Skip to content

vllm_mlx.benchmark

Performance Benchmark for vllm-mlx.

View the complete module source at #L1-L1684.

API details

Each callable below includes its exact signature, type annotations, inputs, defaults, return contract, documented exceptions, implementation source, and parsed docstring sections when the source provides them.

vllm_mlx.benchmark

Performance Benchmark for vllm-mlx.

Measures key performance metrics for LLM and MLLM (Multimodal Language Model) inference: - Time to First Token (TTFT) - Time Per Output Token (TPOT) - Tokens Per Second (TPS) - both input processing and output generation - End-to-End Latency - Throughput - Memory Usage (process and MLX cache) - MLLM: Image resolution performance - MLLM: Video frame count performance

Usage

LLM benchmark

python -m vllm_mlx.benchmark --model mlx-community/Llama-3.2-1B-Instruct-4bit python -m vllm_mlx.benchmark --model mlx-community/Llama-3.2-3B-Instruct-4bit --prompts 10 --max-tokens 256

MLLM image benchmark (auto-detected or use --mllm flag)

python -m vllm_mlx.benchmark --model mlx-community/Qwen3-VL-4B-Instruct-3bit python -m vllm_mlx.benchmark --model mlx-community/Qwen3-VL-4B-Instruct-3bit --mllm --quick

MLLM video benchmark

python -m vllm_mlx.benchmark --model mlx-community/Qwen3-VL-4B-Instruct-3bit --video python -m vllm_mlx.benchmark --model mlx-community/Qwen3-VL-4B-Instruct-3bit --video --video-url https://example.com/video.mp4

vllm_mlx.benchmark.VIDEO_SAMPLE_URLS module-attribute 

VIDEO_SAMPLE_URLS = {'bunny_10s': 'https://test-videos.co.uk/vids/bigbuckbunny/mp4/h264/360/Big_Buck_Bunny_360_10s_1MB.mp4', 'bunny_240p': 'https://docs.evostream.com/sample_content/assets/bunny.mp4', 'sintel_720p': 'https://docs.evostream.com/sample_content/assets/sintel1m720p.mp4'}

vllm_mlx.benchmark.DEFAULT_VIDEO_URL module-attribute 

DEFAULT_VIDEO_URL = VIDEO_SAMPLE_URLS['bunny_10s']

vllm_mlx.benchmark.VLM_TEST_VIDEO_URLS module-attribute 

VLM_TEST_VIDEO_URLS = [VIDEO_SAMPLE_URLS['bunny_10s'], VIDEO_SAMPLE_URLS['bunny_240p']]

vllm_mlx.benchmark.MLLM_PATTERNS module-attribute 

MLLM_PATTERNS = ['-VL-', '-VL/', 'VL-', 'llava', 'LLaVA', 'idefics', 'Idefics', 'paligemma', 'PaliGemma', 'pixtral', 'Pixtral', 'molmo', 'Molmo', 'phi3-vision', 'phi-3-vision', 'cogvlm', 'CogVLM', 'internvl', 'InternVL', 'deepseek-vl', 'DeepSeek-VL']

vllm_mlx.benchmark.MLLM_TEST_IMAGE_URL module-attribute 

MLLM_TEST_IMAGE_URL = 'https://upload.wikimedia.org/wikipedia/commons/thumb/2/26/YellowLabradorLooking_new.jpg/1200px-YellowLabradorLooking_new.jpg'

vllm_mlx.benchmark.MLLM_TEST_IMAGE_URLS module-attribute 

MLLM_TEST_IMAGE_URLS = [MLLM_TEST_IMAGE_URL, 'https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/1200px-Cat03.jpg', 'https://upload.wikimedia.org/wikipedia/commons/thumb/4/47/PNG_transparency_demonstration_1.png/640px-PNG_transparency_demonstration_1.png']

vllm_mlx.benchmark.ResourceMetrics dataclass 

ResourceMetrics(process_memory_gb: float = 0.0, mlx_cache_gb: float = 0.0, mlx_peak_memory_gb: float = 0.0, system_memory_used_gb: float = 0.0, system_memory_total_gb: float = 0.0)

Resource usage metrics during benchmark.

vllm_mlx.benchmark.ResourceMetrics.process_memory_gb class-attribute instance-attribute 

process_memory_gb: float = 0.0

vllm_mlx.benchmark.ResourceMetrics.mlx_cache_gb class-attribute instance-attribute 

mlx_cache_gb: float = 0.0

vllm_mlx.benchmark.ResourceMetrics.mlx_peak_memory_gb class-attribute instance-attribute 

mlx_peak_memory_gb: float = 0.0

vllm_mlx.benchmark.ResourceMetrics.system_memory_used_gb class-attribute instance-attribute 

system_memory_used_gb: float = 0.0

vllm_mlx.benchmark.ResourceMetrics.system_memory_total_gb class-attribute instance-attribute 

system_memory_total_gb: float = 0.0

vllm_mlx.benchmark.ResourceMonitor 

ResourceMonitor()

Monitor system resources during benchmark runs.

Source code in vllm_mlx/benchmark.py 
165
166
167
168
def __init__(self):
    self.samples: list[ResourceMetrics] = []
    self._start_time: float = 0
    self._start_memory: float = 0

vllm_mlx.benchmark.ResourceMonitor.samples instance-attribute 

samples: list[ResourceMetrics] = []

vllm_mlx.benchmark.ResourceMonitor._start_time instance-attribute 

_start_time: float = 0

vllm_mlx.benchmark.ResourceMonitor._start_memory instance-attribute 

_start_memory: float = 0

vllm_mlx.benchmark.ResourceMonitor.start 

start()

Start monitoring.

Source code in vllm_mlx/benchmark.py 
170
171
172
173
174
175
176
def start(self):
    """Start monitoring."""
    self._start_time = time.perf_counter()
    self._start_memory = get_process_memory()

    # Reset MLX peak memory
    reset_mlx_peak_memory()

vllm_mlx.benchmark.ResourceMonitor.sample 

sample() -> ResourceMetrics

Take a resource sample.

Source code in vllm_mlx/benchmark.py 
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
def sample(self) -> ResourceMetrics:
    """Take a resource sample."""
    mlx_info = get_mlx_memory_info()
    sys_used, sys_total = get_system_memory()

    metrics = ResourceMetrics(
        process_memory_gb=get_process_memory(),
        mlx_cache_gb=mlx_info.get("cache_memory_gb", 0.0),
        mlx_peak_memory_gb=mlx_info.get("peak_memory_gb", 0.0),
        system_memory_used_gb=sys_used,
        system_memory_total_gb=sys_total,
    )

    self.samples.append(metrics)
    return metrics

vllm_mlx.benchmark.ResourceMonitor.get_summary 

get_summary() -> ResourceMetrics

Get summary of all samples.

Source code in vllm_mlx/benchmark.py 
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
def get_summary(self) -> ResourceMetrics:
    """Get summary of all samples."""
    if not self.samples:
        return ResourceMetrics()

    # Get peak values
    peak_process = max(s.process_memory_gb for s in self.samples)
    peak_mlx = max(s.mlx_peak_memory_gb for s in self.samples)
    peak_mlx_cache = max(s.mlx_cache_gb for s in self.samples)

    # Get latest system memory
    latest = self.samples[-1]

    return ResourceMetrics(
        process_memory_gb=peak_process,
        mlx_cache_gb=peak_mlx_cache,
        mlx_peak_memory_gb=peak_mlx,
        system_memory_used_gb=latest.system_memory_used_gb,
        system_memory_total_gb=latest.system_memory_total_gb,
    )

vllm_mlx.benchmark.BenchmarkResult dataclass 

BenchmarkResult(prompt: str, prompt_tokens: int, generated_tokens: int, ttft: float, total_time: float, tpot: float = 0.0, generation_tps: float = 0.0, processing_tps: float = 0.0)

Results from a single benchmark run.

vllm_mlx.benchmark.BenchmarkResult.prompt instance-attribute 

prompt: str

vllm_mlx.benchmark.BenchmarkResult.prompt_tokens instance-attribute 

prompt_tokens: int

vllm_mlx.benchmark.BenchmarkResult.generated_tokens instance-attribute 

generated_tokens: int

vllm_mlx.benchmark.BenchmarkResult.ttft instance-attribute 

ttft: float

vllm_mlx.benchmark.BenchmarkResult.total_time instance-attribute 

total_time: float

vllm_mlx.benchmark.BenchmarkResult.tpot class-attribute instance-attribute 

tpot: float = 0.0

vllm_mlx.benchmark.BenchmarkResult.generation_tps class-attribute instance-attribute 

generation_tps: float = 0.0

vllm_mlx.benchmark.BenchmarkResult.processing_tps class-attribute instance-attribute 

processing_tps: float = 0.0

vllm_mlx.benchmark.BenchmarkResult.__post_init__ 

__post_init__()
Source code in vllm_mlx/benchmark.py 
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
def __post_init__(self):
    if self.generated_tokens > 1:
        # TPOT excludes the first token (which is measured by TTFT)
        generation_time = self.total_time - self.ttft
        self.tpot = (
            generation_time / (self.generated_tokens - 1)
            if self.generated_tokens > 1
            else 0
        )
        self.generation_tps = (
            (self.generated_tokens - 1) / generation_time
            if generation_time > 0
            else 0
        )

    # Processing TPS: how fast the prompt was processed (tokens / TTFT)
    if self.ttft > 0:
        self.processing_tps = self.prompt_tokens / self.ttft

vllm_mlx.benchmark.BenchmarkSummary dataclass 

BenchmarkSummary(model_name: str, num_runs: int, total_prompt_tokens: int, total_generated_tokens: int, total_time: float, ttft_mean: float, ttft_min: float, ttft_max: float, ttft_p50: float, ttft_p95: float, tpot_mean: float, tpot_min: float, tpot_max: float, generation_tps_mean: float, generation_tps_max: float, processing_tps_mean: float, latency_mean: float, latency_min: float, latency_max: float, latency_p50: float, latency_p95: float, total_throughput_tps: float, requests_per_second: float, hardware_chip: str = '', hardware_memory_gb: float = 0.0, hardware_bandwidth_gbs: float = 0.0, resources: ResourceMetrics = ResourceMetrics())

Summary statistics across all benchmark runs.

vllm_mlx.benchmark.BenchmarkSummary.model_name instance-attribute 

model_name: str

vllm_mlx.benchmark.BenchmarkSummary.num_runs instance-attribute 

num_runs: int

vllm_mlx.benchmark.BenchmarkSummary.total_prompt_tokens instance-attribute 

total_prompt_tokens: int

vllm_mlx.benchmark.BenchmarkSummary.total_generated_tokens instance-attribute 

total_generated_tokens: int

vllm_mlx.benchmark.BenchmarkSummary.total_time instance-attribute 

total_time: float

vllm_mlx.benchmark.BenchmarkSummary.ttft_mean instance-attribute 

ttft_mean: float

vllm_mlx.benchmark.BenchmarkSummary.ttft_min instance-attribute 

ttft_min: float

vllm_mlx.benchmark.BenchmarkSummary.ttft_max instance-attribute 

ttft_max: float

vllm_mlx.benchmark.BenchmarkSummary.ttft_p50 instance-attribute 

ttft_p50: float

vllm_mlx.benchmark.BenchmarkSummary.ttft_p95 instance-attribute 

ttft_p95: float

vllm_mlx.benchmark.BenchmarkSummary.tpot_mean instance-attribute 

tpot_mean: float

vllm_mlx.benchmark.BenchmarkSummary.tpot_min instance-attribute 

tpot_min: float

vllm_mlx.benchmark.BenchmarkSummary.tpot_max instance-attribute 

tpot_max: float

vllm_mlx.benchmark.BenchmarkSummary.generation_tps_mean instance-attribute 

generation_tps_mean: float

vllm_mlx.benchmark.BenchmarkSummary.generation_tps_max instance-attribute 

generation_tps_max: float

vllm_mlx.benchmark.BenchmarkSummary.processing_tps_mean instance-attribute 

processing_tps_mean: float

vllm_mlx.benchmark.BenchmarkSummary.latency_mean instance-attribute 

latency_mean: float

vllm_mlx.benchmark.BenchmarkSummary.latency_min instance-attribute 

latency_min: float

vllm_mlx.benchmark.BenchmarkSummary.latency_max instance-attribute 

latency_max: float

vllm_mlx.benchmark.BenchmarkSummary.latency_p50 instance-attribute 

latency_p50: float

vllm_mlx.benchmark.BenchmarkSummary.latency_p95 instance-attribute 

latency_p95: float

vllm_mlx.benchmark.BenchmarkSummary.total_throughput_tps instance-attribute 

total_throughput_tps: float

vllm_mlx.benchmark.BenchmarkSummary.requests_per_second instance-attribute 

requests_per_second: float

vllm_mlx.benchmark.BenchmarkSummary.hardware_chip class-attribute instance-attribute 

hardware_chip: str = ''

vllm_mlx.benchmark.BenchmarkSummary.hardware_memory_gb class-attribute instance-attribute 

hardware_memory_gb: float = 0.0

vllm_mlx.benchmark.BenchmarkSummary.hardware_bandwidth_gbs class-attribute instance-attribute 

hardware_bandwidth_gbs: float = 0.0

vllm_mlx.benchmark.BenchmarkSummary.resources class-attribute instance-attribute 

resources: ResourceMetrics = field(default_factory=ResourceMetrics)

vllm_mlx.benchmark.MLLMBenchmarkResult dataclass 

MLLMBenchmarkResult(resolution: str, width: int, height: int, pixels: int, time_seconds: float, tokens_generated: int, tokens_per_second: float, response_preview: str, memory_gb: float = 0.0, mlx_memory_gb: float = 0.0)

