KD trainer w/ logprobs

.github/workflows/tests.yml

@@ -207,7 +207,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.4.1
+            pytorch: 2.5.1
             num_gpus: 1
             axolotl_extras:
     steps:
@@ -253,7 +253,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.4.1
             num_gpus: 1
             axolotl_extras:
     steps:

cicd/tests.py

@@ -59,7 +59,7 @@
 }
 
 N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = modal.gpu.A10G(count=N_GPUS)
+GPU_CONFIG = modal.gpu.L40S(count=N_GPUS)
 
 
 def run_cmd(cmd: str, run_folder: str):

docs/rlhf.qmd

@@ -29,7 +29,7 @@ datasets:
     type: chatml.intel
   - path: argilla/ultrafeedback-binarized-preferences
     split: train
-    type: chatml.argilla
+    type: chatml
 ```
 
 #### IPO

src/axolotl/cli/args.py

@@ -13,6 +13,12 @@ class PreprocessCliArgs:
     debug_num_examples: int = field(default=1)
     prompter: Optional[str] = field(default=None)
     download: Optional[bool] = field(default=True)
+    iterable: Optional[bool] = field(
+        default=None,
+        metadata={
+            "help": "Use IterableDataset for streaming processing of large datasets"
+        },
+    )
 
 
 @dataclass

src/axolotl/cli/main.py

@@ -39,6 +39,8 @@ def preprocess(config: str, **kwargs) -> None:
         kwargs: Additional keyword arguments which correspond to CLI args or `axolotl`
             config options.
     """
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
     from axolotl.cli.preprocess import do_cli
 
     do_cli(config=config, **kwargs)

src/axolotl/cli/preprocess.py

@@ -75,7 +75,10 @@ def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
     )
 
 
-def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
+def do_cli(
+    config: Union[Path, str] = Path("examples/"),
+    **kwargs,
+) -> None:
     """
     Parses `axolotl` config, CLI args, and calls `do_preprocess`.
 

src/axolotl/common/datasets.py

@@ -63,11 +63,17 @@ def load_datasets(
     """
     tokenizer = load_tokenizer(cfg)
     processor = load_processor(cfg, tokenizer=tokenizer) if cfg.processor_type else None
+    preprocess_iterable = (
+        hasattr(cli_args, "iterable")
+        and cli_args.iterable is not None
+        and cli_args.iterable
+    )
 
     train_dataset, eval_dataset, total_num_steps, prompters = prepare_dataset(
         cfg,
         tokenizer,
         processor=processor,
+        preprocess_iterable=preprocess_iterable,
     )
 
     if (