|
| 1 | +# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved. |
| 2 | +# SPDX-License-Identifier: Apache-2.0 |
| 3 | +# |
| 4 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | +# you may not use this file except in compliance with the License. |
| 6 | +# You may obtain a copy of the License at |
| 7 | +# |
| 8 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | +# |
| 10 | +# Unless required by applicable law or agreed to in writing, software |
| 11 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | +# See the License for the specific language governing permissions and |
| 14 | +# limitations under the License. |
| 15 | +"""Export functions for vLLM fakequant.""" |
| 16 | + |
| 17 | +import os |
| 18 | +from pathlib import Path |
| 19 | + |
| 20 | +import torch |
| 21 | +import torch.nn as nn |
| 22 | + |
| 23 | +from modelopt.torch.export.layer_utils import is_quantlinear |
| 24 | +from modelopt.torch.export.model_config import QUANTIZATION_NONE |
| 25 | +from modelopt.torch.quantization.utils import get_quantizer_state_dict |
| 26 | + |
| 27 | + |
| 28 | +def export_hf_vllm_fq_checkpoint( |
| 29 | + model: nn.Module, |
| 30 | + export_dir: Path | str, |
| 31 | +) -> dict[str, torch.Tensor]: |
| 32 | + """Exports the torch model weights and amax values separately. |
| 33 | +
|
| 34 | + This function: |
| 35 | + 1. Extracts amax values for calibration |
| 36 | + 2. Deletes all quantizer parameters from state dict to store only weights in original dtype |
| 37 | +
|
| 38 | + Args: |
| 39 | + model: The quantized model to export |
| 40 | + export_dir: Directory to save the amax values |
| 41 | +
|
| 42 | + Returns: |
| 43 | + post_state_dict: Dict containing quantized weights |
| 44 | + """ |
| 45 | + amax_dict = { |
| 46 | + name + "._amax": param["_amax"].detach().clone().cpu() |
| 47 | + for name, param in get_quantizer_state_dict(model).items() |
| 48 | + if "_amax" in param |
| 49 | + } |
| 50 | + |
| 51 | + # remove quantizer from model |
| 52 | + for _, module in model.named_modules(): |
| 53 | + if is_quantlinear(module): |
| 54 | + delattr(module, "weight_quantizer") |
| 55 | + delattr(module, "input_quantizer") |
| 56 | + delattr(module, "output_quantizer") |
| 57 | + module.export() |
| 58 | + torch.save(amax_dict, f"{export_dir}/quant_amax.pth") |
| 59 | + return model.state_dict() |
| 60 | + |
| 61 | + |
| 62 | +def get_mcore_vllm_fq_quantized_state( |
| 63 | + module: torch.nn.Module, name_to_value: dict, dtype: torch.dtype = torch.bfloat16 |
| 64 | +): |
| 65 | + """Return a state_dict, quantization format, and block_size of the quantized module. |
| 66 | +
|
| 67 | + Args: |
| 68 | + module: The target module to perform real quantization. |
| 69 | + name_to_value: The dictionary to store the quantized state. |
| 70 | + dtype: The default data type. |
| 71 | +
|
| 72 | + Returns: |
| 73 | + Tuple: state dict, quantization format, and block_size of the quantized module. |
| 74 | +
|
| 75 | + """ |
| 76 | + qformat: str = QUANTIZATION_NONE |
| 77 | + block_size = 0 |
| 78 | + |
| 79 | + for name, param in get_quantizer_state_dict(module).items(): |
| 80 | + if "_amax" in param: |
| 81 | + name_to_value[name + "._amax"] = param["_amax"].to(dtype).cpu() |
| 82 | + return name_to_value, qformat, block_size |
| 83 | + |
| 84 | + |
| 85 | +def gather_mcore_vllm_fq_quantized_state_dict( |
| 86 | + state_dict: dict[str, torch.Tensor], save_directory: str | os.PathLike |
| 87 | +): |
| 88 | + """Gather all quantized state dict from all ranks and save them to a file. |
| 89 | +
|
| 90 | + Args: |
| 91 | + state_dict: The state dictionary of the module. |
| 92 | + save_directory: The directory to save the quantized state dict. |
| 93 | +
|
| 94 | + Returns: |
| 95 | + The state dictionary of the module without quantized state. |
| 96 | + """ |
| 97 | + amax_state_dict = { |
| 98 | + k: v.detach().clone().cpu() for k, v in state_dict.items() if k.endswith("_amax") |
| 99 | + } |
| 100 | + |
| 101 | + # Gather all amax dicts to rank 0 |
| 102 | + world_size = torch.distributed.get_world_size() |
| 103 | + rank = torch.distributed.get_rank() |
| 104 | + |
| 105 | + if rank == 0: |
| 106 | + # Rank 0 will collect all amax values |
| 107 | + all_amax_dicts = [None] * world_size |
| 108 | + torch.distributed.gather_object(amax_state_dict, all_amax_dicts, dst=0) |
| 109 | + |
| 110 | + # Merge all amax dicts into one |
| 111 | + merged_amax_dict = {} |
| 112 | + for amax_dict in all_amax_dicts: |
| 113 | + if amax_dict is not None: |
| 114 | + merged_amax_dict.update(amax_dict) |
| 115 | + |
| 116 | + print(f"Total amax entries from all ranks: {len(merged_amax_dict.keys())}") |
| 117 | + torch.save(merged_amax_dict, save_directory + "/quant_amax.pth") |
| 118 | + else: |
| 119 | + # Other ranks just send their amax values |
| 120 | + torch.distributed.gather_object(amax_state_dict, None, dst=0) |
| 121 | + |
| 122 | + torch.distributed.barrier() |
| 123 | + |
| 124 | + # remove amax values from state_dict |
| 125 | + return {k: v for k, v in state_dict.items() if not k.endswith("_amax")} |
0 commit comments