sm121

Qwen3.5-122B-A10B on DGX Spark: 28.3 → 51 tok/s (+80%)

Measured SM121 compatibility recipe for MiniMax H3 FL2VA on one NVIDIA DGX Spark with vLLM-Omni and online FP8.

Serve the home! Inference stack for your Nvidia DGX / RTX Spark (128GB ram) aka the Grace Blackwell AI supercomputer on your desk. Mostly vLLM based for now and single-spark. For the not-so-rich people. Updated once a month. Hobby project

DGX Spark research and tests - containers, benchmarks, and investigation notes for running models on GB10 (SM 12.1)

7.67× LoRA / 8.35× Full FT speedup for Qwen3.5 (0.8B–27B) on NVIDIA DGX Spark — wall-clock parity with rented H100. Lossless within BF16. Three-command interactive wizard handles model picker, data validator, training, and merge.

DeepSeek-V4-Flash DSpark speculative decoding on 2x DGX Spark (GB10/sm_121), 1M context — our fp8 recipe + independent reproduction and cross-build benchmarks of the NVFP4-KV build. Honest, apples-to-apples measurements.

Optimized CUDA kernels for NVIDIA GB10 Blackwell (sm_121, DGX Spark). RMSNorm + GELU. First sm_121 kernel on HuggingFace Kernel Hub.

CUDA++ for the NVIDIA DGX Spark (GB10, sm_121): Aien's own GPU machine-code toolchain, runtime and driver client, without the CUDA toolkit. Other NVIDIA GPUs are not supported yet.

Qwen3.8-Flash-Next as Cogni-Brain on NVIDIA DGX Spark (GB10): HashK GPU PLE + SGLang NEXTN, 36.8 tok/s code, 100/100 tool-eval, 262K context.

Measured serving recipes for DeepSeek-V4.1-Flash on 4x NVIDIA DGX Spark (GB10): 1M context on vLLM (CUDA graphs, vision, tools, DSpark) and a switchless-ring SGLang TP4 lane, plus a cross-project reference table. EN + 中文.

Production runbook for Qwen3.5-122B hybrid INT4+FP8 on NVIDIA DGX Spark GB10 — optimization stack, PD firmware wedge diagnosis, bench results

Pre-built PyTorch wheels and build scripts for NVIDIA DGX Spark (GB10, sm_121, Blackwell, CUDA 13.0, ARM64)

Patches + recipe to deploy festr2/MiMo-V2.5-Pro-NVFP4-MXFP8-attn-TP8 on 8-node DGX Spark sm_121 (Ray + vLLM, TP=8). Fixes the fused-qkv loader bug that mis-slotted Q values as K/V on 7 of 8 ranks.

Optimized SM121 vLLM container and benchmark report for nvidia/diffusiongemma-26B-A4B-it-NVFP4

SM121-optimized vLLM serving of nvidia/GLM-5.3-Flash-NVFP4 on 2x GB10 (TP=2, DFlash2 K=7): patches, click-run container, serve profiles, sanitized evidence

Ornith-1.5-35B NVFP4-W4A16 + BF16 MTP via SGLang on SM12X (SM120/SM121). Click-run serve + SM12X-LLM-BENCH claim row.

Empirical kernel scheduling characterization for NVIDIA GB10 (SM121a). Sweeps GEMM tile configurations, classifies PTX instruction paths, captures hardware telemetry

Build paddlepaddle-gpu from source for NVIDIA DGX Spark and GB10 (Grace Blackwell, aarch64, CUDA 13, sm_121). Prebuilt wheel plus the exact patches upstream is missing.

Docker Compose setup for ComfyUI on NVIDIA DGX Spark (GB10/sm_121) — NGC base image, patched xformers, self-compiled SageAttention, non-root execution

Measured vLLM serving for 1–2 NVIDIA DGX Spark (GB10): TP=2 RoCE flagship, validation ledger, stranger-ready Path A/B onboarding.