dump snapshot location for caching

* bumping version, removing non-osx compatible deps

* updating pylintrc

* fixing linters

* reverting changes

.pre-commit-config.yaml

@@ -23,7 +23,7 @@ repos:
     hooks:
     - id: flake8
 -   repo: https://github.com/PyCQA/pylint
-    rev: v2.17.4
+    rev: v3.3.0
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy

.pylintrc

@@ -1,5 +1,5 @@
 [MASTER]
-init-hook="from pylint.config import find_pylintrc; import os, sys; sys.path.append(os.path.dirname(find_pylintrc()))"
+init-hook="from pylint.config import find_default_config_files; import sys; sys.path.append(next(find_default_config_files()).parent.as_posix())"
 
 [TYPECHECK]
 
@@ -12,3 +12,4 @@ generated-members=numpy.*, torch.*
 disable=missing-function-docstring, line-too-long, import-error,
     too-many-arguments, too-many-locals, too-many-statements, too-many-branches, too-few-public-methods,
     too-many-instance-attributes, fixme, import-outside-toplevel, logging-fstring-interpolation,
+    too-many-positional-arguments, possibly-used-before-assignment

setup.py

@@ -1,4 +1,5 @@
 """setup.py for axolotl"""
+
 import ast
 import os
 import platform
@@ -29,15 +30,29 @@ def parse_requirements():
             elif not is_extras and line and line[0] != "#":
                 # Handle standard packages
                 _install_requires.append(line)
-
     try:
         xformers_version = [req for req in _install_requires if "xformers" in req][0]
         torchao_version = [req for req in _install_requires if "torchao" in req][0]
         autoawq_version = [req for req in _install_requires if "autoawq" in req][0]
-
         if "Darwin" in platform.system():
-            # don't install xformers on MacOS
+            # skip packages not compatible with OSX
-            _install_requires.pop(_install_requires.index(xformers_version))
+            skip_packages = [
+                "bitsandbytes",
+                "triton",
+                "mamba-ssm",
+                "flash-attn",
+                "xformers",
+                "autoawq",
+                "liger-kernel",
+            ]
+            _install_requires = [
+                req
+                for req in _install_requires
+                if re.split(r"[>=<]", req)[0].strip() not in skip_packages
+            ]
+            print(
+                _install_requires, [req in skip_packages for req in _install_requires]
+            )
         else:
             # detect the version of torch already installed
             # and set it so dependencies don't clobber the torch version

src/axolotl/core/trainer_builder.py

@@ -424,11 +424,6 @@ class SchedulerMixin(Trainer):
 
         return self.lr_scheduler
 
-    def _load_optimizer_and_scheduler(self, checkpoint):
-        if not checkpoint and self.args.optimizer_checkpoint is not None:
-            checkpoint = self.args.optimizer_checkpoint
-        return super()._load_optimizer_and_scheduler(checkpoint)
-
 
 class AxolotlTrainer(SchedulerMixin, Trainer):
     """
@@ -1769,10 +1764,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         ] = self.cfg.loraplus_lr_embedding
         training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
         training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
-        if self.cfg.optimizer_checkpoint:
-            training_arguments_kwargs[
-                "optimizer_checkpoint"
-            ] = self.cfg.optimizer_checkpoint
 
         if self.cfg.lr_scheduler in ["one_cycle", "log_sweep"]:
             training_arguments_kwargs["lr_scheduler_type"] = "cosine"

src/axolotl/utils/callbacks/lisa.py

@@ -43,7 +43,7 @@ def lisa_callback_factory(trainer: "AxolotlTrainer"):
                 getattr, self.layers_attribute.split("."), self.trainer.model
             )
             LOG.info(
-                f"LISA will activate {self.n_layers}/{len(layers)} layers ({self.n_layers*100/len(layers)}%) every {self.step_interval} steps"
+                f"LISA will activate {self.n_layers}/{len(layers)} layers ({self.n_layers * 100 / len(layers)}%) every {self.step_interval} steps"
             )
 
         def freeze_all_layers(self):

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -603,8 +603,6 @@ class AxolotlInputConfig(
     strict: Optional[bool] = Field(default=False)
     resume_from_checkpoint: Optional[str] = None
     auto_resume_from_checkpoints: Optional[bool] = None
-    optimizer_checkpoint: Optional[str] = None
-
     resize_token_embeddings_to_32x: Optional[bool] = None
     mean_resizing_embeddings: Optional[bool] = False
 

src/axolotl/utils/data/pretraining.py

@@ -28,8 +28,10 @@ def encode_pretraining(
     )
     # Convert to PyTorch tensors
     input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
+    targets = [torch.tensor(seq) for seq in res["input_ids"]]
     attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
     new_input_ids = []
+    new_labels = []
     new_attention_mask = []
     # Append EOS and PAD tokens to input_ids, and correct attention_mask
     for i, _ in enumerate(input_ids):
@@ -40,22 +42,34 @@ def encode_pretraining(
             ),
             dim=0,
         )
+        targets[i] = torch.cat(
+            (
+                targets[i],
+                torch.tensor([tokenizer.eos_token_id, -100]),
+            ),
+            dim=0,
+        )
         attention_mask[i] = torch.cat((attention_mask[i], torch.tensor([1, 0])), dim=0)
 
