القائمة
يُحدَّد Prompt الذي يراجع المصدر أولًا بشكل افتراضي. يمكنك التبديل إلى أمر مباشر أو تنزيل نسخة محلية.
اقرأ SKILL.md وأي ملفات مرافقة يعرضها SkillsMP قبل أن تقرر التثبيت.
استنادا إلى تصنيف SOC المهني
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
يتجاوز الأمر المباشر Prompt المخصّص للمراجعة. افحص المصدر قبل تشغيله.
npx skills add https://github.com/mindspore-ai/akg --skill verifier-agentيبقى الأمر في سطر واحد. مرّر أفقيًا لمراجعته كاملًا قبل النسخ.
تفضّل نسخة محلية؟ نزّل الملفات المتاحة حاليًا لدى SkillsMP.
عرض SKILL.md
矩阵乘法矩阵乘法 A[M, K] @ B[K, N] = C[M, N]中，大K维度矩阵乘法(K>>M,N)优化：针对M/N较小但K极大(如M=N=256,K=131072)的场景，Split-K切分K维度并行化、Workspace+Reduce替代全局同步，实现显著性能提升
Triton Ascend hard API restrictions and forbidden syntax. MUST-follow rules that apply to every kernel: forbidden control flow (return/break/continue/lambda/while), tensor slice/index restrictions, scalar conversion rules, BLOCK_SIZE upper bound. Violating any of these produces a compile or runtime error on Ascend.
Triton Ascend 性能优化通用策略: BLOCK_SIZE 选择 (1024-2048 for elementwise, must be <65536), grid configuration (use VEC_CORE_NUM / CUBE_CORE_NUM, 2D/3D grid for matmul / conv / reduce, 1D grid + inner loop for elementwise / pointwise), 256B alignment for memory transfers, autotune block-size patterns, fp16 / fp32 precision conversion. Bind via keywords like matmul, elementwise, reduce, block_size, grid, autotune, alignment, fp16, fp32, tile, interleaved-loop, cube-core, vec-core.
| name | verifier-agent |
| description | 验证Agent，负责测试代码正确性和性能profiling |
| category | agent |
| version | 1.5.0 |
| license | MIT |
Verifier Agent负责全方位验证生成的代码，确保：
def test_accuracy(kernel_output, reference_output, rtol=1e-5, atol=1e-8):
"""测试数值精度"""
return np.allclose(kernel_output, reference_output, rtol=rtol, atol=atol)
# GPU计时
start_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()
kernel_function(*args)
end_event.record()
torch.cuda.synchronize()
elapsed_time_ms = start_event.elapsed_time(end_event)
# GFLOPS计算
flops = 2 * M * N * K # MatMul的FLOP数
gflops = (flops / elapsed_time_ms) / 1e6
# 理论带宽 vs 实际带宽
bytes_transferred = (M*K + K*N + M*N) * 4 # float32
bandwidth_gbps = (bytes_transferred / elapsed_time_ms) / 1e6
efficiency = bandwidth_gbps / theoretical_bandwidth
输入: 生成的代码 + 测试用例
↓
步骤1: 编译代码
↓
步骤2: 功能测试（正确性）
├─ 通过 → 步骤3
└─ 失败 → 报告错误，返回Coder
↓
步骤3: 性能测试（Profiling）
↓
步骤4: 资源分析
↓
步骤5: 生成报告
↓
输出: 验证报告 + 性能指标
def generate_test_cases(op_type, input_shapes):
"""生成测试用例"""
test_cases = []
# 1. 正常情况
test_cases.append(generate_normal_case(input_shapes))
# 2. 边界情况
test_cases.extend([
generate_zero_case(input_shapes),
generate_large_case(input_shapes),
generate_small_case(input_shapes),
])
# 3. 特殊情况
test_cases.extend([
generate_nan_case(input_shapes),
generate_inf_case(input_shapes),
])
# 4. 随机情况
for _ in range(10):
test_cases.append(generate_random_case(input_shapes))
return test_cases
def compare_with_baseline(custom_kernel, baseline_impl, inputs):
"""与baseline对比"""
# 测试正确性
custom_output = custom_kernel(*inputs)
baseline_output = baseline_impl(*inputs)
accuracy = test_accuracy(custom_output, baseline_output)
# 测试性能
custom_time = benchmark(custom_kernel, inputs)
baseline_time = benchmark(baseline_impl, inputs)
speedup = baseline_time / custom_time
return {
'accuracy': accuracy,
'custom_time': custom_time,
'baseline_time': baseline_time,
'speedup': speedup
}
验证报告:
算子: matmul
版本: 1.0.0
时间: 2026-01-24 10:30:00
正确性:
状态: PASS
测试用例: 15
通过: 15
失败: 0
精度: 1.2e-6 (相对误差)
性能:
平均延迟: 2.35 ms
标准差: 0.12 ms
吞吐量: 4.2 TFLOPS
带宽利用率: 78%
对比Baseline:
PyTorch: 3.1 ms (1.32x speedup)
cuBLAS: 2.8 ms (1.19x speedup)
资源:
Register/thread: 48
Shared memory/block: 48 KB
Occupancy: 75%
建议:
- ✅ 功能正确，可以部署
-
| 症状 | 可能原因 | 解决方案 |
|---|---|---|
| 数值误差大 | 累加顺序不同 | 使用Kahan求和 |
| 偶发错误 | Race condition | 添加同步 |
| 边界错误 | 索引越界 | 添加边界检查 |
| NaN输出 | 除零/log(0) | 添加epsilon |
| 症状 | 可能原因 | 解决方案 |
|---|---|---|
| 慢于baseline | 内存访问未合并 | 调整访问模式 |
| 低occupancy | Register压力大 | 减少局部变量 |
| 低带宽利用 | 计算访存比低 | 增加计算密度 |
| Bank conflict | Shared memory布局 | Padding避免冲突 |
ncu --set full -o profile kernel_program
nsys profile --stats=true python script.py
rocprof --stats python script.py
with torch.profiler.profile(
activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
record_shapes=True,
profile_memory=True,
with_stack=True
) as prof:
kernel(*inputs)
print(prof.key_averages().table(sort_by="cuda_time_total"))
verifier_agent:
mode: standard # fast, standard, strict
accuracy:
rtol: 1e-5
atol: 1e-8
test_nan: true
test_inf: true
performance:
warmup_runs: 3
measure_runs: 10
use_percentile: 50 # 中位数
profiling:
enable: true
tool: nsight-compute
metrics:
- sm_efficiency
- memory_throughput
- achieved_occupancy
from skill_system import SkillLoader, SkillRegistry
# 加载verifier-agent skill
loader = SkillLoader()
skills = loader.load_from_directory(Path("./skills"))
registry = SkillRegistry()
registry.register_batch(skills)
verifier_skill = registry.get("verifier-agent")
# 使用skill内容指导验证
def verify_kernel(kernel, inputs, expected_outputs):
"""根据skill指导进行验证"""
# 正确性测试
outputs = kernel(*inputs)
accuracy_pass = all(
np.allclose(out, exp, rtol=1e-5, atol=1e-8)
for out, exp in zip(outputs, expected_outputs)
)
# 性能测试
times = []
for _ in range(10):
start = time.time()
kernel(*inputs)
torch.cuda.synchronize()
times.append(time.time() - start)
avg_time = np.mean(times)
std_time = np.std(times)
return {
'accuracy': accuracy_pass,
'avg_time': avg_time,
'std_time': std_time
}