Result from a single MLLM benchmark run.

vllm_mlx.benchmark.MLLMBenchmarkResult.resolution instance-attribute 

resolution: str

vllm_mlx.benchmark.MLLMBenchmarkResult.width instance-attribute 

width: int

vllm_mlx.benchmark.MLLMBenchmarkResult.height instance-attribute 

height: int

vllm_mlx.benchmark.MLLMBenchmarkResult.pixels instance-attribute 

pixels: int

vllm_mlx.benchmark.MLLMBenchmarkResult.time_seconds instance-attribute 

time_seconds: float

vllm_mlx.benchmark.MLLMBenchmarkResult.tokens_generated instance-attribute 

tokens_generated: int

vllm_mlx.benchmark.MLLMBenchmarkResult.tokens_per_second instance-attribute 

tokens_per_second: float

vllm_mlx.benchmark.MLLMBenchmarkResult.response_preview instance-attribute 

response_preview: str

vllm_mlx.benchmark.MLLMBenchmarkResult.memory_gb class-attribute instance-attribute 

memory_gb: float = 0.0

vllm_mlx.benchmark.MLLMBenchmarkResult.mlx_memory_gb class-attribute instance-attribute 

mlx_memory_gb: float = 0.0

vllm_mlx.benchmark.VideoBenchmarkResult dataclass 

VideoBenchmarkResult(config_name: str, fps: float, max_frames: int, frames_extracted: int, video_duration: float, time_seconds: float, prompt_tokens: int, completion_tokens: int, tokens_per_second: float, response_preview: str, memory_gb: float = 0.0, mlx_memory_gb: float = 0.0)

Result from a single video benchmark run.

vllm_mlx.benchmark.VideoBenchmarkResult.config_name instance-attribute 

config_name: str

vllm_mlx.benchmark.VideoBenchmarkResult.fps instance-attribute 

fps: float

vllm_mlx.benchmark.VideoBenchmarkResult.max_frames instance-attribute 

max_frames: int

vllm_mlx.benchmark.VideoBenchmarkResult.frames_extracted instance-attribute 

frames_extracted: int

vllm_mlx.benchmark.VideoBenchmarkResult.video_duration instance-attribute 

video_duration: float

vllm_mlx.benchmark.VideoBenchmarkResult.time_seconds instance-attribute 

time_seconds: float

vllm_mlx.benchmark.VideoBenchmarkResult.prompt_tokens instance-attribute 

prompt_tokens: int

vllm_mlx.benchmark.VideoBenchmarkResult.completion_tokens instance-attribute 

completion_tokens: int

vllm_mlx.benchmark.VideoBenchmarkResult.tokens_per_second instance-attribute 

tokens_per_second: float

vllm_mlx.benchmark.VideoBenchmarkResult.response_preview instance-attribute 

response_preview: str

vllm_mlx.benchmark.VideoBenchmarkResult.memory_gb class-attribute instance-attribute 

memory_gb: float = 0.0

vllm_mlx.benchmark.VideoBenchmarkResult.mlx_memory_gb class-attribute instance-attribute 

mlx_memory_gb: float = 0.0

vllm_mlx.benchmark.reset_mlx_peak_memory 

reset_mlx_peak_memory()

Reset MLX peak memory counter.

Source code in vllm_mlx/benchmark.py 
83
84
85
86
87
88
89
90
91
92
93
94
95
def reset_mlx_peak_memory():
    """Reset MLX peak memory counter."""
    if not HAS_MLX:
        return

    try:
        # Use new API (mx.*) if available, fallback to deprecated (mx.metal.*)
        if hasattr(mx, "reset_peak_memory"):
            mx.reset_peak_memory()
        else:
            mx.metal.reset_peak_memory()
    except Exception:
        pass

vllm_mlx.benchmark.get_mlx_memory_info 

get_mlx_memory_info(reset_peak: bool = True) -> dict

Get MLX memory usage information.

Parameters:

  • reset_peak (bool, default: True ) –

    If True, reset peak memory counter after reading.

Source code in vllm_mlx/benchmark.py 
 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
def get_mlx_memory_info(reset_peak: bool = True) -> dict:
    """Get MLX memory usage information.

    Args:
        reset_peak: If True, reset peak memory counter after reading.
    """
    if not HAS_MLX:
        return {}

    try:
        # Use new API (mx.*) if available, fallback to deprecated (mx.metal.*)
        if hasattr(mx, "get_cache_memory"):
            cache_memory = mx.get_cache_memory()
            peak_memory = mx.get_peak_memory()
            active_memory = (
                mx.get_active_memory() if hasattr(mx, "get_active_memory") else 0
            )
        else:
            # Fallback for older MLX versions
            cache_memory = mx.metal.get_cache_memory()
            peak_memory = mx.metal.get_peak_memory()
            active_memory = (
                mx.metal.get_active_memory()
                if hasattr(mx.metal, "get_active_memory")
                else 0
            )

        info = {
            "cache_memory_gb": cache_memory / (1024**3),
            "peak_memory_gb": peak_memory / (1024**3),
            "active_memory_gb": active_memory / (1024**3),
        }
        # Reset peak for next measurement if requested
        if reset_peak:
            reset_mlx_peak_memory()
        return info
    except Exception:
        return {}

vllm_mlx.benchmark.get_process_memory 

get_process_memory() -> float

Get current process memory usage in GB.

Source code in vllm_mlx/benchmark.py 
138
139
140
141
142
143
144
145
146
147
def get_process_memory() -> float:
    """Get current process memory usage in GB."""
    if not HAS_PSUTIL:
        return 0.0

    try:
        process = psutil.Process()
        return process.memory_info().rss / (1024**3)
    except Exception:
        return 0.0

vllm_mlx.benchmark.get_system_memory 

get_system_memory() -> tuple[float, float]

Get system memory (used, total) in GB.

Source code in vllm_mlx/benchmark.py 
150
151
152
153
154
155
156
157
158
159
def get_system_memory() -> tuple[float, float]:
    """Get system memory (used, total) in GB."""
    if not HAS_PSUTIL:
        return 0.0, 0.0

    try:
        mem = psutil.virtual_memory()
        return mem.used / (1024**3), mem.total / (1024**3)
    except Exception:
        return 0.0, 0.0

vllm_mlx.benchmark.calculate_percentile 

calculate_percentile(data: list, percentile: float) -> float

Calculate percentile from a list.

Source code in vllm_mlx/benchmark.py 
318
319
320
321
322
323
324
325
def calculate_percentile(data: list, percentile: float) -> float:
    """Calculate percentile from a list."""
    if not data:
        return 0.0
    sorted_data = sorted(data)
    index = int(len(sorted_data) * percentile / 100)
    index = min(index, len(sorted_data) - 1)
    return sorted_data[index]

vllm_mlx.benchmark.benchmark_single_prompt 

benchmark_single_prompt(model, tokenizer, prompt: str, max_tokens: int = 256, temperature: float = 0.7) -> Optional[BenchmarkResult]

Benchmark a single prompt with detailed timing.

Parameters:

  • model –

    The loaded MLX model

  • tokenizer –

    The tokenizer

  • prompt (str) –

    The prompt to benchmark

  • max_tokens (int, default: 256 ) –

    Maximum tokens to generate

  • temperature (float, default: 0.7 ) –

    Sampling temperature

Returns:

Source code in vllm_mlx/benchmark.py 
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
def benchmark_single_prompt(
    model,
    tokenizer,
    prompt: str,
    max_tokens: int = 256,
    temperature: float = 0.7,
) -> Optional[BenchmarkResult]:
    """
    Benchmark a single prompt with detailed timing.

    Args:
        model: The loaded MLX model
        tokenizer: The tokenizer
        prompt: The prompt to benchmark
        max_tokens: Maximum tokens to generate
        temperature: Sampling temperature

    Returns:
        BenchmarkResult with timing metrics
    """
    from mlx_lm import stream_generate
    from mlx_lm.sample_utils import make_sampler

    try:
        # Tokenize the prompt to count tokens
        prompt_tokens = tokenizer.encode(prompt)
        prompt_token_count = len(prompt_tokens)

        # Create sampler
        sampler = make_sampler(temp=temperature)

        # Start timing
        start_time = time.perf_counter()
        ttft = None
        token_count = 0

        # Generate tokens using stream_generate
        for response in stream_generate(
            model,
            tokenizer,
            prompt,
            max_tokens=max_tokens,
            sampler=sampler,
        ):
            if ttft is None:
                ttft = time.perf_counter() - start_time
            token_count += 1

        total_time = time.perf_counter() - start_time

        if ttft is None:
            ttft = total_time

        return BenchmarkResult(
            prompt=prompt[:50] + "..." if len(prompt) > 50 else prompt,
            prompt_tokens=prompt_token_count,
            generated_tokens=token_count,
            ttft=ttft,
            total_time=total_time,
        )

    except Exception as e:
        print(f"Error during benchmark: {e}")
        import traceback

        traceback.print_exc()
        return None

vllm_mlx.benchmark.run_benchmark 

run_benchmark(model_name: str, num_prompts: int = 5, max_tokens: int = 256, temperature: float = 0.7, warmup_runs: int = 1) -> Optional[BenchmarkSummary]

Run the full benchmark suite.

Parameters:

  • model_name (str) –

    HuggingFace model name or local path

  • num_prompts (int, default: 5 ) –

    Number of prompts to test

  • max_tokens (int, default: 256 ) –

    Maximum tokens per generation

  • temperature (float, default: 0.7 ) –

    Sampling temperature

  • warmup_runs (int, default: 1 ) –

    Number of warmup runs before measuring

Returns:

Source code in vllm_mlx/benchmark.py 
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
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
def run_benchmark(
    model_name: str,
    num_prompts: int = 5,
    max_tokens: int = 256,
    temperature: float = 0.7,
    warmup_runs: int = 1,
) -> Optional[BenchmarkSummary]:
    """
    Run the full benchmark suite.

