powderluv

powderluv

Visualization of the MobileNet-v3-large shapes (ordered similarly by decreasing time percentage, so the most important shapes come first).

powderluv

# Training data consist of X=(str, str), y=float:
from sklearn.model_selection import train_test_split
X = [
    ["Hello World!", "Good morning!"],
    ["It is raining", "It is cold"],
    ["Beautiful city beside mountain", "Quiet street in downtown area"],
    ["AI is the future", "AI is just a tool"],
    ["This application is great", "software is the problem"],
    ["Hello World!", "Good morning!"],
    ["It is raining", "It is cold"],

powderluv

# Stick this in C:\Users\<login>\OneDrive\Documents\PowerShell\
# Test-Path -Path $PROFILE
# New-Item -ItemType File -Path $PROFILE -Force
# code $PROFILE
$vswhere = "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe"
# Check if vswhere exists, if not you may need to adjust the path or install the VS installer
if (-not (Test-Path $vswhere)) {
    Write-Host "vswhere.exe not found at the expected location."
    # Provide a link to the official vswhere documentation if needed
}

powderluv

hipblaslt-bench --api_method c -m 768 -n 77 -k 768 --lda 768 --ldb 768 --ldc 768 --ldd 768  --stride_a 0 --stride_b 0 --stride_c 0 --stride_d 0  --alpha 1.000000 --beta 0.000000 --transA T --transB N --batch_count 1 --scaleA 0 --scaleB 0 --bias_vector --bias_source d --a_type f32_r --b_type f32_r --c_type f32_r --d_type f32_r --scale_type f32_r --bias_type f32_r   --compute_type f32_r --algo_method index --solution_index 1270 --activation_type none --any_stride --rotating 0 --cold_iters 0 --iters 0 

powderluv

Launch docker:
rocm/sgl-dev:v0.5.8.post1-rocm720-mi30x-20260211-preview


Inside container -
 
Install Sglang from source by following -
https://docs.sglang.io/platforms/amd_gpu.html#install-from-source
 
Install transformer from source -

powderluv

Split-K Matmul Transpiler Bug: Deep Investigation

Overview

The GFX1250→GFX942 cross-family ISA transpiler passes 18-19/20 tests. The remaining failure is the matmul_splitk test, where the inner loop exits prematurely and non-deterministically. This document details the exhaustive investigation.

The Bug

The split-K matmul uses two kernels:

• matmul_splitk_compute (236 GFX12 instructions): Each workgroup computes a partial matmul for a chunk of K. Uses blockIdx.y for split index.

powderluv

gfx1201 (RDNA4) lite:: GPU bring-up — macOS eGPU / Linux / Windows

Experimental effort to run ROCm + PyTorch on a gfx1201 GPU (AMD Radeon RX 9070 XT / AI PRO R9700, RDNA4, PCI 0x1002:0x7551) through a shared, firmware-light lite:: ROCr backend that programs the GPU's compute queue directly from userspace — on three OSes that lack the usual ROCm/KFD kernel path:

• macOS (Apple Silicon) — AMD eGPU over Thunderbolt, via a DriverKit DEXT (no kernel driver).
• Linux (x86) — amdgpu_lite minimal kernel shim + a userspace bring-up.
• Windows — userspace D3DKMTEscape backend over the production amdgpu_wddm KMD; compute shader dispatch working on gfx1201.