init v0.23.0

Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>

examples/quantization/llm-compressor/w4a8_dynamic_moe.py

@@ -0,0 +1,58 @@
+from llmcompressor import oneshot
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+MODEL_ID = "Qwen/Qwen3-30B-A3B-Instruct-2507"
+
+# Load model.
+model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+
+recipe = """
+quant_stage:
+    quant_modifiers:
+        QuantizationModifier:
+            ignore: ["lm_head", "re:.*mlp.gate$"]
+            config_groups:
+                group_0:
+                    weights:
+                        num_bits: 8
+                        type: int
+                        strategy: channel
+                        dynamic: false
+                        symmetric: true
+                    input_activations:
+                        num_bits: 8
+                        type: int
+                        strategy: token
+                        dynamic: true
+                        symmetric: true
+                    targets: ["re:.*self_attn.k_proj.*", "re:.*self_attn.o_proj.*",
+                        "re:.*self_attn.q_proj.*", "re:.*self_attn.v_proj.*"]
+                group_1:
+                    weights:
+                        num_bits: 4
+                        type: int
+                        strategy: group
+                        group_size: 128
+                        dynamic: false
+                        symmetric: true
+                    input_activations:
+                        num_bits: 8
+                        type: int
+                        strategy: token
+                        dynamic: true
+                        symmetric: true
+                    targets: ["re:.*down_proj.*", "re:.*gate_proj.*", "re:.*up_proj.*"]
+"""
+
+# Apply quantization.
+oneshot(
+    model=model,
+    recipe=recipe,
+    trust_remote_code_model=True,
+)
+
+# Save to disk in compressed-tensors format.
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-W4A8"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)

examples/quantization/llm-compressor/w8a8_int8.py

@@ -0,0 +1,150 @@
+import os
+
+import torch
+from compressed_tensors.quantization import QuantizationArgs, QuantizationScheme, QuantizationStrategy, QuantizationType
+from datasets import load_dataset
+from llmcompressor import oneshot
+from llmcompressor.modifiers.awq import AWQModifier
+from llmcompressor.modifiers.quantization import GPTQModifier, QuantizationModifier
+from transformers import (
+    AutoConfig,
+    AutoModelForCausalLM,
+    AutoTokenizer,
+)
+
+W8A8_W_cha_A_ten_static_symmetric = {
+    "group_0": QuantizationScheme(
+        targets=["Linear"],
+        weights=QuantizationArgs(
+            num_bits=8, type=QuantizationType.INT, strategy=QuantizationStrategy.CHANNEL, symmetric=True, dynamic=False
+        ),
+        input_activations=QuantizationArgs(
+            num_bits=8, type=QuantizationType.INT, strategy=QuantizationStrategy.TENSOR, symmetric=True, dynamic=False
+        ),
+    ),
+}
+
+# supported modifiers
+MODIFIER_DICT = {
+    "PTQ": QuantizationModifier,
+    "AWQ": AWQModifier,
+    "GPTQ": GPTQModifier,
+}
+
+# supported schemes
+SCHEMES_DICT = {
+    "W8A8_W_cha_A_ten_static_symmetric": W8A8_W_cha_A_ten_static_symmetric,
+}
+
+MODEL_DICT = {
+    "qwen3": AutoModelForCausalLM,
+}
+
+TOKENIZER_DICT = {
+    "qwen3": AutoTokenizer,
+}
+
+
+def load_environment_variables():
+    env_vars = {
+        "model_path": "Qwen/Qwen3-32B",
+        "export_path": "/llm-compressor/export/GPTQ/W8A8_W_cha_A_ten_static_symmetric",
+        "modifier": "GPTQ",
+        "schemes": "W8A8_W_cha_A_ten_static_symmetric",
+        "calib_prompt_path": "HuggingFaceH4/ultrachat_200k",
+    }
+
+    # verify export model path
+    if env_vars["export_path"] is None:
+        env_vars["export_path"] = env_vars["model_path"].rstrip("/") + "-" + env_vars["modifier"]
+        if env_vars["schemes"] is not None:
+            env_vars["export_path"] += "-" + env_vars["schemes"]
+    os.makedirs(env_vars["export_path"], exist_ok=True)
+
+    return env_vars
+
+
+def load_calibration_text_dataset(calib_prompt_path, tokenizer):
+    # Load dataset
+    for f in os.listdir(calib_prompt_path):
+        print(f)
+    if any(f.lower().endswith(".jsonl") for f in os.listdir(calib_prompt_path)):
+        ds = load_dataset("json", data_dir=calib_prompt_path, split="validation")
+    elif any(f.lower().endswith(".parquet") for f in os.listdir(calib_prompt_path)):
+        ds = load_dataset("parquet", data_dir=calib_prompt_path, split="train[:512]")
+    else:
+        raise ValueError("Unsupported calibration file format: {}".format(calib_prompt_path.split(".")[-1]))
+
+    # Preprocess dataset
+    def preprocess(example):
+        if tokenizer.chat_template is not None:
+            return {"text": tokenizer.apply_chat_template(example["messages"], tokenize=False)}
+        else:
+            return {"text": example["messages"]}
+
+    # Tokenize inputs
+    def tokenize(sample):
+        return tokenizer(
+            sample["text"],
+            add_special_tokens=False,
+        )
+
+    ds = ds.map(preprocess)
+    ds = ds.map(tokenize, remove_columns=ds.column_names)
+    return ds
+
+
+# Define a oneshot data collator for multimodal inputs.
+def data_collator(batch):
+    assert len(batch) == 1
+    return {
+        key: torch.tensor(value, dtype=torch.bfloat16 if key == "pixel_values" else torch.long)
+        for key, value in batch[0].items()
+    }
+
+
+def quantize_model(model, env_vars, dataset_dict=None):
+    # since the MoE gate layers are sensitive to quantization, we add them to the ignore
+    # list so they remain at full precision
+    ignore = ["lm_head", "re:.*mlp.down_proj"]
+
+    # define a llmcompressor recipe
+    recipe = [
+        MODIFIER_DICT[env_vars["modifier"]](
+            config_groups=SCHEMES_DICT[env_vars["schemes"]],
+            ignore=ignore,
+        ),
+    ]
+
+    # quantize the model
+    oneshot(
+        model=model,
+        dataset=dataset_dict,
+        recipe=recipe,
+        trust_remote_code_model=True,
+    )
+
+
+def save_quantized_model(model, tokenizer, save_path, save_compressed=False):
+    model.save_pretrained(save_path, save_compressed=save_compressed)
+    tokenizer.save_pretrained(save_path)
+
+
+if __name__ == "__main__":
+    # get environment variables
+    env_vars = load_environment_variables()
+
+    # support model type list
+    config = AutoConfig.from_pretrained(env_vars["model_path"], trust_remote_code=True)
+    model_type = config.model_type
+
+    model = MODEL_DICT[model_type].from_pretrained(env_vars["model_path"], torch_dtype="auto", trust_remote_code=True)
+    tokenizer = TOKENIZER_DICT[model_type].from_pretrained(env_vars["model_path"], trust_remote_code=True)
+
+    ds = load_calibration_text_dataset(env_vars["calib_prompt_path"], tokenizer)
+
+    # Quantize the model
+    quantize_model(model, env_vars, ds)
+
+    # save the quantized model
+    save_quantized_model(model, tokenizer, env_vars["export_path"], True)

