用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/Aradotso/trending-skills --skill flash-moe-inference命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
AI agent skill for using deepsec, the agent-powered security vulnerability scanner for large codebases
Open-source Luau/Lua script IDE and executor built as a Windows Forms C# application with Monaco editor integration
Use Claude Code's autonomous agent loop with DeepSeek V4 Pro, OpenRouter, or any Anthropic-compatible backend at up to 17x lower cost.
基于 SOC 职业分类
正在显示 SKILL.md
| name | flash-moe-inference |
| description | Run 397B parameter Mixture-of-Experts LLMs on a MacBook using pure C/Metal with SSD streaming |
| triggers | ["run a large language model on my laptop","stream expert weights from SSD","flash moe inference engine","run Qwen3.5 397B on Mac","mixture of experts on Apple Silicon","metal inference engine for large models","quantized MoE inference macOS","run 209GB model on MacBook"] |
Skill by ara.so — Daily 2026 Skills collection.
Flash-MoE is a pure C/Objective-C/Metal inference engine that runs Qwen3.5-397B-A17B (397B parameter Mixture-of-Experts) on a MacBook Pro with 48GB RAM at 4.4+ tokens/second. It streams 209GB of expert weights from NVMe SSD on demand — no Python, no ML frameworks, just C, Objective-C, and hand-tuned Metal shaders.
# Clone the repo
git clone https://github.com/danveloper/flash-moe
cd flash-moe/metal_infer
# Build everything
make
# Verify build artifacts
ls infer chat main
The Makefile compiles infer.m, chat.m, main.m with Metal shader compilation for shaders.metal.
# From the metal_infer/ directory
# Point to your downloaded Qwen3.5-397B safetensors directory
python3 extract_weights.py /path/to/Qwen3.5-397B-A17B-Instruct/
# Produces:
# model_weights.bin (~5.5GB, mmap'd at runtime)
# model_weights.json (tensor manifest)
# vocab.bin (vocabulary)
# tokenizer.bin (BPE tokenizer data)
# From repo root
python3 repack_experts.py /path/to/Qwen3.5-397B-A17B-Instruct/ metal_infer/packed_experts/
# Produces packed_experts/ directory (~209GB)
# Each expert is a separate file: layer_XX_expert_YYYY.bin
# Convert 4-bit experts to 2-bit (saves ~89GB, 120GB total)
python3 metal_infer/repack_experts_2bit.py \
metal_infer/packed_experts/ \
metal_infer/packed_experts_2bit/
cd metal_infer
# 4-bit inference (production quality, tool calling works)
./infer --prompt "Explain quantum computing" --tokens 100
# 2-bit inference (faster, breaks JSON/tool calling)
./infer --prompt "Explain quantum computing" --tokens 100 --2bit
# Per-layer timing breakdown
./infer --prompt "Hello" --tokens 20 --timing
./chat
# Opens TUI with full tool calling support
# Uses 4-bit experts by default
./main
# Runs pure expert forward-pass benchmark
# Reports tokens/sec without attention overhead
flash-moe/
├── paper/
│ └── flash_moe.pdf # Full technical paper
├── metal_infer/
│ ├── infer.m # Complete inference engine (~7000 lines)
│ ├── shaders.metal # Metal compute kernels (~1200 lines)
│ ├── chat.m # Interactive chat TUI
│ ├── tokenizer.h # Single-header C BPE tokenizer (449 lines)
│ ├── main.m # MoE-only benchmark
│ ├── Makefile
│ ├── extract_weights.py # Safetensors → model_weights.bin
│ ├── repack_experts_2bit.py # 4-bit → 2-bit requantization
│ ├── train_predictor.py # Expert routing prediction analysis
│ ├── model_weights.bin # Non-expert weights (mmap'd)
│ ├── model_weights.json # Tensor manifest
│ ├── vocab.bin
│ ├── tokenizer.bin
│ ├── packed_experts/ # 4-bit expert files (209GB)
│ └── packed_experts_2bit/ # 2-bit expert files (120GB, optional)
├── repack_experts.py # 4-bit expert packing from safetensors
├── progress.py # Results visualization
└── results.tsv # Experiment log
The model has 60 transformer layers:
CMD3(prev) → CMD1: attention projections + delta-net [1.22ms GPU]
→ CPU: flush results [0.01ms CPU]
→ CMD2: o_proj + norm + routing + shared [0.55ms GPU]
→ CPU: softmax + topK routing [0.003ms]
→ I/O: parallel pread K=4 experts [2.41ms SSD]
→ CMD3: expert forward + combine + norm [0.04ms encode, DEFERRED]
The shaders.metal file contains hand-written kernels. Key kernels:
// 4-bit dequantized matrix-vector multiply (FMA-optimized)
// Key insight: fma(nibble, scale*x, bias*x) instead of (nibble*scale + bias)*x
// Pre-compute scale*x and bias*x to fuse dequant+multiply in one FMA instruction
kernel void matvec_4bit_fma(
device const uint8_t* weights [[buffer(0)]],
device const float* scales [[buffer(1)]],
device const float* biases [[buffer(2)]],
device const float* x [[buffer(3)]],
device float* out [[buffer(4)]],
uint tid [[thread_position_in_threadgroup]],
uint gid [[threadgroup_position_in_grid]])
{
// ... tiled SIMD-reduced FMA kernel
// 12% faster than naive (nibble * scale + bias) * x
}
// Fused SwiGLU activation
kernel void swiglu(device float* gate [[buffer(0)]],
device const float* up [[buffer(1)]],
uint gid [[thread_position_in_grid]])
{
float g = gate[gid];
gate[gid] = (g / (1.0f + exp(-g))) * up[gid];
}
// RMS normalization (two-pass)
kernel void rms_norm_pass1(...) // sum of squares reduction
kernel void rms_norm_pass2(...) // apply normalization
// GPU RoPE (fused with Q deinterleave and K normalization)
kernel void rope_qk(...)