    Args:
        model_name: HuggingFace model name or local path
        num_prompts: Number of prompts to test
        max_tokens: Maximum tokens per generation
        temperature: Sampling temperature
        warmup_runs: Number of warmup runs before measuring

    Returns:
        BenchmarkSummary with aggregate statistics
    """
    from vllm_mlx.utils.tokenizer import load_model_with_fallback
    from vllm_mlx.optimizations import detect_hardware

    # Detect hardware
    hw = detect_hardware()

    # Test prompts of varying lengths: 3 short, 3 medium, 4 long
    prompts = [
        # Short prompts (~5-15 tokens) - 3 prompts
        "Hello, how are you?",
        "What is 2+2?",
        "Say hello in Spanish.",
        # Medium prompts (~30-60 tokens) - 3 prompts
        "What is the capital of France and why is it historically significant? Include some interesting facts about the city.",
        "Write a Python function to calculate fibonacci numbers using memoization. Explain how it works.",
        "Explain the difference between a list and a tuple in Python. When should you use each one?",
        # Long prompts (100+ tokens) - 4 prompts
        """Explain quantum computing in comprehensive detail. You should cover all of the following topics thoroughly:
        1. What are qubits and how do they fundamentally differ from classical bits in traditional computing?
        2. What is quantum superposition and how does it enable parallel computation?
        3. What is quantum entanglement and why is it crucial for quantum algorithms?
        4. What are the most promising potential applications of quantum computing in cryptography, drug discovery, and optimization?
        5. What are the current hardware limitations, error correction challenges, and decoherence problems?
        6. Compare the approaches of IBM, Google, and other major players in quantum computing research.""",
        """Write a comprehensive and detailed guide to building a production-ready REST API with Python Flask. Include all of the following sections with code examples:
        1. Setting up the project structure with blueprints, configuration management, and environment variables
        2. Creating routes and endpoints following RESTful conventions with proper HTTP methods
        3. Handling JSON requests and responses with validation using marshmallow or pydantic
        4. Adding authentication and authorization with JWT tokens and role-based access control
        5. Implementing error handling best practices with custom exception handlers
        6. Writing comprehensive tests with pytest including unit tests and integration tests
        7. Setting up logging, monitoring, and API documentation with Swagger/OpenAPI""",
        """Describe the complete process of photosynthesis in plants with scientific detail. Your explanation should cover:
        1. The light-dependent reactions that occur in the thylakoid membrane, including photosystems I and II
        2. The electron transport chain and chemiosmosis for ATP synthesis
        3. The Calvin cycle (light-independent reactions) and the process of carbon fixation by RuBisCO
        4. The role of chlorophyll a, chlorophyll b, and accessory pigments like carotenoids
        5. How environmental factors like light intensity, CO2 concentration, and temperature affect photosynthesis rate
        6. The importance of photosynthesis for life on Earth and its role in the carbon cycle
        7. C3, C4, and CAM photosynthesis adaptations in different plant species""",
        """You are a senior software architect with 15 years of experience. Design a complete microservices architecture for a large-scale e-commerce platform that handles millions of users. Your design should include:
        1. Service breakdown with detailed responsibilities: User service (authentication, profiles, preferences), Product catalog service (search, filtering, recommendations), Inventory service (stock management, warehouses), Order service (cart, checkout, order history), Payment service (multiple providers, refunds), Notification service (email, SMS, push)
        2. Database choices for each service with justification (PostgreSQL vs MongoDB vs Redis)
        3. Inter-service communication patterns: synchronous REST/gRPC vs asynchronous message queues
        4. API gateway design with rate limiting, authentication, and request routing
        5. Caching strategy with Redis for sessions, product data, and search results
        6. Message queue architecture with RabbitMQ or Kafka for event-driven communication
        7. Kubernetes deployment with horizontal pod autoscaling, health checks, and rolling updates
        8. CI/CD pipeline with GitHub Actions, testing stages, and blue-green deployments""",
    ]

    # Use only the requested number of prompts
    test_prompts = (prompts * ((num_prompts // len(prompts)) + 1))[:num_prompts]

    print(f"\n{'='*60}")
    print("vllm-mlx Performance Benchmark")
    print(f"{'='*60}")

    # Hardware info table
    hw_table = [
        ["Model", model_name],
        ["Hardware", f"{hw.chip_name} ({hw.total_memory_gb:.0f} GB)"],
        ["Memory Bandwidth", f"{hw.memory_bandwidth_gbs} GB/s"],
        ["GPU Cores", hw.gpu_cores],
        ["Prompts", num_prompts],
        ["Max Tokens", max_tokens],
        ["Temperature", temperature],
    ]
    print(tabulate(hw_table, tablefmt="plain"))
    print(f"{'='*60}\n")

    # Initialize resource monitor
    monitor = ResourceMonitor()
    monitor.start()

    # Load model
    print(f"Loading model: {model_name}...")
    load_start = time.perf_counter()
    model, tokenizer = load_model_with_fallback(model_name)
    load_time = time.perf_counter() - load_start
    print(f"Model loaded in {load_time:.2f}s\n")

    # Show prompt length distribution
    prompt_lengths = [len(tokenizer.encode(p)) for p in test_prompts]
    short = sum(1 for length in prompt_lengths if length < 20)
    medium = sum(1 for length in prompt_lengths if 20 <= length < 100)
    long_p = sum(1 for length in prompt_lengths if length >= 100)
    print("Prompt Distribution:")
    dist_data = [
        ["Short (<20 tokens)", short],
        ["Medium (20-100)", medium],
        ["Long (100+)", long_p],
        ["Total input tokens", sum(prompt_lengths)],
    ]
    print(tabulate(dist_data, tablefmt="plain"))
    print()

    # Warmup runs
    if warmup_runs > 0:
        print(f"Running {warmup_runs} warmup run(s)...")
        for i in range(warmup_runs):
            benchmark_single_prompt(
                model, tokenizer, "Hello, how are you?", max_tokens=20
            )

        # Show model memory after warmup (MLX uses lazy evaluation)
        mlx_info = get_mlx_memory_info(reset_peak=True)
        if mlx_info and mlx_info.get("peak_memory_gb", 0) > 0:
            print(f"Model memory: {mlx_info['peak_memory_gb']:.2f} GB (MLX peak)")
        print("Warmup complete.\n")

    # Main benchmark runs
    results: list[BenchmarkResult] = []
    overall_start = time.perf_counter()

    run_data = []
    for i, prompt in enumerate(test_prompts, 1):
        result = benchmark_single_prompt(
            model, tokenizer, prompt, max_tokens, temperature
        )

        if result:
            results.append(result)
            run_data.append(
                [
                    i,
                    result.prompt_tokens,
                    result.generated_tokens,
                    f"{result.ttft*1000:.1f}",
                    f"{result.generation_tps:.1f}",
                ]
            )
            # Sample resources after each run
            monitor.sample()

    overall_time = time.perf_counter() - overall_start

    # Print per-run results table
    print("Per-Run Results:")
    print(
        tabulate(
            run_data,
            headers=["Run", "Input", "Output", "TTFT (ms)", "Gen TPS"],
            tablefmt="simple",
        )
    )
    print()

    if not results:
        print("No successful benchmark runs!")
        return None

    # Calculate summary statistics
    ttfts = [r.ttft for r in results]
    tpots = [r.tpot for r in results if r.tpot > 0]
    latencies = [r.total_time for r in results]
    gen_tps = [r.generation_tps for r in results if r.generation_tps > 0]
    proc_tps = [r.processing_tps for r in results if r.processing_tps > 0]

    total_prompt_tokens = sum(r.prompt_tokens for r in results)
    total_generated_tokens = sum(r.generated_tokens for r in results)
    total_tokens = total_prompt_tokens + total_generated_tokens

    summary = BenchmarkSummary(
        model_name=model_name,
        num_runs=len(results),
        total_prompt_tokens=total_prompt_tokens,
        total_generated_tokens=total_generated_tokens,
        total_time=overall_time,
        ttft_mean=statistics.mean(ttfts),
        ttft_min=min(ttfts),
        ttft_max=max(ttfts),
        ttft_p50=calculate_percentile(ttfts, 50),
        ttft_p95=calculate_percentile(ttfts, 95),
        tpot_mean=statistics.mean(tpots) if tpots else 0,
        tpot_min=min(tpots) if tpots else 0,
        tpot_max=max(tpots) if tpots else 0,
        generation_tps_mean=statistics.mean(gen_tps) if gen_tps else 0,
        generation_tps_max=max(gen_tps) if gen_tps else 0,
        processing_tps_mean=statistics.mean(proc_tps) if proc_tps else 0,
        latency_mean=statistics.mean(latencies),
        latency_min=min(latencies),
        latency_max=max(latencies),
        latency_p50=calculate_percentile(latencies, 50),
        latency_p95=calculate_percentile(latencies, 95),
        total_throughput_tps=total_tokens / overall_time,
        requests_per_second=len(results) / overall_time,
        hardware_chip=hw.chip_name,
        hardware_memory_gb=hw.total_memory_gb,
        hardware_bandwidth_gbs=hw.memory_bandwidth_gbs,
        resources=monitor.get_summary(),
    )

    return summary

vllm_mlx.benchmark.is_mllm_model 

is_mllm_model(model_name: str) -> bool

Check if model name indicates a multimodal language model.

Source code in vllm_mlx/benchmark.py 
651
652
653
654
655
656
657
def is_mllm_model(model_name: str) -> bool:
    """Check if model name indicates a multimodal language model."""
    model_lower = model_name.lower()
    for pattern in MLLM_PATTERNS:
        if pattern.lower() in model_lower:
            return True
    return False

vllm_mlx.benchmark.download_test_image 

download_test_image(url: str, timeout: int = 30) -> Image

Download image from URL and return PIL Image.

Source code in vllm_mlx/benchmark.py 
677
678
679
680
681
682
683
684
def download_test_image(url: str, timeout: int = 30) -> Image.Image:
    """Download image from URL and return PIL Image."""
    headers = {
        "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36"
    }
    response = requests.get(url, timeout=timeout, headers=headers)
    response.raise_for_status()
    return Image.open(io.BytesIO(response.content))

vllm_mlx.benchmark.resize_image 

resize_image(img: Image, width: int, height: int) -> Image

Resize image to specified dimensions.

Source code in vllm_mlx/benchmark.py 
687
688
689
def resize_image(img: Image.Image, width: int, height: int) -> Image.Image:
    """Resize image to specified dimensions."""
    return img.resize((width, height), Image.Resampling.LANCZOS)

vllm_mlx.benchmark.image_to_base64 

image_to_base64(img: Image, format: str = 'JPEG') -> str

Convert PIL Image to base64 data URL.

Source code in vllm_mlx/benchmark.py 
692
693
694
695
696
697
698
699
700
701
702
703
704
705
def image_to_base64(img: Image.Image, format: str = "JPEG") -> str:
    """Convert PIL Image to base64 data URL."""
    if img.mode == "RGBA":
        background = Image.new("RGB", img.size, (255, 255, 255))
        background.paste(img, mask=img.split()[3])
        img = background
    elif img.mode != "RGB":
        img = img.convert("RGB")

    buffer = io.BytesIO()
    img.save(buffer, format=format, quality=85)
    b64 = base64.b64encode(buffer.getvalue()).decode()
    mime = "image/jpeg" if format == "JPEG" else "image/png"
    return f"data:{mime};base64,{b64}"

vllm_mlx.benchmark.benchmark_mllm_resolution 

benchmark_mllm_resolution(model, processor, config, base_image: Image, width: int, height: int, max_tokens: int = 256, warmup: bool = False) -> MLLMBenchmarkResult

Run MLLM benchmark for a specific resolution.

Source code in vllm_mlx/benchmark.py 
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
def benchmark_mllm_resolution(
    model,
    processor,
    config,
    base_image: Image.Image,
    width: int,
    height: int,
    max_tokens: int = 256,
    warmup: bool = False,
) -> MLLMBenchmarkResult:
    """Run MLLM benchmark for a specific resolution."""
    from mlx_vlm import generate
    from mlx_vlm.prompt_utils import apply_chat_template

    # Reset MLX peak memory before this run
    reset_mlx_peak_memory()

    # Resize image
    img = resize_image(base_image, width, height)

    # Save to temp file for mlx_vlm
    import tempfile

    temp_file = tempfile.NamedTemporaryFile(suffix=".jpg", delete=False)
    img.save(temp_file.name, "JPEG", quality=85)
    image_path = temp_file.name

    resolution_name = f"{width}x{height}"
    pixels = width * height

    if not warmup:
        print(f"  {resolution_name:>10} | {pixels:>12,} |", end=" ", flush=True)

    # Apply chat template
    prompt = "What animal is in this image? Describe it briefly."
    try:
        formatted_prompt = apply_chat_template(
            processor,
            config,
            prompt,
            num_images=1,
        )
    except Exception:
        formatted_prompt = prompt

    # Generate
    start_time = time.perf_counter()
    result = generate(
        model,
        processor,
        formatted_prompt,
        [image_path],
        max_tokens=max_tokens,
        temp=0.7,
        verbose=False,
    )
    elapsed = time.perf_counter() - start_time

    # Extract text
    if hasattr(result, "text"):
        text = result.text
        tokens = getattr(result, "generation_tokens", len(text.split()))
    else:
        text = str(result)
        tokens = len(text.split())

    tps = tokens / elapsed if elapsed > 0 else 0

    # Get memory metrics
    mlx_info = get_mlx_memory_info()
    process_mem = get_process_memory()

    if not warmup:
        mem_str = f"{mlx_info.get('peak_memory_gb', 0):.1f} GB" if mlx_info else "-"
        print(f"{elapsed:>6.2f}s | {tokens:>6} | {tps:>10.1f} tok/s | {mem_str}")

    # Cleanup
    import os

    os.unlink(image_path)

    return MLLMBenchmarkResult(
        resolution=resolution_name,
        width=width,
        height=height,
        pixels=pixels,
        time_seconds=elapsed,
        tokens_generated=tokens,
        tokens_per_second=tps,
        response_preview=text[:150] + "..." if len(text) > 150 else text,
        memory_gb=process_mem,
        mlx_memory_gb=mlx_info.get("peak_memory_gb", 0.0),
    )

vllm_mlx.benchmark.run_mllm_benchmark 

run_mllm_benchmark(model_name: str, quick: bool = False, max_tokens: int = 256, warmup_runs: int = 1) -> list[MLLMBenchmarkResult]

Run MLLM benchmark across multiple image resolutions.

