slow fsdp1 test

docs/cli.qmd

@@ -210,8 +210,6 @@ axolotl lm-eval config.yml
 Configuration options:
 
 ```yaml
-lm_eval_model: # model to evaluate (local or hf path)
-
 # List of tasks to evaluate
 lm_eval_tasks:
   - arc_challenge
@@ -220,7 +218,7 @@ lm_eval_batch_size: # Batch size for evaluation
 output_dir: # Directory to save evaluation results
 ```
 
-See [LM Eval Harness integration docs](https://docs.axolotl.ai/docs/custom_integrations.html#language-model-evaluation-harness-lm-eval) for full configuration details.
+See [LM Eval Harness](https://github.com/EleutherAI/lm-evaluation-harness) for more details.
 
 ### delinearize-llama4
 

requirements.txt

@@ -2,10 +2,10 @@
 
 # START section of dependencies that don't install on Darwin/MacOS
 bitsandbytes==0.49.1
-triton>=3.0.0
+triton>=3.4.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
-liger-kernel==0.6.4
+liger-kernel==0.6.5
 # END section
 
 packaging==26.0

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -104,7 +104,7 @@ class CutCrossEntropyPlugin(BasePlugin):
 
     def patch_llama_like(
         self,
-        model_type_to_patch: str,
+        model_type: str,
     ) -> None:
         """
         Generic patch for model architectures with causal lm similar to llama
@@ -112,10 +112,7 @@ class CutCrossEntropyPlugin(BasePlugin):
         from cut_cross_entropy.transformers.patch import PATCH_FNS
 
         def patch_generic(
-            maybe_model,
-            patch_options,
-            remote_model_id: str | None,
-            model_type: str,
+            maybe_model, patch_options, model_type: str, remote_model_id: str | None
         ):
             import cut_cross_entropy.transformers.llama
             from cut_cross_entropy.transformers.llama import cce_forward
@@ -139,13 +136,11 @@ class CutCrossEntropyPlugin(BasePlugin):
                     f"Error: {str(e)}"
                 ) from e
 
-        if model_type_to_patch not in PATCH_FNS:
+        if model_type not in PATCH_FNS:
             LOG.warning_once(
-                "Setting up generic cce patch for model type: %s", model_type_to_patch
+                "Setting up generic cce patch for model type: %s", model_type
             )
             LOG.warning_once(
-                f"Generic Cut Cross Entropy + {model_type_to_patch} support is experimental and may not work as expected."
-            )
-            PATCH_FNS[model_type_to_patch] = partial(
-                patch_generic, model_type=model_type_to_patch
+                f"Generic Cut Cross Entropy + {model_type} support is experimental and may not work as expected."
             )
+            PATCH_FNS[model_type] = partial(patch_generic, model_type=model_type)

src/axolotl/integrations/kernels/README.md

@@ -1,44 +0,0 @@
-# Kernels Integration
-
-MoE (Mixture of Experts) kernels speed up training for MoE layers and reduce VRAM costs. In transformers v5, `batched_mm` and `grouped_mm` were integrated as built-in options via the `experts_implementation` config kwarg:
-
-```python
-class ExpertsInterface(GeneralInterface):
-    _global_mapping = {
-        "batched_mm": batched_mm_experts_forward,
-        "grouped_mm": grouped_mm_experts_forward,
-    }
-```
-
-In our custom integration, we add support for **ScatterMoE**, which is even more efficient and faster than `grouped_mm`.
-
-## Usage
-
-Add the following to your axolotl YAML config:
-
-```yaml
-plugins:
-  - axolotl.integrations.kernels.KernelsPlugin
-
-use_kernels: true
-use_scattermoe: true
-```
-
-**Important:** Setting `experts_implementation` is incompatible with `use_scattermoe`.
-
-## How It Works
-
-The `KernelsPlugin` runs before model loading and:
-
-1. Registers the ScatterMoE kernel from the [`axolotl-ai-co/scattermoe`](https://huggingface.co/axolotl-ai-co/scattermoe) Hub repo.
-2. Patches the model's `SparseMoeBlock` forward method with the optimized ScatterMoE implementation.
-
-This works for any MoE model in transformers that uses a `SparseMoeBlock` class (Mixtral, Qwen2-MoE, OLMoE, etc.).
-
-## Limitations
-
-ScatterMoE uses a softmax -> topk routing, so results may be different for some model arch as baseline (GPT-OSS, GLM_MOE_DSA).
-
-## Note on MegaBlocks
-
-We tested [MegaBlocks](https://huggingface.co/kernels-community/megablocks) but were unable to ensure numerical accuracy, so we did not integrate it. It was also incompatible with many newer model architectures in transformers.

