[Transform] QuIP Modifier

compress_model.py

@@ -0,0 +1,60 @@
+# python3 compress_model.py --model_id meta-llama/Llama-3.2-1B-Instruct --transform_type random-hadamard
+import argparse
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import QuantizationModifier
+from llmcompressor.modifiers.transform import SpinQuantModifier
+from llmcompressor.utils import dispatch_for_generation
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_id", type=str, help="Model stub to compress")
+    parser.add_argument("--transform_type", type=str, default=None, help="Type of transform used in SpinQuantModifier")
+    parser.add_argument("--scheme", type=str, default=None, help="Quantization scheme (e.g. W4A16)")
+    return parser.parse_args()
+
+if __name__ == "__main__":
+    args = parse_args()
+
+    # Select model and load it.
+    MODEL_ID = args.model_id
+    model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype="auto")
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+
+    # Select number of samples. 512 samples is a good place to start.
+    # Increasing the number of samples can improve accuracy.
+    NUM_CALIBRATION_SAMPLES = 512
+    MAX_SEQUENCE_LENGTH = 2048
+
+    # Configure the quantization algorithm to run.
+    recipe = []
+    if args.transform_type:
+        recipe.append(SpinQuantModifier(rotations=["R1", "R2"], transform_type=args.transform_type))
+
+    if args.scheme:
+        recipe.append(QuantizationModifier(targets="Linear", scheme=args.scheme, ignore=["lm_head"]))
+
+    # Apply algorithms.
+    oneshot(
+        model=model,
+        recipe=recipe,
+        dataset="ultrachat_200k",
+        splits={"calibration": f"train_sft[:{NUM_CALIBRATION_SAMPLES}]"},
+        max_seq_length=MAX_SEQUENCE_LENGTH,
+        num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    )
+
+    # Confirm generations of the quantized model look sane.
+    print("\n\n")
+    print("========== SAMPLE GENERATION ==============")
+    dispatch_for_generation(model)
+    input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
+    output = model.generate(input_ids, max_new_tokens=100)
+    print(tokenizer.decode(output[0]))
+    print("==========================================\n\n")
+
+    # Save to disk compressed.
+    SAVE_DIR = MODEL_ID.split("/")[1] + f"-{args.transform_type}-{args.scheme}"
+    model.save_pretrained(SAVE_DIR, save_compressed=True)
+    tokenizer.save_pretrained(SAVE_DIR)

examples/transform/quip_example.py

@@ -0,0 +1,87 @@
+from datasets import load_dataset
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import QuantizationModifier
+from llmcompressor.modifiers.transform import QuIPModifier
+from llmcompressor.utils import dispatch_for_generation
+
+# Select model and load it.
+MODEL_ID = "meta-llama/Llama-3.1-8B-Instruct"
+
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID,
+    torch_dtype="auto",
+)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+
+# Select calibration dataset.
+DATASET_ID = "HuggingFaceH4/ultrachat_200k"
+DATASET_SPLIT = "train_sft"
+
+# Select number of samples. 512 samples is a good place to start.
+# Increasing the number of samples can improve accuracy.
+NUM_CALIBRATION_SAMPLES = 512
+MAX_SEQUENCE_LENGTH = 2048
+
+# Load dataset and preprocess.
+ds = load_dataset(DATASET_ID, split=f"{DATASET_SPLIT}[:{NUM_CALIBRATION_SAMPLES}]")
+ds = ds.shuffle(seed=42)
+
+
+def preprocess(example):
+    return {
+        "text": tokenizer.apply_chat_template(
+            example["messages"],
+            tokenize=False,
+        )
+    }
+
+
+ds = ds.map(preprocess)
+
+
+# Tokenize inputs.
+def tokenize(sample):
+    return tokenizer(
+        sample["text"],
+        padding=False,
+        max_length=MAX_SEQUENCE_LENGTH,
+        truncation=True,
+        add_special_tokens=False,
+    )
+
+
+ds = ds.map(tokenize, remove_columns=ds.column_names)
+
+# Configure the quantization algorithm to run.
+#   * apply spinquant transforms to model in order to make quantization easier
+#   * quantize the weights to 4 bit with GPTQ with a group size 128
+recipe = [
+    QuIPModifier(transform_type="random-hadamard"),
+    QuantizationModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"]),
+]
+
+# Apply algorithms.
+oneshot(
+    model=model,
+    recipe=recipe,
+    dataset=ds,
+    max_seq_length=MAX_SEQUENCE_LENGTH,
+    num_calibration_samples=NUM_CALIBRATION_SAMPLES,
+    pipeline="datafree",
+)
+
+# Confirm generations of the quantized model look sane.
+print("\n\n")
+print("========== SAMPLE GENERATION ==============")
+dispatch_for_generation(model)
+input_ids = tokenizer("Hello my name is", return_tensors="pt").input_ids.to("cuda")
+output = model.generate(input_ids, max_new_tokens=100)
+print(tokenizer.decode(output[0]))
+print("==========================================\n\n")
+
+# Save to disk compressed.
+SAVE_DIR = MODEL_ID.split("/")[1] + "-quip-w4a16"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)

