File size: 9,940 Bytes
3a464db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import logging
import random
import pytest
import torch
import torch.distributed as dist
from transformers import AutoTokenizer, AutoConfig

from vllm import distributed
from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config, get_current_vllm_config

from dinfer.model import LLaDAModelLM, LLaDAMoeModelLM
from dinfer import BlockWiseDiffusionLLM, ThresholdParallelDecoder, HierarchyDecoder
from dinfer import DiffusionLLMServing, SamplingParams
from dinfer.decoding.utils import BlockIteratorFactory

LLADA_MODEL_PATH = "/mnt/infra/myx/models/LLaDA-1.5/"
MOE_MODEL_PATH = '/mnt/infra/dulun.dl/models/LLaDA-MoE/fusemoe/step45567_converted_hf_fusemoe'

def get_prompts(tokenizer, mask_id, device, num=1):
    prompt = "Lily can run 12 kilometers per hour for 4 hours. After that, she can run 6 kilometers per hour. How many kilometers can she run in 8 hours? "
    m = [{"role": "user", "content": prompt}, ]
    prompt = tokenizer.apply_chat_template(m, add_generation_prompt=True, tokenize=False)
    input_ids1 = torch.tensor(tokenizer(prompt)['input_ids']).to(device).unsqueeze(0)
    len1 = input_ids1.shape[1]

    if num == 2:
        prompt = "Lily can run 12 kilometers per hour for 4 hours. How many kilometers can she run in 4 hours? "
        m = [{"role": "user", "content": prompt}, ]
        prompt = tokenizer.apply_chat_template(m, add_generation_prompt=True, tokenize=False)
        input_ids2 = torch.tensor(tokenizer(prompt)['input_ids']).to(device).unsqueeze(0)
        len2 = input_ids2.shape[1]
        ret = torch.zeros(2, max(len1, len2), dtype=input_ids1.dtype)
        ret[0, 0:len1] = input_ids1
        ret[1, 0:len2] = input_ids2
    else:
        ret = input_ids1

    return ret


@pytest.fixture(scope="function")
def setup_llada_reference():
    """
    Sets up the standard LLaDA model to generate a ground-truth reference 
    before running the server.
    """
    # 1. GPU Selection
    if 'PYTEST_XDIST_WORKER' in os.environ:
        worker_num = int(os.environ['PYTEST_XDIST_WORKER'].replace('gw', ''))
        gpu_id = worker_num % torch.cuda.device_count()
    else:
        gpu_id = 0
    
    device = torch.device(gpu_id)
    torch.cuda.set_device(gpu_id)
    print(f"[test_serving] Initializing LLaDA Reference on GPU {gpu_id}")

    # 2. Init Distributed Env (Required for VLLM internals)
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = str(40000 + random.randint(0, 1000) + gpu_id)
    
    try:
        distributed.init_distributed_environment(1, 0, 'env://', 0, 'nccl')
    except (RuntimeError, AssertionError) as e:
        print(f"Distributed environment already initialized: {e}")
    
    try:
        distributed.initialize_model_parallel(1, backend='nccl')
    except (RuntimeError, AssertionError) as e:
        print(f"Model parallel already initialized: {e}")

    # 3. Load Model
    with set_current_vllm_config(VllmConfig()):
        config = AutoConfig.from_pretrained(LLADA_MODEL_PATH, trust_remote_code=True, local_files_only=True)
        config.flash_attention = True
        config.train_max_sequence_length = 4096
        
        model = LLaDAModelLM.from_pretrained(LLADA_MODEL_PATH, torch_dtype=torch.bfloat16, config=config).eval()
        model = model.to(device)
        
        decoder = ThresholdParallelDecoder(gpu_id, threshold=0.9, use_float64=True)
        tokenizer = AutoTokenizer.from_pretrained(LLADA_MODEL_PATH, trust_remote_code=True, local_files_only=True)
    
    input_ids = get_prompts(tokenizer, mask_id=126336, device=device)
    
    yield model, decoder, tokenizer, input_ids, device
    
    # 4. Cleanup
    print(f"[test_serving] Cleaning up LLaDA Reference on GPU {gpu_id}")
    del model
    del decoder
    torch.cuda.empty_cache()
    try:
        distributed.destroy_model_parallel()
        distributed.destroy_distributed_environment()
    except:
        pass

@pytest.fixture(scope="function")
def setup_moe_reference():
    """
    Sets up the MoE model to generate a ground-truth reference 
    before running the server.
    """
    if 'PYTEST_XDIST_WORKER' in os.environ:
        worker_num = int(os.environ['PYTEST_XDIST_WORKER'].replace('gw', ''))
        gpu_id = worker_num % torch.cuda.device_count()
    else:
        gpu_id = 0
    
    device = torch.device(gpu_id)
    torch.cuda.set_device(gpu_id)
    print(f"[test_serving] Initializing MoE Reference on GPU {gpu_id}")
    
