Source code for tensorrt_llm.quantization.quantize_by_modelopt

# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Adapted from examples/quantization/hf_ptq.py
"""

import copy
import json
import os
import random
import sys
import time

import numpy as np
import safetensors
import torch
from accelerate.hooks import remove_hook_from_module
from datasets import load_dataset
from safetensors.torch import load_file, save_file
from torch.utils.data import DataLoader
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

from .._utils import release_gc, str_dtype_to_torch
from ..logger import logger
from ..mapping import Mapping
from .mode import QuantAlgo

EMPTY_CFG = {
    "quant_cfg": {
        "*weight_quantizer": {
            "enable": False,
        },
        "*input_quantizer": {
            "enable": False
        },
        "*lm_head*": {
            "enable": False
        },
        "*output_layer*": {
            "enable": False
        },
        "default": {
            "enable": False
        },
    },
    "algorithm": "max",
}

KV_CACHE_CFG = {
    "*.query_key_value.output_quantizer": {
        "num_bits": 8,
        "axis": None,
        "enable": True
    },
    "*.Wqkv.output_quantizer": {
        "num_bits": 8,
        "axis": None,
        "enable": True
    },
    "*.W_pack.output_quantizer": {
        "num_bits": 8,
        "axis": None,
        "enable": True
    },
    "*.c_attn.output_quantizer": {
        "num_bits": 8,
        "axis": None,
        "enable": True
    },
    "*.k_proj.output_quantizer": {
        "num_bits": 8,
        "axis": None,
        "enable": True
    },
    "*.v_proj.output_quantizer": {
        "num_bits": 8,
        "axis": None,
        "enable": True
    },
}


def quant_cfg_choices():
    import modelopt.torch.quantization as mtq
    QUANT_CFG_CHOICES = {
        "int8_sq": mtq.INT8_SMOOTHQUANT_CFG,
        "fp8": mtq.FP8_DEFAULT_CFG,
        "int4_awq": mtq.INT4_AWQ_CFG,
        "w4a8_awq": mtq.W4A8_AWQ_BETA_CFG,
        "int8_wo": EMPTY_CFG,
        "int4_wo": EMPTY_CFG,
        "full_prec": EMPTY_CFG,
    }
    return QUANT_CFG_CHOICES


MODEL_NAME_PATTERN_MAP = {
    "GPT2": "gpt2",
    "Xverse": "llama",
    "Llama": "llama",
    "Mistral": "llama",
    "GPTJ": "gptj",
    "FalconForCausalLM": "falcon",
    "RWForCausalLM": "falcon",
    "baichuan": "baichuan",
    "MPT": "mpt",
    "Bloom": "bloom",
    "ChatGLM": "chatglm",
    "QWen": "qwen",
    "Gemma2": "gemma2",
    "Gemma": "gemma",
    "MixtralForCausalLM": "llama",
    "NemotronForCausalLM": "nemotron",
    "GPTBigCodeForCausalLM": "gpt_bigcode",
    "ArcticForCausalLM": "llama",
    "Phi3SmallForCausalLM": "phi3small",
    "Phi3ForCausalLM": "phi3",
    "Starcoder2ForCausalLM": "gptnext",
    "GPTBigCodeForCausalLM": "gptnext",
    "GLM": "glm",
    "Exaone": "exaone",
    "DeciLMForCausalLM": "deci",
    "DeepseekForCausalLM": "deepseek",
}


class _CustomDataset(torch.utils.data.Dataset):

    def __init__(self, encodings):
        self.encodings = encodings

    def __getitem__(self, idx):
        item = {
            key: val[idx].clone().detach().requires_grad_(False)
            for key, val in self.encodings.items()
        }
        return item

    def __len__(self):
        return len(self.encodings["input_ids"])


def get_tokenizer(ckpt_path, max_seq_length=2048, model_type=None):
    logger.info(f"Initializing tokenizer from {ckpt_path}")
    tokenizer = AutoTokenizer.from_pretrained(
        ckpt_path,
        model_max_length=max_seq_length,
        padding_side="left",
        trust_remote_code=True,
    )

    if tokenizer.pad_token is None:
        if model_type and model_type == "qwen":
            # qwen use token id 151643 as pad and eos tokens
            tokenizer.eos_token = tokenizer.convert_ids_to_tokens(151643)
            tokenizer.pad_token = tokenizer.convert_ids_to_tokens(151643)
        else:
            tokenizer.pad_token = tokenizer.eos_token
    assert tokenizer.pad_token is not None, f"Pad token for {model_type} cannot be set!"

