File size: 9,532 Bytes
64ddf8d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 | #!/usr/bin/env python3
import re
import pandas as pd
import torch
from hqq.core.quantize import Quantizer as hQuant
from safetensors.torch import save_file as safe_save
from scipy.stats import kurtosis
from torch import uint8
from lm_quant_toolkit.eval.common import calc_bits
from lm_quant_toolkit.utils.safetensors import get_tensor, get_tensor_dual
def load_weight(matrix_name, base_dir):
m = f"{matrix_name}.weight"
return get_tensor(m, base_dir)
def load_weight_dual(prefix, base_dir, st_file):
o = f"{prefix}.weight"
q = f"{prefix}.qweight"
return get_tensor_dual(o, q, base_dir, st_file)
def dequantize(wq, meta):
# Zero/Scale packed together
if "zero_scale" in meta:
zero_scale = meta["zero_scale"]
if zero_scale.dtype == uint8:
meta["zero_q"], meta["scale_q"] = zero_scale[0], zero_scale[1]
else:
meta["zero"], meta["scale"] = zero_scale[0], zero_scale[1]
if meta["quant_zero"]:
meta["zero"] = hQuant.dequantize(meta["zero_q"], meta["meta_zero"])
if meta["quant_scale"]:
meta["scale"] = hQuant.dequantize(meta["scale_q"], meta["meta_scale"])
return hQuant.dequantize(wq, meta)
def restore_weight(matrix, state_dict):
key = matrix
if key in state_dict:
m_dikt = state_dict[key]
if "meta" in m_dikt:
meta_dict = m_dikt["meta"]
meta_scale_dict = meta_dict.get("meta_scale", None)
b1 = meta_dict["nbits"]
g1 = meta_dict["group_size"]
b2 = meta_scale_dict["nbits"] if meta_scale_dict else 8
g2 = meta_scale_dict["group_size"] if meta_scale_dict else 128
quant_config = {
"b1": b1,
"g1": g1,
"b2": b2,
"g2": g2,
}
wq = dequantize(m_dikt["W_q"], meta_dict)
return wq, quant_config
else:
return None, None
else:
return None, None
def save_compare_pair(
base_dir, quant_base_dir, quant_cfg, model_id, layers, output_dir
):
file_path = f"{quant_base_dir}/{model_id}-{quant_cfg}-hqq/qmodel.pt"
state_dict = torch.load(file_path, map_location="cpu")
tensors = {}
metadata = {}
# walk thru the linear layers
# for each layer
# for each linear module
# load the original weight
# load the quantized weight and dequantized
# save the two matrix into a combined safetensors
for layer in range(layers):
matricies = [
f"model.layers.{layer}.mlp.down_proj",
f"model.layers.{layer}.mlp.gate_proj",
f"model.layers.{layer}.mlp.up_proj",
f"model.layers.{layer}.self_attn.k_proj",
f"model.layers.{layer}.self_attn.o_proj",
f"model.layers.{layer}.self_attn.q_proj",
f"model.layers.{layer}.self_attn.v_proj",
]
for matrix in matricies:
wq, quant_cfg = restore_weight(matrix, state_dict)
if wq is None:
# skip unquantized matrix
continue
wo = load_weight(matrix, base_dir)
tensors[f"{matrix}.weight"] = wo
tensors[f"{matrix}.qweight"] = wq
metadata[f"{matrix}.quant_cfg.b1"] = str(quant_cfg["b1"])
metadata[f"{matrix}.quant_cfg.b2"] = str(quant_cfg["b2"])
metadata[f"{matrix}.quant_cfg.g1"] = str(quant_cfg["g1"])
metadata[f"{matrix}.quant_cfg.g2"] = str(quant_cfg["g2"])
output_fp = f"{output_dir}/{model_id}-cmp.safetensors"
safe_save(tensors, output_fp, metadata=metadata)
def compare_pair(model_id, layers, output_dir):
st_file = f"{output_dir}/{model_id}-cmp.safetensors"
for layer in range(layers):
matricies = [
f"model.layers.{layer}.mlp.down_proj",
f"model.layers.{layer}.mlp.gate_proj",
f"model.layers.{layer}.mlp.up_proj",
f"model.layers.{layer}.self_attn.k_proj",
f"model.layers.{layer}.self_attn.o_proj",
f"model.layers.{layer}.self_attn.q_proj",
f"model.layers.{layer}.self_attn.v_proj",
]
for matrix in matricies:
wo, wq = load_weight_dual(matrix, output_dir, st_file)
diff = torch.norm(wo - wq).item()
