r9700

LLM speculative inference server for heterogeneous hardware & consumer GPUs

The fastest vLLM build for AMD R9700's. For the latest updates join the Launch80 discord server. I am on X now also https://x.com/bkuyper

Working on the fastest vllm path for AMD Radeon AI PRO R9700

WHIRL (Windows HIP Inference for RDNA LLMs): native Windows C++/HIP LLM inference engine for the AMD Ryzen AI Max+ 395 (Radeon 8060S) and Radeon AI PRO R9700. Unzip and run, only the AMD driver.

vLLM for AMD RDNA4 (gfx1201): Radeon AI PRO R9700 & RX 9070 XT — native MXFP4 linear kernel + spec-decode verify fix, carried as rebased branches (fork-carry model)

The unmodified upstream Linux amdgpu + amdkfd driver running on macOS: AMD Radeon GPUs over Thunderbolt on Apple Silicon, inside a DriverKit extension.

Reproducible diagnostics, patches, and validation evidence for AMD GPU Direct P2P on Intel X99 / Broadwell-EP, with RCCL and SGLang/Qwen3.8 case studies.

App to handle one GPU to multiple models/element with queuing and log tracking all with a local hosted web dashboard

Experimental ROCm 10 compatibility build for Qwen3.8-Flash-Next on 4x AMD R9700

Minimal AITER LDS and ROCr idle-CPU fixes for Radeon AI PRO R9700 on vLLM 0.29.0 / ROCm 7.2.3

Verified dual-R9700 deployment recipe, configs, scripts, and benchmarks for Dyluhn/R9V Qwen3.8 Flash-Next

HyperLoom R9700 R&D hub for reproducible AMD/ROCm agentic-infrastructure experiments with explicit evidence and truth boundaries.

Field notes: Radeon AI PRO R9700 (RDNA4, gfx1201) + ROCm for local LLMs. Measured on 4 cards, traps included.

Instrumented AMD R9700 inference with numerical alignment, qualified performance backports, persistent chat snapshots and Pi integration. Downstream of Radiance.

HyperLoom R9700 experiment: live A/B validation path for AMD local inference. Not a standalone product.

HyperLoom R9700 experiment: provider bridge validation against the local AMD execution fabric. Not a standalone product.

Performance benchmark for local LLM inference: llama.cpp, any GGUF, one GPU or several (ROCm). Records the conditions, not just the numbers.

HyperLoom R9700 probe: minimal agent execution smoke test on the AMD/local runtime. Not a standalone product.

HyperLoom R9700 experiment: live runner validation for local AMD workloads and evidence capture. Not a standalone product.

HyperLoom R9700 experiment: minimal container/runtime baseline for reproducible AMD/ROCm validation. Not a standalone product.