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
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
|
#!/usr/bin/env python3
"""
Benchmark inference speed: Transformers vs vLLM.
This script helps diagnose the 100x slowdown issue by comparing:
1. Raw transformers inference (current implementation)
2. vLLM server inference (target implementation)
Usage:
# First, start vLLM server:
# CUDA_VISIBLE_DEVICES=0 vllm serve /path/to/model --port 8003
# Then run benchmark:
python benchmark_inference.py --mode both --n 20
python benchmark_inference.py --mode vllm --url http://localhost:8003/v1 --n 50
python benchmark_inference.py --mode transformers --model /path/to/model --n 10
"""
import argparse
import json
import time
import sys
from pathlib import Path
from typing import List, Dict, Any
from dataclasses import dataclass
# Add paths
sys.path.insert(0, str(Path(__file__).parent.parent))
sys.path.insert(0, str(Path(__file__).parent.parent.parent / "src"))
@dataclass
class BenchmarkResult:
mode: str
n_requests: int
total_time_s: float
avg_latency_ms: float
min_latency_ms: float
max_latency_ms: float
throughput_req_per_s: float
throughput_conv_per_hr: float # Estimated conversations per hour
errors: int
def benchmark_transformers(
model_path: str,
n_requests: int = 10,
device: str = "cuda:0",
) -> BenchmarkResult:
"""Benchmark raw transformers inference."""
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
print(f"Loading model from {model_path}...")
load_start = time.time()
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
model_path,
torch_dtype=torch.bfloat16,
device_map=device,
)
if tokenizer.pad_token_id is None:
tokenizer.pad_token = tokenizer.eos_token
load_time = time.time() - load_start
print(f"Model loaded in {load_time:.1f}s")
# Test prompt (simulating a typical user simulator turn)
test_messages = [
{"role": "system", "content": "You are a user simulator. Output JSON with reasoning, draft_answer, should_terminate, and response fields."},
{"role": "user", "content": "The agent said: 'Hello, how can I help you today?' Respond as the user."},
]
prompt = tokenizer.apply_chat_template(test_messages, tokenize=False, add_generation_prompt=True)
latencies = []
errors = 0
print(f"Running {n_requests} inference requests...")
start_time = time.time()
for i in range(n_requests):
try:
req_start = time.time()
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)
inputs = {k: v.to(model.device) for k, v in inputs.items()}
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=256,
do_sample=True,
temperature=0.7,
top_p=0.9,
eos_token_id=tokenizer.eos_token_id,
pad_token_id=tokenizer.pad_token_id,
)
# Decode output
input_len = inputs["input_ids"].shape[1]
gen_ids = outputs[0][input_len:]
response = tokenizer.decode(gen_ids, skip_special_tokens=True)
latency_ms = (time.time() - req_start) * 1000
latencies.append(latency_ms)
if (i + 1) % 5 == 0:
print(f" Completed {i + 1}/{n_requests}, last latency: {latency_ms:.0f}ms")
except Exception as e:
errors += 1
print(f" Error on request {i + 1}: {e}")
total_time = time.time() - start_time
if not latencies:
return BenchmarkResult(
mode="transformers",
n_requests=n_requests,
total_time_s=total_time,
avg_latency_ms=0,
min_latency_ms=0,
max_latency_ms=0,
throughput_req_per_s=0,
throughput_conv_per_hr=0,
errors=errors,
)
avg_latency = sum(latencies) / len(latencies)
# Estimate: ~10 turns per conversation, so conv/hr = (req/s) * 3600 / 10
throughput = len(latencies) / total_time
conv_per_hr = throughput * 3600 / 10
return BenchmarkResult(
mode="transformers",
n_requests=n_requests,
total_time_s=total_time,
avg_latency_ms=avg_latency,
min_latency_ms=min(latencies),
max_latency_ms=max(latencies),
throughput_req_per_s=throughput,
throughput_conv_per_hr=conv_per_hr,
errors=errors,
)
def benchmark_vllm(
base_url: str = "http://localhost:8003/v1",
n_requests: int = 10,
concurrent: bool = False,
n_workers: int = 4,
) -> BenchmarkResult:
"""Benchmark vLLM server inference."""