examples/quantization/llm-compressor/w8a8_int8_dynamic.py

@@ -0,0 +1,82 @@
+from datasets import load_dataset
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import GPTQModifier
+from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
+from llmcompressor.utils import dispatch_for_generation
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# 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 algorithms. In this case, we:
+#   * apply SmoothQuant to make the activations easier to quantize
+#   * quantize the weights to int8 with GPTQ (static per channel)
+#   * quantize the activations to int8 (dynamic per token)
+recipe = [
+    SmoothQuantModifier(smoothing_strength=0.8),
+    GPTQModifier(targets="Linear", scheme="W8A8", ignore=["lm_head"]),
+]
+
+# Apply algorithms and save to output_dir
+oneshot(
+    model=model,
+    dataset=ds,
+    recipe=recipe,
+    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("npu")
+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.rstrip("/").split("/")[-1] + "-W8A8-Dynamic-Per-Token"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)

examples/quantization/llm-compressor/w8a8_int8_dynamic_moe.py

@@ -0,0 +1,26 @@
+import torch
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import QuantizationModifier
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+MODEL_ID = "Qwen/Qwen3-30B-A3B-Instruct-2507"
+
+model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype=torch.bfloat16, trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+
+recipe = QuantizationModifier(
+    targets="Linear",
+    scheme="INT8",
+    ignore=["lm_head", "re:.*mlp.gate$"],
+)
+
+oneshot(
+    model=model,
+    recipe=recipe,
+    trust_remote_code_model=True,
+)
+
+# Save to disk in compressed-tensors format.
+SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-INT8_W8A8"
+model.save_pretrained(SAVE_DIR, save_compressed=True)
+tokenizer.save_pretrained(SAVE_DIR)