    return tokenizer


def _get_vila_model(model_dir):
    sys.path.append(model_dir + "/../VILA")
    from llava.model import LlavaLlamaConfig, LlavaLlamaModel  # noqa
    from transformers import AutoModel
    model = AutoModel.from_pretrained(
        model_dir,
        device_map='auto',
        trust_remote_code=True,
    )
    return model.llm


def get_hf_config(ckpt_path):
    if "mpt" in ckpt_path:
        # MPT-7B cannot get initialized from AutoConfig
        from transformers import MptConfig
        return MptConfig.from_pretrained(ckpt_path)
    else:
        return AutoConfig.from_pretrained(ckpt_path, trust_remote_code=True)


def _get_llava_qwen_model(model_dir, dtype, device):
    if "hf" in model_dir:
        from transformers import LlavaOnevisionForConditionalGeneration
        model = LlavaOnevisionForConditionalGeneration.from_pretrained(
            model_dir, torch_dtype=dtype, device_map=device)
        model = model.language_model
    else:
        from llava.model.builder import load_pretrained_model
        _, model, _, _ = load_pretrained_model(model_dir,
                                               None,
                                               'llava_qwen',
                                               torch_dtype=dtype,
                                               device_map=device)
    return model


def get_model(ckpt_path: str, dtype: str = 'bfloat16', device: str = 'cuda'):
    logger.info(f"Initializing model from {ckpt_path}")
    # Note: VILA model is not in public HF model zoo yet. We need to explicitly import from the git repo
    hf_config = get_hf_config(ckpt_path)
    torch_dtype = str_dtype_to_torch(dtype)

    model_cls = AutoModelForCausalLM
    if hf_config.model_type == "llava":
        from transformers import LlavaForConditionalGeneration
        model_cls = LlavaForConditionalGeneration
    elif hf_config.model_type == "mpt":
        from transformers import MptForCausalLM
        model_cls = MptForCausalLM
    if "vila" in ckpt_path:
        model = _get_vila_model(ckpt_path)
    elif "llava-onevision-qwen2" in ckpt_path:
        model = _get_llava_qwen_model(ckpt_path, dtype, device)
    elif hf_config.model_type == "glm":
        from transformers import AutoModelForSeq2SeqLM
        model = AutoModelForSeq2SeqLM.from_pretrained(ckpt_path,
                                                      device_map="cuda",
                                                      torch_dtype=torch_dtype,
                                                      trust_remote_code=True)
    else:
        model = model_cls.from_pretrained(
            ckpt_path,
            device_map="auto" if device != "cpu" else "cpu",
            torch_dtype="auto",
            trust_remote_code=True)
        if hf_config.model_type in ["llava", "internvl_chat"]:
            model = model.language_model
    model.eval()

    model_dtype = next(model.parameters()).dtype
    if torch_dtype != model_dtype:
        logger.info(
            f"[TensorRT-LLM][WARNING] The manually set model data type is {dtype}, "
            f"but the data type of the HuggingFace model is {model_dtype}.")

    return model


def get_model_type(model):
    for k, v in MODEL_NAME_PATTERN_MAP.items():
        if k.lower() in type(model).__name__.lower():
            return v
    return None


def get_calib_dataloader(dataset_name_or_dir="cnn_dailymail",
                         tokenizer=None,
                         batch_size=1,
                         calib_size=512,
                         block_size=512,
                         device=None,
                         include_labels=False):
    logger.info("Loading calibration dataset")
    if dataset_name_or_dir == "pileval":
        dataset = load_dataset(
            "json",
            data_files="https://the-eye.eu/public/AI/pile/val.jsonl.zst",
            split="train")
        dataset = dataset["text"][:calib_size]
    elif "cnn_dailymail" in dataset_name_or_dir:
        dataset = load_dataset(
            dataset_name_or_dir,
            name="3.0.0",
            split="train",
        )
        dataset = dataset["article"][:calib_size]
    elif os.path.isdir(dataset_name_or_dir):
        logger.info(
            f"Recognized local dataset repo {dataset_name_or_dir} for calibration; "
            "assuming the calibration data are in the train split and text column."
        )
        dataset = load_dataset(dataset_name_or_dir, split="train")
        dataset = dataset["text"][:calib_size]
    else:
        raise NotImplementedError(
            f"Unsupported dataset name or local repo directory: {dataset_name_or_dir}."
        )