src/llmcompressor/modeling/__init__.py

@@ -1,3 +1,4 @@
 # flake8: noqa
 
+from .fuse import *
 from .prepare import *

src/llmcompressor/modeling/fuse.py

@@ -6,55 +6,58 @@
     get_execution_device,
     update_offload_parameter,
 )
+from transformers.models.llama.modeling_llama import LlamaRMSNorm
 
-__all__ = ["center_embeddings", "fuse_norm_linears"]
+__all__ = ["normalize_embedding", "fuse_norm_linears"]
 
 
 PRECISION = torch.float64
 
 
-def center_embeddings(embedding: torch.nn.Module):
+def normalize_embedding(embedding: torch.nn.Module):
     """
-    Shift each embedding to have a mean of zero
+    Normalize each embedding to have a mean of zero
 
     :param embedding: embedding module containing embeddings to center
     """
-    if not hasattr(embedding, "weight"):
-        raise ValueError(f"Cannot fuse norm of type {type(embedding)}")
+    if isinstance(embedding, (torch.nn.Embedding)):
+        with align_module_device(embedding):
+            weight_dtype = embedding.weight.dtype
+            weight = embedding.weight.to(PRECISION)
+            new_weight = weight - weight.mean(dim=-1, keepdim=True)
+            new_weight = new_weight.to(weight_dtype)
 
-    with align_module_device(embedding):
-        weight_dtype = embedding.weight.dtype
-        weight = embedding.weight.to(PRECISION)
-        new_weight = weight - weight.mean(dim=-1, keepdim=True)
-        new_weight = new_weight.to(weight_dtype)
+        update_offload_parameter(embedding, "weight", new_weight)
 
-    update_offload_parameter(embedding, "weight", new_weight)
+    else:
+        raise ValueError(f"Cannot normalize embedding of type {type(embedding)}")
 
 
 def fuse_norm_linears(norm: torch.nn.Module, linears: Iterable[torch.nn.Linear]):
     """
-    Fuse the scaling operation of norm layer into subsequent linear layers.
-    This useful for ensuring transform invariance between norm and linear layers.
+    Fuse a norm layer into subsequent linear layers. This useful for ensuring transform
+    invariance between norm and linear layers.
 
-    Note that unitary transforms (rotation) commute with normalization, but not scaling
+    Note that a model cannot be properly trained after its norms have been fused
 
     :param norm: norm layer whose weight will be fused into subsequent linears
     :param linears: linear layers which directly follow the norm layer
     """
-    if not hasattr(norm, "weight"):
+    if isinstance(norm, (torch.nn.RMSNorm, LlamaRMSNorm, torch.nn.LayerNorm)):
+        for linear in linears:
+            # NOTE: spinquant does this op in float64
+            exec_device = get_execution_device(norm)
+            with align_module_device(norm, exec_device), align_module_device(
+                linear, exec_device
+            ):
+                weight_dtype = linear.weight.dtype
+                new_weight = linear.weight.to(PRECISION) * norm.weight.to(PRECISION)
+                new_weight = new_weight.to(weight_dtype)
+
+            update_offload_parameter(linear, "weight", new_weight)
+
+        new_norm_weight = torch.ones_like(norm.weight, device="cpu")
+        update_offload_parameter(norm, "weight", new_norm_weight)
+
+    else:
         raise ValueError(f"Cannot fuse norm of type {type(norm)}")
-
-    for linear in linears:
-        # NOTE: spinquant does this op in float64
-        exec_device = get_execution_device(norm)
-        with align_module_device(norm, exec_device), align_module_device(
-            linear, exec_device
-        ):
-            weight_dtype = linear.weight.dtype
-            new_weight = linear.weight.to(PRECISION) * norm.weight.to(PRECISION)
-            new_weight = new_weight.to(weight_dtype)
-
-        update_offload_parameter(linear, "weight", new_weight)
-
-    new_norm_weight = torch.ones_like(norm.weight, device="cpu")
-    update_offload_parameter(norm, "weight", new_norm_weight)

