ita9naiwa

import torch
import triton
import triton.language as tl

@triton.jit
def mxfp_matmul(
    a_ptr, b_ptr, output_ptr,
    a_scale, b_scale,
    M, N, K,
    stride_scale: tl.constexpr,

ita9naiwa

#!/usr/bin/env python3

import torch
import triton
import triton.language as tl

@triton.jit
def scaled_dot_kernel(
    # Pointers to matrices
    a_ptr, b_ptr, output_ptr,

ita9naiwa

// simple.cu - Direct PTX Tensor Core GEMM using mma.m16n8k16
// Demonstrates: cp.async, ldmatrix, mma instruction, 8-row interleaved output
// Target: Ampere+ GPUs (sm_80+)

#include <cuda_fp16.h>
#include <cstdio>
#include <cmath>

// ============================================================================
// CPU Reference Implementation

ita9naiwa

in_len=1024; out_len=1024; batch_size=128;
model=meta-llama/Llama-3.1-8B-Instruct

VLLM_USE_V1=1 VLLM_DISABLE_COMPILE_CACHE=1 TRITON_PRINT_AUTOTUNING=1 python3 vllm_benchmark.py --input-len $in_len --output-len $out_len --model $model --dtype float16 --batch-size $batch_size --num_iters_warmup 5 --num_iters 5