Add training callback to send predictions to WandB table (#521)

* WIP Add training callback to send predictions to WandB table

* WIP improve wandb table reporting callback

* WIP improve wandb table reporting callback (cont)

* Add VSCode launching for debugging

* Add tiny llama example

* WIP attempt to improve post-eval prediction generation for table

* WIP attempt to improve post-eval prediction generation for table - part 2

* WIP batch generation

* WIP attempt to handle sample_packing using position_ids for wandb prediction table

* WIP add code for debugging

* Fix sample_packing support for wandb prediction table

* Clean up code for PR review

* Add eval_table_size, eval_table_max_new_tokens configs & clean up code

* Clean up PR, delete VSCode config, add tiny-llama example

* Add eval_table_size, eval_table_max_new_tokens documentation. Fix linting/formatting

README.md

@@ -534,6 +534,9 @@ eval_steps: # leave empty to eval at each epoch
 save_total_limit: # checkpoints saved at a time
 max_steps:
 
+eval_table_size: # approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
+eval_table_max_new_tokens: # total number of tokens generated for predictions sent to wandb. Default is 128
+
 # save model as safetensors (require safetensors package)
 save_safetensors:
 

examples/llama-2/lora.yml

@@ -56,6 +56,8 @@ flash_attention: true
 
 warmup_steps: 10
 eval_steps: 20
+eval_table_size: 5
+eval_table_max_new_tokens: 128
 save_steps:
 debug:
 deepspeed:

examples/llama-2/qlora.yml

@@ -58,6 +58,7 @@ flash_attention: true
 
 warmup_steps: 10
 eval_steps: 20
+eval_table_size: 5
 save_steps:
 debug:
 deepspeed:

examples/llama-2/tiny-llama.yml

@@ -0,0 +1,69 @@
+base_model: PY007/TinyLlama-1.1B-step-50K-105b
+base_model_config: PY007/TinyLlama-1.1B-step-50K-105b
+
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 4096
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+eval_steps: 20
+eval_table_size: 5
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -193,7 +193,7 @@ def flashattn_forward(
         # only on first autoregressive step q,k,v have same seqlen
         is_causal = key_states.shape == query_states.shape
 
-    if cu_seqlens is not None and max_seqlen is not None:
+    if cu_seqlens is not None and max_seqlen is not None and cu_seqlens.dim() == 1:
         # special handling using sample packing
         qkv = torch.stack(
             [query_states, key_states, value_states], dim=2
@@ -261,6 +261,8 @@ def flashattn_forward(
                 if attention_mask is not None
                 else None,
             )
+            if q_unpad.dtype != kv_unpad.dtype:
+                kv_unpad = kv_unpad.to(q_unpad.dtype)
             output_unpad = flash_attn_varlen_kvpacked_func(
                 q_unpad,
                 kv_unpad,

src/axolotl/utils/callbacks.py

@@ -11,10 +11,13 @@ import numpy as np
 import pandas as pd
 import torch
 import torch.distributed as dist
+import wandb
 from datasets import load_dataset
 from optimum.bettertransformer import BetterTransformer
 from tqdm import tqdm
 from transformers import (
+    GenerationConfig,
+    Trainer,
     TrainerCallback,
     TrainerControl,
     TrainerState,
@@ -323,3 +326,191 @@ def bench_eval_callback_factory(trainer, tokenizer):
                 metrics[key] = val
 
