File size: 5,532 Bytes
7d51e61 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 | # Benchmark System
## Overview
๋ฒค์น๋งํฌ ์์คํ
์ ์ปค์คํ
CUDA ์ปค๋์ forward/backward ์ฑ๋ฅ์ naive PyTorch ๊ตฌํ ๋๋น ์ธก์ ํ๋ค. Triton์ `do_bench`๋ฅผ ์ฌ์ฉํ๋ฉฐ, ์ ํ๋ ๊ฒ์ฆ(correctness check) ํ ์ฑ๋ฅ์ ์ธก์ ํ๋ค.
## Directory Structure
```
benchmarks/
โโโ run_cases.py # CLI ์ง์
์
โโโ common/
โ โโโ bench_framework.py # ๋ฒค์น๋งํฌ ์ ํธ๋ฆฌํฐ (Triton perf_report ๊ธฐ๋ฐ)
โ โโโ diff_engine.py # ์ ํ๋ ๊ฒ์ฆ ์์ง (DiffCase ABC)
โโโ cases/ # ๋ฒค์น๋งํฌ ์ผ์ด์ค ๊ตฌํ
โ โโโ rms.py # RMSNorm
โ โโโ add_rms.py # Fused Add + RMSNorm
โ โโโ poly.py # PolyNorm
โ โโโ mul_poly.py # Fused Mul + PolyNorm
โ โโโ grouped_mul_poly.py # Grouped MoE Fused Mul + PolyNorm
โโโ benchmark.yaml # Kubeflow ๋ฒค์น๋งํฌ job config
โโโ test.yaml # Kubeflow ํ
์คํธ job config
โโโ plots/ # ์์ฑ๋ ํ๋กฏ ๊ฒฐ๊ณผ
โโโ results/ # ํ์์คํฌํ๋ณ ๋ฒค์น๋งํฌ ๊ฒฐ๊ณผ
```
## Usage
```bash
python benchmarks/run_cases.py --case <CASE> [OPTIONS]
```
### Arguments
| Argument | Default | Choices | Description |
|----------|---------|---------|-------------|
| `--case` | (ํ์) | `rms`, `add_rms`, `poly`, `mul_poly`, `grouped_mul_poly` | ๋ฒค์น๋งํฌ ์ผ์ด์ค |
| `--dtype` | `bf16` | `fp16`, `bf16`, `fp32`, `all` | ๋ฐ์ดํฐ ํ์
|
| `--save-path` | `./configs/` | ๊ฒฝ๋ก | ๊ฒฐ๊ณผ ์ถ๋ ฅ ๋๋ ํ ๋ฆฌ |
| `--plot` | false | - | ํ๋กฏ ์์ฑ ๋ชจ๋ |
| `--profile` | false | - | Chrome trace ํ๋กํ์ผ๋ง ๋ด๋ณด๋ด๊ธฐ |
### Examples
```bash
# bf16 ๊ธฐ๋ณธ ๋ฒค์น๋งํฌ
python benchmarks/run_cases.py --case grouped_mul_poly
# ๋ชจ๋ dtype + ํ๋กํ์ผ๋ง
python benchmarks/run_cases.py --case mul_poly --dtype all --profile --save-path ./results
# ํ๋กฏ๋ง ์์ฑ
python benchmarks/run_cases.py --case rms --plot --save-path ./plots
```
## Benchmark Cases
๊ฐ ์ผ์ด์ค๋ `DiffCase` ABC๋ฅผ ๊ตฌํํ๋ฉฐ, naive(PyTorch ์ฐธ์กฐ)์ CUDA ์ปค๋์ ๋น๊ตํ๋ค.
| Case | Naive | CUDA | Inputs |
|------|-------|------|--------|
| `rms` | `torch.nn.RMSNorm` | `activation.layers.RMSNorm` | x, weight, eps |
| `add_rms` | custom `FusedAddRMSNorm` | `activation.layers.FusedAddRMSNorm` | x, residual, weight, eps |
| `poly` | custom `PolyNorm` (x^3, x^2, x ์กฐํฉ) | `activation.layers.PolyNorm` | x, weight(3), bias(1), eps |
| `mul_poly` | custom `FusedMulPolyNorm` | `activation.layers.FusedMulPolyNorm` | x, mul, weight(3), bias, eps |
| `grouped_mul_poly` | `fused_mul_grouped_poly_norm_ref` | `fused_mul_grouped_poly_norm` | x, mul, weight(num_experts, 3), bias, offsets |
`grouped_mul_poly`๋ ์ถ๊ฐ๋ก `compiled`(torch.compile๋ naive)์ `compiled_cuda`(torch.compile๋ CUDA) provider๋ ์ธก์ ํ๋ค.