Parameters:

  • model_name (str) –

    HuggingFace model name

  • quick (bool, default: False ) –

    If True, test only 4 resolutions

  • max_tokens (int, default: 256 ) –

    Max tokens to generate

  • warmup_runs (int, default: 1 ) –

    Number of warmup runs

Returns:

Source code in vllm_mlx/benchmark.py 
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
def run_mllm_benchmark(
    model_name: str,
    quick: bool = False,
    max_tokens: int = 256,
    warmup_runs: int = 1,
) -> list[MLLMBenchmarkResult]:
    """
    Run MLLM benchmark across multiple image resolutions.

    Args:
        model_name: HuggingFace model name
        quick: If True, test only 4 resolutions
        max_tokens: Max tokens to generate
        warmup_runs: Number of warmup runs

    Returns:
        List of MLLMBenchmarkResult
    """
    from mlx_vlm import load
    from mlx_vlm.utils import load_config
    from vllm_mlx.optimizations import detect_hardware

    # Detect hardware
    hw = detect_hardware()

    # Define resolutions
    if quick:
        resolutions = [
            (224, 224),
            (448, 448),
            (768, 768),
            (1024, 1024),
        ]
    else:
        resolutions = [
            (224, 224),
            (336, 336),
            (448, 448),
            (512, 512),
            (672, 672),
            (768, 768),
            (896, 896),
            (1024, 1024),
            (1280, 720),
            (1920, 1080),
        ]

    print(f"\n{'='*70}")
    print("vllm-mlx MLLM Performance Benchmark")
    print(f"{'='*70}")

    # Info table
    info_table = [
        ["Model", model_name],
        ["Hardware", f"{hw.chip_name} ({hw.total_memory_gb:.0f} GB)"],
        ["Test Image", "Yellow Labrador (Wikimedia Commons)"],
        ["Resolutions", len(resolutions)],
        ["Max Tokens", max_tokens],
    ]
    print(tabulate(info_table, tablefmt="plain"))
    print(f"{'='*70}\n")

    # Load model
    print(f"Loading MLLM model: {model_name}...")
    load_start = time.perf_counter()
    model, processor = load(model_name)
    config = load_config(model_name)
    load_time = time.perf_counter() - load_start
    print(f"Model loaded in {load_time:.2f}s\n")

    # Download test image
    print("Downloading test image...")
    try:
        base_image = download_test_image(MLLM_TEST_IMAGE_URL)
        print(f"  Original size: {base_image.size[0]}x{base_image.size[1]}\n")
    except Exception as e:
        print(f"Error downloading image: {e}")
        return []

    # Warmup
    if warmup_runs > 0:
        print(f"Running {warmup_runs} warmup run(s)...")
        for _ in range(warmup_runs):
            benchmark_mllm_resolution(
                model, processor, config, base_image, 224, 224, max_tokens, warmup=True
            )

        # Show model memory after warmup (MLX uses lazy evaluation)
        mlx_info = get_mlx_memory_info(reset_peak=True)
        if mlx_info and mlx_info.get("peak_memory_gb", 0) > 0:
            print(f"Model memory: {mlx_info['peak_memory_gb']:.2f} GB (MLX peak)")
        print("Warmup complete.\n")

    # Run benchmarks
    print("-" * 80)
    print(
        f"  {'Resolution':>10} | {'Pixels':>12} | {'Time':>6} | {'Tokens':>6} | {'Speed':>14} | {'Memory':>8}"
    )
    print("-" * 80)

    results = []
    for width, height in resolutions:
        try:
            result = benchmark_mllm_resolution(
                model, processor, config, base_image, width, height, max_tokens
            )
            results.append(result)
        except Exception as e:
            print(f"  Error at {width}x{height}: {e}")

    return results

vllm_mlx.benchmark.print_mllm_summary 

print_mllm_summary(results: list[MLLMBenchmarkResult], model_name: str)

Print MLLM benchmark summary.

Source code in vllm_mlx/benchmark.py 
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
def print_mllm_summary(results: list[MLLMBenchmarkResult], model_name: str):
    """Print MLLM benchmark summary."""
    if not results:
        print("No results to display.")
        return

    print(f"\n{'='*80}")
    print("MLLM BENCHMARK RESULTS")
    print(f"{'='*80}\n")

    # Results table
    table_data = []
    for r in results:
        table_data.append(
            [
                r.resolution,
                f"{r.pixels:,}",
                f"{r.time_seconds:.2f}s",
                r.tokens_generated,
                f"{r.tokens_per_second:.1f}",
                (
                    f"{r.pixels / r.time_seconds / 1000:.1f}K"
                    if r.time_seconds > 0
                    else "N/A"
                ),
                f"{r.mlx_memory_gb:.2f}" if r.mlx_memory_gb > 0 else "-",
            ]
        )

    headers = [
        "Resolution",
        "Pixels",
        "Time",
        "Tokens",
        "Tok/s",
        "Pixels/s",
        "Mem (GB)",
    ]
    print(tabulate(table_data, headers=headers, tablefmt="simple"))

    # Summary stats
    total_time = sum(r.time_seconds for r in results)
    total_tokens = sum(r.tokens_generated for r in results)
    avg_tps = total_tokens / total_time if total_time > 0 else 0
    peak_memory = max(r.mlx_memory_gb for r in results) if results else 0

    print("-" * 80)
    print(f"Total Time:      {total_time:.2f}s")
    print(f"Total Tokens:    {total_tokens}")
    print(f"Average Tok/s:   {avg_tps:.1f}")
    if peak_memory > 0:
        print(f"Peak Memory:     {peak_memory:.2f} GB")

    fastest = min(results, key=lambda r: r.time_seconds)
    slowest = max(results, key=lambda r: r.time_seconds)

    print(f"\nFastest:  {fastest.resolution} ({fastest.time_seconds:.2f}s)")
    print(f"Slowest:  {slowest.resolution} ({slowest.time_seconds:.2f}s)")
    print(f"Slowdown: {slowest.time_seconds / fastest.time_seconds:.1f}x")
    print(f"{'='*80}")

vllm_mlx.benchmark.create_test_video 

create_test_video(duration: float = 10.0, fps: float = 30.0, width: int = 640, height: int = 480) -> str

Create a synthetic test video with colored frames and text.

Source code in vllm_mlx/benchmark.py 
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
def create_test_video(
    duration: float = 10.0,
    fps: float = 30.0,
    width: int = 640,
    height: int = 480,
) -> str:
    """Create a synthetic test video with colored frames and text."""
    temp_file = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
    temp_file.close()

    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
    out = cv2.VideoWriter(temp_file.name, fourcc, fps, (width, height))

    total_frames = int(duration * fps)

    scenes = [
        ((255, 0, 0), "Blue Scene"),
        ((0, 255, 0), "Green Scene"),
        ((0, 0, 255), "Red Scene"),
        ((255, 255, 0), "Cyan Scene"),
        ((255, 0, 255), "Magenta Scene"),
        ((0, 255, 255), "Yellow Scene"),
    ]

    frames_per_scene = total_frames // len(scenes)

    for i in range(total_frames):
        frame = np.zeros((height, width, 3), dtype=np.uint8)
        scene_idx = min(i // frames_per_scene, len(scenes) - 1)
        color, scene_name = scenes[scene_idx]
        frame[:] = color

        cv2.putText(
            frame,
            scene_name,
            (width // 4, height // 2 - 30),
            cv2.FONT_HERSHEY_SIMPLEX,
            1.5,
            (255, 255, 255),
            3,
        )
        cv2.putText(
            frame,
            f"Frame {i}/{total_frames}",
            (width // 4, height // 2 + 30),
            cv2.FONT_HERSHEY_SIMPLEX,
            1.0,
            (255, 255, 255),
            2,
        )

        out.write(frame)

    out.release()
    return temp_file.name

vllm_mlx.benchmark.download_video 

download_video(url: str, timeout: int = 120) -> str

Download video from URL and return local path.

Source code in vllm_mlx/benchmark.py 
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
def download_video(url: str, timeout: int = 120) -> str:
    """Download video from URL and return local path."""
    headers = {"User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7)"}

    print(f"  Downloading video from: {url[:60]}...")
    response = requests.get(url, timeout=timeout, headers=headers, stream=True)
    response.raise_for_status()

    temp_file = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
    for chunk in response.iter_content(chunk_size=8192):
        temp_file.write(chunk)
    temp_file.close()

    file_size = Path(temp_file.name).stat().st_size
    print(f"  Downloaded: {file_size / 1024 / 1024:.1f} MB")

    return temp_file.name

vllm_mlx.benchmark.get_video_info 

get_video_info(video_path: str) -> dict

Get information about a video file.

Source code in vllm_mlx/benchmark.py 
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
def get_video_info(video_path: str) -> dict:
    """Get information about a video file."""
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        return {"error": "Cannot open video"}

    info = {
        "path": video_path,
        "total_frames": int(cap.get(cv2.CAP_PROP_FRAME_COUNT)),
        "fps": cap.get(cv2.CAP_PROP_FPS) or 30.0,
        "width": int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)),
        "height": int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)),
    }
    info["duration"] = info["total_frames"] / info["fps"] if info["fps"] > 0 else 0

    cap.release()
    return info

vllm_mlx.benchmark.benchmark_video_config 

benchmark_video_config(model, video_path: str, fps: float, max_frames: int, config_name: str, video_info: dict, max_tokens: int = 150, warmup: bool = False) -> VideoBenchmarkResult

Run a single video benchmark configuration.

Source code in vllm_mlx/benchmark.py 
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
def benchmark_video_config(
    model,
    video_path: str,
    fps: float,
    max_frames: int,
    config_name: str,
    video_info: dict,
    max_tokens: int = 150,
    warmup: bool = False,
) -> VideoBenchmarkResult:
    """Run a single video benchmark configuration."""

    # Reset MLX peak memory before this run
    reset_mlx_peak_memory()

    if not warmup:
        print(f"  {config_name:>25} |", end=" ", flush=True)

    start_time = time.perf_counter()

    output = model.generate(
        prompt="Describe what happens in this video. What do you see?",
        videos=[video_path],
        video_fps=fps,
        video_max_frames=max_frames,
        max_tokens=max_tokens,
        temperature=0.7,
    )

    elapsed = time.perf_counter() - start_time

    prompt_tokens = output.prompt_tokens
    completion_tokens = output.completion_tokens
    tps = completion_tokens / elapsed if elapsed > 0 else 0

    # Estimate frames extracted
    duration = video_info["duration"]
    frames_from_fps = int(duration * fps)
    frames_extracted = min(frames_from_fps, max_frames, video_info["total_frames"])

    # Get memory metrics
    mlx_info = get_mlx_memory_info()
    process_mem = get_process_memory()

    if not warmup:
        mem_str = f"{mlx_info.get('peak_memory_gb', 0):.1f} GB" if mlx_info else "-"
        print(
            f"{frames_extracted:>2} frames | {elapsed:>5.2f}s | {completion_tokens:>3} tok | {tps:>6.1f} tok/s | {mem_str}"
        )

    return VideoBenchmarkResult(
        config_name=config_name,
        fps=fps,
        max_frames=max_frames,
        frames_extracted=frames_extracted,
        video_duration=duration,
        time_seconds=elapsed,
        prompt_tokens=prompt_tokens,
        completion_tokens=completion_tokens,
        tokens_per_second=tps,
        response_preview=(
            output.text[:100] + "..." if len(output.text) > 100 else output.text
        ),
        memory_gb=process_mem,
        mlx_memory_gb=mlx_info.get("peak_memory_gb", 0.0),
    )

vllm_mlx.benchmark.run_video_benchmark 

run_video_benchmark(model_name: str, video_url: str = None, video_path: str = None, quick: bool = False, max_tokens: int = 150, warmup_runs: int = 1) -> list[VideoBenchmarkResult]

Run video benchmark across multiple frame configurations.

