avifenesh

# Phase 4a — FP16 storage / FP32 accumulate GPU kernels for Qwen.
#
# Mix: existing qwen_gpu_primitives tile/block_reduce shape (proven 15/15
# FP32 pass) + handover FP16-V1 contract + EXL2 split matvec/matmul (decode
# vs prefill) + selective fused projections (Phase 2 arena groups gate/up,
# q/k/v contiguous so kernels see one weight pointer per fused group).
#
# Storage = DType.float16. Accumulate = Float32 (numerical stability through
# K reduction on attention/FFN dims of 4096-12288).

avifenesh

llama.cpp MoE CPU-offload prompt benchmark harness

This is the small llama-server harness I used to compare MoE CPU-offload behavior across baseline and patched llama.cpp builds.

The benchmark is intentionally narrow: it measures sequential chat-completion latency and token throughput while exercising the CPU-MoE host-to-device expert staging path. It is not a model-quality eval, and the prompt strings are synthetic systems-engineering workload data used only to reproduce timing.

avifenesh

{
    "mcpServers": {
        "filesystem": {
            "command": "npx",
            "args": [
                "-y",
                "@modelcontextprotocol/server-filesystem",
                "/home/ubuntu"
            ],
            "disabled": false,

avifenesh

Valkey GLIDE AI Coding Assistant Instructions

Project Overview

Multi-language Valkey/Redis client with Rust core (glide-core) and language wrappers. Uses protobuf for cross-language communication and FFI/UDS for performance.

Architecture

Core Components

• glide-core/: Rust core (protocol, connections, business logic)

avifenesh

#!/bin/bash

#remove comment to print the lines running.
#set -x

set -e

valkey_version=
should_clone_repo=false
valkey_clone_git=https://github.com/valkey-io/valkey.git