    batch_encoded = tokenizer.batch_encode_plus(dataset,
                                                return_tensors="pt",
                                                padding=True,
                                                truncation=True,
                                                max_length=block_size)

    if device:
        batch_encoded = batch_encoded.to(device)

    if include_labels:
        # Labels are needed when backward is called in the model.
        # The labels should be a shifted version of the input_ids.
        # However, we should not shift the input_ids here since the labels are shifted by
        # Huggingface models during loss calculation as shown here -
        # https://github.com/huggingface/transformers/blob/7f79a97399bb52aad8460e1da2f36577d5dccfed/src/transformers/models/llama/modeling_llama.py#L1093-L1095
        batch_encoded["labels"] = torch.where(
            batch_encoded["attention_mask"] > 0.5, batch_encoded["input_ids"],
            -100)
        batch_encoded = _CustomDataset(batch_encoded)
    else:
        # For backward compatibility, if labels are not needed, we only return input_ids.
        batch_encoded = _CustomDataset(
            {"input_ids": batch_encoded["input_ids"]})

    calib_dataloader = DataLoader(batch_encoded,
                                  batch_size=batch_size,
                                  shuffle=False)

    return calib_dataloader


def quantize_model(model, quant_cfg, calib_dataloader, batch_size, qformat,
                   auto_quantize_bits):
    import modelopt.torch.quantization as mtq

    # NOTE: for ModelOpt v0.19 release
    # calibrate_loop = dataset_utils.create_forward_loop(
    #     calib_dataloader, dataloader=calib_dataloader)

    def calibrate_loop():
        if calib_dataloader is None:
            return
        with torch.no_grad():
            low_mem_mode = False
            for _, data in enumerate(calib_dataloader):
                batch_size = data[list(data.keys())[0]].shape[0]
                if batch_size == 1:
                    model(**data)
                elif not low_mem_mode:
                    # Try running the forward once.
                    # If output memory, we try running inference with split input tensors
                    try:
                        model(**data)
                    except torch.OutOfMemoryError:
                        print(
                            "Warning: torch.OutOfMemoryError detected, try reducing the batch size..."
                        )
                        low_mem_mode = True

                if low_mem_mode:
                    split_data_1 = {
                        key: data[key][:batch_size // 2, ...]
                        for key in data
                    }
                    model(**split_data_1)

                    split_data_2 = {
                        key: data[key][batch_size // 2:, ...]
                        for key in data
                    }
                    model(**split_data_2)

    QUANT_CFG_CHOICES = {
        "int8": "INT8_DEFAULT_CFG",
        "int8_sq": "INT8_SMOOTHQUANT_CFG",
        "fp8": "FP8_DEFAULT_CFG",
        "int4_awq": "INT4_AWQ_CFG",
        "w4a8_awq": "W4A8_AWQ_BETA_CFG",
    }

    logger.info("Starting quantization...")
    start_time = time.time()
    if auto_quantize_bits:
        logger.info("Starting mixed precision quantization...")
        model, search_history = mtq.auto_quantize(
            model,
            data_loader=calib_dataloader,
            loss_func=lambda output, batch: output.loss,
            constraints={"effective_bits": auto_quantize_bits},
            forward_step=lambda model, batch: model(**batch),
            quantization_formats=[
                QUANT_CFG_CHOICES[item] for item in qformat.split(",")
            ] + [None],
            num_calib_steps=len(calib_dataloader),
            num_score_steps=min(
                len(calib_dataloader), 128 // batch_size
            ),  # Limit the number of score steps to avoid long calibration time
            verbose=True,
        )
        mtq.print_quant_summary(model)

