Fix: adding magistral fsdp config, fixing not eval with test_datasets, handle mllama attention (#2789) [skip ci]

* feat: add fsdp config for magistral

* fix: add mllama self attention handling for lora kernels

* fix: no eval if val_set_size 0 despite having test_datasets

* fix: add note for cce for vlm in newer model

src/axolotl/core/builders/base.py

@@ -380,8 +380,8 @@ class TrainerBuilderBase(abc.ABC):
         )
 
         # eval_strategy and eval_steps
-        if not self.eval_dataset or self.cfg.val_set_size == 0:
-            # do not eval if no eval_dataset or val_set_size=0
+        if not self.eval_dataset and self.cfg.val_set_size == 0:
+            # do not eval if no eval_dataset and val_set_size=0
             training_args_kwargs["eval_strategy"] = "no"
         elif self.cfg.eval_steps:
             training_args_kwargs["eval_strategy"] = "steps"

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -24,6 +24,14 @@ pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transform
 
 ## Usage
 
+**NOTE**: If you are training a VLM model, please use older version of Axolotl as upstream has applied a major VLM refactor, and our patches have not been updated yet.
+
+```bash
+git checkout 787880215b3ab32ccaf81c1b2e9588c6f3e6e764
+
+pip3 install --no-build-isolation -e .
+```
+
 ```yaml
 plugins:
   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

src/axolotl/monkeypatch/lora_kernels.py

@@ -145,6 +145,11 @@ def get_attention_cls_from_config(cfg: DictDefault) -> Type[nn.Module]:
 
         return Qwen2Attention
 
+    if model_type == "mllama":
+        from transformers.models.mllama.modeling_mllama import MllamaTextSelfAttention
+
+        return MllamaTextSelfAttention
+
     try:
         # Dynamically import the module and attention class
         module_path = f"transformers.models.{model_type}.modeling_{model_type}"