abatilo

DFlash2 executable seed ideas

These ideas target the immutable SPECULATIVE_DECODING_V1 Truth workload on Qwen/Qwen3-8B. A DFlash2 artifact uses an eight-token verify block, temperature 0, concurrency 16, the published two-tap group-16 dynamic causal convolutions around every attention and MLP sublayer, and the published rank-256 top-16 adjacent-pair candidate selector. The authorized numeric comparator is Evaluation 3HnQfcMqXmLCFdIzsU0mkXnXNo0, whose aggregate mean acceptance length is 3.826854272156405.

The architecture and inference contract come from z-lab/dflash commit 07ebd93db9f472af339b644bb70221ad8428328a, especially dflash/model.py, and SGLang serving commit 6228c78dd28368babedaa24d264266912c934b32. Upstream publishes no DFlash2 training loop or selector loss. The DFlash2 control therefore adapts the pinned TorchSpec DFlash trainer without claiming training-recipe reproduction. D-PACE comes from arXiv 2605.18810v1; its detached position weights apply only to the full-vocabulary un

abatilo

Evaluator-compatible Amestris seed ideas

These ideas are bounded by the immutable SPECULATIVE_DECODING_V1 Truth profile: Qwen3DSparkModel, DSPARK block size 7, temperature 0, concurrency 16, and the exact 16-workload aggregate. The authorized comparator is Evaluation 3HnQfcMqXmLCFdIzsU0mkXnXNo0, whose aggregate mean acceptance length is 3.826854272156405.

Evaluate fixed-profile position-adaptive acceptance loss

• Kind: LOSS
• Rationale: Apply position-adaptive acceptance credit; its effect on fixed-profile aggregate mean acceptance length is unknown.
• Falsifier: Reject when fixed-profile aggregate mean acceptance length is less than or equal to 3.826854272156405.
• Estimated H100 seconds: 158400

abatilo

Polyglot Coding Prompts for genai-bench

100 real-world coding prompts extracted from Aider's polyglot-benchmark (originally sourced from Exercism). Each prompt includes problem instructions and a starter code stub.

Prompt lengths range from ~66 to ~3000 tokens, covering problems like cipher implementations, data structures, parsers, and reactive systems.

Quick Start

# Download the prompts

abatilo

#!/usr/bin/env python3
"""
Hugging Face Transformers Distributed Training Example for Spin

Trains a GPT-2 model on WikiText-2 using distributed data parallel.
Designed to run on multiple nodes via Spin's SyncSet orchestration.
"""

import os

abatilo

#!/usr/bin/env python3
"""
Tiny Shakespeare DDP Training Example for Spin

Minimal causal-masked Transformer trained on Tiny Shakespeare using PyTorch DDP.
Uses character-level tokenization (~65 unique characters) for simplicity.
Designed to run on multiple nodes via Spin's SyncSet orchestration.

Usage with spinctl:
    # Push this script to your SyncSet pods

abatilo

How to use channels across different processes in Go

A couple of code samples to show how a named pipe can be used to extend Go's channel paradigm for use between different processes running on a system.

• interprocess1.go details a single byte channel.
• interprocess2.go details a channel that passes slices of bytes.

Note that opening a write channel will return two channels -

abatilo

FROM ubuntu:22.04

# Install sshd and xz to unzip the s6-overlay tarball
RUN <<EOF
apt-get update;
apt-get install -yq openssh-server xz-utils;
EOF

# Install s6-overlay
ADD https://github.com/just-containers/s6-overlay/releases/download/v3.1.6.2/s6-overlay-noarch.tar.xz /tmp

abatilo

export ZSH="$HOME/.oh-my-zsh"
ZSH_THEME="pygmalion"

plugins=(
  mise
  fzf
  git
  kubectl
)

abatilo

package gogptindex

import (
	"context"
	_ "embed"
	"encoding/json"
	"fmt"
	"net/http"
	"os"
	"os/signal"

abatilo

apiVersion: apps.kruise.io/v1alpha1
kind: ImagePullJob
metadata:
  name: sudokurace
  namespace: kruise-system
spec:
  image: '911907402684.dkr.ecr.us-west-2.amazonaws.com/sudokurace:'  # Tag set by kustomization.yaml
  pullSecrets:
  # Must match https://gist.github.com/abatilo/6b287265d541d06da567893c1522999f#file-imagepulljob-tf-L3
  - kruise-ecr-token