kurt_peason = kurtosis(
wo.numpy(), axis=None, fisher=False, bias=True, nan_policy="omit"
)
kurt_fisher = kurtosis(
wo.numpy(), axis=None, fisher=True, bias=True, nan_policy="omit"
)
# print(f"{matrix} FNorm Diff: {diff:.5f} Kurtosis: {kurt:.2f}")
print(f"{matrix},{diff:.5f},{kurt_fisher:.3f},{kurt_peason:.3f}")
def is_linear_module(key):
self_attns = ["q_proj", "v_proj", "k_proj", "o_proj"]
mlps = ["gate_proj", "up_proj", "down_proj"]
modules = self_attns + mlps
for module in modules:
if module in key:
return True
return False
def extract_quant_config(base_dir, model_id, config, algo="hqq"):
file_path = f"{base_dir}/{model_id}-{config}-{algo}/qmodel.pt"
dikt = torch.load(file_path, map_location="cpu")
quant_configs = {}
mem_fp16_all_total = 0
mem_all_total = 0
mem_quant_total = 0
param_quant_total = 0
param_all_total = 0
# search quantized linear module with meta
for key in dikt.keys():
m_dikt = dikt[key]
if is_linear_module(key):
if "meta" in m_dikt:
meta_dict = m_dikt["meta"]
meta_scale_dict = meta_dict.get("meta_scale", None)
shape = meta_dict["shape"]
b1 = meta_dict["nbits"]
g1 = meta_dict["group_size"]
b2 = meta_scale_dict["nbits"] if meta_scale_dict else 8
g2 = meta_scale_dict["group_size"] if meta_scale_dict else 128
param_count = shape[0] * shape[1]
param_quant_total += param_count
param_all_total += param_count
memmb = (
(b1 + 2 * b2 / g1 + 32 / (g1 * g2)) * param_count / 8 / 1024 / 1024
)
mem_fp16_all_total += param_count * 2 / 1024 / 1024
mem_quant_total += memmb
mem_all_total += memmb
quant_configs[key] = {
"b1": b1,
"g1": g1,
"b2": b2,
"g2": g2,
"memmb": memmb,
"param_cnt": param_count,
}
else:
w = m_dikt["weight"]
mem_all_total += w.numel() * 2 / 1024 / 1024
mem_fp16_all_total += w.numel() * 2 / 1024 / 1024
param_all_total += w.numel()
return (
quant_configs,
mem_quant_total,
mem_all_total,
mem_fp16_all_total,
param_quant_total,
param_all_total,
)
def get_mem_usage_df(model_ids, confs, base_dir):
dikts = []
for model_id in model_ids:
for conf in confs:
configs, mem_quant_total, mem_all_total, mem_fp16_all_total, _, _ = (
extract_quant_config(base_dir, model_id, conf)
)
dikt = {
"model": model_id.split("/")[1],
"config": conf,
"mem_quant_total": mem_quant_total,
"mem_all_total": mem_all_total,
"mem_fp16_all_total": mem_fp16_all_total,
}
dikts.append(dikt)
df = pd.DataFrame(dikts)
return df
def dump_quant_allocation(
quant_dir, model_ids, confs, csv_fp="mxq-allot.csv", attempts=None, algo="mxq"
):
dikt = []
pat = re.compile(r"model\.layers\.(\d+)\.(.+)")
pat2 = re.compile(r"b(\d+)g(\d+)")
loops = ["mxq1"] if attempts is None else attempts
for attempt in loops:
snapshot_dir = f"{quant_dir}/{attempt}"
for model_id in model_ids:
for conf in confs:
m2 = re.match(pat2, conf)
if m2:
b1 = int(m2.group(1))
g1 = int(m2.group(2))
bit_budget = round(calc_bits(b1, g1), 2)
else:
bit_budget = conf.replace("_", ".")
(
configs,
mem_quant_total,
mem_all_total,
mem_fp16_all_total,
param_quant_total,
param_all_total,
) = extract_quant_config(snapshot_dir, model_id, conf, algo=algo)
for key, val in configs.items():
matcher = re.match(pat, key)
if matcher:
layer = matcher.group(1)
module = matcher.group(2)
val["model"] = model_id.split("/")[1]
val["layer"] = layer
val["module"] = module
val["bit_budget"] = bit_budget
val["params_quant_tot"] = param_quant_total
val["params_all_tot"] = param_all_total
dikt.append(val)
columns = [
"model",
"module",
"layer",
"memmb",
"param_cnt",
"bit_budget",
"b1",
"g1",
"b2",
"g2",
"params_quant_tot",
"params_all_tot",
]
df = pd.DataFrame(dikt)
df.to_csv(csv_fp, index=False, columns=columns)
|