feat: LoRA kernel support for bias, dropout, dora, embeddings (#3528) [skip ci]

* feat: LoRA kernel support for bias, dropout, dora, embeddings

* chore: lint

* chore: lint

* address PR feedback, add regression tests, add fsdp2 tests for lora kernels

* update tests for new sigs

* update tests now that bias and dropout are supported

tests/e2e/kernels/test_lora.py

@@ -102,7 +102,7 @@ def mock_proj():
 def test_get_lora_parameters(mock_proj):
     """Tests get_lora_parameters function"""
     # Test with LoRA enabled
-    W, b, _, A, B, s = get_lora_parameters(mock_proj)
+    W, b, _, A, B, s, *_ = get_lora_parameters(mock_proj)
 
     assert isinstance(W, torch.Tensor)
     assert W.shape == (128, 64)
@@ -113,13 +113,13 @@ def test_get_lora_parameters(mock_proj):
 
     # Test with LoRA disabled
     mock_proj.disable_adapters = True
-    W, b, _, A, B, s = get_lora_parameters(mock_proj)
+    W, b, _, A, B, s, *_ = get_lora_parameters(mock_proj)
     assert A is None and B is None and s is None
 
     # Test with merged state
     mock_proj.disable_adapters = False
     mock_proj.merged = True
-    W, b, _, A, B, s = get_lora_parameters(mock_proj)
+    W, b, _, A, B, s, *_ = get_lora_parameters(mock_proj)
     assert A is None and B is None and s is None
 
 

tests/e2e/multigpu/test_fsdp2_lora_kernels.py

@@ -0,0 +1,120 @@
+"""Test LoRA kernels under FSDP2 multi-GPU training.
+
+Verifies that lora_qkv_kernel, lora_o_kernel, lora_mlp_kernel, and
+lora_embedding_kernel work correctly with FSDP2 sharding, including
+with bias, dropout, and DoRA enabled.
+"""
+
+from pathlib import Path
+
+import yaml
+from accelerate.test_utils import execute_subprocess_async
+from transformers.testing_utils import get_torch_dist_unique_port
+
+from axolotl.utils.dict import DictDefault
+
+from tests.e2e.utils import require_torch_2_7_0
+
+AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
+
+
+def _run_training(temp_dir, cfg):
+    """Write config and launch multi-GPU training."""
+    Path(temp_dir).mkdir(parents=True, exist_ok=True)
+    with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+        fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+    execute_subprocess_async(
+        [
+            "axolotl",
+            "train",
+            str(Path(temp_dir) / "config.yaml"),
+            "--num-processes",
+            "2",
+            "--main-process-port",
+            f"{get_torch_dist_unique_port()}",
+        ]
+    )
+
+
+def _base_lora_fsdp2_config(temp_dir, **overrides):
+    """Base config for LoRA + FSDP2 + kernel tests."""
+    cfg = {
+        "base_model": "Qwen/Qwen3-0.6B",
+        "sequence_len": 512,
+        "val_set_size": 0.0,
+        "datasets": [
+            {
+                "path": "tatsu-lab/alpaca",
+                "type": "alpaca",
+                "split": "train[:1%]",
+            },
+        ],
+        "adapter": "lora",
+        "lora_r": 8,
+        "lora_alpha": 16,
+        "lora_target_linear": True,
+        "num_epochs": 1,
+        "max_steps": 3,
+        "micro_batch_size": 1,
+        "gradient_accumulation_steps": 1,
+        "output_dir": temp_dir,
+        "learning_rate": 1e-4,
+        "optimizer": "adamw_torch_fused",
+        "lr_scheduler": "cosine",
+        "flash_attention": True,
+        "bf16": True,
+        "fsdp_version": 2,
+        "fsdp_config": {
+            "offload_params": False,
+            "cpu_ram_efficient_loading": False,
+            "transformer_layer_cls_to_wrap": "Qwen3DecoderLayer",
+            "state_dict_type": "FULL_STATE_DICT",
+            "auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+            "reshard_after_forward": True,
+        },
+        # Enable all LoRA kernels
+        "lora_mlp_kernel": True,
+        "lora_qkv_kernel": True,
+        "lora_o_kernel": True,
+        "lora_embedding_kernel": True,
+        "save_safetensors": True,
+    }
+    cfg.update(overrides)
+    return DictDefault(cfg)
+
+
+class TestFSDP2LoRAKernels:
+    """Test LoRA kernels under FSDP2."""
+
+    @require_torch_2_7_0
+    def test_lora_kernels_basic(self, temp_dir):
+        """Basic LoRA + kernels + FSDP2: no dropout, no bias, no DoRA."""
+        cfg = _base_lora_fsdp2_config(temp_dir)
+        _run_training(temp_dir, cfg)
+        assert (Path(temp_dir) / "adapter_model.safetensors").exists()
+
+    @require_torch_2_7_0
+    def test_lora_kernels_with_dropout(self, temp_dir):
+        """LoRA kernels + dropout + FSDP2."""
+        cfg = _base_lora_fsdp2_config(temp_dir, lora_dropout=0.1)
+        _run_training(temp_dir, cfg)
+        assert (Path(temp_dir) / "adapter_model.safetensors").exists()
+
+    @require_torch_2_7_0
+    def test_lora_kernels_with_dora(self, temp_dir):
+        """LoRA kernels + DoRA + FSDP2."""
+        cfg = _base_lora_fsdp2_config(temp_dir, peft_use_dora=True)
+        _run_training(temp_dir, cfg)
+        assert (Path(temp_dir) / "adapter_model.safetensors").exists()
+
+    @require_torch_2_7_0
+    def test_lora_kernels_with_dora_and_dropout(self, temp_dir):
+        """LoRA kernels + DoRA + dropout + FSDP2."""
+        cfg = _base_lora_fsdp2_config(
+            temp_dir,
+            peft_use_dora=True,
+            lora_dropout=0.05,
+        )
+        _run_training(temp_dir, cfg)
+        assert (Path(temp_dir) / "adapter_model.safetensors").exists()