        # We need to explicitly calibrate for kv cache quantization
        enable_kv_cache_quantization = "int8" not in qformat
        if enable_kv_cache_quantization:
            mtq.set_quantizer_by_cfg(
                model,
                quant_cfg={
                    "*output_quantizer": {
                        "num_bits": (4, 3),
                        "axis": None,
                        "enable": True
                    }
                },
            )
            # Lets calibrate only the output quantizer this time. Let's disable all other quantizers.
            with mtq.set_quantizer_by_cfg_context(model, {
                    "*": {
                        "enable": False
                    },
                    "*output_quantizer": {
                        "enable": True
                    }
            }):
                mtq.calibrate(model,
                              algorithm="max",
                              forward_loop=calibrate_loop)
    else:
        mtq.quantize(model, quant_cfg, forward_loop=calibrate_loop)
    end_time = time.time()
    logger.info(
        "Quantization done. Total time used: {:.2f} s.".format(end_time -
                                                               start_time))
    return model


def combine_medusa_weight(tp_size, pp_size, base_model_output_dir,
                          num_medusa_heads, num_medusa_layers, max_draft_len,
                          medusa_hidden_act, medusa_model_dir,
                          quant_medusa_head):

    with open(f"{medusa_model_dir}/config.json", "r") as fp:
        medusa_config = json.load(fp)

    num_medusa_heads_from_config = medusa_config.get('medusa_num_heads',
                                                     num_medusa_heads)
    num_medusa_layers = medusa_config.get('medusa_num_layers',
                                          num_medusa_layers)
    if num_medusa_heads is None:
        num_medusa_heads = num_medusa_heads_from_config

    assert max_draft_len > 0, "should have max_draft_len > 0"

    world_size = tp_size * pp_size
    # Process for each rank
    for rank in range(world_size):
        mapping = Mapping(world_size=world_size,
                          rank=rank,
                          tp_size=tp_size,
                          pp_size=pp_size)
        # 1. Load medusa weight for each rank
        from tensorrt_llm.models.medusa.weight import load_medusa_hf
        medusa_weights = load_medusa_hf(medusa_path=medusa_model_dir,
                                        num_medusa_heads=num_medusa_heads,
                                        num_medusa_layers=num_medusa_layers,
                                        mapping=mapping,
                                        dtype="float16")
        # 2. Load base model safetensors (after quant)
        base_model_weights = load_file(
            f"{base_model_output_dir}/rank{rank}.safetensors")

        # 3. Combine and save weight
        base_model_weights.update(medusa_weights)
        save_file(base_model_weights,
                  f"{base_model_output_dir}/rank{rank}.safetensors")

    # 4. Add medusa config into config.json
    with open(f"{base_model_output_dir}/config.json", 'r') as f:
        base_model_config = json.load(f)
        f.close()

    with open(f"{base_model_output_dir}/config.json", 'w') as f:
        base_model_config['architecture'] = "MedusaForCausalLM"
        base_model_config['quantization']['exclude_modules'] = [
            'lm_head',
            '*router',
            '*vocab_embedding',
            '*position_embedding',
            '*block_embedding',
        ]
        if not quant_medusa_head:
            base_model_config['quantization']['exclude_modules'].append(
                '*medusa_heads*')

        base_model_config['max_draft_len'] = max_draft_len
        base_model_config['num_medusa_heads'] = num_medusa_heads
        base_model_config['num_medusa_layers'] = num_medusa_layers
        json.dump(base_model_config, f, indent=4)

    torch.cuda.empty_cache()
    logger.info("Combine medusa heads' weight, done.")


