Hermann-SW

#!/usr/bin/env python3
""" multiple Instinct MI50 real time GPU% graphing """
import glob
import os
import time
from collections import deque
import matplotlib.pyplot as plt
from matplotlib import animation

def is_amd_vega20(ddir):

Hermann-SW

≥239 tokens/s prompt processing
57.76 tokens/s token generation

hermann@7600x:~/llama.cpp$ ./bench5
ggml_cuda_init: found 5 ROCm devices (Total VRAM: 98224 MiB):
  Device 0: AMD Radeon Graphics, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64, VRAM: 32752 MiB
  Device 1: AMD Instinct MI50/MI60, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64, VRAM: 16368 MiB
  Device 2: AMD Instinct MI50/MI60, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64, VRAM: 16368 MiB
 Device 3: AMD Instinct MI50/MI60, gfx906:sramecc+:xnack- (0x906), VMM: no, Wave Size: 64, VRAM: 16368 MiB

Hermann-SW

prompt:

this code works, please explain:

Q=get_tqf(n1);
if(type(Q)!="t_MAT",return([]));
assert(qfeval(Q,x)==n1);
assert(matdet(Q)==1);

G=qflllgram(Q,flag=1);

Hermann-SW

Your intuition on the second point is spot-on: if $n$ is a product of many medium primes $3 < p_i < 2^{32}$, dropping only the single largest prime factor $q_r$ can leave the remaining product $\prod_{j=1}^{r-1} q_j$ as large as $2^{59}$, which completely breaks any naive fencepost mask inside 64 bits.

What rescues the 64-bit hint is a see-saw trade-off between the bit-length of the dropped largest prime ($\ge 64/r$ bits) and the number of LSB/MSB bits stripped from the remaining $r - 1$ odd primes ($r - 1$ bits). Below is the complete mathematical proof for Scheme B (uint64_t + uint16_t), the exact proof and construction of the 64-bit hint H given $n$, and the $O(1)$ incremental pushprime(p) / inc_last_prime_exponent(p) state machine.

---

1. Deep Dive & Proof of Bounds for Scheme B (uint64_t W + uint16_t shape_id)

In Scheme B, the 64-bit word W chains the odd prime factors $3 \le q_1 \le q_2 \le \dots \le q_r$ (sorted with multiplicity) by overlapping the MSB 1 of $q_{j-1}$ with th

Hermann-SW

#include <iostream>
#include <vector>
#include <string>
#include <stdexcept>
#include <chrono>
#include <gmpxx.h>
#include <gmp.h>

class FactorFoundException : public std::exception {
public:

Hermann-SW

sage(cmd) = eval(externstr(Str("sage -q -c \"", cmd, "\""))[1]);

ellintpoints(E) = {
  sage(Str("print([[int(P[0]), int(P[1])] for P in EllipticCurve(",
       E[1..5],").integral_points(both_signs=True)])"));
}

as=eval(getenv("as"));
E=ellinit(as); 
print("   E[1..5]=",E[1..5]);

Hermann-SW

sage(cmd) = eval(externstr(Str("sage -q -c \"", cmd, "\""))[1]);

ellintpoints(E) = {
  sage(Str("print([[int(P[0]), int(P[1])] for P in EllipticCurve(",
       E[1..5],").integral_points()])"));
}

xy=eval(getenv("xy"));
[y1,x1]=xy;
[x2,y2]=xy;x2*=-1;

Hermann-SW

#include <iostream>
#include <cstdint>
#include <gmpxx.h>
#include <cassert>
#include <omp.h>

// Single-round Miller-Rabin test using GMP modular exponentiation
inline bool miller_rabin_test(uint32_t n, uint32_t a, uint32_t d, int s) {
    mpz_class base(a), exp(d), mod(n), x;
    

Hermann-SW

This Python script became obsolete with this new repo:  
https://github.com/Hermann-SW/graphviz_latex_pipeline
    
Because that repo's texdot.py compiles the LaTeX in each node or edge separately, with the real LaTeX.

You can still access this gist's Python code by opening the file at earlier revisions.

Hermann-SW

/*
   f=gfx900_multi_precision_bench
   hipcc -O3 --offload-arch=gfx900 $f.cpp -o $f
*/
#include <hip/hip_runtime.h>
#include <hip/hip_fp16.h>
#include <iostream>
#include <iomanip>
#include <vector>