from utils.vllm_client import VLLMClient
client = VLLMClient(base_url=base_url)
# Check health
if not client.health_check():
print(f"ERROR: vLLM server at {base_url} is not responding")
return BenchmarkResult(
mode="vllm",
n_requests=n_requests,
total_time_s=0,
avg_latency_ms=0,
min_latency_ms=0,
max_latency_ms=0,
throughput_req_per_s=0,
throughput_conv_per_hr=0,
errors=n_requests,
)
print(f"vLLM server healthy: {client.get_model_info()}")
# Test messages
test_messages = [
{"role": "system", "content": "You are a user simulator. Output JSON with reasoning, draft_answer, should_terminate, and response fields."},
{"role": "user", "content": "The agent said: 'Hello, how can I help you today?' Respond as the user."},
]
latencies = []
errors = 0
print(f"Running {n_requests} inference requests (concurrent={concurrent})...")
start_time = time.time()
if concurrent:
from concurrent.futures import ThreadPoolExecutor, as_completed
with ThreadPoolExecutor(max_workers=n_workers) as executor:
futures = [
executor.submit(client.chat, test_messages, 256, 0.7)
for _ in range(n_requests)
]
for i, future in enumerate(as_completed(futures)):
try:
result = future.result()
latencies.append(result["latency_ms"])
if (i + 1) % 10 == 0:
print(f" Completed {i + 1}/{n_requests}")
except Exception as e:
errors += 1
print(f" Error: {e}")
else:
for i in range(n_requests):
try:
result = client.chat(test_messages, 256, 0.7)
latencies.append(result["latency_ms"])
if (i + 1) % 5 == 0:
print(f" Completed {i + 1}/{n_requests}, last latency: {result['latency_ms']:.0f}ms")
except Exception as e:
errors += 1
print(f" Error on request {i + 1}: {e}")
total_time = time.time() - start_time
if not latencies:
return BenchmarkResult(
mode="vllm" + ("_concurrent" if concurrent else ""),
n_requests=n_requests,
total_time_s=total_time,
avg_latency_ms=0,
min_latency_ms=0,
max_latency_ms=0,
throughput_req_per_s=0,
throughput_conv_per_hr=0,
errors=errors,
)
avg_latency = sum(latencies) / len(latencies)
throughput = len(latencies) / total_time
conv_per_hr = throughput * 3600 / 10
return BenchmarkResult(
mode="vllm" + ("_concurrent" if concurrent else ""),
n_requests=n_requests,
total_time_s=total_time,
avg_latency_ms=avg_latency,
min_latency_ms=min(latencies),
max_latency_ms=max(latencies),
throughput_req_per_s=throughput,
throughput_conv_per_hr=conv_per_hr,
errors=errors,
)
def benchmark_full_conversation(
vllm_url_70b: str,
vllm_url_8b: str,
n_conversations: int = 5,
max_turns: int = 10,
) -> Dict[str, Any]:
"""
Benchmark a full multi-turn conversation with user simulator and agent.
This simulates the actual experiment loop.
"""
from utils.vllm_client import VLLMClient, VLLMUserSimulator, VLLMAgentAdapter
user_client = VLLMClient(base_url=vllm_url_70b)
agent_client = VLLMClient(base_url=vllm_url_8b)
if not user_client.health_check():
print(f"ERROR: 70B server at {vllm_url_70b} not responding")
return {"error": "70B server not available"}
if not agent_client.health_check():
print(f"ERROR: 8B server at {vllm_url_8b} not responding")
return {"error": "8B server not available"}
print(f"Running {n_conversations} full conversations (max {max_turns} turns each)...")
conversation_times = []
total_turns = 0
start_time = time.time()
for conv_idx in range(n_conversations):
conv_start = time.time()
# Create user simulator
user_sim = VLLMUserSimulator(
problem="What is 2 + 2? Explain your reasoning step by step.",
user_persona="A student learning math",
user_preferences="- I prefer step-by-step explanations\n- Always show your work",
vllm_client=user_client,
)
# Create agent
agent = VLLMAgentAdapter(
vllm_client=agent_client,
system_prompt="You are a helpful math tutor. Explain concepts clearly."
