfg.model_type and "rwforcausallm" in cfg.model_type.lower())
|
| 103 |
)
|
| 104 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 105 |
log_gpu_memory_usage(LOG, "baseline", cfg.device)
|
| 106 |
|
| 107 |
|
|
|
|
| 82 |
cfg.is_llama_derived_model = (
|
| 83 |
(hasattr(model_config, "model_type") and model_config.model_type == "llama")
|
| 84 |
or cfg.is_llama_derived_model
|
| 85 |
+
or "llama" in cfg.base_model.lower()
|
| 86 |
or (cfg.model_type and "llama" in cfg.model_type.lower())
|
| 87 |
)
|
| 88 |
|
|
|
|
| 98 |
]
|
| 99 |
)
|
| 100 |
or cfg.is_falcon_derived_model
|
| 101 |
+
or "falcon" in cfg.base_model.lower()
|
| 102 |
or (cfg.model_type and "rwforcausallm" in cfg.model_type.lower())
|
| 103 |
)
|
| 104 |
|
| 105 |
+
cfg.is_mistral_derived_model = (
|
| 106 |
+
(
|
| 107 |
+
hasattr(model_config, "model_type")
|
| 108 |
+
and model_config.model_type
|
| 109 |
+
in [
|
| 110 |
+
"mistral",
|
| 111 |
+
]
|
| 112 |
+
)
|
| 113 |
+
or cfg.is_mistral_derived_model
|
| 114 |
+
or "mistral" in cfg.base_model.lower()
|
| 115 |
+
or (cfg.model_type and "mistral" in cfg.model_type.lower())
|
| 116 |
+
)
|
| 117 |
+
|
| 118 |
log_gpu_memory_usage(LOG, "baseline", cfg.device)
|
| 119 |
|
| 120 |
|