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = str(50000 + random.randint(0, 1000) + gpu_id)
    
    try:
        distributed.init_distributed_environment(1, 0, 'env://', 0, 'nccl')
    except (RuntimeError, AssertionError) as e:
        print(f"Distributed environment already initialized: {e}")
    
    try:
        distributed.initialize_model_parallel(1, backend='nccl')
    except (RuntimeError, AssertionError) as e:
        print(f"Model parallel already initialized: {e}")
    
    parallel_config = ParallelConfig(enable_expert_parallel=True)
    with set_current_vllm_config(VllmConfig(parallel_config=parallel_config)):
        model_config = AutoConfig.from_pretrained(MOE_MODEL_PATH, trust_remote_code=True, local_files_only=True)
        model = LLaDAMoeModelLM(config=model_config).eval()
        model.load_weights(MOE_MODEL_PATH, torch_dtype=torch.bfloat16)
        model = model.to(device)
    
    decoder = ThresholdParallelDecoder(0, threshold=0.9, mask_id=156895, eos_id=156892, use_float64=True)
    tokenizer = AutoTokenizer.from_pretrained(MOE_MODEL_PATH, trust_remote_code=True, local_files_only=True)
    input_ids = get_prompts(tokenizer, mask_id=156895, device=device)
    
    yield model, decoder, tokenizer, input_ids, device
    
    print(f"[test_serving] Cleaning up MoE Reference on GPU {gpu_id}")
    del model
    del decoder
    torch.cuda.empty_cache()
    try:
        distributed.destroy_model_parallel()
        distributed.destroy_distributed_environment()
    except:
        pass


def test_llada_server(setup_llada_reference):
    model, decoder, tokenizer, input_ids, device = setup_llada_reference
    print('test serving of standard diffusion LLaDA')
    
    # 1. Generate Reference Result
    params = SamplingParams(temperature=0, threshold=0.9, mask_id=126336, eos_id=126081, early_stop=True, cache='', cont_weight=0, enable_torch_compile=True, use_bd=False)
    dllm = BlockWiseDiffusionLLM(model, decoder, BlockIteratorFactory(), early_stop=True)
    res1 = dllm.generate(input_ids, gen_length=256, block_length=32).cpu()
    
    del model
    torch.cuda.empty_cache()

    # 2. Test Serving: DP == 1 and TPEP == 1
    print('Test serving: DP == 1 and TPEP == 1')
    llm = DiffusionLLMServing(model=LLADA_MODEL_PATH, model_type='llada', backend='vllm', sample_params=params, num_gpus=1, server_port=random.randint(40000, 50000))
    
    try:
        res = llm.generate(input_ids, gen_length=256, block_length=32)
    finally:
        llm.stop_serving()
        
    assert res.shape == res1.shape
    res1 = res1.to(res.device)
    assert torch.all(res == res1)

def test_moe_server(setup_moe_reference):
    print('test serving of diffusion-MOE')
    model, decoder, tokenizer, input_ids, device = setup_moe_reference
    params = SamplingParams(temperature=0, threshold=0.9, mask_id=156895, eos_id=156892, early_stop=True, cache='', cont_weight=0, enable_torch_compile=False, use_bd=False)

    # 1. Generate Reference Result
    # setup EP context for reference generation
    parallel_config = ParallelConfig(enable_expert_parallel=True)
    with set_current_vllm_config(VllmConfig(parallel_config=parallel_config)):
        dllm = BlockWiseDiffusionLLM(model, decoder, BlockIteratorFactory(), early_stop=True)
        res1 = dllm.generate(input_ids, gen_length=256, block_length=32).cpu()
    
    # Release reference model memory
    del model
    torch.cuda.empty_cache()

    # 2. Test Serving: DP == 1 and TPEP == 1
    print('Test serving: DP == 1 and TPEP == 1')
    llm = DiffusionLLMServing(model=MOE_MODEL_PATH, model_type='llada-moe', backend='vllm', sample_params=params, num_gpus=1, server_port=random.randint(50000, 60000))
    try:
        res = llm.generate(input_ids, gen_length=256, block_length=32)
        assert res.shape == res1.shape
        res1 = res1.to(res.device)
        assert torch.all(res == res1)
    finally:
        llm.stop_serving()

    # 3. Test Serving: DP == 2 and TPEP == 1
    input_ids2 = torch.cat([input_ids, input_ids])
    print('Test serving: DP == 2 and TPEP == 1')
    llm = DiffusionLLMServing(model=MOE_MODEL_PATH, model_type='llada-moe', backend='vllm', sample_params=params, num_gpus=2, dp_size=2, tpep_size=1, server_port=random.randint(50000, 60000))
    try:
        res2 = llm.generate(input_ids2, gen_length=256, block_length=32)
        # Remove EOS and padding tokens before comparison
        assert torch.all(res2[0][res2[0] != 156892] == res[0][res[0] != 156892])
    finally:
        llm.stop_serving()

    # 4. Test Serving: DP == 2 and TPEP == 2
    print('Test serving: DP == 2 and TPEP == 2 (2 GPUs)')
    llm = DiffusionLLMServing(model=MOE_MODEL_PATH, model_type='llada-moe', backend='vllm', sample_params=params, num_gpus=2, dp_size=1, tpep_size=2, server_port=random.randint(50000, 60000))
    try:
        res = llm.generate(input_ids, gen_length=256, block_length=32)
    finally:
        llm.stop_serving()

    print('Test serving: DP == 2 and TPEP == 2 (4 GPUs)')
    input_ids2 = torch.cat([input_ids, input_ids])
    llm = DiffusionLLMServing(model=MOE_MODEL_PATH, model_type='llada-moe', backend='vllm', sample_params=params, num_gpus=4, dp_size=2, tpep_size=2, server_port=random.randint(40000, 50000))
    try:
        res2 = llm.generate(input_ids2, gen_length=256, block_length=32)
        assert torch.all(res2[0][res2[0] != 156892] == res[0][res[0] != 156892])
    finally:
        llm.stop_serving()