feat(profiling): GPU kernel profiler and memory analyzer

[m96-chan]

Overview

Built-in profiling tools for kernel performance and memory analysis.

Features

Kernel Profiler

import pygpukit as gk
from pygpukit.profiling import Profiler

with Profiler() as p:
    result = gk.ops.matmul(a, b)
    gk.ops.softmax(result)

p.summary()
# Kernel          | Time (ms) | TFLOPS | Memory (GB/s)
# gemm_fp8_bf16   | 0.42      | 312.5  | 1250.0
# softmax         | 0.08      | -      | 890.0

p.export('profile.json')  # Chrome trace format

Memory Analyzer

from pygpukit.profiling import MemoryProfiler

with MemoryProfiler() as mp:
    model = QwenModel.from_safetensors(path)
    output = model.generate(prompt)

mp.peak_memory()      # 12.5 GB
mp.current_memory()   # 8.2 GB
mp.allocation_trace() # List of allocations
mp.fragmentation()    # 0.15 (15% fragmented)

Timeline View

from pygpukit.profiling import timeline

with timeline('inference.json'):
    for _ in range(100):
        model.forward(input)
# Open in chrome://tracing

Implementation

• CUDA Events timing wrapper
• cupti integration (optional)
• Memory tracking hooks
• Chrome trace export
• Summary statistics

[m96-chan]

Closed via PR #175