Parameters:

  • model_name (str) –

    HuggingFace MLLM model name

  • video_url (str, default: None ) –

    URL to download video from

  • video_path (str, default: None ) –

    Local video file path

  • quick (bool, default: False ) –

    If True, test only 3 configurations

  • max_tokens (int, default: 150 ) –

    Max tokens to generate

  • warmup_runs (int, default: 1 ) –

    Number of warmup runs

Returns:

Source code in vllm_mlx/benchmark.py 
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
def run_video_benchmark(
    model_name: str,
    video_url: str = None,
    video_path: str = None,
    quick: bool = False,
    max_tokens: int = 150,
    warmup_runs: int = 1,
) -> list[VideoBenchmarkResult]:
    """
    Run video benchmark across multiple frame configurations.

    Args:
        model_name: HuggingFace MLLM model name
        video_url: URL to download video from
        video_path: Local video file path
        quick: If True, test only 3 configurations
        max_tokens: Max tokens to generate
        warmup_runs: Number of warmup runs

    Returns:
        List of VideoBenchmarkResult
    """
    from vllm_mlx.models.mllm import MLXMultimodalLM
    from vllm_mlx.optimizations import detect_hardware

    # Detect hardware
    hw = detect_hardware()

    # Define configurations: (name, fps, max_frames)
    if quick:
        configs = [
            ("4 frames @ 1fps", 1.0, 4),
            ("8 frames @ 2fps", 2.0, 8),
            ("16 frames @ 2fps", 2.0, 16),
        ]
    else:
        configs = [
            # Varying frame counts (from 2 to 64 frames)
            # Note: 96+ frames causes GPU timeout on most hardware
            ("2 frames @ 0.5fps", 0.5, 2),
            ("4 frames @ 1fps", 1.0, 4),
            ("6 frames @ 1fps", 1.0, 6),
            ("8 frames @ 2fps", 2.0, 8),
            ("12 frames @ 2fps", 2.0, 12),
            ("16 frames @ 2fps", 2.0, 16),
            ("24 frames @ 4fps", 4.0, 24),
            ("32 frames @ 4fps", 4.0, 32),
            ("48 frames @ 8fps", 8.0, 48),
            ("64 frames @ 8fps", 8.0, 64),
        ]

    print(f"\n{'='*70}")
    print("vllm-mlx Video Performance Benchmark")
    print(f"{'='*70}")

    # Info table
    info_table = [
        ["Model", model_name],
        ["Hardware", f"{hw.chip_name} ({hw.total_memory_gb:.0f} GB)"],
        ["Configurations", len(configs)],
        ["Max Tokens", max_tokens],
    ]
    print(tabulate(info_table, tablefmt="plain"))
    print(f"{'='*70}\n")

    # Load model
    print(f"Loading MLLM model: {model_name}...")
    load_start = time.perf_counter()
    model = MLXMultimodalLM(model_name)
    model.load()
    load_time = time.perf_counter() - load_start
    print(f"Model loaded in {load_time:.2f}s\n")

    # Get or create video
    if video_path and Path(video_path).exists():
        print(f"Using local video: {video_path}")
    elif video_url:
        video_path = download_video(video_url)
    else:
        print("Downloading default test video (Big Buck Bunny 10s)...")
        video_path = download_video(DEFAULT_VIDEO_URL)

    video_info = get_video_info(video_path)
    print(
        f"\nVideo: {video_info['width']}x{video_info['height']}, "
        f"{video_info['duration']:.1f}s, {video_info['fps']:.1f} fps, "
        f"{video_info['total_frames']} frames\n"
    )

    # Warmup
    if warmup_runs > 0:
        print(f"Running {warmup_runs} warmup run(s)...")
        for _ in range(warmup_runs):
            benchmark_video_config(
                model, video_path, 1.0, 4, "warmup", video_info, max_tokens, warmup=True
            )

        # Show model memory after warmup (MLX uses lazy evaluation)
        mlx_info = get_mlx_memory_info(reset_peak=True)
        if mlx_info and mlx_info.get("peak_memory_gb", 0) > 0:
            print(f"Model memory: {mlx_info['peak_memory_gb']:.2f} GB (MLX peak)")
        print("Warmup complete.\n")

    # Run benchmarks
    print("-" * 85)
    print(
        f"  {'Configuration':>25} | {'Frames':>6} | {'Time':>6} | {'Tokens':>4} | {'Speed':>10} | {'Memory':>8}"
    )
    print("-" * 85)

    results = []
    for config_name, fps, max_frames in configs:
        try:
            result = benchmark_video_config(
                model, video_path, fps, max_frames, config_name, video_info, max_tokens
            )
            results.append(result)
        except Exception as e:
            print(f"  Error with {config_name}: {e}")

    return results

vllm_mlx.benchmark.print_video_summary 

print_video_summary(results: list[VideoBenchmarkResult], model_name: str)

Print video benchmark summary.

Source code in vllm_mlx/benchmark.py 
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
def print_video_summary(results: list[VideoBenchmarkResult], model_name: str):
    """Print video benchmark summary."""
    if not results:
        print("No results to display.")
        return

    print(f"\n{'='*85}")
    print("VIDEO BENCHMARK RESULTS")
    print(f"{'='*85}\n")

    # Results table
    table_data = []
    for r in sorted(results, key=lambda x: x.frames_extracted):
        table_data.append(
            [
                r.config_name,
                r.frames_extracted,
                f"{r.time_seconds:.2f}s",
                r.completion_tokens,
                f"{r.tokens_per_second:.1f}",
                f"{r.mlx_memory_gb:.2f}" if r.mlx_memory_gb > 0 else "-",
            ]
        )

    headers = ["Configuration", "Frames", "Time", "Tokens", "Tok/s", "Mem (GB)"]
    print(tabulate(table_data, headers=headers, tablefmt="simple"))

    # Summary stats
    total_time = sum(r.time_seconds for r in results)
    total_tokens = sum(r.completion_tokens for r in results)
    avg_tps = total_tokens / total_time if total_time > 0 else 0
    peak_memory = max(r.mlx_memory_gb for r in results) if results else 0

    print("-" * 85)
    print(f"Total Time:      {total_time:.2f}s")
    print(f"Total Tokens:    {total_tokens}")
    print(f"Average Tok/s:   {avg_tps:.1f}")
    if peak_memory > 0:
        print(f"Peak Memory:     {peak_memory:.2f} GB")

    fastest = min(results, key=lambda r: r.time_seconds)
    slowest = max(results, key=lambda r: r.time_seconds)
    most_frames = max(results, key=lambda r: r.frames_extracted)

    print(
        f"\nFastest:     {fastest.config_name} ({fastest.time_seconds:.2f}s, {fastest.tokens_per_second:.1f} tok/s)"
    )
    print(
        f"Slowest:     {slowest.config_name} ({slowest.time_seconds:.2f}s, {slowest.tokens_per_second:.1f} tok/s)"
    )
    print(
        f"Most Frames: {most_frames.config_name} ({most_frames.frames_extracted} frames)"
    )
    print(f"{'='*85}")

vllm_mlx.benchmark.print_summary 

print_summary(summary: BenchmarkSummary)

Print a formatted summary of benchmark results using tabulate.

Source code in vllm_mlx/benchmark.py 
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
def print_summary(summary: BenchmarkSummary):
    """Print a formatted summary of benchmark results using tabulate."""
    print(f"\n{'='*60}")
    print("BENCHMARK RESULTS")
    print(f"{'='*60}\n")

    # Overview table
    overview_data = [
        ["Model", summary.model_name],
        ["Hardware", f"{summary.hardware_chip} ({summary.hardware_memory_gb:.0f} GB)"],
        ["Total Runs", summary.num_runs],
        ["Input Tokens", f"{summary.total_prompt_tokens:,}"],
        ["Output Tokens", f"{summary.total_generated_tokens:,}"],
        ["Total Time", f"{summary.total_time:.2f}s"],
    ]
    print(tabulate(overview_data, tablefmt="plain"))
    print()

    # Main metrics table
    metrics_data = [
        [
            "TTFT (Time to First Token)",
            f"{summary.ttft_mean*1000:.1f} ms",
            f"{summary.ttft_p95*1000:.1f} ms",
        ],
        [
            "TPOT (Time Per Output Token)",
            f"{summary.tpot_mean*1000:.2f} ms",
            f"{summary.tpot_max*1000:.2f} ms",
        ],
        [
            "Generation Speed",
            f"{summary.generation_tps_mean:.1f} tok/s",
            f"{summary.generation_tps_max:.1f} tok/s",
        ],
        ["Processing Speed", f"{summary.processing_tps_mean:.1f} tok/s", "-"],
        [
            "Latency (per request)",
            f"{summary.latency_mean:.2f}s",
            f"{summary.latency_p95:.2f}s",
        ],
    ]
    print("Performance Metrics:")
    print(
        tabulate(
            metrics_data,
            headers=["Metric", "Mean", "P95/Max"],
            tablefmt="simple",
        )
    )
    print()

    # Throughput table
    throughput_data = [
        ["Total Throughput", f"{summary.total_throughput_tps:.1f} tok/s"],
        ["Requests/Second", f"{summary.requests_per_second:.2f} req/s"],
    ]
    print("Throughput:")
    print(tabulate(throughput_data, tablefmt="plain"))
    print()

    # Resource metrics
    res = summary.resources
    if res.process_memory_gb > 0 or res.mlx_peak_memory_gb > 0:
        print("Resource Usage:")
        resource_data = []

        if res.process_memory_gb > 0:
            resource_data.append(
                ["Process Memory (peak)", f"{res.process_memory_gb:.2f} GB"]
            )

        if res.mlx_peak_memory_gb > 0:
            resource_data.append(
                ["MLX Peak Memory", f"{res.mlx_peak_memory_gb:.2f} GB"]
            )

        if res.mlx_cache_gb > 0:
            resource_data.append(["MLX Cache Memory", f"{res.mlx_cache_gb:.2f} GB"])

        if res.system_memory_total_gb > 0:
            used_pct = (res.system_memory_used_gb / res.system_memory_total_gb) * 100
            resource_data.append(
                [
                    "System Memory",
                    f"{res.system_memory_used_gb:.1f} / {res.system_memory_total_gb:.0f} GB ({used_pct:.0f}%)",
                ]
            )

        print(tabulate(resource_data, tablefmt="plain"))
        print()

    print(f"{'='*60}")

vllm_mlx.benchmark.main 

main()

Run the benchmark.

