Configurable embeddings upcast (#2621)

* fsdp embeddings should be float32 per comment

* patch peft to not upcast everything

* add tabs back to code check

* fix import

* add configurable option and fix check

* add check for dtypes

* move embeddings test to patch dir

* fix test

* fix comment and logic

docs/config.qmd

@@ -32,6 +32,8 @@ tokenizer_legacy:
 resize_token_embeddings_to_32x:
 # Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
 shrink_embeddings:
+# Optional[bool] Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs
+embeddings_skip_upcast:
 # Whether to load the model with randomly initialized weights. Useful for
 # pre-training a model from scratch or debugging purposes.
 random_init_weights:

src/axolotl/monkeypatch/peft/utils.py

@@ -0,0 +1,78 @@
+"""
+Patch prepare_model_for_kbit_training to not upcast everything
+"""
+
+import inspect
+import logging
+
+import peft
+
+import axolotl
+from axolotl.monkeypatch.utils import detab_code
+
+LOG = logging.getLogger(__name__)
+
+ORIGINAL_PREPARE_CODE = """
+        for param in model.parameters():
+            if (
+                (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
+            ) and param.__class__.__name__ != "Params4bit":
+                param.data = param.data.to(torch.float32)
+"""
+
+PATCHED_PREPARE_CODE = """
+        for name, param in model.named_parameters():
+            if (
+                (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
+            ) and param.__class__.__name__ != "Params4bit" and all(embed_name not in name for embed_name in ["embed_tokens", "lm_head"]):
+                param.data = param.data.to(torch.float32)
+"""
+
+
+def get_peft_prep_code() -> str:
+    prepare = inspect.getsource(peft.utils.other.prepare_model_for_kbit_training)
+    return prepare
+
+
+def check_peft_prep_code_is_patchable() -> bool:
+    prep_code = get_peft_prep_code()
+    prep_code, _ = detab_code(prep_code)
+    return ORIGINAL_PREPARE_CODE in prep_code
+
+
+def patch_peft_prep_code():
+    """
+    monkeypatch create_accelerator_and_postprocess so it checks for additional kwargs
+    """
+
+    try:
+        prep_code = get_peft_prep_code()
+    except OSError:
+        return
+    peft.utils.other._original_create_accelerator_and_postprocess = (  # pylint: disable=protected-access
+        prep_code
+    )
+    prep_code, _ = detab_code(prep_code)
+    if ORIGINAL_PREPARE_CODE not in prep_code:
+        return
+
+    prep_code = prep_code.replace(ORIGINAL_PREPARE_CODE, PATCHED_PREPARE_CODE)
+    prep_code = prep_code.replace(
+        "def prepare_model_for_kbit_training(",
+        "def fixed_prepare_model_for_kbit_training(",
+        1,
+    )
+
+    items_to_import = []
+    for item in dir(peft.utils.other):
+        if item in prep_code:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from peft.utils.other import (" + ", ".join(x for x in items_to_import) + ")",
+        globals(),
+    )
+    exec(prep_code, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching prepare_model_for_kbit_training to allow for overrides")
+    peft.utils.other.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821
+    axolotl.utils.models.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821

src/axolotl/utils/models.py

@@ -566,6 +566,11 @@ class ModelLoader:
 
             patch_accelerate_fsdp_utils()
 
+        if self.cfg.adapter and self.cfg.embeddings_skip_upcast:
+            from axolotl.monkeypatch.peft.utils import patch_peft_prep_code
+
+            patch_peft_prep_code()
+
         if self.cfg.flex_attention:
             from axolotl.monkeypatch.attention.flex_attn import (
                 patch_flex_make_mask,
@@ -1185,7 +1190,7 @@ class ModelLoader:
                 ],
             )
 
-    def prepare_model(self, qlora_fsdp) -> None:
+    def prepare_model(self, qlora_fsdp: bool) -> None:
         skip_prepare_model_for_kbit_training = False
         if self.cfg.model_config_type == "qwen" and self.cfg.adapter == "lora":
             # Qwen doesn't play nicely with LoRA if this is enabled
@@ -1315,7 +1320,10 @@ class ModelLoader:
         # make sure these are fp32 per Ramesh et al. (2021)
         embedding_modules = get_linear_embedding_layers(self.cfg.model_config_type)
         if not self.cfg.fsdp:
-            # FSDP doesn't like mixed Float and BFloat16
+            # we don't run this during FSDP because this will leave mixed
+            # float and bfloat16 dtypes in the model which FSDP doesn't like
+            if self.cfg.load_in_4bit and self.cfg.embeddings_skip_upcast:
+                embedding_modules = []
             self.convert_embedding_modules_dtype(
                 embedding_modules,
                 dist_dtype=torch.float32,

src/axolotl/utils/schemas/config.py

@@ -82,6 +82,7 @@ class AxolotlInputConfig(
     mean_resizing_embeddings: bool | None = False
     # optionally shrink the embeddings when the tokenizer vocab size is smaller
     shrink_embeddings: bool | None = None
+    embeddings_skip_upcast: bool | None = None
 
     rl: RLType | None = None
     trl: TRLConfig | None = Field(

tests/e2e/patched/test_peft_embeddings.py

@@ -0,0 +1,63 @@
+"""
+Test case for handling embeddings when using peft
+"""
+
+import torch
+
+from axolotl.train import setup_model_and_tokenizer
+from axolotl.utils.config import normalize_config, validate_config
+from axolotl.utils.dict import DictDefault
+
+
+class TestLlamaPeftEmbeddings:
+    """
+    test class for handling embeddings when using peft
+    """
+
+    def test_peft_embeddings_upcast(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_target_linear": True,
+                "trust_remote_code": True,
+                "sequence_len": 512,
+                "val_set_size": 0.01,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 2,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "sample_packing": False,
+                "bf16": "auto",
+                "save_safetensors": True,
+                "embeddings_skip_upcast": True,
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+
+        model, _, _, _ = setup_model_and_tokenizer(cfg)
+
+        # Check if the embeddings are upcast correctly
+        # only embed_tokens is a parameter that may be upcast
+        assert model.base_model.model.model.embed_tokens.weight.dtype == torch.bfloat16
+        assert model.base_model.model.lm_head.weight.dtype == torch.bfloat16