Add a chat data preprocessing script (#1239)

* Add a chat data preprocessing script

* add EOT at end of a chat

* update README.md

* apply pre-commit

---------

Co-authored-by: Quentin Anthony <[email protected]>

.github/workflows/pull_request.yml

@@ -52,7 +52,7 @@ jobs:
       - name: install pytest
         run: python3 -m pip install pytest pytest-forked pyyaml requests wandb
       - name: install torch
-        run: python3 -m pip install torch 
+        run: python3 -m pip install torch
       - name: install requirements
         run: pip install -r requirements/requirements.txt
       - name: Run Tests

megatron/data/helpers.cpp

@@ -428,9 +428,9 @@ py::array build_mapping_impl(const py::array_t<int64_t>& docs_,
                         }
 
                     }  // for (auto sent_index=sent_index_first; ...
-                }      // if (num_remain_sent > 1) {
-            }          // for (int doc=0; doc < num_docs; ++doc) {
-        }              // for (int epoch=0; epoch < num_epochs; ++epoch) {
+                }  // if (num_remain_sent > 1) {
+            }  // for (int doc=0; doc < num_docs; ++doc) {
+        }  // for (int epoch=0; epoch < num_epochs; ++epoch) {
 
         if (!second) {
             if (verbose) {
@@ -660,9 +660,9 @@ py::array build_blocks_mapping_impl(const py::array_t<int64_t>& docs_,
                             num_sent = 0;
                         }
                     }  // for (auto sent_index=sent_index_first; ...
-                }      // if (num_remain_sent > 1) {
-            }          // for (int doc=0; doc < num_docs; ++doc) {
-        }              // for (int epoch=0; epoch < num_epochs; ++epoch) {
+                }  // if (num_remain_sent > 1) {
+            }  // for (int doc=0; doc < num_docs; ++doc) {
+        }  // for (int epoch=0; epoch < num_epochs; ++epoch) {
 
         if (!second) {
             if (verbose) {

megatron/model/norms.py

@@ -24,6 +24,7 @@ def get_norm(neox_args):
         eps = neox_args.layernorm_epsilon
         if neox_args.layernorm_fusion:
             from .fused_layer_norm import MixedFusedLayerNorm
+
             norm = MixedFusedLayerNorm
         else:
             norm = LayerNorm

megatron/neox_arguments/arguments.py

@@ -794,7 +794,9 @@ def calculate_batch_parameters(
 
         # either none of the three parameters are provided or just gradient_accumulation_step is provided
         else:
-            assert False, "Either train_batch_size or train_micro_batch_size_per_gpu needs to be provided"
+            assert (
+                False
+            ), "Either train_batch_size or train_micro_batch_size_per_gpu needs to be provided"
         return int(train_batch), int(micro_batch), int(grad_acc)
 
     @staticmethod
@@ -1098,8 +1100,8 @@ def calculate_derived(self):
         if "flash" in self.attention_config:
             _flash_version = packaging.version.Version(version("flash-attn"))
             if self.sliding_window_width is not None:
-                assert (
-                    _flash_version >= packaging.version.Version("2.3.0")
+                assert _flash_version >= packaging.version.Version(
+                    "2.3.0"
                 ), f"Flash-Attention version ({str(_flash_version)}) must be >= 2.3.0 to support sliding window attention."
             if self.pos_emb == "alibi":
                 if not _flash_version >= packaging.version.Version("2.4.0.post1"):

tools/datasets/README.md

@@ -93,6 +93,57 @@ output data:
   --dataset-impl {lazy,cached,mmap}
                         Dataset implementation to use. Default: mmap
 
+runtime:
+  --workers WORKERS     Number of worker processes to launch
+  --log-interval LOG_INTERVAL
+                        Interval between progress updates
+```
+## `preprocess_data_with_chat_template.py`
+Similar, but uses huggingface's [chat templates](https://huggingface.co/docs/transformers/main/en/chat_templating) to
+tokenize the data to support multiturn and more complicated use cases.
+
+N.B. If using this, you  **must** specify your data when training/finetuning with the following configs
+```json
+"train_data_paths": ["train_documents"],
+"test_data_paths": ["test_documents"],
+"valid_data_paths": ["test_documents"],
+"label_data_paths": ["label_documents"]
+```
+
+the `"data_path"` option will not work with `"label_data_paths"`.
+
+
+```
+usage: preprocess_data_with_chat_template.py [-h] --input INPUT [--jsonl-keys JSONL_KEYS [JSONL_KEYS ...]] [--no-mask]
+                                             [--generation-role GENERATION_ROLE] [--only-last] [--num-docs NUM_DOCS]
+                                             --tokenizer-path TOKENIZER_PATH [--ftfy] --output-prefix OUTPUT_PREFIX
+                                             [--dataset-impl {lazy,cached,mmap}] [--workers WORKERS]
+                                             [--log-interval LOG_INTERVAL]
+
+options:
+  -h, --help            show this help message and exit
+
+input data:
+  --input INPUT         Path to input jsonl files or lmd archive(s) - if using multiple archives, put them in a comma separated list
+  --jsonl-keys JSONL_KEYS [JSONL_KEYS ...]
+                        space separate listed of keys to extract from jsonl. Default: text
+  --no-mask             If set, this will not mask any tokens in the input data.
+  --generation-role GENERATION_ROLE
+                        The role of the model generating the chat, usually 'assistant'. Default: assistant
+  --only-last           If set, this will mask everything except the last turn in the chat.
+  --num-docs NUM_DOCS   Optional: Number of documents in the input data (if known) for an accurate progress bar.
+
+tokenizer:
+  --tokenizer-path TOKENIZER_PATH
+                        Path to HF Tokenizer.
+  --ftfy                Use ftfy to clean text
+
+output data:
+  --output-prefix OUTPUT_PREFIX
+                        Path to binary output file without suffix
+  --dataset-impl {lazy,cached,mmap}
+                        Dataset implementation to use. Default: mmap
+
 runtime:
   --workers WORKERS     Number of worker processes to launch
   --log-interval LOG_INTERVAL