add: support mxfp4 axo (#3375)

* mxfp4 axo

* import lint

* test for qat mxfp4

* config for mxfp4

* add qat:

* pass base config

* MXFakeQuantizeConfig

* lint

* tune config so it fits in 32GB VRAM

---------

Co-authored-by: Wing Lian <wing@axolotl.ai>

tests/e2e/test_qat.py

@@ -8,6 +8,8 @@ from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
+from axolotl.utils.schemas.enums import TorchAOQuantDType
+from axolotl.utils.schemas.quantization import QATConfig, validate_ao_dtype
 
 from .utils import check_model_output_exists, check_tensorboard
 
@@ -130,3 +132,32 @@ class TestQATLlama:
             loss_threshold,
             "Train Loss (%s) is too high",
         )
+
+
+class TestMXFP4Schema:
+    """Test MXFP4 schema validation"""
+
+    def test_validate_mxfp4_dtype(self):
+        result = validate_ao_dtype("mxfp4")
+        assert result == TorchAOQuantDType.mxfp4
+
+    def test_qat_config_with_mxfp4(self):
+        """Test QATConfig accepts mxfp4 weight_dtype"""
+        config = QATConfig(
+            weight_dtype="mxfp4",
+            group_size=32,
+            quantize_embedding=False,
+        )
+        assert config.weight_dtype == TorchAOQuantDType.mxfp4
+        assert config.group_size == 32
+
+    def test_qat_config_mxfp4_invalid_group_size(self):
+        """Test that invalid group_size raises appropriate error during quantization"""
+        # Note: Schema validation doesn't check group_size compatibility,
+        # that happens in get_quantization_config
+        config = QATConfig(
+            weight_dtype="mxfp4",
+            group_size=16,  # Invalid for mxfp4, but schema allows it
+        )
+        assert config.group_size == 16  # Schema accepts it
+        # Actual validation happens at runtime in get_quantization_config

tests/e2e/test_quantization.py

@@ -5,6 +5,7 @@ Tests for axolotl.utils.quantization
 import pytest
 import torch
 from torch import nn
+from torchao.prototype.qat import MXFakeQuantizeConfig
 from torchao.quantization import LinearActivationQuantizedTensor
 from torchao.quantization.qat.embedding import FakeQuantizedEmbedding
 from torchao.quantization.qat.linear import FakeQuantizedLinear
@@ -117,6 +118,21 @@ class TestQuantization:
         config = get_quantization_config(weight_dtype, activation_dtype, group_size)
         assert isinstance(config, expected_type)
 
+    @require_torch_2_8_0
+    @requires_sm_ge_100
+    def test_get_ptq_config_mxfp4(self):
+        config = get_quantization_config(TorchAOQuantDType.mxfp4, None, 32)
+        assert isinstance(config, MXFakeQuantizeConfig)
+        assert config.block_size == 32
+
+    @require_torch_2_8_0
+    @requires_sm_ge_100
+    def test_get_ptq_config_mxfp4_invalid_group_size(self):
+        with pytest.raises(
+            ValueError, match="MXFP4 quantization must use a block_size"
+        ):
+            get_quantization_config(TorchAOQuantDType.mxfp4, None, 16)
+
     @requires_cuda_ge_8_9
     @require_torch_2_8_0
     def test_get_ptq_config_int4_weight_only(self):
@@ -262,6 +278,35 @@ class TestQuantization:
                 else:
                     assert child.activation_fake_quantizer is None
 
+    @pytest.mark.parametrize(
+        "weight_dtype,activation_dtype,group_size,quantize_embedding",
+        [
+            (TorchAOQuantDType.mxfp4, None, 32, False),
+            (TorchAOQuantDType.mxfp4, None, 32, True),
+        ],
+    )
+    @require_torch_2_8_0
+    @requires_sm_ge_100
+    def test_prepare_model_for_qat_mxfp4(
+        self, model, weight_dtype, activation_dtype, group_size, quantize_embedding
+    ):
+        prepare_model_for_qat(
+            model,
+            weight_dtype,
+            group_size,
+            activation_dtype,
+            quantize_embedding,
+        )
+
+        if quantize_embedding:
+            assert isinstance(model.model.embed_tokens, FakeQuantizedEmbedding)
+            assert hasattr(model.model.embed_tokens, "weight_fake_quantizer")
+
+        for child in list(model.children()):
+            if isinstance(child, torch.nn.Linear):
+                assert isinstance(child, FakeQuantizedLinear)
+                assert hasattr(child, "weight_fake_quantizer")
+
     @require_torch_2_8_0
     @requires_cuda_ge_8_9
     def test_convert_qat_model(self, model):