[docs] def quantize_and_export(*, model_dir, device, calib_dataset, dtype, qformat, kv_cache_dtype, calib_size, batch_size, calib_max_seq_length, awq_block_size, output_dir, tp_size, pp_size, seed, tokenizer_max_seq_length, num_medusa_heads=None, num_medusa_layers=None, max_draft_len=None, medusa_hidden_act=None, medusa_model_dir=None, quant_medusa_head=None, auto_quantize_bits=None): ''' Load model from the model_dir, call Modelopt to quantize the model, and then export the quantized model as TRT-LLM checkpoint ''' try: import modelopt # noqa except ImportError as e: logger.error( "Failed to import modelopt, pls check the Modelopt installation. Currently it is known to be unsupported on Windows OS" ) raise e from modelopt.torch.export import export_tensorrt_llm_checkpoint from tensorrt_llm.models.convert_utils import infer_dtype if not torch.cuda.is_available(): raise EnvironmentError("GPU is required for inference.") random.seed(seed) np.random.seed(seed) # Check that only one quantization format is provided for non auto_quant case if not auto_quantize_bits: assert (len(qformat.split(",")) == 1 ), "Quantization supports only one quantization format." hf_config = get_hf_config(model_dir) dtype = infer_dtype(dtype, getattr(hf_config, 'torch_dtype', None)) model = get_model(model_dir, dtype, device=device) model_type = get_model_type(model) if "vila" in model_dir: tokenizer = get_tokenizer(model_dir + "/llm", max_seq_length=tokenizer_max_seq_length, model_type=model_type) else: tokenizer = get_tokenizer(model_dir, max_seq_length=tokenizer_max_seq_length, model_type=model_type) if qformat in ["full_prec", "int8_wo", "int4_wo" ] and kv_cache_dtype is None: logger.info(f"No quantization applied, export {dtype} model") else: if "awq" in qformat: if calib_size > 32: logger.info( f"AWQ calibration could take longer with calib_size = {calib_size}, Using" " calib_size=32 instead") calib_size = 32 logger.info( "\nAWQ calibration could take longer than other calibration methods. Please" " increase the batch size to speed up the calibration process. Batch size can be" " set by adding the argument --batch_size <batch_size> to the command line.\n" ) quant_cfg = None if not auto_quantize_bits: if qformat in quant_cfg_choices(): quant_cfg = quant_cfg_choices()[qformat] else: raise ValueError(f"Unsupported quantization format: {qformat}") if "awq" in qformat: quant_cfg = copy.deepcopy(quant_cfg_choices()[qformat]) weight_quantizer = quant_cfg["quant_cfg"]["*weight_quantizer"] if isinstance(weight_quantizer, list): weight_quantizer = weight_quantizer[0] if awq_block_size: weight_quantizer["block_sizes"][-1] = awq_block_size # Coarser optimal scale search seems to resolve the overflow in TRT-LLM for some models if "w4a8_awq" == qformat and model_type in ["gemma", "mpt"]: quant_cfg["algorithm"] = { "method": "awq_lite", "alpha_step": 1 } if kv_cache_dtype is not None: if kv_cache_dtype == "fp8": for value in KV_CACHE_CFG.values(): value.update({"num_bits": (4, 3)}) # type: ignore quant_cfg["quant_cfg"].update(KV_CACHE_CFG) # type: ignore # Gemma 7B has accuracy regression using alpha 1. We set 0.5 instead. if model_type == "gemma" and "int8_sq" in qformat: quant_cfg["algorithm"] = {"method": "smoothquant", "alpha": 0.5} calib_dataloader = get_calib_dataloader( dataset_name_or_dir=calib_dataset, tokenizer=tokenizer, batch_size=batch_size, calib_size=calib_size, block_size=calib_max_seq_length, device=device, include_labels=auto_quantize_bits is not None, ) model = quantize_model(model, quant_cfg, calib_dataloader, batch_size, qformat, auto_quantize_bits) with torch.inference_mode(): if model_type is None: logger.info( f"Unknown model type {type(model).__name__}. Continue exporting..." ) model_type = f"unknown:{type(model).