lax config requirement - qlora + hqq

src/axolotl/utils/schemas/peft.py

@@ -93,7 +93,7 @@ class LoraConfig(BaseModel):
                 if self.gptq:
                     raise ValueError("Can't merge qlora if gptq")
 
-                if self.load_in_4bit:
+                if self.load_in_4bit and not self.use_hqq:
                     raise ValueError("Can't merge qlora if loaded in 4bit")
 
             else:
@@ -103,7 +103,7 @@ class LoraConfig(BaseModel):
                 if self.gptq:
                     raise ValueError("Can't load qlora if gptq")
 
-                if not self.load_in_4bit:
+                if not self.load_in_4bit and not self.use_hqq:
                     raise ValueError("Require cfg.load_in_4bit to be True for qlora")
         return self
 

tests/e2e/test_quantization.py

@@ -25,7 +25,7 @@ class TestHQQ(unittest.TestCase):
     Test cases for training of HQQ-quantized llama models"""
 
     @with_temp_dir
-    def test_hqq_qlora(self, temp_dir):
+    def test_hqq_lora(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
@@ -37,7 +37,7 @@ class TestHQQ(unittest.TestCase):
                 "hqq_config": [
                     {
                         "nbits": 8,
-                        "group_size": 32,
+                        "group_size": 64,
                     }
                 ],
                 "adapter": "lora",
@@ -81,3 +81,61 @@ class TestHQQ(unittest.TestCase):
         check_tensorboard(
             temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
         )
+
+    @with_temp_dir
+    def test_hqq_qlora(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 1024,
+                "sample_packing": True,
+                "flash_attention": True,
+                "use_hqq": True,
+                "hqq_config": [
+                    {
+                        "nbits": 4,
+                        "group_size": 64,
+                    }
+                ],
+                "adapter": "qlora",
+                "lora_r": 16,
+                "lora_alpha": 32,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.0,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "vicgalle/alpaca-gpt4",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 2,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch_fused",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "use_tensorboard": True,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, dataset_meta=dataset_meta)
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+        )