Source code in vllm_mlx/benchmark.py 
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
def main():
    """Run the benchmark."""
    parser = argparse.ArgumentParser(
        description="vllm-mlx Performance Benchmark (LLM, MLLM Image & Video)",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Examples:
    # LLM benchmark
    vllm-mlx-bench --model mlx-community/Llama-3.2-1B-Instruct-4bit
    vllm-mlx-bench --model mlx-community/Llama-3.2-3B-Instruct-4bit --prompts 10

    # MLLM image benchmark (auto-detected)
    vllm-mlx-bench --model mlx-community/Qwen3-VL-4B-Instruct-3bit
    vllm-mlx-bench --model mlx-community/Qwen3-VL-4B-Instruct-3bit --quick

    # MLLM video benchmark
    vllm-mlx-bench --model mlx-community/Qwen3-VL-4B-Instruct-3bit --video
    vllm-mlx-bench --model mlx-community/Qwen3-VL-4B-Instruct-3bit --video --quick
    vllm-mlx-bench --model mlx-community/Qwen3-VL-4B-Instruct-3bit --video --video-url https://example.com/video.mp4

    # Force MLLM mode
    vllm-mlx-bench --model custom-vision-model --mllm
        """,
    )
    parser.add_argument(
        "--model",
        type=str,
        required=True,
        help="Model name (HuggingFace model name or local path)",
    )
    parser.add_argument(
        "--prompts",
        type=int,
        default=5,
        help="Number of prompts to benchmark for LLM (default: 5)",
    )
    parser.add_argument(
        "--max-tokens",
        type=int,
        default=256,
        help="Maximum tokens to generate per prompt (default: 256)",
    )
    parser.add_argument(
        "--temperature",
        type=float,
        default=0.7,
        help="Sampling temperature (default: 0.7)",
    )
    parser.add_argument(
        "--warmup",
        type=int,
        default=1,
        help="Number of warmup runs (default: 1)",
    )
    parser.add_argument(
        "--output",
        type=str,
        default=None,
        help="Output file for JSON results",
    )
    parser.add_argument(
        "--mllm",
        action="store_true",
        help="Force MLLM benchmark mode (auto-detected by default)",
    )
    parser.add_argument(
        "--quick",
        action="store_true",
        help="Quick benchmark with fewer configurations",
    )
    # Video benchmark arguments
    parser.add_argument(
        "--video",
        action="store_true",
        help="Run video benchmark instead of image benchmark (for MLLM models)",
    )
    parser.add_argument(
        "--video-url",
        type=str,
        default=None,
        help="URL of video to use for benchmark (default: Big Buck Bunny 10s)",
    )
    parser.add_argument(
        "--video-path",
        type=str,
        default=None,
        help="Local path to video file for benchmark",
    )

    args = parser.parse_args()

    # Determine if MLLM model
    run_mllm = args.mllm or is_mllm_model(args.model)

    if args.video:
        # Video Benchmark
        results = run_video_benchmark(
            model_name=args.model,
            video_url=args.video_url,
            video_path=args.video_path,
            quick=args.quick,
            max_tokens=args.max_tokens,
            warmup_runs=args.warmup,
        )

        if results:
            print_video_summary(results, args.model)

            # Save to JSON if requested
            if args.output:
                with open(args.output, "w") as f:
                    json.dump(
                        {
                            "type": "video",
                            "model": args.model,
                            "test_video": args.video_url
                            or args.video_path
                            or "Big Buck Bunny 10s",
                            "results": [
                                {
                                    "config_name": r.config_name,
                                    "fps": r.fps,
                                    "max_frames": r.max_frames,
                                    "frames_extracted": r.frames_extracted,
                                    "video_duration": r.video_duration,
                                    "time_seconds": r.time_seconds,
                                    "prompt_tokens": r.prompt_tokens,
                                    "completion_tokens": r.completion_tokens,
                                    "tokens_per_second": r.tokens_per_second,
                                    "response_preview": r.response_preview,
                                }
                                for r in results
                            ],
                        },
                        f,
                        indent=2,
                    )
                print(f"\nResults saved to: {args.output}")

    elif run_mllm:
        # MLLM Image Benchmark
        results = run_mllm_benchmark(
            model_name=args.model,
            quick=args.quick,
            max_tokens=args.max_tokens,
            warmup_runs=args.warmup,
        )

        if results:
            print_mllm_summary(results, args.model)

            # Save to JSON if requested
            if args.output:
                with open(args.output, "w") as f:
                    json.dump(
                        {
                            "type": "mllm_image",
                            "model": args.model,
                            "test_image": "Yellow Labrador (Wikimedia Commons)",
                            "results": [
                                {
                                    "resolution": r.resolution,
                                    "width": r.width,
                                    "height": r.height,
                                    "pixels": r.pixels,
                                    "time_seconds": r.time_seconds,
                                    "tokens_generated": r.tokens_generated,
                                    "tokens_per_second": r.tokens_per_second,
                                    "response_preview": r.response_preview,
                                }
                                for r in results
                            ],
                        },
                        f,
                        indent=2,
                    )
                print(f"\nResults saved to: {args.output}")
    else:
        # LLM Benchmark
        summary = run_benchmark(
            model_name=args.model,
            num_prompts=args.prompts,
            max_tokens=args.max_tokens,
            temperature=args.temperature,
            warmup_runs=args.warmup,
        )

        if summary:
            print_summary(summary)

            # Save to JSON if requested
            if args.output:
                with open(args.output, "w") as f:
                    json.dump(
                        {
                            "type": "llm",
                            "model": summary.model_name,
                            "hardware": {
                                "chip": summary.hardware_chip,
                                "memory_gb": summary.hardware_memory_gb,
                                "bandwidth_gbs": summary.hardware_bandwidth_gbs,
                            },
                            "num_runs": summary.num_runs,
                            "total_prompt_tokens": summary.total_prompt_tokens,
                            "total_generated_tokens": summary.total_generated_tokens,
                            "total_time_seconds": summary.total_time,
                            "ttft_ms": {
                                "mean": summary.ttft_mean * 1000,
                                "min": summary.ttft_min * 1000,
                                "max": summary.ttft_max * 1000,
                                "p50": summary.ttft_p50 * 1000,
                                "p95": summary.ttft_p95 * 1000,
                            },
                            "tpot_ms": {
                                "mean": summary.tpot_mean * 1000,
                                "min": summary.tpot_min * 1000,
                                "max": summary.tpot_max * 1000,
                            },
                            "tokens_per_second": {
                                "generation_mean": summary.generation_tps_mean,
                                "generation_max": summary.generation_tps_max,
                                "processing_mean": summary.processing_tps_mean,
                                "total_throughput": summary.total_throughput_tps,
                            },
                            "latency_seconds": {
                                "mean": summary.latency_mean,
                                "min": summary.latency_min,
                                "max": summary.latency_max,
                                "p50": summary.latency_p50,
                                "p95": summary.latency_p95,
                            },
                            "requests_per_second": summary.requests_per_second,
                        },
                        f,
                        indent=2,
                    )
                print(f"\nResults saved to: {args.output}")

Complete contract reference

Expand any definition for its exact inputs, annotations, defaults, return contract, directly raised exceptions, source-grounded behavior, and immutable line link. This section includes private and nested definitions that ordinary API generators omit.

vllm_mlx.benchmark.ResourceMetrics · class 
vllm_mlx.benchmark.ResourceMetrics(process_memory_gb: float = 0.0, mlx_cache_gb: float = 0.0, mlx_peak_memory_gb: float = 0.0, system_memory_used_gb: float = 0.0, system_memory_total_gb: float = 0.0)

Resource usage metrics during benchmark.

Parameters

Name Type Required Default Description
process_memory_gb float no 0.0 Optional constructor field; defaults to 0.0.
mlx_cache_gb float no 0.0 Optional constructor field; defaults to 0.0.
mlx_peak_memory_gb float no 0.0 Optional constructor field; defaults to 0.0.
system_memory_used_gb float no 0.0 Optional constructor field; defaults to 0.0.
system_memory_total_gb float no 0.0 Optional constructor field; defaults to 0.0.

Returns

  • Constructs: vllm_mlx.benchmark.ResourceMetrics

Exceptions and behavior

Class ResourceMetrics declares 0 direct member(s). No direct raise statement appears in this definition.

View source #L72-L80.

vllm_mlx.benchmark.reset_mlx_peak_memory · function 
vllm_mlx.benchmark.reset_mlx_peak_memory() -> not annotated

Reset MLX peak memory counter.

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated
  • Direct return expressions: None

Exceptions and behavior

Function reset_mlx_peak_memory calls hasattr, mx.reset_peak_memory, mx.metal.reset_peak_memory; returns None. No direct raise statement appears in this definition.

View source #L83-L95.

vllm_mlx.benchmark.get_mlx_memory_info · function 
vllm_mlx.benchmark.get_mlx_memory_info(reset_peak: bool = True) -> dict

Get MLX memory usage information.

Parameters

Name Type Required Default Description
reset_peak bool no True If True, reset peak memory counter after reading.

Returns

  • Type: dict
  • Direct return expressions: {}; info

Exceptions and behavior

Function get_mlx_memory_info calls hasattr, mx.get_cache_memory, mx.get_peak_memory, mx.get_active_memory; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L98-L135.

vllm_mlx.benchmark.get_process_memory · function 
vllm_mlx.benchmark.get_process_memory() -> float

Get current process memory usage in GB.

Parameters

This callable has no explicit inputs.

Returns

  • Type: float
  • Direct return expressions: 0.0; process.memory_info().rss / 1024 ** 3

Exceptions and behavior

Function get_process_memory calls psutil.Process, process.memory_info; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L138-L147.

vllm_mlx.benchmark.get_system_memory · function 
vllm_mlx.benchmark.get_system_memory() -> tuple[float, float]

Get system memory (used, total) in GB.

Parameters

This callable has no explicit inputs.

Returns

  • Type: tuple[float, float]
  • Direct return expressions: (0.0, 0.0); (mem.used / 1024 ** 3, mem.total / 1024 ** 3)

Exceptions and behavior

Function get_system_memory calls psutil.virtual_memory; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L150-L159.

vllm_mlx.benchmark.ResourceMonitor · class 
vllm_mlx.benchmark.ResourceMonitor()

Monitor system resources during benchmark runs.

Parameters

This callable has no explicit inputs.

Returns

  • Constructs: vllm_mlx.benchmark.ResourceMonitor

Exceptions and behavior

Class ResourceMonitor declares 4 direct member(s). No direct raise statement appears in this definition.

View source #L162-L213.

vllm_mlx.benchmark.ResourceMonitor.__init__ · method 
vllm_mlx.benchmark.ResourceMonitor.__init__() -> not annotated

Method ResourceMonitor.__init__ updates self.samples, self._start_time, self._start_memory.

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated

Exceptions and behavior

Method ResourceMonitor.__init__ updates self.samples, self._start_time, self._start_memory. No direct raise statement appears in this definition.

View source #L165-L168.

vllm_mlx.benchmark.ResourceMonitor.start · method 
vllm_mlx.benchmark.ResourceMonitor.start() -> not annotated

Start monitoring.

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated

Exceptions and behavior

Method ResourceMonitor.start updates self._start_time, self._start_memory; calls time.perf_counter, get_process_memory, reset_mlx_peak_memory. No direct raise statement appears in this definition.

View source #L170-L176.

vllm_mlx.benchmark.ResourceMonitor.sample · method 
vllm_mlx.benchmark.ResourceMonitor.sample() -> ResourceMetrics

Take a resource sample.

Parameters

This callable has no explicit inputs.

Returns

  • Type: ResourceMetrics
  • Direct return expressions: metrics

Exceptions and behavior

Method ResourceMonitor.sample calls get_mlx_memory_info, get_system_memory, ResourceMetrics, get_process_memory; returns metrics. No direct raise statement appears in this definition.

View source #L178-L192.

vllm_mlx.benchmark.ResourceMonitor.get_summary · method 
vllm_mlx.benchmark.ResourceMonitor.get_summary() -> ResourceMetrics

Get summary of all samples.

Parameters

This callable has no explicit inputs.

Returns

  • Type: ResourceMetrics
  • Direct return expressions: ResourceMetrics(); ResourceMetrics(process_memory_gb=peak_process, mlx_cache_gb=peak_mlx_cache, mlx_peak_memory_gb=peak_mlx, system_memory…

Exceptions and behavior

Method ResourceMonitor.get_summary calls ResourceMetrics, max; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L194-L213.

vllm_mlx.benchmark.BenchmarkResult · class 
vllm_mlx.benchmark.BenchmarkResult(prompt: str, prompt_tokens: int, generated_tokens: int, ttft: float, total_time: float, tpot: float = 0.0, generation_tps: float = 0.0, processing_tps: float = 0.0)

Results from a single benchmark run.

Parameters

Name Type Required Default Description
prompt str yes none Required constructor field.
prompt_tokens int yes none Required constructor field.
generated_tokens int yes none Required constructor field.
ttft float yes none Required constructor field.
total_time float yes none Required constructor field.
tpot float no 0.0 Optional constructor field; defaults to 0.0.
generation_tps float no 0.0 Optional constructor field; defaults to 0.0.
processing_tps float no 0.0 Optional constructor field; defaults to 0.0.

Returns

  • Constructs: vllm_mlx.benchmark.BenchmarkResult

Exceptions and behavior

Class BenchmarkResult declares 1 direct member(s). No direct raise statement appears in this definition.

View source #L235-L268.

vllm_mlx.benchmark.BenchmarkResult.__post_init__ · method 
vllm_mlx.benchmark.BenchmarkResult.__post_init__() -> not annotated

Method BenchmarkResult.__post_init__ updates self.tpot, self.generation_tps, self.processing_tps.