     return BenchEvalCallback
+
+
+def log_prediction_callback_factory(trainer: Trainer, tokenizer):
+    class LogPredictionCallback(TrainerCallback):
+        """Callback to log prediction values during each evaluation"""
+
+        def __init__(self, cfg):
+            self.cfg = cfg
+            self.logged = False
+
+        def on_evaluate(
+            self,
+            args: AxolotlTrainingArguments,  # pylint: disable=unused-argument
+            state: TrainerState,
+            control: TrainerControl,
+            train_dataloader,  # pylint: disable=unused-argument
+            eval_dataloader,
+            **kwargs,  # pylint: disable=unused-argument
+        ):
+            eval_table_size = self.cfg.eval_table_size
+
+            if eval_table_size <= 0:
+                return control
+
+            trainer.model.eval()
+            device = torch.device(self.cfg.device)
+
+            # pylint: disable=duplicate-code
+            generation_config = GenerationConfig(
+                max_new_tokens=self.cfg.eval_table_max_new_tokens,
+                bos_token_id=tokenizer.bos_token_id,
+                eos_token_id=tokenizer.eos_token_id,
+                pad_token_id=tokenizer.pad_token_id,
+                do_sample=False,
+                use_cache=True,
+                return_dict_in_generate=True,
+                output_attentions=False,
+                output_hidden_states=False,
+                output_scores=False,
+            )
+
+            def logits_to_tokens(logits) -> str:
+                probabilities = torch.softmax(logits, dim=-1)
+                # Get the predicted token ids (the ones with the highest probability)
+                predicted_token_ids = torch.argmax(probabilities, dim=-1)
+                return predicted_token_ids
+
+            def find_ranges(lst):
+                ranges = []
+                start = 0
+                for i in range(1, len(lst)):
+                    if lst[i] == 0:
+                        ranges.append((start, i - 1))
+                        start = i
+                end = len(lst) - 1
+                ranges.append((start, end))
+                return ranges
+
+            def log_table_from_dataloader(name: str, table_dataloader):
+                table = wandb.Table(
+                    columns=[
+                        "id",
+                        "Prompt",
+                        "Correct Completion",
+                        "Predicted Completion (model.generate)",
+                        "Predicted Completion (trainer.prediction_step)",
+                    ]
+                )
+                row_index = 0
+
+                for batch in tqdm(table_dataloader):
+                    if row_index > eval_table_size:
+                        break
+
+                    batch_labels = batch["labels"].to(device)
+                    batch_input_ids = batch["input_ids"].to(device)
+
+                    if "position_ids" in batch:
+                        batch_pos_ids = batch["position_ids"].tolist()
+                    else:
+                        batch_pos_ids = [None] * len(batch["input_ids"])
+
+                    (_, batch_logits, _) = trainer.prediction_step(
+                        trainer.model,
+                        batch,
+                        prediction_loss_only=False,
+                    )
+
+                    prompt_token_ids_list = []
+                    pred_step_token_ids_list = []
+                    completion_token_ids_list = []
+
+                    for input_ids_all, labels_all, pos_ids, logits in zip(
+                        batch_input_ids,
+                        batch_labels,
+                        batch_pos_ids,
+                        batch_logits,
+                    ):
+                        if pos_ids is None:
+                            pos_ranges = [(0, len(input_ids_all) - 1)]
+                        else:
+                            pos_ranges = find_ranges(pos_ids)
+
+                        for pos_range in pos_ranges:
+                            start, end = pos_range
+                            if start == end:
+                                continue
+
+                            input_ids = input_ids_all[start : end + 1]
+                            labels = labels_all[start : end + 1]
+
+                            tokens_without_loss = labels == IGNORE_INDEX
+                            tokens_with_loss = labels != IGNORE_INDEX
+                            tokens_exclude_padding = input_ids != tokenizer.pad_token_id
+                            prompt_token_includes = (
+                                tokens_without_loss & tokens_exclude_padding
+                            )
+
+                            prompt_token_ids = input_ids[prompt_token_includes]
+                            prompt_token_ids_list.append(prompt_token_ids)
+
+                            completion_token_ids = input_ids[tokens_with_loss]
+                            completion_token_ids_list.append(completion_token_ids)
+
+                            pred_step_token_ids = logits_to_tokens(
+                                logits[start : end + 1]
+                            )[tokens_with_loss]
+                            pred_step_token_ids_list.append(pred_step_token_ids)
+
+                    prompt_texts = tokenizer.batch_decode(
+                        prompt_token_ids_list, skip_special_tokens=True
+                    )
+                    completion_texts = tokenizer.batch_decode(
+                        completion_token_ids_list, skip_special_tokens=True
+                    )
+                    pred_step_texts = tokenizer.batch_decode(
+                        pred_step_token_ids_list, skip_special_tokens=True
+                    )
+
+                    with torch.no_grad():
+                        prompt_encoding = tokenizer(
+                            prompt_texts, padding=True, return_tensors="pt"
+                        ).to(self.cfg.device)
+                        predictions = trainer.model.generate(
+                            **prompt_encoding, generation_config=generation_config
+                        )
+
+                    prediction_all_tokens = predictions["sequences"].cpu().tolist()
+                    prediction_without_prompt_tokens_list = []
+                    for prompt_token_ids, prediction_tokens in zip(
+                        prompt_token_ids_list, prediction_all_tokens
+                    ):
+                        prediction_without_prompt_tokens = prediction_tokens[
+                            len(prompt_token_ids) :
+                        ]
+                        prediction_without_prompt_tokens_list.append(
+                            prediction_without_prompt_tokens
+                        )
+
+                    predicted_texts = tokenizer.batch_decode(
+                        prediction_without_prompt_tokens_list, skip_special_tokens=True
+                    )
+
+                    for (
+                        prompt_text,
+                        completion_text,
+                        prediction_text,
+                        pred_step_text,
+                    ) in zip(
+                        prompt_texts, completion_texts, predicted_texts, pred_step_texts
+                    ):
+                        table.add_data(
+                            row_index,
+                            prompt_text,
+                            completion_text,
+                            prediction_text,
+                            pred_step_text,
+                        )
+                        row_index += 1
+
+                wandb.run.log({f"{name} - Predictions vs Ground Truth": table})
+
+            if is_main_process():
+                log_table_from_dataloader("Eval", eval_dataloader)
+
+            return control
+
+    return LogPredictionCallback