src/axolotl/integrations/lm_eval/README.md

@@ -6,12 +6,6 @@ See https://github.com/EleutherAI/lm-evaluation-harness
 
 ## Usage
 
-There are two ways to use the LM Eval integration:
-
-### 1. Post-Training Evaluation
-
-When training with the plugin enabled, evaluation runs automatically after training completes:
-
 ```yaml
 plugins:
   - axolotl.integrations.lm_eval.LMEvalPlugin
@@ -22,50 +16,9 @@ lm_eval_tasks:
   - arc_easy
 
 lm_eval_batch_size: # Batch size for evaluation
-
-# Directory to save evaluation results.
-# The final model is loaded from this directory
-# unless specified otherwise (see below)
-output_dir:
+output_dir: # Directory to save evaluation results
 ```
 
-Run training as usual:
-```bash
-axolotl train config.yml
-```
-
-### 2. Standalone CLI Evaluation
-
-Evaluate any model directly without training:
-
-```yaml
-lm_eval_model: meta-llama/Llama-2-7b-hf
-
-plugins:
-  - axolotl.integrations.lm_eval.LMEvalPlugin
-
-lm_eval_tasks:
-  - gsm8k
-  - hellaswag
-  - arc_easy
-
-lm_eval_batch_size: 8
-output_dir: ./outputs
-```
-
-Run evaluation:
-```bash
-axolotl lm-eval config.yml
-```
-
-## Model Selection Priority
-
-The model to evaluate is selected in the following priority order:
-
-1. **`lm_eval_model`** - Explicit model path or HuggingFace repo (highest priority)
-2. **`hub_model_id`** - Trained model pushed to HuggingFace Hub
-3. **`output_dir`** - Local checkpoint directory containing trained model weights
-
 ## Citation
 
 ```bib

src/axolotl/integrations/lm_eval/__init__.py

@@ -5,7 +5,7 @@ Module for the Plugin for LM Eval Harness
 import subprocess  # nosec
 
 from axolotl.integrations.base import BasePlugin
-from axolotl.integrations.lm_eval.cli import build_lm_eval_command, get_model_path
+from axolotl.integrations.lm_eval.cli import build_lm_eval_command
 
 from .args import LMEvalArgs as LMEvalArgs
 
@@ -29,7 +29,7 @@ class LMEvalPlugin(BasePlugin):
                 wandb_project=cfg.wandb_project,
                 wandb_entity=cfg.wandb_entity,
                 wandb_name=cfg.wandb_name,
-                model=get_model_path(cfg),
+                model=cfg.lm_eval_model or cfg.hub_model_id,
             ):
                 subprocess.run(  # nosec
                     lm_eval_args,

src/axolotl/integrations/lm_eval/cli.py

@@ -13,21 +13,6 @@ import yaml
 from axolotl.utils.dict import DictDefault
 
 
-def get_model_path(cfg: DictDefault) -> str | None:
-    """
-    Determine which model path to use for evaluation.
-
-    Priority order (highest to lowest):
-    1. lm_eval_model - Explicit model path override
-    2. hub_model_id - Model pushed to HuggingFace Hub
-    3. None - Falls back to output_dir in build_lm_eval_command
-
-    Returns:
-        Model path string or None to use output_dir fallback
-    """
-    return cfg.lm_eval_model or cfg.hub_model_id or None
-
-
 def build_lm_eval_command(
     tasks: list[str],
     bfloat16=True,
@@ -123,7 +108,7 @@ def lm_eval(config: str, cloud: Optional[str] = None):
             wandb_project=cfg.wandb_project,
             wandb_entity=cfg.wandb_entity,
             wandb_name=cfg.wandb_name,
-            model=get_model_path(cfg),
+            model=cfg.lm_eval_model or cfg.hub_model_id,
             revision=cfg.revision,
             apply_chat_template=cfg.apply_chat_template,
             fewshot_as_multiturn=cfg.fewshot_as_multiturn,

tests/e2e/multigpu/test_fsdp1.py

@@ -186,6 +186,7 @@ class TestFSDP1:
 
         verify_training_success(temp_dir)
 
+    @pytest.mark.skip(reason="slow test, deprecate fsdp1 asap")
     def test_dpo_fft(self, temp_dir):
         cfg = DictDefault(
             {