__name__}" export_path = output_dir start_time = time.time() # Move meta tensor back to device before exporting. remove_hook_from_module(model, recurse=True) QUANT_ALGO = { "int8": "INT8", "int8_sq": "W8A8_SQ_PER_CHANNEL", "fp8": "FP8", "int4_awq": "W4A16_AWQ", "w4a8_awq": "W4A8_AWQ", } export_tensorrt_llm_checkpoint( model, model_type, getattr(torch, dtype), export_dir=export_path, inference_tensor_parallel=tp_size, inference_pipeline_parallel=pp_size, ) with open(f"{export_path}/config.json", "r") as f: tensorrt_llm_config = json.load(f) tensorrt_llm_config["model_type"] = model_type # Workaround for wo quantization if qformat in ["int8_wo", "int4_wo", "full_prec"]: if qformat == "int8_wo": tensorrt_llm_config["quantization"][ "quant_algo"] = QuantAlgo.W8A16 elif qformat == "int4_wo": tensorrt_llm_config["quantization"][ "quant_algo"] = QuantAlgo.W4A16 else: tensorrt_llm_config["quantization"]["quant_algo"] = None # HF uses rope_scaling while tensorrt_llm uses rotary_scaling if hasattr( model.config, "rope_scaling") and "rotary_scaling" not in tensorrt_llm_config: tensorrt_llm_config["rotary_scaling"] = getattr( model.config, "rope_scaling") with open(f"{export_path}/config.json", "w") as f: json.dump(tensorrt_llm_config, f, indent=4) # Workaround for Modelopt 0.9.x fp8_kv_cache knob issue if qformat == 'fp8' and kv_cache_dtype is None: with open(f"{export_path}/config.json", "r") as f: tensorrt_llm_config = json.load(f) tensorrt_llm_config["quantization"]["kv_cache_quant_algo"] = None with open(f"{export_path}/config.json", "w") as f: json.dump(tensorrt_llm_config, f, indent=4) # Workaround for share_embedding_table if pp_size == 1: with safetensors.safe_open(f"{export_path}/rank0.safetensors", framework='pt', device='cpu') as f: share_embedding_table = 'lm_head.weight' not in f.keys() if share_embedding_table: with open(f"{export_path}/config.json", "r") as f: tensorrt_llm_config = json.load(f) tensorrt_llm_config["share_embedding_table"] = True with open(f"{export_path}/config.json", "w") as f: json.dump(tensorrt_llm_config, f, indent=4) # Workaround for qwen version if model_type == 'qwen': with open(f"{export_path}/config.json", "r") as f: tensorrt_llm_config = json.load(f) qwen_config = AutoConfig.from_pretrained(model_dir, trust_remote_code=True) try: from transformers import LlavaOnevisionConfig if isinstance(qwen_config, LlavaOnevisionConfig): qwen_config = qwen_config.text_config except: pass tensorrt_llm_config["qwen_type"] = qwen_config.model_type if qwen_config.model_type == "qwen2": tensorrt_llm_config["norm_epsilon"] = qwen_config.rms_norm_eps tensorrt_llm_config["rotary_base"] = qwen_config.rope_theta tensorrt_llm_config[ "intermediate_size"] = qwen_config.intermediate_size with open(f"{export_path}/config.json", "w") as f: json.dump(tensorrt_llm_config, f, indent=4) # Set rotary parameters correctly for chatglm. if model_type == 'chatglm': rotary_base = 10000.0 rotary_embedding_scaling = None chatglm_config = AutoConfig.from_pretrained(model_dir, trust_remote_code=True) chatglm_version = tensorrt_llm_config['chatglm_version'] rope_ratio = tensorrt_llm_config.get('rope_ratio', 1.0) if chatglm_version == 'chatglm2': if rope_ratio > 1: rotary_embedding_scaling = { 'type': 'linear', 'factor': rope_ratio } elif chatglm_version == 'chatglm3': rotary_base *= rope_ratio with open(f"{export_path}/config.json", "r") as f: tensorrt_llm_config = json.load(f) tensorrt_llm_config['rotary_base'] = rotary_base tensorrt_llm_config['rotary_scaling'] = rotary_embedding_scaling tensorrt_llm_config['rotary_pct'] = 0.5 with open(f"{export_path}/config.json", "w") as f: json.dump(tensorrt_llm_config, f, indent=4) # Workaround for combining medusa head # TODO: move these integration into modelopt to avoid redundant reading and writing if medusa_model_dir is not None: combine_medusa_weight(tp_size, pp_size, export_path, num_medusa_heads, num_medusa_layers, max_draft_len, medusa_hidden_act, medusa_model_dir, quant_medusa_head) end_time = time.time() logger.info( "Quantized model exported to {} \nTotal time used {:.2f} s.".format( export_path, end_time - start_time)) # Need to delete the model and release memory explicitly; # otherwise torch may retain its GPU memory until a delayed GC running, # which reduces the available GPU memory for subsequent stages. del model release_gc()