src/llmcompressor/modifiers/transform/__init__.py

@@ -0,0 +1,3 @@
+# flake8: noqa
+
+from .quip import QuIPModifier

src/llmcompressor/modifiers/transform/quip/__init__.py

@@ -0,0 +1,3 @@
+# flake8: noqa
+
+from .base import *

src/llmcompressor/modifiers/transform/quip/base.py

@@ -0,0 +1,135 @@
+from typing import List, Literal, Optional, Union
+
+from compressed_tensors.transform import (
+    TransformArgs,
+    TransformConfig,
+    TransformScheme,
+    apply_transform_config,
+)
+from pydantic import Field, ValidationInfo, field_validator
+
+from llmcompressor.core import Event, EventType, State
+from llmcompressor.modifiers import Modifier
+
+__all__ = ["QuIPModifier"]
+
+
+class QuIPModifier(Modifier):
+    """
+    Implements the transforms according to
+    [QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks](https://arxiv.org/pdf/2402.04396)  # noqa: E501
+    [QuIP: 2-Bit Quantization of Large Language Models With Guarantees](https://arxiv.org/abs/2307.13304)  # noqa: E501
+
+    Transforms (rotations) are extra layers added to a model which reduce the accuracy
+    loss induced by quantization. This is achieved through "rotating" weights and
+    activations into a space with a smaller dynamic range of values, thus decreasing
+    the range of scales required for quantization.
+
+    QuIP and QuIP# apply transforms to every linear layer, two of which are fused into
+    the model weights and two of which remain as online rotations computed at runtime.
+
+    :param transform_type: The type of transform to apply to the model.
+        `"hadamard"` has the least performance cost but only supports sizes which are
+        powers of power of two.
+        `"random-hadamard"` has more performance cost, but supports a much larger set of
+            sizes.
+        `"random-matrix"` has the greatest performance cost, but supports any size
+    :param randomize: If true, create distinct transforms for each application
+    :param learnable: If true, attach gradients to transform weights for training
+    :param ignore: Modules to ignore when attaching transforms
+    :param transform_config: Optional transform config for overriding provided arguments
+    """
+
+    transform_type: Literal["hadamard", "random-hadamard", "random-matrix"] = Field(
+        default="hadamard", exclude=True
+    )
+    randomize: bool = Field(default=False, exclude=True)
+    learnable: bool = Field(default=False, exclude=True)
+    ignore: Union[str, List[str]] = Field(default="lm_head", exclude=True)
+
+    # optional override for more fine-grained control
+    # also included in recipe serialization
+    transform_config: Optional[TransformConfig] = Field(default=None, repr=False)
+
+    @field_validator("randomize", "learnable", mode="before")
+    def validate_not_implemented(cls, value, info: ValidationInfo):
+        if value:
+            raise NotImplementedError(f"{info.field_name} is not supported right now")
+        return value
+
+    def on_initialize(self, state: State, **kwargs) -> bool:
+        if self.transform_config is not None:
+            return True
+
+        self.transform_config = self._create_config()
+        return True
+
+    def on_start(self, state: State, event: Event, **kwargs):
+        self.started_ = True
+
+        apply_transform_config(state.model, self.transform_config)
+
+    def on_event(self, state: State, event: Event, **kwargs):
+        if event.type_ == EventType.CALIBRATION_EPOCH_START:
+            if not self.started_:
+                self.on_start(state, None)
+
+        elif event.type_ == EventType.SEQUENTIAL_EPOCH_END:
+            pass
+
+        elif event.type_ == EventType.CALIBRATION_EPOCH_END:
+            if not self.ended_:
+                self.on_end(state, None)
+
+    def on_end(self, state: State, event: Event, **kwargs):
+        self.ended_ = True
+
+    def on_finalize(self, state: State, **kwargs) -> bool:
+        if not self.ended_:
+            self.on_end(state, None)
+
+        return True
+
+    def _create_config(self) -> TransformConfig:
+        return TransformConfig(
+            config_groups={
+                "v": TransformScheme(
+                    type=self.transform_type,
+                    apply=[
+                        TransformArgs(
+                            targets=["Linear"],
+                            location="input",  # non-mergable
+                            ignore=self.ignore,
+                        ),
+                        TransformArgs(
+                            targets=["Linear"],
+                            location="weight_input",
+                            # location="input",
+                            inverse=True,
+                            ignore=self.ignore,
+                        ),
+                    ],
+                    randomize=self.randomize,
+                    requires_grad=self.learnable,
+                ),
+                "u": TransformScheme(
+                    type=self.transform_type,
+                    apply=[
+                        TransformArgs(
+                            targets=["Linear"],
+                            location="weight_output",
+                            # location="output",
+                            ignore=self.ignore,
+                        ),
+                        TransformArgs(
+                            targets=["Linear"],
+                            location="output",  # non-mergable
+                            inverse=True,
+                            ignore=self.ignore,
+                        ),
+                    ],
+                    randomize=self.randomize,
+                    requires_grad=self.learnable,
+                ),
+            }
+        )