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated

Exceptions and behavior

Method BenchmarkResult.__post_init__ updates self.tpot, self.generation_tps, self.processing_tps. No direct raise statement appears in this definition.

View source #L251-L268.

vllm_mlx.benchmark.BenchmarkSummary · class 
vllm_mlx.benchmark.BenchmarkSummary(model_name: str, num_runs: int, total_prompt_tokens: int, total_generated_tokens: int, total_time: float, ttft_mean: float, ttft_min: float, ttft_max: float, ttft_p50: float, ttft_p95: float, tpot_mean: float, tpot_min: float, tpot_max: float, generation_tps_mean: float, generation_tps_max: float, processing_tps_mean: float, latency_mean: float, latency_min: float, latency_max: float, latency_p50: float, latency_p95: float, total_throughput_tps: float, requests_per_second: float, hardware_chip: str = '', hardware_memory_gb: float = 0.0, hardware_bandwidth_gbs: float = 0.0, resources: ResourceMetrics = field(default_factory=ResourceMetrics))

Summary statistics across all benchmark runs.

Parameters

Name Type Required Default Description
model_name str yes none Required constructor field.
num_runs int yes none Required constructor field.
total_prompt_tokens int yes none Required constructor field.
total_generated_tokens int yes none Required constructor field.
total_time float yes none Required constructor field.
ttft_mean float yes none Required constructor field.
ttft_min float yes none Required constructor field.
ttft_max float yes none Required constructor field.
ttft_p50 float yes none Required constructor field.
ttft_p95 float yes none Required constructor field.
tpot_mean float yes none Required constructor field.
tpot_min float yes none Required constructor field.
tpot_max float yes none Required constructor field.
generation_tps_mean float yes none Required constructor field.
generation_tps_max float yes none Required constructor field.
processing_tps_mean float yes none Required constructor field.
latency_mean float yes none Required constructor field.
latency_min float yes none Required constructor field.
latency_max float yes none Required constructor field.
latency_p50 float yes none Required constructor field.
latency_p95 float yes none Required constructor field.
total_throughput_tps float yes none Required constructor field.
requests_per_second float yes none Required constructor field.
hardware_chip str no '' Optional constructor field; defaults to ''.
hardware_memory_gb float no 0.0 Optional constructor field; defaults to 0.0.
hardware_bandwidth_gbs float no 0.0 Optional constructor field; defaults to 0.0.
resources ResourceMetrics no field(default_factory=ResourceMetrics) Optional constructor field; defaults to field(default_factory=ResourceMetrics).

Returns

  • Constructs: vllm_mlx.benchmark.BenchmarkSummary

Exceptions and behavior

Class BenchmarkSummary declares 0 direct member(s). No direct raise statement appears in this definition.

View source #L272-L315.

vllm_mlx.benchmark.calculate_percentile · function 
vllm_mlx.benchmark.calculate_percentile(data: list, percentile: float) -> float

Calculate percentile from a list.

Parameters

Name Type Required Default Description
data list yes none Required positional or keyword input.
percentile float yes none Required positional or keyword input.

Returns

  • Type: float
  • Direct return expressions: 0.0; sorted_data[index]

Exceptions and behavior

Function calculate_percentile calls sorted, int, len, min; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L318-L325.

vllm_mlx.benchmark.benchmark_single_prompt · function 
vllm_mlx.benchmark.benchmark_single_prompt(model, tokenizer, prompt: str, max_tokens: int = 256, temperature: float = 0.7) -> Optional[BenchmarkResult]

Benchmark a single prompt with detailed timing.

Parameters

Name Type Required Default Description
model not annotated yes none The loaded MLX model
tokenizer not annotated yes none The tokenizer
prompt str yes none The prompt to benchmark
max_tokens int no 256 Maximum tokens to generate
temperature float no 0.7 Sampling temperature

Returns

  • Type: Optional[BenchmarkResult]
  • Direct return expressions: BenchmarkResult(prompt=prompt[:50] + '...' if len(prompt) > 50 else prompt, prompt_tokens=prompt_token_count, generated…; None

Exceptions and behavior

Function benchmark_single_prompt calls tokenizer.encode, len, make_sampler, time.perf_counter; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L328-L394.

vllm_mlx.benchmark.run_benchmark · function 
vllm_mlx.benchmark.run_benchmark(model_name: str, num_prompts: int = 5, max_tokens: int = 256, temperature: float = 0.7, warmup_runs: int = 1) -> Optional[BenchmarkSummary]

Run the full benchmark suite.

Parameters

Name Type Required Default Description
model_name str yes none HuggingFace model name or local path
num_prompts int no 5 Number of prompts to test
max_tokens int no 256 Maximum tokens per generation
temperature float no 0.7 Sampling temperature
warmup_runs int no 1 Number of warmup runs before measuring

Returns

  • Type: Optional[BenchmarkSummary]
  • Direct return expressions: None; summary

Exceptions and behavior

Function run_benchmark calls detect_hardware, len, print, tabulate; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L397-L610.

vllm_mlx.benchmark.is_mllm_model · function 
vllm_mlx.benchmark.is_mllm_model(model_name: str) -> bool

Check if model name indicates a multimodal language model.

Parameters

Name Type Required Default Description
model_name str yes none Required positional or keyword input.

Returns

  • Type: bool
  • Direct return expressions: True; False

Exceptions and behavior

Function is_mllm_model calls model_name.lower, pattern.lower; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L651-L657.

vllm_mlx.benchmark.MLLMBenchmarkResult · class 
vllm_mlx.benchmark.MLLMBenchmarkResult(resolution: str, width: int, height: int, pixels: int, time_seconds: float, tokens_generated: int, tokens_per_second: float, response_preview: str, memory_gb: float = 0.0, mlx_memory_gb: float = 0.0)

Result from a single MLLM benchmark run.

Parameters

Name Type Required Default Description
resolution str yes none Required constructor field.
width int yes none Required constructor field.
height int yes none Required constructor field.
pixels int yes none Required constructor field.
time_seconds float yes none Required constructor field.
tokens_generated int yes none Required constructor field.
tokens_per_second float yes none Required constructor field.
response_preview str yes none Required constructor field.
memory_gb float no 0.0 Optional constructor field; defaults to 0.0.
mlx_memory_gb float no 0.0 Optional constructor field; defaults to 0.0.

Returns

  • Constructs: vllm_mlx.benchmark.MLLMBenchmarkResult

Exceptions and behavior

Class MLLMBenchmarkResult declares 0 direct member(s). No direct raise statement appears in this definition.

View source #L661-L674.

vllm_mlx.benchmark.download_test_image · function 
vllm_mlx.benchmark.download_test_image(url: str, timeout: int = 30) -> Image.Image

Download image from URL and return PIL Image.

Parameters

Name Type Required Default Description
url str yes none Required positional or keyword input.
timeout int no 30 Optional positional or keyword input; defaults to 30.

Returns

  • Type: Image.Image
  • Direct return expressions: Image.open(io.BytesIO(response.content))

Exceptions and behavior

Function download_test_image calls requests.get, response.raise_for_status, Image.open, io.BytesIO; returns Image.open(io.BytesIO(response.content)). No direct raise statement appears in this definition.

View source #L677-L684.

vllm_mlx.benchmark.resize_image · function 
vllm_mlx.benchmark.resize_image(img: Image.Image, width: int, height: int) -> Image.Image

Resize image to specified dimensions.

Parameters

Name Type Required Default Description
img Image.Image yes none Required positional or keyword input.
width int yes none Required positional or keyword input.
height int yes none Required positional or keyword input.

Returns

  • Type: Image.Image
  • Direct return expressions: img.resize((width, height), Image.Resampling.LANCZOS)

Exceptions and behavior

Function resize_image calls img.resize; returns img.resize((width, height), Image.Resampling.LANCZOS). No direct raise statement appears in this definition.

View source #L687-L689.

vllm_mlx.benchmark.image_to_base64 · function 
vllm_mlx.benchmark.image_to_base64(img: Image.Image, format: str = 'JPEG') -> str

Convert PIL Image to base64 data URL.

Parameters

Name Type Required Default Description
img Image.Image yes none Required positional or keyword input.
format str no 'JPEG' Optional positional or keyword input; defaults to 'JPEG'.

Returns

  • Type: str
  • Direct return expressions: f'data:{mime};base64,{b64}'

Exceptions and behavior

Function image_to_base64 calls Image.new, background.paste, img.split, img.convert; returns f'data:{mime};base64,{b64}'. No direct raise statement appears in this definition.

View source #L692-L705.

vllm_mlx.benchmark.benchmark_mllm_resolution · function 
vllm_mlx.benchmark.benchmark_mllm_resolution(model, processor, config, base_image: Image.Image, width: int, height: int, max_tokens: int = 256, warmup: bool = False) -> MLLMBenchmarkResult

Run MLLM benchmark for a specific resolution.

Parameters

Name Type Required Default Description
model not annotated yes none Required positional or keyword input.
processor not annotated yes none Required positional or keyword input.
config not annotated yes none Required positional or keyword input.
base_image Image.Image yes none Required positional or keyword input.
width int yes none Required positional or keyword input.
height int yes none Required positional or keyword input.
max_tokens int no 256 Optional positional or keyword input; defaults to 256.
warmup bool no False Optional positional or keyword input; defaults to False.

Returns

  • Type: MLLMBenchmarkResult
  • Direct return expressions: MLLMBenchmarkResult(resolution=resolution_name, width=width, height=height, pixels=pixels, time_seconds=elapsed, tokens…

Exceptions and behavior

Function benchmark_mllm_resolution calls reset_mlx_peak_memory, resize_image, tempfile.NamedTemporaryFile, img.save; returns MLLMBenchmarkResult(resolution=resolution_name, width=width, height=height, pixels=pixels, time_seconds=elapsed, tokens…. No direct raise statement appears in this definition.

View source #L708-L800.

vllm_mlx.benchmark.run_mllm_benchmark · function 
vllm_mlx.benchmark.run_mllm_benchmark(model_name: str, quick: bool = False, max_tokens: int = 256, warmup_runs: int = 1) -> list[MLLMBenchmarkResult]

Run MLLM benchmark across multiple image resolutions.

Parameters

Name Type Required Default Description
model_name str yes none HuggingFace model name
quick bool no False If True, test only 4 resolutions
max_tokens int no 256 Max tokens to generate
warmup_runs int no 1 Number of warmup runs

Returns

  • Type: list[MLLMBenchmarkResult]
  • Direct return expressions: []; results

Exceptions and behavior

Function run_mllm_benchmark calls detect_hardware, print, len, tabulate; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L803-L913.

vllm_mlx.benchmark.print_mllm_summary · function 
vllm_mlx.benchmark.print_mllm_summary(results: list[MLLMBenchmarkResult], model_name: str) -> not annotated

Print MLLM benchmark summary.

Parameters

Name Type Required Default Description
results list[MLLMBenchmarkResult] yes none Required positional or keyword input.
model_name str yes none Required positional or keyword input.

Returns

  • Type: not annotated
  • Direct return expressions: None

Exceptions and behavior

Function print_mllm_summary calls print, table_data.append, tabulate, sum; returns None. No direct raise statement appears in this definition.

View source #L916-L975.

vllm_mlx.benchmark.VideoBenchmarkResult · class 
vllm_mlx.benchmark.VideoBenchmarkResult(config_name: str, fps: float, max_frames: int, frames_extracted: int, video_duration: float, time_seconds: float, prompt_tokens: int, completion_tokens: int, tokens_per_second: float, response_preview: str, memory_gb: float = 0.0, mlx_memory_gb: float = 0.0)

Result from a single video benchmark run.

Parameters

Name Type Required Default Description
config_name str yes none Required constructor field.
fps float yes none Required constructor field.
max_frames int yes none Required constructor field.
frames_extracted int yes none Required constructor field.
video_duration float yes none Required constructor field.
time_seconds float yes none Required constructor field.
prompt_tokens int yes none Required constructor field.
completion_tokens int yes none Required constructor field.
tokens_per_second float yes none Required constructor field.
response_preview str yes none Required constructor field.
memory_gb float no 0.0 Optional constructor field; defaults to 0.0.
mlx_memory_gb float no 0.0 Optional constructor field; defaults to 0.0.

Returns

  • Constructs: vllm_mlx.benchmark.VideoBenchmarkResult

Exceptions and behavior

Class VideoBenchmarkResult declares 0 direct member(s). No direct raise statement appears in this definition.

