File size: 8,672 Bytes
e0be88b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 |
# Copyright 2024 The HuggingFace Team. All rights reserved.
#
# 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.
import gc
import unittest
from transformers import AutoModelForCausalLM, AutoTokenizer, HqqConfig
from transformers.testing_utils import (
backend_empty_cache,
require_accelerate,
require_hqq,
require_torch_gpu,
require_torch_multi_gpu,
slow,
torch_device,
)
from transformers.utils import is_hqq_available, is_torch_available
if is_torch_available():
import torch
if is_hqq_available():
from hqq.core.quantize import HQQBackend, HQQLinear
class HQQLLMRunner:
def __init__(self, model_id, quant_config, compute_dtype, device, cache_dir=None):
self.model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=compute_dtype,
device_map=device,
quantization_config=quant_config,
low_cpu_mem_usage=True,
cache_dir=cache_dir,
)
self.tokenizer = AutoTokenizer.from_pretrained(model_id, cache_dir=cache_dir)
self.device = self.model.device
HQQLinear.set_backend(HQQBackend.PYTORCH)
def cleanup():
backend_empty_cache(torch_device)
gc.collect()
def check_hqqlayer(test_module, hqq_layer, batch_size=1, context_size=1024):
# Test HQQ layer
W_dequant = hqq_layer.dequantize() # Reconstructed weights
inputs = (
torch.randn(
(batch_size, context_size, hqq_layer.meta["shape"][1]),
device=hqq_layer.device,
dtype=hqq_layer.compute_dtype,
)
/ 10.0
)
with torch.no_grad():
outputs = hqq_layer(inputs)
test_module.assertEqual(outputs.shape[-1], W_dequant.shape[0])
test_module.assertEqual(outputs.dtype, hqq_layer.compute_dtype)
del W_dequant, inputs, outputs
cleanup()
def check_forward(test_module, model, batch_size=1, context_size=1024):
# Test forward pass
with torch.no_grad():
out = model(torch.zeros([batch_size, context_size], device=model.device, dtype=torch.int32)).logits
test_module.assertEqual(out.shape[0], batch_size)
test_module.assertEqual(out.shape[1], context_size)
cleanup()
MODEL_ID = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
@require_torch_gpu
@require_hqq
class HqqConfigTest(unittest.TestCase):
def test_to_dict(self):
"""
Makes sure the config format is properly set
"""
quantization_config = HqqConfig()
hqq_orig_config = quantization_config.to_dict()
self.assertEqual(quantization_config.quant_config, hqq_orig_config["quant_config"])
@slow
@require_torch_gpu
@require_accelerate
@require_hqq
class HQQTest(unittest.TestCase):
def tearDown(self):
cleanup()
def test_fp16_quantized_model(self):
"""
Simple LLM model testing fp16
"""
quant_config = HqqConfig(nbits=8, group_size=64)
hqq_runner = HQQLLMRunner(
model_id=MODEL_ID, quant_config=quant_config, compute_dtype=torch.float16, device=torch_device
)
check_hqqlayer(self, hqq_runner.model.model.layers[0].self_attn.v_proj)
check_forward(self, hqq_runner.model)
@slow
@require_torch_gpu
@require_torch_multi_gpu
@require_accelerate
@require_hqq
class HQQTestMultiGPU(unittest.TestCase):
def tearDown(self):
cleanup()
def test_fp16_quantized_model_multipgpu(self):
"""
Simple LLM model testing fp16 with multi-gpu
"""
quant_config = HqqConfig(nbits=8, group_size=64)
hqq_runner = HQQLLMRunner(
model_id=MODEL_ID, quant_config=quant_config, compute_dtype=torch.float16, device="auto"
)
check_hqqlayer(self, hqq_runner.model.model.layers[0].self_attn.v_proj)
check_forward(self, hqq_runner.model)
@slow
@require_torch_gpu
@require_accelerate
@require_hqq