     # Concatenate tokens so that their lengths are less than max_tokens
     buffer_input_ids = torch.tensor([], dtype=torch.long)
+    buffer_labels = torch.tensor([], dtype=torch.long)
     buffer_attention_mask = torch.tensor([], dtype=torch.long)
 
-    for ids, mask in zip(input_ids, attention_mask):
+    for ids, labels, mask in zip(input_ids, targets, attention_mask):
         if buffer_input_ids.numel() == max_tokens:
             new_input_ids.append(buffer_input_ids)
+            new_labels.append(buffer_labels)
             new_attention_mask.append(buffer_attention_mask)
             buffer_input_ids = torch.tensor([], dtype=torch.long)
+            buffer_labels = torch.tensor([], dtype=torch.long)
             buffer_attention_mask = torch.tensor([], dtype=torch.long)
             buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
+            buffer_labels = torch.cat((buffer_labels, labels), dim=0)
             buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
         elif buffer_input_ids.numel() + ids.numel() <= max_tokens:
             buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
+            buffer_labels = torch.cat((buffer_labels, labels), dim=0)
             buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
         else:
             buffer_input_ids = torch.cat(
@@ -69,6 +83,17 @@ def encode_pretraining(
                 ),
                 dim=0,
             )
+            buffer_labels = torch.cat(
+                (
+                    buffer_labels,
+                    torch.full(
+                        (max_tokens - buffer_labels.numel(),),
+                        -100,
+                        dtype=torch.long,
+                    ),
+                ),
+                dim=0,
+            )
             buffer_attention_mask = torch.cat(
                 (
                     buffer_attention_mask,
@@ -81,11 +106,14 @@ def encode_pretraining(
                 dim=0,
             )
             new_input_ids.append(buffer_input_ids)
+            new_labels.append(buffer_labels)
             new_attention_mask.append(buffer_attention_mask)
             buffer_input_ids = torch.tensor([], dtype=torch.long)
+            buffer_labels = torch.tensor([], dtype=torch.long)
             buffer_attention_mask = torch.tensor([], dtype=torch.long)
 
             buffer_input_ids = torch.cat((buffer_input_ids, ids), dim=0)
+            buffer_labels = torch.cat((buffer_labels, labels), dim=0)
             buffer_attention_mask = torch.cat((buffer_attention_mask, mask), dim=0)
 
     if buffer_input_ids.numel() > 0:  # for any leftover tokens
@@ -101,6 +129,17 @@ def encode_pretraining(
                 ),
                 dim=0,
             )
+            buffer_labels = torch.cat(
+                (
+                    buffer_labels,
+                    torch.full(
+                        (max_tokens - buffer_labels.numel(),),
+                        -100,
+                        dtype=torch.long,
+                    ),
+                ),
+                dim=0,
+            )
             buffer_attention_mask = torch.cat(
                 (
                     buffer_attention_mask,
@@ -113,11 +152,12 @@ def encode_pretraining(
                 dim=0,
             )
         new_input_ids.append(buffer_input_ids)
+        new_labels.append(buffer_labels)
         new_attention_mask.append(buffer_attention_mask)
 
     ret = {
         "input_ids": [seq.tolist() for seq in new_input_ids],
-        "labels": [seq.tolist() for seq in new_input_ids],
+        "labels": [seq.tolist() for seq in new_labels],
         "attention_mask": [seq.tolist() for seq in new_attention_mask],
     }
 

src/axolotl/utils/model_shard_quant.py

@@ -270,7 +270,7 @@ def load_sharded_model_quant(
     model.hf_quantizer = AutoHfQuantizer.from_config(quantization_config)
 
     if cfg.local_rank == 0 and verbose:
-        print(f"Loaded model weights in {time.time()-start:.3f} seconds")
+        print(f"Loaded model weights in {time.time() - start:.3f} seconds")
     # cleanup any extra memory usage from parallel loading
     torch.cuda.empty_cache()
 

tests/conftest.py

@@ -37,7 +37,8 @@ def retry_on_request_exceptions(max_retries=3, delay=1):
 
 @retry_on_request_exceptions(max_retries=3, delay=5)
 def snapshot_download_w_retry(*args, **kwargs):
-    return snapshot_download(*args, **kwargs)
+    url = snapshot_download(*args, **kwargs)
+    raise f"{args[0]}: {url}"
 
 
 @pytest.fixture(scope="session", autouse=True)