## Execution Flow
1. **์ ํ๋ ๊ฒ์ฆ** - 3๊ฐ config์ ๋ํด `calculate_diff()` ์คํ
- `(bs=2, sl=128, hidden=4096)`
- `(bs=8, sl=4096, hidden=1280)`
- `(bs=1, sl=32768, hidden=1280)`
- forward/backward ๋ชจ๋ `atol=1e-2, rtol=1e-2`๋ก ๋น๊ต
2. **๋ฒค์น๋งํฌ ์คํ** - dtype๋ณ๋ก forward/backward ์ฑ๋ฅ ์ธก์
3. **๊ฒฐ๊ณผ ์ ์ฅ** - CSV ํ์ผ (๋ฐ ์ ํ์ ์ผ๋ก ํ๋กฏ/trace)
## Configuration Ranges
**Standard cases** (rms, add_rms, poly, mul_poly):
- Batch sizes: 1, 2, 4, 8
- Sequence lengths: 1024, 2048, 4096, 8192
- Hidden dims: 2048, 4096
**Grouped case** (grouped_mul_poly):
- Total tokens: 1024 ~ 65536 (bs x sl)
- Hidden dim: 1280 (๊ณ ์ )
- Experts: 48 per rank
`--plot` ๋ชจ๋์์๋ `bs=1`๋ก ๊ณ ์ ํ๊ณ seq_len๋ง sweepํ๋ค.
## Output
### CSV
`{save_path}/{case}/{dtype}/` ๋๋ ํ ๋ฆฌ์ ์ ์ฅ:
- `{case}-{dtype}-fwd-perf.csv` - forward ๊ฒฐ๊ณผ
- `{case}-{dtype}-bwd-perf.csv` - backward ๊ฒฐ๊ณผ
์ปฌ๋ผ: `dim`, `batch_size`, `seq_len`, `Naive (us)`, `Compiled (us)`, `Cuda (us)`, `SpeedUp (us)`
### Chrome Trace (`--profile`)
`{save_path}/{case}/{dtype}/traces/` ๋๋ ํ ๋ฆฌ์ JSON ํ์์ผ๋ก ์ ์ฅ. `chrome://tracing`์์ ๋ก๋ํ์ฌ GPU ํ์๋ผ์ธ์ ๋ถ์ํ ์ ์๋ค.
ํ์ผ๋ช
ํจํด: `trace_{fwd|bwd}_{naive|compiled|cuda|compiled_cuda}_N{total_tokens}.json`
### Plot (`--plot`)
Speedup ๋น๊ต ํ๋กฏ ์์ฑ. Geometric mean์ผ๋ก ์ ์ฒด speedup์ ์ง๊ณํ๋ค.
## Framework Internals
### bench_framework.py
Triton์ `perf_report`/`Benchmark`๋ฅผ ์ฌ์ฉํ๋ 4๊ฐ ํฉํ ๋ฆฌ ํจ์:
- `make_fwd_benchmark_for_case()` - forward ๋ฒค์น๋งํฌ (CSV)
- `make_bwd_benchmark_for_case()` - backward ๋ฒค์น๋งํฌ (CSV)
- `make_fwd_benchmark_plot_for_case()` - forward ํ๋กฏ
- `make_bwd_benchmark_plot_for_case()` - backward ํ๋กฏ
ํ์ด๋ฐ์ `triton.testing.do_bench()`๋ก ์ธก์ ํ๋ฉฐ, ms ๋จ์๋ฅผ us๋ก ๋ณํํ๋ค (`time_unit_scale=1000`).
### diff_engine.py
`DiffCase` ABC ์ธํฐํ์ด์ค:
- `build_inputs(bs, sl, dim)` - ์
๋ ฅ ํ
์ ์์ฑ
- `make_naive()` / `make_cuda()` - ๊ตฌํ์ฒด ์์ฑ
- `forward(module, inputs)` - forward ์คํ
- `grad_inputs(inputs)` - gradient ๋์ ํ
์ ๋ฐํ
`calculate_diff()`๊ฐ naive์ CUDA ์์ชฝ์ forward output + backward gradient๋ฅผ `torch.testing.assert_close()`๋ก ๋น๊ตํ๋ค.
## Kubeflow Integration
`benchmark.yaml`๋ก ํด๋ฌ์คํฐ์์ ๋ฒค์น๋งํฌ๋ฅผ ์คํํ ์ ์๋ค:
- triton, matplotlib, pandas ์ค์น
- C++ extension ๋น๋ (`setup.py`)
- GPU warmup (100 iterations matmul)
- ๊ฒฐ๊ณผ๋ฅผ `benchmarks/results/{YY_MM_DD_HH_MM}/`์ ์ ์ฅ
|