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

@@ -222,9 +222,9 @@ def test_model_specific_activation(model_name, expected_activation):
 
 
 def test_kernel_patch_conditions():
-    """Test various conditions that should prevent kernel patching."""
+    """Test that kernels ARE patched even with dropout and bias (now supported)."""
     test_configs = [
-        # Dropout prevents patching
+        # Dropout — kernels now support this
         {
             "peft_type": "LORA",
             "task_type": "CAUSAL_LM",
@@ -234,7 +234,7 @@ def test_kernel_patch_conditions():
             "lora_dropout": 0.1,
             "bias": "none",
         },
-        # Bias prevents patching
+        # Bias — kernels now support this
         {
             "peft_type": "LORA",
             "task_type": "CAUSAL_LM",
@@ -252,13 +252,14 @@ def test_kernel_patch_conditions():
         model = PeftModelForCausalLM(model, peft_config)
         cfg = DictDefault({"lora_mlp_kernel": True})
 
-        # Should not patch
         patched_model = apply_lora_kernel_patches(model, cfg)
         layer = patched_model.model.model.layers[0].mlp
 
-        # Verify no patches applied
-        assert layer.forward.__func__ is not apply_lora_mlp_swiglu
-        assert layer.forward.__func__ is not apply_lora_mlp_geglu
+        # Verify patches ARE applied (dropout and bias are now supported)
+        assert (
+            layer.forward.__func__ is apply_lora_mlp_swiglu
+            or layer.forward.__func__ is apply_lora_mlp_geglu
+        )
 
 
 def test_kernel_config_options():
@@ -511,7 +512,7 @@ def test_kernel_training_integration_auto_enable(temp_dir):
 
 
 def test_kernel_training_integration_dropout_non_zero(temp_dir):
-    """Test model loading with dropout non-zero should not patch."""
+    """Test model loading with dropout non-zero DOES patch (now supported)."""
 
     from axolotl.cli.utils import load_model_and_tokenizer
 
@@ -546,31 +547,18 @@ def test_kernel_training_integration_dropout_non_zero(temp_dir):
     # Load config
     cfg = load_cfg(str(path))
 
-    # Get original attention class
-    attention_cls = get_attention_cls_from_config(cfg)
-
-    # Store original state before patching
-    original_forward_method = attention_cls.forward
-
     # Load model
     model, tokenizer, _ = load_model_and_tokenizer(cfg=cfg)
 
-    # We call modelloader as that's where the patches are applied
-    # despite the fact that we're not using it to load the model
     model_loader = ModelLoader(cfg, tokenizer)
 
-    # Apply patch
+    # Apply patches — should succeed even with dropout > 0
     model_loader.patch_manager._apply_self_attention_lora_patch()
-
-    # Verify patch was not applied
-    assert attention_cls.forward == original_forward_method
-
-    # Apply apply_lora_kernel_patches
     model_loader.patch_manager._apply_lora_kernel_patch(model)
 
-    # Verify patch was not applied
+    # Verify patches WERE applied (dropout is now supported by kernels)
     layers = get_layers(model)
     for layer in layers:
         for self_attn in find_self_attn_in_layer(layer):
-            assert not hasattr(self_attn, "apply_qkv")
-            assert not hasattr(self_attn, "apply_o")
+            assert hasattr(self_attn, "apply_qkv")
+            assert hasattr(self_attn, "apply_o")