dhurandhar 0.1.0

Model-agnostic edge deployment analysis framework for memory-constrained devices

dhurandhar — धुरंधर

dhura (धुर, burden) + dhara (धर, one who bears)

"Bearer of burdens" — a framework for deploying large multimodal models on memory-constrained edge devices where they have no right to survive.

---

What it does

dhurandhar is a model-agnostic edge deployment analysis framework. Given a model architecture and a target device, it answers the questions that matter before you ship:

Module	Question answered
PLE Analysis	What is the peak live memory footprint at context length N?
Device Feasibility	Can this model run resident, mmap, or not at all on this device?
TurboQuant Sweep	What is the quality / memory tradeoff at 2/3/4/6/8-bit KV compression?
RotorQuant Comparison	TurboQuant vs RotorQuant — quality vs arithmetic cost?
Mmap Profiler	What is the real mmap throughput and peak RSS on this host?

All five analyses are exposed as a CLI, a Python API, and a 5-tab Gradio dashboard.

---

Supported models (built-in)

Slug	Architecture	Params
gemma4-e2b	Gemma4 (sliding-window + global attention)	2B
qwen2.5-0.5b	Qwen2.5 GQA	0.5B
qwen2.5-1.5b	Qwen2.5 GQA	1.5B
granite-3.3-2b	IBM Granite MHA	2B
llama-3.2-1b	Llama3 GQA	1B

Any model can be added via a simple YAML profile — no code required.

---

Install

# Core (analysis + CLI)
uv add dhurandhar

# With interactive dashboard
uv add "dhurandhar[dashboard]"

# With HuggingFace Hub auto-profile derivation
uv add "dhurandhar[hf]"

# Everything
uv add "dhurandhar[all]"

---

Quickstart

# PLE memory breakdown for Gemma4 E2B at 4096 context
dhurandhar-ple-analyze --model gemma4-e2b --context 4096

# Device feasibility across all registered devices
dhurandhar-device-check --model gemma4-e2b

# TurboQuant quality sweep
dhurandhar-turbo-sweep --model gemma4-e2b --residual-bits 2,3,4,6,8

# Codec comparison
dhurandhar-compare-codecs --model gemma4-e2b --head-dim 256 --seq-len 2048

# Launch 5-tab dashboard
dhurandhar-dashboard

Python API

from dhurandhar.models import get_profile
from dhurandhar.devices import get_device

model  = get_profile("gemma4-e2b")
device = get_device("pixel-8")

print(f"KV cache at 4096 ctx: {model.kv_cache_bytes(4096) / 1e6:.1f} MB")
print(f"Device available RAM: {device.available_ram_gb:.1f} GB")

Custom model via YAML

# my_model.yaml
name: my-custom-2b
param_count_b: 2.0
weight_bytes: 4000000000
num_layers: 32
num_attention_layers: 32
num_kv_heads: 8
head_dim: 128
architecture_family: llama

dhurandhar-ple-analyze --model my_model.yaml --context 2048

---

Status

v0.1.0 — model/device registries stable, analysis modules and dashboard landing in v0.1.x point releases.

---

License

Apache 2.0