Skip to content

Instantly share code, notes, and snippets.

View stellaraccident's full-sized avatar

Stella Laurenzo stellaraccident

260 followers · 8 following
  • Amd
  • Seattle Washington, US
View GitHub Profile
@stellaraccident
stellaraccident / wheel-split-report.md
Last active March 10, 2026 02:08
PyTorch ROCm wheel split report — torch 2.10.0+rocm7.1 (v2: hierarchy-aware bundling, xnack collapse, METADATA extras/variants)

PyTorch ROCm Wheel Split Report

Split of torch-2.10.0+rocm7.1-cp313-cp313-manylinux_2_28_x86_64.whl using rocm_kpack.tools.split_python_wheels.

Summary

Metric Value
Input wheel (.whl) 5.1 GB
Host wheel (.whl) 431 MB
@stellaraccident
stellaraccident / therock-shared-python-for-nuitka.md
Last active January 29, 2026 21:15
TheRock shared libpython for Nuitka/rocprofiler-compute

TheRock Already Provides Shared libpython for Nuitka/Embedding

Summary

Yes, TheRock already builds Python with --enable-shared for embedding use cases. The infrastructure exists and is used by rocgdb. rocprofiler-compute just needs to wire into it.

Current Infrastructure

1. install_shared_pythons.sh builds Python with shared library support

@stellaraccident
stellaraccident / install-rocm-asan-artifacts.md
Created January 29, 2026 18:16
Installing ROCm from TheRock ASAN CI Artifacts
@stellaraccident
stellaraccident / mixtral_compile_test.mlir
Created January 28, 2026 08:16
Mixtral compilation failure: expand_shape dimension inference through moe_ffn_block
#map = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map1 = affine_map<(d0, d1) -> (d1, d0)>
#map2 = affine_map<(d0, d1) -> (d0, d1)>
#map3 = affine_map<(d0, d1) -> (d1)>
#map4 = affine_map<(d0, d1, d2) -> (d2, d0)>
#map5 = affine_map<(d0, d1, d2) -> (d1, d2)>
#map6 = affine_map<(d0, d1, d2) -> (d0, d2)>
#map7 = affine_map<(d0, d1, d2) -> (d2, d0, d1)>
#map8 = affine_map<(d0, d1, d2) -> (d1, d2, d0)>
#map9 = affine_map<(d0, d1, d2) -> (d2)>
@stellaraccident
stellaraccident / quartz-design.md
Created January 26, 2026 20:58
Quartz Design Document: ROCm CI/CD Dashboard & Orchestration

Quartz Design Document: ROCm CI/CD Dashboard & Orchestration

Executive Summary

This document provides architectural guidance for "Quartz" - a PyTorch HUD-like system for ROCm downstream CI/CD orchestration. The junior engineer's instinct to start with status.json is understandable but insufficient for the stated requirements. A database-first approach is correct.

Requirements Recap

@stellaraccident
stellaraccident / purrfect-crunching-glade.md
Created January 26, 2026 04:10
MoE components plan for frank-models (iree-link composition)

Plan: Add MoE Components (mul_mat_id, moe_ffn_block)

Overview

Extract MoE primitives from /develop/ai-no-fluff/kb/ben/moe_f32_parameterized.mlir:

  • mul_mat_id - Expert-selected matrix multiplication (gather + batch_matmul)
  • moe_ffn_block - Full MoE FFN block composing routing, expert compute, weighted sum

Key challenge: moe_ffn_block depends on mul_mat_id and swiglu. Need systematic composition without manual inlining.

@stellaraccident
stellaraccident / gist:4d3b5d24077b17ebfb8137bf3ad8135f
Last active January 22, 2026 01:18
Why merge commits are better than pushing hundreds of commits to main

Why Merge Commits Beat Pushing Hundreds of Commits to Main

The Comparison

Merge Commit Individual Commits on Main
One atomic integration point 500 commits sprawled on main
git revert -m1 <merge> undoes everything Good luck reverting
git bisect can skip the whole merge Bisect walks through 500 commits
Main history is readable Main history is chaos
@stellaraccident
stellaraccident / gist:f31fee06c22e6d172335f2ae881d7ff5
Last active January 8, 2026 04:54
TheRock: Developer builds from PRs and PyTorch packages

TheRock: Developer Builds from PRs and PyTorch Packages

Claude Code Prompt: Read the docs and workflows in therock using subagents and tell me how to make developer builds from a PR and get packages w/ pytorch.

Triggering PR Builds

Label-based (simplest)

Add labels to your PR:

@stellaraccident
stellaraccident / procedure.md
Last active January 12, 2022 21:36
IREE LLVM Integration Procedure

Notes on integrating LLVM from the OSS side

Strategy 1: Sync everything to a Google/TensorFlow commit

cd ~/src
git clone git clone https://github.com/tensorflow/tensorflow.git
git clone https://github.com/tensorflow/mlir-hlo.git
@stellaraccident
stellaraccident / 0_original.mlir
Created December 16, 2021 04:48
Const prop
module @aqt_matmul {
iree_input.global private @_params$0 = dense<[[0.000000e+00, 5.003000e+02, 1.000600e+03], [1500.8999, 2.001200e+03, 2.501500e+03], [3001.7998, 3502.09985, 4.002400e+03], [4502.69971, 5.003000e+03, 5.503300e+03], [6003.59961, 6503.8999, 7004.1997], [7.504500e+03, 8004.7998, 8.505100e+03]]> : tensor<6x3xf32>
iree_input.global private @_params$1 = dense<5.000000e+00> : tensor<f32>
func @compute_native(%arg0: tensor<5x6xf32>) -> tensor<5x3xf32> {
%0 = iree_input.global.load @_params$0 : tensor<6x3xf32>
%1 = iree_input.global.load @_params$1 : tensor<f32>
%2 = call @main(%0, %1, %arg0) : (tensor<6x3xf32>, tensor<f32>, tensor<5x6xf32>) -> tensor<5x3xf32>
return %2 : tensor<5x3xf32>
}
func private @main(%arg0: tensor<6x3xf32>, %arg1: tensor<f32>, %arg2: tensor<5x6xf32>) -> tensor<5x3xf32> {