[Transform] Spinquant with R1 and R2

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/llama3_example.py

@@ -0,0 +1,86 @@
+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 SpinQuantModifier
+from llmcompressor.utils import dispatch_for_generation
+
+# Select model and load it.
+MODEL_ID = "meta-llama/Meta-Llama-3-8B-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 = [
+    SpinQuantModifier(rotations=["R1", "R2"], transform_type="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,
+)
+
+# 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] + "-transformed-w4a16"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)

src/llmcompressor/entrypoints/utils.py

@@ -20,7 +20,7 @@
 from llmcompressor.pytorch.model_load.helpers import parse_dtype
 from llmcompressor.transformers.sparsification.compressed_tensors_utils import (
     modify_save_pretrained,
-    patch_tied_tensors_bug,
+    untie_word_embeddings,
 )
 from llmcompressor.transformers.utils.helpers import (
     detect_last_checkpoint,
@@ -61,7 +61,8 @@ def pre_process(model_args: "ModelArguments"):
         )
 
     # untie tie_word_embeddings weights
-    patch_tied_tensors_bug(model_args.model)
+    if not model_args.tie_word_embeddings:
+        untie_word_embeddings(model_args.model)
 
     # wrap model.save_pretrained
     modify_save_pretrained(model_args.model)
@@ -143,7 +144,6 @@ def initialize_model_from_path(
         cache_dir=model_args.cache_dir,
         revision=model_args.model_revision,
         use_auth_token=True if model_args.use_auth_token else None,
-        tie_word_embeddings=model_args.tie_word_embeddings,
         trust_remote_code=model_args.trust_remote_code_model,
     )
 
@@ -156,7 +156,6 @@ def initialize_model_from_path(
             AutoConfig.from_pretrained(
                 model_args.distill_teacher,
                 use_auth_token=True if model_args.use_auth_token else None,
-                tie_word_embeddings=model_args.tie_word_embeddings,
                 trust_remote_code=model_args.trust_remote_code_model,
             )
             if model_args.distill_teacher

src/llmcompressor/modeling/__init__.py

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

src/llmcompressor/modeling/fuse.py

@@ -0,0 +1,58 @@
+from typing import Iterable
+
+import torch
+from compressed_tensors import (
+    align_module_device,
+    get_execution_device,
+    update_offload_parameter,
+)
+from transformers.models.llama.modeling_llama import LlamaRMSNorm
+
+__all__ = ["normalize_embedding", "fuse_norm_linears"]
+
+
+PRECISION = torch.float64
+
+
+def normalize_embedding(embedding: torch.nn.Module):
+    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)
+
+        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 a norm layer into subsequent linear layers. This useful for ensuring transform
+    invariance between norm and linear layers.
+
+    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 isinstance(norm, (torch.nn.RMSNorm, LlamaRMSNorm)):
+        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)}")

src/llmcompressor/modifiers/transform/__init__.py

@@ -0,0 +1,3 @@
+# flake8: noqa
+
+from .spinquant import SpinQuantModifier

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

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