View source #L984-L999.

vllm_mlx.benchmark.create_test_video · function 
vllm_mlx.benchmark.create_test_video(duration: float = 10.0, fps: float = 30.0, width: int = 640, height: int = 480) -> str

Create a synthetic test video with colored frames and text.

Parameters

Name Type Required Default Description
duration float no 10.0 Optional positional or keyword input; defaults to 10.0.
fps float no 30.0 Optional positional or keyword input; defaults to 30.0.
width int no 640 Optional positional or keyword input; defaults to 640.
height int no 480 Optional positional or keyword input; defaults to 480.

Returns

  • Type: str
  • Direct return expressions: temp_file.name

Exceptions and behavior

Function create_test_video calls tempfile.NamedTemporaryFile, temp_file.close, cv2.VideoWriter_fourcc, cv2.VideoWriter; returns temp_file.name. No direct raise statement appears in this definition.

View source #L1002-L1056.

vllm_mlx.benchmark.download_video · function 
vllm_mlx.benchmark.download_video(url: str, timeout: int = 120) -> str

Download video from URL and return local path.

Parameters

Name Type Required Default Description
url str yes none Required positional or keyword input.
timeout int no 120 Optional positional or keyword input; defaults to 120.

Returns

  • Type: str
  • Direct return expressions: temp_file.name

Exceptions and behavior

Function download_video calls print, requests.get, response.raise_for_status, tempfile.NamedTemporaryFile; returns temp_file.name. No direct raise statement appears in this definition.

View source #L1059-L1075.

vllm_mlx.benchmark.get_video_info · function 
vllm_mlx.benchmark.get_video_info(video_path: str) -> dict

Get information about a video file.

Parameters

Name Type Required Default Description
video_path str yes none Required positional or keyword input.

Returns

  • Type: dict
  • Direct return expressions: {'error': 'Cannot open video'}; info

Exceptions and behavior

Function get_video_info calls cv2.VideoCapture, cap.isOpened, int, cap.get; has 2 explicit return paths. No direct raise statement appears in this definition.

View source #L1078-L1094.

vllm_mlx.benchmark.benchmark_video_config · function 
vllm_mlx.benchmark.benchmark_video_config(model, video_path: str, fps: float, max_frames: int, config_name: str, video_info: dict, max_tokens: int = 150, warmup: bool = False) -> VideoBenchmarkResult

Run a single video benchmark configuration.

Parameters

Name Type Required Default Description
model not annotated yes none Required positional or keyword input.
video_path str yes none Required positional or keyword input.
fps float yes none Required positional or keyword input.
max_frames int yes none Required positional or keyword input.
config_name str yes none Required positional or keyword input.
video_info dict yes none Required positional or keyword input.
max_tokens int no 150 Optional positional or keyword input; defaults to 150.
warmup bool no False Optional positional or keyword input; defaults to False.

Returns

  • Type: VideoBenchmarkResult
  • Direct return expressions: VideoBenchmarkResult(config_name=config_name, fps=fps, max_frames=max_frames, frames_extracted=frames_extracted, video_…

Exceptions and behavior

Function benchmark_video_config calls reset_mlx_peak_memory, print, time.perf_counter, model.generate; returns VideoBenchmarkResult(config_name=config_name, fps=fps, max_frames=max_frames, frames_extracted=frames_extracted, video_…. No direct raise statement appears in this definition.

View source #L1097-L1162.

vllm_mlx.benchmark.run_video_benchmark · function 
vllm_mlx.benchmark.run_video_benchmark(model_name: str, video_url: str = None, video_path: str = None, quick: bool = False, max_tokens: int = 150, warmup_runs: int = 1) -> list[VideoBenchmarkResult]

Run video benchmark across multiple frame configurations.

Parameters

Name Type Required Default Description
model_name str yes none HuggingFace MLLM model name
video_url str no None URL to download video from
video_path str no None Local video file path
quick bool no False If True, test only 3 configurations
max_tokens int no 150 Max tokens to generate
warmup_runs int no 1 Number of warmup runs

Returns

  • Type: list[VideoBenchmarkResult]
  • Direct return expressions: results

Exceptions and behavior

Function run_video_benchmark calls detect_hardware, print, len, tabulate; returns results. No direct raise statement appears in this definition.

View source #L1165-L1285.

vllm_mlx.benchmark.print_video_summary · function 
vllm_mlx.benchmark.print_video_summary(results: list[VideoBenchmarkResult], model_name: str) -> not annotated

Print video benchmark summary.

Parameters

Name Type Required Default Description
results list[VideoBenchmarkResult] yes none Required positional or keyword input.
model_name str yes none Required positional or keyword input.

Returns

  • Type: not annotated
  • Direct return expressions: None

Exceptions and behavior

Function print_video_summary calls print, sorted, table_data.append, tabulate; returns None. No direct raise statement appears in this definition.

View source #L1288-L1341.

vllm_mlx.benchmark.print_summary · function 
vllm_mlx.benchmark.print_summary(summary: BenchmarkSummary) -> not annotated

Print a formatted summary of benchmark results using tabulate.

Parameters

Name Type Required Default Description
summary BenchmarkSummary yes none Required positional or keyword input.

Returns

  • Type: not annotated

Exceptions and behavior

Function print_summary calls print, tabulate, resource_data.append. No direct raise statement appears in this definition.

View source #L1349-L1441.

vllm_mlx.benchmark.main · function 
vllm_mlx.benchmark.main() -> not annotated

Run the benchmark.

Parameters

This callable has no explicit inputs.

Returns

  • Type: not annotated

Exceptions and behavior

Function main calls argparse.ArgumentParser, parser.add_argument, parser.parse_args, is_mllm_model. No direct raise statement appears in this definition.

View source #L1444-L1680.

Complete symbol map

This map also includes private definitions and nested helpers. The signature column exposes every explicit input even when an internal helper has no dedicated parameter prose.

Symbol Kind Signature and inputs What it does Source
ResourceMetrics class ResourceMetrics(process_memory_gb: float = 0.0, mlx_cache_gb: float = 0.0, mlx_peak_memory_gb: float = 0.0, system_memory_used_gb: float = 0.0, system_memory_total_gb: float = 0.0) Resource usage metrics during benchmark. #L72-L80
reset_mlx_peak_memory function reset_mlx_peak_memory() -> not annotated Reset MLX peak memory counter. #L83-L95
get_mlx_memory_info function get_mlx_memory_info(reset_peak: bool = True) -> dict Get MLX memory usage information. #L98-L135
get_process_memory function get_process_memory() -> float Get current process memory usage in GB. #L138-L147
get_system_memory function get_system_memory() -> tuple[float, float] Get system memory (used, total) in GB. #L150-L159
ResourceMonitor class ResourceMonitor() Monitor system resources during benchmark runs. #L162-L213
ResourceMonitor.__init__ method ResourceMonitor.__init__() -> not annotated Method ResourceMonitor.__init__ updates self.samples, self._start_time, self._start_memory. #L165-L168
ResourceMonitor.start method ResourceMonitor.start() -> not annotated Start monitoring. #L170-L176
ResourceMonitor.sample method ResourceMonitor.sample() -> ResourceMetrics Take a resource sample. #L178-L192
ResourceMonitor.get_summary method ResourceMonitor.get_summary() -> ResourceMetrics Get summary of all samples. #L194-L213
BenchmarkResult class BenchmarkResult(prompt: str, prompt_tokens: int, generated_tokens: int, ttft: float, total_time: float, tpot: float = 0.0, generation_tps: float = 0.0, processing_tps: float = 0.0) Results from a single benchmark run. #L235-L268
BenchmarkResult.__post_init__ method BenchmarkResult.__post_init__() -> not annotated Method BenchmarkResult.__post_init__ updates self.tpot, self.generation_tps, self.processing_tps. #L251-L268
BenchmarkSummary class BenchmarkSummary(model_name: str, num_runs: int, total_prompt_tokens: int, total_generated_tokens: int, total_time: float, ttft_mean: float, ttft_min: float, ttft_max: float, ttft_p50: float, ttft_p95: float, tpot_mean: float, tpot_min: float, tpot_max: float, generation_tps_mean: float, generation_tps_max: float, processing_tps_mean: float, latency_mean: float, latency_min: float, latency_max: float, latency_p50: float, latency_p95: float, total_throughput_tps: float, requests_per_second: float, hardware_chip: str = '', hardware_memory_gb: float = 0.0, hardware_bandwidth_gbs: float = 0.0, resources: ResourceMetrics = field(default_factory=ResourceMetrics)) Summary statistics across all benchmark runs. #L272-L315
calculate_percentile function calculate_percentile(data: list, percentile: float) -> float Calculate percentile from a list. #L318-L325
benchmark_single_prompt function benchmark_single_prompt(model, tokenizer, prompt: str, max_tokens: int = 256, temperature: float = 0.7) -> Optional[BenchmarkResult] Benchmark a single prompt with detailed timing. #L328-L394
run_benchmark function run_benchmark(model_name: str, num_prompts: int = 5, max_tokens: int = 256, temperature: float = 0.7, warmup_runs: int = 1) -> Optional[BenchmarkSummary] Run the full benchmark suite. #L397-L610
is_mllm_model function is_mllm_model(model_name: str) -> bool Check if model name indicates a multimodal language model. #L651-L657
MLLMBenchmarkResult class MLLMBenchmarkResult(resolution: str, width: int, height: int, pixels: int, time_seconds: float, tokens_generated: int, tokens_per_second: float, response_preview: str, memory_gb: float = 0.0, mlx_memory_gb: float = 0.0) Result from a single MLLM benchmark run. #L661-L674
download_test_image function download_test_image(url: str, timeout: int = 30) -> Image.Image Download image from URL and return PIL Image. #L677-L684
resize_image function resize_image(img: Image.Image, width: int, height: int) -> Image.Image Resize image to specified dimensions. #L687-L689
image_to_base64 function image_to_base64(img: Image.Image, format: str = 'JPEG') -> str Convert PIL Image to base64 data URL. #L692-L705
benchmark_mllm_resolution function benchmark_mllm_resolution(model, processor, config, base_image: Image.Image, width: int, height: int, max_tokens: int = 256, warmup: bool = False) -> MLLMBenchmarkResult Run MLLM benchmark for a specific resolution. #L708-L800
run_mllm_benchmark function run_mllm_benchmark(model_name: str, quick: bool = False, max_tokens: int = 256, warmup_runs: int = 1) -> list[MLLMBenchmarkResult] Run MLLM benchmark across multiple image resolutions. #L803-L913
print_mllm_summary function print_mllm_summary(results: list[MLLMBenchmarkResult], model_name: str) -> not annotated Print MLLM benchmark summary. #L916-L975
VideoBenchmarkResult class VideoBenchmarkResult(config_name: str, fps: float, max_frames: int, frames_extracted: int, video_duration: float, time_seconds: float, prompt_tokens: int, completion_tokens: int, tokens_per_second: float, response_preview: str, memory_gb: float = 0.0, mlx_memory_gb: float = 0.0) Result from a single video benchmark run. #L984-L999
create_test_video function create_test_video(duration: float = 10.0, fps: float = 30.0, width: int = 640, height: int = 480) -> str Create a synthetic test video with colored frames and text. #L1002-L1056
download_video function download_video(url: str, timeout: int = 120) -> str Download video from URL and return local path. #L1059-L1075
get_video_info function get_video_info(video_path: str) -> dict Get information about a video file. #L1078-L1094
benchmark_video_config function benchmark_video_config(model, video_path: str, fps: float, max_frames: int, config_name: str, video_info: dict, max_tokens: int = 150, warmup: bool = False) -> VideoBenchmarkResult Run a single video benchmark configuration. #L1097-L1162
run_video_benchmark function run_video_benchmark(model_name: str, video_url: str = None, video_path: str = None, quick: bool = False, max_tokens: int = 150, warmup_runs: int = 1) -> list[VideoBenchmarkResult] Run video benchmark across multiple frame configurations. #L1165-L1285
print_video_summary function print_video_summary(results: list[VideoBenchmarkResult], model_name: str) -> not annotated Print video benchmark summary. #L1288-L1341
print_summary function print_summary(summary: BenchmarkSummary) -> not annotated Print a formatted summary of benchmark results using tabulate. #L1349-L1441
main function main() -> not annotated Run the benchmark. #L1444-L1680