Merge remote-tracking branch 'origin/mosaic' into gptj

.gitignore

@@ -1,3 +1,4 @@
+*.pkl
 ckpts*
 .deepspeed_env
 *.jsonl

configs/train/finetune_gptj.yaml

@@ -2,14 +2,14 @@
 model_name: "EleutherAI/gpt-j-6B"
 tokenizer_name: "EleutherAI/gpt-j-6B"
 gradient_checkpointing: true
-save_name: "nomic-ai/gpt4all-gptj-multinode-deepspeed"
+save_name: "nomic-ai/gpt4all-mosaic"
 
 # dataset
 streaming: false
 num_proc: 64
-dataset_path: "data_multiplus" 
+dataset_path: "nomic-ai/turbo-500k-multi"
 max_length: 1024
-batch_size: 32
+batch_size: 8
 
 # train dynamics
 lr: 2.0e-5
@@ -23,7 +23,7 @@ output_dir: "ckpts/gpt4all-gptj-multinode"
 checkpoint: null
 lora: false
 warmup_steps: 500
-num_epochs: 4 
+num_epochs: 2 
 
 # logging
 wandb: true

configs/train/finetune_gptj_lora.yaml

@@ -0,0 +1,33 @@
+# model/tokenizer
+model_name: "EleutherAI/gpt-j-6b"
+tokenizer_name: "EleutherAI/gpt-j-6b"
+gradient_checkpointing: false
+save_name: "nomic-ai/gpt4all-mosaic"
+
+# dataset
+streaming: false
+num_proc: 64
+dataset_path: "nomic-ai/turbo-500k-multi"
+max_length: 1024
+batch_size: 4
+
+# train dynamics
+lr: 2.0e-5
+min_lr: 0 
+weight_decay: 0.0
+eval_every: 500
+eval_steps: 105
+save_every: 500
+log_grads_every: 500
+output_dir: "ckpts/gpt4all-gptj-multinode"
+checkpoint: null
+lora: true
+warmup_steps: 500
+num_epochs: 2 
+
+# logging
+wandb: true
+wandb_entity: zanussbaum
+wandb_project_name: mosaic
+seed: 42
+

configs/train/finetune_lora.yaml

@@ -7,12 +7,14 @@ save_name: "nomic-ai/gpt4all-lora-multi-turn"
 # dataset
 streaming: false
 num_proc: 64
-dataset_path: "data_multiturn"
+dataset_path: "nomic-ai/turbo-500k-multi"
 max_length: 1024
 batch_size: 4
 
 # train dynamics
 lr: 5.0e-5
+min_lr: 0
+weight_decay: 0.0
 eval_every: 2000
 eval_steps: 100
 save_every: 2000

data.py

@@ -9,10 +9,6 @@ from transformers import DefaultDataCollator
 
 def tokenize_inputs(config, tokenizer, examples):
     max_length = config["max_length"]
-    # ignore bos
-    newline_tokens = tokenizer("\n", return_tensors="pt")["input_ids"][0]
-    if newline_tokens[0] == tokenizer.bos_token_id:
-        newline_tokens = newline_tokens[1:]
 
     # hacky backward compatible
     different_eos = tokenizer.eos_token != "</s>"
@@ -22,7 +18,7 @@ def tokenize_inputs(config, tokenizer, examples):
             if response.count("</s>") > 0:
                 response = response.replace("</s>", tokenizer.eos_token)
 
-        prompt_len = len(tokenizer(prompt, return_tensors="pt")["input_ids"][0])
+        prompt_len = len(tokenizer(prompt + "\n", return_tensors="pt")["input_ids"][0])
 
         # hack if our prompt is super long
         # we need to include some labels so we arbitrarily trunacate at max_length // 2
@@ -33,7 +29,7 @@ def tokenize_inputs(config, tokenizer, examples):
             new_len = min(max_length // 2, len(prompt) // 2)
             prompt = prompt[:new_len]
             # get new prompt length
-            prompt_len = tokenizer(prompt, return_tensors="pt", max_length=max_length // 2, truncation=True).input_ids.ne(tokenizer.pad_token_id).sum().item()
+            prompt_len = tokenizer(prompt + "\n", return_tensors="pt", max_length=max_length // 2, truncation=True).input_ids.ne(tokenizer.pad_token_id).sum().item()
 
         assert prompt_len <= max_length // 2, f"prompt length {prompt_len} exceeds max length {max_length}"
 
@@ -41,11 +37,13 @@ def tokenize_inputs(config, tokenizer, examples):
                                  truncation=True, max_length=max_length, return_tensors="pt")["input_ids"].squeeze()
 
         labels = input_tokens.clone()
-        labels[:prompt_len + len(newline_tokens)] = -100
+        labels[:prompt_len] = -100
         if len(labels) < max_length:
             # pad to max_length with -100
             labels = torch.cat([labels, torch.full((max_length - len(labels),), -100)])
 
+        assert (labels == -100).sum() < len(labels), f"Labels are all -100, something wrong. prompt length {prompt_len} exceeds max length {max_length}" 
+        
         if (labels == -100).sum() == len(labels) - 1:
             print(prompt)
             print(response)

train.py

@@ -1,8 +1,6 @@
 import os
-from transformers import AutoModelForCausalLM, AutoTokenizer, get_scheduler
+from transformers import AutoModelForCausalLM, AutoTokenizer, get_scheduler, LlamaForCausalLM
-from transformers.trainer_pt_utils import get_parameter_names
 import torch
-import torch.nn as nn
 from torch.optim import AdamW
 from argparse import ArgumentParser
 from read import read_config
@@ -45,7 +43,7 @@ def train(accelerator, config):
     accelerator.print(f"Using {accelerator.num_processes} GPUs")
 
     tokenizer = AutoTokenizer.from_pretrained(config['tokenizer_name'], model_max_length=config['max_length'])
-    # llama has no pad token, set it to new token
+    # if no pad token, set it to eos
     if tokenizer.pad_token is None:
         tokenizer.pad_token = tokenizer.eos_token
 
@@ -76,21 +74,9 @@ def train(accelerator, config):
         else DummyOptim
     )
 
-    no_decay = ["bias", "LayerNorm.weight"]
-    optimizer_grouped_parameters = [
-        {
-            "params": [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)],
-            "weight_decay": config["weight_decay"],
-        },
-        {
-            "params": [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)],
-            "weight_decay": 0.0,
-        },
-    ]
-
     # karpathy doesn't decay embeddding, maybe we should exclude
     # https://github.com/karpathy/minGPT/commit/bbbdac74fa9b2e55574d70056163ffbae42310c1#diff-2075fa9c224b395be5bda85544dd36572b59c76c54562819eadadbf268602834R157s
-    optimizer = optimizer_cls(optimizer_grouped_parameters, lr=config["lr"])
+    optimizer = optimizer_cls(model.parameters(), lr=config["lr"], weight_decay=config["weight_decay"])
 
     if accelerator.state.deepspeed_plugin is not None:
         gradient_accumulation_steps = accelerator.state.deepspeed_plugin.deepspeed_config[