philipturner

import Metal

// MARK: - Usage

// Usage:
//
// 1) Install Xcode from the Mac App Store
//
// 2) From the command line:

philipturner

 20:47:34 ▶ metal ▶ 12⎘ ▶ $ ▶ ./bin/main -m ../models/7B/ggml-model-q4_0.bin --export
main: build = 652 (b252acb)
main: seed  = 1685900854
llama.cpp: loading model from ../models/7B/ggml-model-q4_0.bin
llama_model_load_internal: format     = ggjt v3 (latest)
llama_model_load_internal: n_vocab    = 32000
llama_model_load_internal: n_ctx      = 512
llama_model_load_internal: n_embd     = 4096
llama_model_load_internal: n_mult     = 256
llama_model_load_internal: n_head     = 32

philipturner

//
//  main.swift
//  CalculateLLaMA
//
//  Created by Philip Turner on 6/8/23.
//

import Foundation
import QuartzCore

philipturner

//
//  ContentView.swift
//  SIMDFuturesA15
//
//  Created by Philip Turner on 6/9/23.
//

import SwiftUI
import Foundation
import QuartzCore

philipturner

//
//  main.swift
//  TestMFAGEMM
//
//  Created by Philip Turner on 6/21/23.
//

import Accelerate
import Metal

philipturner

Simulation results, ranked from highest to lowest latency. Generated using https://gist.github.com/philipturner/d408351d68b5b1701bb651d4542e26e6

Raw data is private, but there's an older, publicly available substitute at https://gist.github.com/philipturner/94e7c5094915f23438440d49da823c9d

The statistics for MFA Winograd are speculative and eventual performance may be less. For example, Winograd may not be finished this summer, and/or upon publication of the code.

---

System:

• 32-core Apple 7 GPU, 1.296 GHz

philipturner

REMARK   5
REMARK   5 Created 2023/06/04 01:23:03 AM UTC
REMARK   5 with NanoEngineer-1 version 1.1.1.14 nanoengineer-1.com
REMARK   5
REMARK   6
REMARK   6 This file generally complies with the PDB format version 2.3
REMARK   6 Notes:
REMARK   6
REMARK   6 - Sets of atoms are treated as PDB chains terminated with TER
REMARK   6   records. Since the number of atom sets can exceed the 95 possible

philipturner

TODO: Put this into the GitHub repository instead, along with data from the Google Sheet.

Attention Heads Scaling with Model Size

Model	size	$n_{heads}$	$d_{head}$	size $/ n_{heads}^2$	size $/ n_{heads}^4$
GPT-3 Small	125M	12	64	0.9M
GPT-3 Medium	350M	16	64	1.4M
GPT-3 Large	760M	16	96	3.0M
GPT-3 XL	1.3B	24	128	2.3M

philipturner

__attribute__((__always_inline__))
inline static double _kmeans1d_cost(double* cumsum, double* cumsum2, int i, int j)
{
  if (j < i)
    return 0;
  double mu = (cumsum[j + 1] - cumsum[i]) / (j - i + 1);
  double result = cumsum2[j + 1] - cumsum2[i];
  result += (j - i + 1) * (mu * mu);
  result -= (2 * mu) * (cumsum[j + 1] - cumsum[i]);
  return result;

philipturner

//
//  main.swift
//  KMeans
//
//  Created by Philip Turner on 6/29/23.
//

import Metal

example_smawk()