Amrendra Singh IxMxAMAR

Hi, I'm IxMxAMAR

I like making generative models run faster, and run on hardware people actually own. That means custom attention kernels for AMD's RX 9070 XT, ComfyUI node packs, one-click RunPod images, and tools for running LLMs locally. On the side, I write hypervisors.

🏆 Flagship Engineering Highlights

🚀 SageAttention-RDNA4

Quantized Attention Kernels for AMD RDNA4 (gfx1201 / RX 9070 & 9070 XT)

Custom HIP Attention Kernels: fp16 and bf16 kernels for head_dim 128 on RDNA4 (int8 Q·K in fp16, fp8 in bf16), plus a HIP quantization step, so nothing compiles on first use.
Faster than the gfx12 Port: 1.5–2.0× faster than SageAttention PR #368 on causal attention and 1.06–1.3× on non-causal in fp16 (2.2× and 1.3× in bf16), measured over the full call.
Fused Q/K/V Views: Serves the strided q/k/v views that fused-projection video models send, without copies. Run end to end in Krea2 and MiniMax-H3.
Drop-in for ComfyUI: Prebuilt Windows wheels for PyTorch 2.13 and 2.14, picked up by --use-sage-attention. Latest: v0.5.0, with a 6× faster first call.

Stack: ROCm / HIP • C++ • PyTorch • RDNA4 gfx1201 • int8 / fp8

🛡️ Aria

Type-1 Bare-Metal Hypervisor with In-Place OS Virtualization

UEFI Bootloader Architecture: Boots directly from UEFI firmware prior to OS initialization, taking full hardware control before Windows boots.
In-Place Windows Virtualization: Dynamically transforms the running Windows kernel into a guest VM without rebooting or disconnecting hardware.
Nested VMX Engine: Full Intel VT-x hardware-assisted virtualization with comprehensive VMCS configuration and nested virtualization support.
Advanced EPT & Shadowing: Extended Page Tables for transparent physical memory virtualization, stealth page hiding, and VMCALL dispatch control.

Stack: C • x86_64 ASM • UEFI / EDK2 • Intel VT-x • EPT Paging

🛠️ Technical Stack & Expertise

💻 Low-Level Systems & Hypervisors

⚡ AI Acceleration & GPU Compute

☁️ Cloud & Infrastructure

🔧 Languages & Applied Tooling

⚡ GPU & Performance

Project	What it is
SageAttention-RDNA4	SageAttention for the RX 9070 / 9070 XT (`gfx1201`), with custom HIP attention kernels for fp16 and bf16. 1.06–2.2× faster than SageAttention PR #368's gfx12 port over the full call, handling fused q/k/v views for video models, and a drop-in for ComfyUI.
bitsandbytes	Accessible large language models via k-bit quantization for PyTorch with ROCm / AMD GPU acceleration enhancements.
comfyui-mossdype	DyPE-M dynamic position extrapolation for MOSS-Image — render past the 4MP training range without retraining.

🧩 ComfyUI Node Ecosystem

Project	What it is
ComfyUI-API-Toolkit	Kling AI, ElevenLabs, and Google Gemini in one extensible package: 68 production nodes.
ComfyUI-Utility-MegaPack	Comprehensive utility pack featuring 156 operations condensed across 7 nodes, with 11 custom visual graph themes.
ComfyUI-IxMxAMAR-Masterworks	12 production-ready masterwork workflow templates orchestrating multimodal generation pipelines end-to-end.
ComfyUI-NanoBanana2	Full-parameter Google Gemini API node pack for ComfyUI: text, vision, high-res image generation, audio, music, video, and embeddings.
ComfyUI-NanoBanana-FaceSwap	Zero-shot face and head replacement powered by Gemini multimodal semantic image editing.
ComfyUI-ElevenLabs-Pro	Full-featured ElevenLabs in ComfyUI: voice cloning, low-latency TTS, speech-to-speech, sound effects, and generative music.
ComfyUI-Kling-Direct	Native Kling AI API integration for ComfyUI: video, image, and audio generation using your own API key.

☁️ Cloud & RunPod

Project	What it is
ComfyUI-Ultimate	Batteries-included ComfyUI Docker image for RunPod: 29 curated node packs, SageAttention and FlashAttention, pinned dependencies, CI-gated builds.
ComfyUI-Ultimate-ST	All-in-one RunPod image combining ComfyUI-Ultimate generation with SillyTavern AI chat.
runpod-wan22-serverless	Serverless Wan 2.2 14B text-to-video and image-to-video worker endpoints for RunPod, with desktop client integration.
ComfyUI-Serverless-FaceSwap	Scalable ComfyUI face-swap workflows served as an on-demand RunPod serverless API.
rigma-runpod	Hardware-tuned RunPod pod template, Dockerfile, and container startup spec for instant Rigma LLM deployments.

🖥️ Local AI Tooling

Project	What it is
rigma	Hardware-aware local LLM deployment: `rigma up` probes your GPU and RAM and serves the best-tuned model setup for that exact machine.
rigma-registry	Community registry of verified, benchmarked model + quant + flag combinations for Rigma hardware-aware deployments.
raggity	Local-first search and Q&A over notes, docs, and PDFs: hybrid retrieval, cross-encoder reranking, and verified citations powered by Claude Agent SDK (CLI & FastAPI).
LoRA-Dataset-Forge	Two photos in, a captioned, aspect-bucketed LoRA training set out, with review step, Gemini face validation, and one-click export.

🛠️ Systems & Hypervisors

Project	What it is
Aria	A Type-1 hypervisor in C that boots through UEFI and virtualizes a running Windows in place, with a full nested VMX engine and EPT paging.
Visor	Personal hypervisor experiment in Rust: UEFI loader, Windows kernel driver, VMCALL dispatch, and EPT page hiding.
barevisor	Personal fork of tandasat's barevisor in Rust with experimental modifications and stealth enhancements.
SimpleVisor	Modified SimpleVisor, ionescu007's minimal Intel VT-x hypervisor in C with custom telemetry and hooks.

🔬 Research & Innovation

Project	What it is
Research-and-Innovation	A living compilation of research papers and ideas organized by domain. First entry: The x16 Ceiling (PCIe memory bandwidth and host-to-device streaming saturation).

🐍 Contribution Activity

📊 Activity & Statistics

GitHub Stats GitHub Streak Top Languages

Navigation Menu