def unwrap_model(model, module_instances=None): # Reference: https://github.com/NVIDIA/Megatron-LM/blob/core_r0.8.0/megatron/training/utils.py from megatron.core import DistributedDataParallel as DDP from megatron.core.transformer.module import Float16Module if module_instances is None: module_instances = (DDP, Float16Module) return_list = True if not isinstance(model, list): model = [model] return_list = False unwrapped_model = [] for model_module in model: while isinstance(model_module, module_instances): model_module = model_module.module unwrapped_model.append(model_module) if not return_list: return unwrapped_model[0] return unwrapped_model def get_nemo_calib_dataloader(dataset_name_or_dir="cnn_dailymail", batch_size=64, calib_size=512, max_sequence_length=512): if dataset_name_or_dir == "pileval": dataset = load_dataset( "json", data_files="https://the-eye.eu/public/AI/pile/val.jsonl.zst", split="train") text_column = "text" elif "wikitext" in dataset_name_or_dir: dataset = load_dataset(dataset_name_or_dir, "wikitext-103-v1", split="train") text_column = "text" elif "cnn_dailymail" in dataset_name_or_dir: dataset = load_dataset(dataset_name_or_dir, name="3.0.0", split="train") text_column = "article" elif os.path.isdir(dataset_name_or_dir): logger.info( f"Recognized local dataset repo {dataset_name_or_dir} for calibration; " "assuming the calibration data are in the train split and text column." ) dataset = load_dataset(dataset_name_or_dir, split="train") text_column = "text" else: raise NotImplementedError( f"Unsupported dataset name or local repo directory: {dataset_name_or_dir}." ) calib_size = max(min(len(dataset), calib_size), batch_size) for i in range(calib_size // batch_size): batch = dataset[i * batch_size:(i + 1) * batch_size][text_column] for j in range(len(batch)): batch[j] = batch[j][:max_sequence_length] yield batch def quantize_nemo_and_export(*, nemo_ckpt_path, decoder_type, calib_dataset, calib_tp_size, calib_pp_size, dtype, qformat, kv_cache_dtype, calib_size, batch_size, calib_max_seq_length, awq_block_size, output_dir, tp_size, pp_size, seed): try: import modelopt # noqa except ImportError as e: logger.error( "Failed to import modelopt, pls check the modelopt installation. Currently it is known to be unsupported on Windows OS" ) raise e import modelopt.torch.quantization as mtq from megatron.core import parallel_state from megatron.core.transformer.module import Float16Module from modelopt.torch.export import export_tensorrt_llm_checkpoint from modelopt.torch.utils import print_rank_0 from nemo.collections.nlp.models.language_modeling.megatron_gpt_model import \ MegatronGPTModel from nemo.collections.nlp.modules.common.text_generation_strategy import \ GPTModelTextGenerationStrategy from nemo.collections.nlp.parts.nlp_overrides import ( NLPDDPStrategy, NLPSaveRestoreConnector) from nemo.utils.model_utils import load_config, save_artifacts from omegaconf.omegaconf import open_dict from pytorch_lightning.trainer.trainer import Trainer if not torch.cuda.is_available(): raise EnvironmentError("GPU is required for the inference.") random.seed(seed) np.random.seed(seed) model_cfg = load_config(nemo_ckpt_path) # dtype is used for non-quantized layers supported_dtype = ["auto", "float16", "bfloat16"] assert dtype in supported_dtype, f"{dtype} not supported. Supported dtypes are {supported_dtype}" if dtype == 'auto': dtype = model_cfg.get('precision', None) if dtype is None: dtype = 'float16' elif 'bf16' in dtype or 'bfloat16' in dtype: dtype = 'bfloat16' else: dtype = 'float16' logger.info(f"Specified dtype 'auto'; inferred dtype {dtype!r}.") torch_dtype = getattr(torch, dtype) with open_dict(model_cfg): model_cfg.activations_checkpoint_method = None model_cfg.activations_checkpoint_granularity = None model_cfg.tensor_model_parallel_size = calib_tp_size model_cfg.pipeline_model_parallel_size = calib_pp_size model_cfg.sequence_parallel = False # Only custom modelopt spec is supported for PTQ: this custom spec is largely based on local Megatron-LM # layer definitions to avoid Transformer Engine implementations that are currently not supported. model_cfg.name = "modelopt" # trainer required for restoring model parallel models trainer_config = { 'devices': calib_tp_size * calib_pp_size, 'num_nodes': 1, 'accelerator': 'gpu', 'logger': False, 'precision': model_cfg.precision, 'enable_checkpointing': False, } trainer = Trainer(strategy=NLPDDPStrategy(), **trainer_config) connector = NLPSaveRestoreConnector() model = MegatronGPTModel.restore_from( restore_path=nemo_ckpt_path, trainer=trainer, override_config_path=model_cfg, save_restore_connector=connector, ) model.freeze() print_rank_0(model) # Have to turn off activations_checkpoint_method