class HQQTestBias(unittest.TestCase):
def tearDown(self):
cleanup()
def test_fp16_quantized_model(self):
"""
Simple LLM model testing fp16 with bias
"""
quant_config = HqqConfig(nbits=8, group_size=64)
hqq_runner = HQQLLMRunner(
model_id="facebook/opt-125m", quant_config=quant_config, compute_dtype=torch.float16, device=torch_device
)
check_hqqlayer(self, hqq_runner.model.model.decoder.layers[0].self_attn.v_proj)
check_forward(self, hqq_runner.model)
def test_save_and_load_quantized_model(self):
"""
Test saving and loading a quantized model with bias
"""
import tempfile
quant_config = HqqConfig(nbits=8, group_size=64)
hqq_runner = HQQLLMRunner(
model_id="facebook/opt-125m", quant_config=quant_config, compute_dtype=torch.float16, device=torch_device
)
input_tensor = torch.zeros((1, 8), dtype=torch.int32, device=torch_device)
# Get reference logits
with torch.no_grad():
logits_ref = hqq_runner.model.forward(input_tensor).logits
with tempfile.TemporaryDirectory() as tmpdirname:
hqq_runner.model.save_pretrained(tmpdirname)
del hqq_runner.model
backend_empty_cache(torch_device)
model_loaded = AutoModelForCausalLM.from_pretrained(
tmpdirname, torch_dtype=torch.float16, device_map=torch_device
)
with torch.no_grad():
logits_loaded = model_loaded.forward(input_tensor).logits
self.assertEqual((logits_loaded - logits_ref).abs().mean().item(), 0)
@slow
@require_torch_gpu
@require_accelerate
@require_hqq
class HQQSerializationTest(unittest.TestCase):
def tearDown(self):
cleanup()
def test_model_serialization(self):
"""
Simple HQQ LLM save/load test
"""
quant_config = HqqConfig(nbits=4, group_size=64)
hqq_runner = HQQLLMRunner(
model_id=MODEL_ID, quant_config=quant_config, compute_dtype=torch.float16, device=torch_device
)
input_tensor = torch.zeros((1, 8), dtype=torch.int32, device=torch_device)
with torch.no_grad():
logits_ref = hqq_runner.model.forward(input_tensor).logits
# Save
saved_model_id = "quant_model"
hqq_runner.model.save_pretrained(saved_model_id)
# Remove old model
del hqq_runner.model
backend_empty_cache(torch_device)
# Load and check if the logits match
model_loaded = AutoModelForCausalLM.from_pretrained(
"quant_model", torch_dtype=torch.float16, device_map=torch_device, low_cpu_mem_usage=True
)
with torch.no_grad():
logits_loaded = model_loaded.forward(input_tensor).logits
self.assertEqual((logits_loaded - logits_ref).abs().mean().item(), 0)
def test_model_serialization_dynamic_quant_with_skip(self):
"""
Simple HQQ LLM save/load test with dynamic quant
"""
q4_config = {"nbits": 4, "group_size": 64}
q3_config = {"nbits": 3, "group_size": 64}
quant_config = HqqConfig(
dynamic_config={
"self_attn.q_proj": q4_config,
"self_attn.k_proj": q4_config,
"self_attn.v_proj": q4_config,
"self_attn.o_proj": q4_config,
"mlp.gate_proj": q3_config,
"mlp.up_proj": q3_config,
},
skip_modules=["lm_head", "down_proj"],
)
hqq_runner = HQQLLMRunner(
model_id=MODEL_ID, quant_config=quant_config, compute_dtype=torch.float16, device=torch_device
)
model = hqq_runner.model
input_tensor = torch.zeros((1, 8), dtype=torch.int32, device=torch_device)
with torch.no_grad():
model.forward(input_tensor).logits
self.assertEqual(isinstance(model.model.layers[1].mlp.down_proj, torch.nn.Linear), True)
self.assertEqual(model.model.layers[1].self_attn.v_proj.quant_config["weight_quant_params"]["nbits"], 4)
self.assertEqual(model.model.layers[1].mlp.gate_proj.quant_config["weight_quant_params"]["nbits"], 3)
|