gradient accumulation tests, embeddings w pad_token fix, smaller models (#2059)

* add more test cases for gradient accumulation and fix zero3

* swap out for smaller model

* fix missing return

* fix missing pad_token in config

* support concurrency for multigpu testing

* cast empty deepspeed to empty string for zero3 check

* fix temp_dir as fixture so parametrize works properly

* fix test file for multigpu evals

* don't use default

* don't use default for fsdp_state_dict_type

* don't use llama tokenizer w smollm

* also automatically cancel multigpu for concurrency

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

@@ -1291,6 +1291,25 @@ class AxolotlInputConfig(
             )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def warn_qlora_zero3_w_use_reentrant(cls, data):
+        if (
+            data.get("adapter") == "qlora"
+            and data.get("gradient_checkpointing_kwargs", {})
+            and data.get("gradient_checkpointing_kwargs", {}).get("use_reentrant")
+            is False
+            and "zero3" in data.get("deepspeed", "")
+        ):
+            # may result in:
+            # torch.utils.checkpoint.CheckpointError: torch.utils.checkpoint:
+            # Recomputed values for the following tensors have different metadata
+            # than during the forward pass.
+            LOG.warning(
+                "qlora + zero3 with use_reentrant: false may result in a CheckpointError about recomputed values"
+            )
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_val_w_test_datasets(cls, data):

src/axolotl/utils/models.py

@@ -238,6 +238,7 @@ def load_tokenizer(cfg):
                         x in cfg.lora_modules_to_save for x in lora_modules_to_save
                     )
                 )
+                and k != "pad_token"
             ):
                 lora_modules_to_save = ", ".join(
                     [f"`{x}`" for x in lora_modules_to_save]