for inference try: model.model.module.language_model.encoder.activations_checkpoint_method = None except AttributeError: pass # Check whether the DDP is initialized if parallel_state.is_unitialized(): def dummy(): return if model.trainer.strategy.launcher is not None: model.trainer.strategy.launcher.launch(dummy, trainer=model.trainer) model.trainer.strategy.setup_environment() inference_config = { 'greedy': False, 'top_k': 0, 'top_p': 0.9, 'temperature': 1.0, 'add_BOS': True, 'tokens_to_generate': 30, 'all_probs': False, 'repetition_penalty': 1.2, 'min_tokens_to_generate': 0, 'compute_logprob': False, 'batch_size': batch_size, 'max_context_length': calib_max_seq_length, 'strategy': GPTModelTextGenerationStrategy(model), } model.set_inference_config(inference_config) if qformat in ["full_prec", "int8_wo", "int4_wo" ] and kv_cache_dtype is None: print_rank_0(f"No quantization applied, export {dtype} model") else: if "awq" in qformat: if calib_size > 32: print_rank_0( "AWQ calibration could take longer with calib_size =" f" {calib_size}, Using calib_size=32 instead") calib_size = 32 print_rank_0( "\nAWQ calibration could take longer than other calibration methods. Please" " increase the batch size to speed up the calibration process. Batch size can be" " set by adding the argument inference.batch_size=<batch_size> to the command" " line.\n") dataloader = get_nemo_calib_dataloader( dataset_name_or_dir=calib_dataset, batch_size=batch_size, calib_size=calib_size, max_sequence_length=calib_max_seq_length, ) # =================== Start Quantization ==================== if qformat in quant_cfg_choices(): quant_cfg = quant_cfg_choices()[qformat] else: raise ValueError(f"Unsupported quantization format: {qformat}") if "awq" in qformat: quant_cfg = copy.deepcopy(quant_cfg_choices()[qformat]) weight_quantizer = quant_cfg["quant_cfg"][ "*weight_quantizer"] # type: ignore if isinstance(weight_quantizer, list): weight_quantizer = weight_quantizer[0] weight_quantizer["block_sizes"][-1] = awq_block_size if kv_cache_dtype is not None: if kv_cache_dtype == "fp8": for value in KV_CACHE_CFG.values(): value.update({"num_bits": (4, 3)}) # type: ignore quant_cfg["quant_cfg"].update(KV_CACHE_CFG) # type: ignore print_rank_0(quant_cfg) # Always turn on FP8 kv cache to save memory footprint. # For int8_sq, we use int8 kv cache. # TODO: Investigate why enabling FP8 kv cache will cause accuracy regressions for nemotron. # quant_cfg["quant_cfg"]["*output_quantizer"] = { # type: ignore[index] # "num_bits": 8 if args.qformat == "int8_sq" else (4, 3), # "axis": None, # "enable": args.decoder_type != "gptnext", # } dataloader = [data for data in dataloader] def forward_loop(model): for i, batch in enumerate(dataloader): print_rank_0(f"Calibrating batch {i}") model.predict_step(batch, i) start_time = time.time() model = mtq.quantize(model, quant_cfg, forward_loop) # type: ignore[arg-type] end_time = time.time() tot_time = end_time - start_time tput = calib_size / tot_time print_rank_0( f"Quantization done. Total time used {tot_time}s. Throughput {tput} samples/s" ) # =================== End Quantization ====================== if decoder_type == "gptnext": # We found squared_relu may have an under-calibration problem. # Clamp the scaling_factor with a min threshold to avoid under-calibration. maxbound = 0 if qformat == "fp8": maxbound = 448 elif qformat == "int8_sq": maxbound = 127 model = mtq.postprocess_amax( model, "*input_quantizer", lambda amax: torch.clamp(amax, min=0.01 * maxbound)) if torch.distributed.get_rank() == 0: mtq.print_quant_summary(model) if model_cfg.megatron_amp_O2: model.model = unwrap_model(model.model, Float16Module) start_time = time.time() export_tensorrt_llm_checkpoint( model, decoder_type, torch_dtype, export_dir=output_dir, inference_tensor_parallel=tp_size, inference_pipeline_parallel=pp_size, ) end_time = time.time() print_rank_0( f"Model config exported to: {output_dir}. Total time used {end_time - start_time}s" ) if torch.distributed.get_rank() == 0: save_artifacts(model, output_dir, use_abspath=True) # Need to delete the model and release memory explicitly; # otherwise torch may retain its GPU memory until a delayed GC running, # which reduces the available GPU memory for subsequent stages. del model release_gc()