// MoE combine + residual + sigmoid gate (fused)
kernel void moe_combine_residual(...)
The core innovation — loading only K=4 active experts per layer from SSD:
// Parallel expert loading using GCD dispatch groups
// From infer.m (conceptual pattern)
dispatch_group_t group = dispatch_group_create();
dispatch_queue_t ioQueue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
for (int k = 0; k < K_EXPERTS; k++) {
int expert_id = top_k_indices[k];
dispatch_group_async(group, ioQueue, ^{
// Each expert: ~6.75MB at 4-bit
char path[256];
snprintf(path, sizeof(path),
"packed_experts/layer_%02d_expert_%04d.bin",
layer, expert_id);
int fd = open(path, O_RDONLY);
// pread() — non-blocking, OS page cache handles LRU
pread(fd, expert_buffer[k], expert_size, 0);
close(fd);
});
}
dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
// GPU compute follows — serial pipeline is hardware-optimal on Apple Silicon
Why pread() not mmap(): mmap incurs per-page fault overhead on cold data (~5x slower). Direct pread() with OS page cache achieves ~71% hit rate naturally.
The recurrence update uses Accelerate BLAS — 64% faster than scalar:
// GatedDeltaNet state update per head (conceptual pattern)
// state: 128×128 float matrix, 64 heads
// From infer.m
#import <Accelerate/Accelerate.h>
for (int h = 0; h < 64; h++) {
float* S = state + h * 128 * 128; // 128×128 state matrix
float* q = Q + h * 128;
float* k = K + h * 128;
float* v = V + h * 128;
// β·(k⊗v) outer product update
// cblas_sger: S += beta * (k ⊗ v)
cblas_sger(CblasRowMajor, 128, 128,
beta[h], k, 1, v, 1, S, 128);
// Decay: S = alpha * S
cblas_sscal(128 * 128, alpha[h], S, 1);
// Output: o = S @ q
cblas_sgemv(CblasRowMajor, CblasNoTrans,
128, 128, 1.0f, S, 128, q, 1, 0.0f,
output + h * 128, 1);
}
\name\ instead of "name")F_NOCACHE flag to avoid page cache thrashing| Approach | Why it fails |
|---|---|
mmap() expert files | Per-page fault overhead: 5x slower than pread() |
dispatch_io | dispatch_data management overhead: -70% |
F_RDADVISE prefetch | SSD DMA + GPU share memory controller — concurrent access: -73% GPU speed |
| Custom Metal LRU cache | GPU memory pressure: -38% vs OS page cache |
| LZ4 expert compression | Decompress overhead > warm cache savings: -13% |
| Temporal expert prediction | 25% hit rate, wastes SSD bandwidth: -18% |
| Speculative early routing | Cache pollution: -38% |
| MTP speculative decoding | MoE I/O scales per-token (unlike dense models): break-even |
| Spin-poll GPU wait | CPU thermal throttle competes with GPU: -23% |
| Parallel SSD + GPU overlap | Unified memory controller arbitration: net negative |
Key principle: On Apple Silicon, GPU DMA and SSD DMA share the same memory controller. The serial pipeline (GPU → SSD → GPU) is hardware-optimal.
# Ensure Xcode CLI tools are installed
xcode-select --install
# Check Metal compiler is available
xcrun -sdk macosx metal --version
The engine is designed to use ~6GB active:
model_weights.bin (mmap'd, read-only)If you see OOM, check for other processes consuming unified memory:
sudo memory_pressure
vm_stat
# Check SSD speed — needs ~17GB/s for target performance
# Run with timing to identify bottleneck
./infer --prompt "Hello" --tokens 5 --timing
# Verify packed_experts/ is on internal SSD, not external drive
diskutil info /
# Default paths expected by infer.m:
# metal_infer/packed_experts/ (4-bit)
# metal_infer/packed_experts_2bit/ (2-bit)
# Ensure you're running from metal_infer/ directory
cd metal_infer
./infer --prompt "test"
Use 4-bit, not 2-bit. The 2-bit quantization corrupts quote characters in JSON output, making tool calling unreliable. Always use the default 4-bit configuration for agentic workloads.
The engine explicitly manages all allocations:
model_weights.bin is mmap'd read-only — kernel manages pages