)
# Run conversation
conversation = [{"role": "assistant", "content": "How can I help you today?"}]
for turn in range(max_turns):
# User turn
user_response = user_sim.generate_user_response(conversation)
if user_response is None:
break
conversation.append({"role": "user", "content": user_response["response"]})
if user_response.get("should_terminate", False):
break
# Agent turn
agent_response = agent.generate_response(user_response["response"])
conversation.append({"role": "assistant", "content": agent_response["response"]})
total_turns += 1
conv_time = time.time() - conv_start
conversation_times.append(conv_time)
print(f" Conversation {conv_idx + 1}/{n_conversations}: {len(conversation)} messages, {conv_time:.1f}s")
total_time = time.time() - start_time
return {
"n_conversations": n_conversations,
"total_turns": total_turns,
"total_time_s": total_time,
"avg_conv_time_s": sum(conversation_times) / len(conversation_times) if conversation_times else 0,
"throughput_conv_per_hr": n_conversations / total_time * 3600,
"throughput_turns_per_hr": total_turns / total_time * 3600,
}
def print_results(results: List[BenchmarkResult]):
"""Print benchmark results in a nice table."""
print("\n" + "=" * 80)
print("BENCHMARK RESULTS")
print("=" * 80)
print(f"\n{'Mode':<20} {'Requests':<10} {'Avg Latency':<12} {'Throughput':<15} {'Conv/hr':<12} {'Errors':<8}")
print("-" * 80)
for r in results:
print(f"{r.mode:<20} {r.n_requests:<10} {r.avg_latency_ms:>8.0f}ms {r.throughput_req_per_s:>10.2f}/s {r.throughput_conv_per_hr:>8.0f} {r.errors:<8}")
print("-" * 80)
# Compare speedup
if len(results) >= 2:
transformers_result = next((r for r in results if r.mode == "transformers"), None)
vllm_result = next((r for r in results if "vllm" in r.mode and r.throughput_req_per_s > 0), None)
if transformers_result and vllm_result and transformers_result.throughput_req_per_s > 0:
speedup = vllm_result.throughput_req_per_s / transformers_result.throughput_req_per_s
print(f"\nvLLM speedup over transformers: {speedup:.1f}x")
# Target comparison
target_conv_per_hr = 2000
for r in results:
if r.throughput_conv_per_hr > 0:
ratio = r.throughput_conv_per_hr / target_conv_per_hr
status = "✓" if ratio >= 0.5 else "✗"
print(f"{status} {r.mode}: {r.throughput_conv_per_hr:.0f} conv/hr ({ratio:.1%} of paper's 2000 conv/hr)")
def main():
parser = argparse.ArgumentParser(description="Benchmark inference speed")
parser.add_argument("--mode", choices=["transformers", "vllm", "both", "conversation"], default="vllm")
parser.add_argument("--model", type=str, default="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct",
help="Model path for transformers benchmark")
parser.add_argument("--url", type=str, default="http://localhost:8003/v1",
help="vLLM server URL")
parser.add_argument("--url-70b", type=str, default="http://localhost:8004/v1",
help="vLLM server URL for 70B model (user simulator)")
parser.add_argument("--url-8b", type=str, default="http://localhost:8003/v1",
help="vLLM server URL for 8B model (agent)")
parser.add_argument("-n", type=int, default=20, help="Number of requests")
parser.add_argument("--concurrent", action="store_true", help="Run vLLM benchmark with concurrent requests")
parser.add_argument("--device", type=str, default="cuda:0", help="Device for transformers")
args = parser.parse_args()
results = []
if args.mode == "conversation":
# Full conversation benchmark
conv_results = benchmark_full_conversation(
args.url_70b,
args.url_8b,
n_conversations=args.n,
)
print("\n" + "=" * 80)
print("FULL CONVERSATION BENCHMARK")
print("=" * 80)
print(json.dumps(conv_results, indent=2))
if "throughput_conv_per_hr" in conv_results:
target = 2000
actual = conv_results["throughput_conv_per_hr"]
print(f"\nTarget: {target} conv/hr (paper)")
print(f"Actual: {actual:.0f} conv/hr ({actual/target:.1%} of target)")
else:
if args.mode in ["transformers", "both"]:
print("\n" + "=" * 40)
print("TRANSFORMERS BENCHMARK")
print("=" * 40)
result = benchmark_transformers(args.model, args.n, args.device)
results.append(result)
if args.mode in ["vllm", "both"]:
print("\n" + "=" * 40)
print("vLLM BENCHMARK (sequential)")
print("=" * 40)
result = benchmark_vllm(args.url, args.n, concurrent=False)
results.append(result)
if args.concurrent:
print("\n" + "=" * 40)
print("vLLM BENCHMARK (concurrent)")
print("=" * 40)
result = benchmark_vllm(args.url, args.n, concurrent=True, n_workers=4)
results.append(result)
print_results(results)
if __name__ == "__main__":
main()
|