src/axolotl/utils/config.py

@@ -48,6 +48,8 @@ def normalize_config(cfg):
     )
     cfg.world_size = int(os.environ.get("WORLD_SIZE", 1))
     cfg.local_rank = int(os.environ.get("LOCAL_RANK", 0))
+    cfg.eval_table_size = cfg.eval_table_size or 0
+    cfg.eval_table_max_new_tokens = cfg.eval_table_max_new_tokens or 128
     choose_device(cfg)
     cfg.ddp = cfg.ddp if cfg.ddp is not None else cfg.world_size != 1
     if cfg.ddp:

src/axolotl/utils/models.py

@@ -296,10 +296,10 @@ def load_model(
     if (
         hasattr(model.config, "max_position_embeddings")
         and model.config.max_position_embeddings
-        and cfg.sequence_len >= model.config.max_position_embeddings
+        and cfg.sequence_len > model.config.max_position_embeddings
     ):
         LOG.warning(
-            f"increasing model.config.max_position_embeddings to {cfg.sequence_len}"
+            f"increasing model.config.max_position_embeddings from {model.config.max_position_embeddings} to {cfg.sequence_len}"
         )
         model.config.max_position_embeddings = cfg.sequence_len
 

src/axolotl/utils/trainer.py

@@ -30,6 +30,7 @@ from axolotl.utils.callbacks import (
     SaveBetterTransformerModelCallback,
     SavePeftModelCallback,
     bench_eval_callback_factory,
+    log_prediction_callback_factory,
 )
 from axolotl.utils.collators import DataCollatorForSeq2Seq
 from axolotl.utils.dataloader import MultipackDistributedDataloader
@@ -703,6 +704,10 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         **trainer_kwargs,
     )
 
+    if cfg.use_wandb and cfg.eval_table_size > 0:
+        LogPredictionCallback = log_prediction_callback_factory(trainer, tokenizer)
+        trainer.add_callback(LogPredictionCallback(cfg))
+
     if cfg.do_bench_eval:
         trainer.add_callback(bench_eval_callback_factory(trainer, tokenizer))