sm75

Veda sparse attention for ComfyUI's MiniMax-H3, with a hand-written CUDA kernel for NVIDIA Turing (sm_75) so it runs on RTX 2080 Ti / GTX 16-series and any compute-capability 7.5 card. Upstream Veda only supports SM80+. Measured on one 2080 Ti: 4.16x at 768x768x124, 1.83x at 512x512x124.

FlashAttention v1 forward pass in CUDA for NVIDIA Turing (SM75)

Compiler MVP that detects Transformer fusion patterns, generates optimized CUDA kernels with WMMA Tensor Cores, and executes them on real GPU hardware — 10.5 TFLOPs on RTX 2070, correctness validated against PyTorch.

Backward pass CUDA kernels for fused GEMM+Bias+GeLU on SM75 (Turing). Float4 vectorized WMMA kernels validated against PyTorch autograd. 3.1x faster than autograd at M=1024. 24/24 tests passing.

MX450 (TU117) has no Tensor Core in silicon — a reproducible reverse-engineering case study: Falcon microcode, register-level proofs, and why every HMMA library is 18x slower. 三语文档。

Reproducible FreeToken MoE expert-offload compatibility lab and benchmarks for NVIDIA Turing sm_75.

RTX 2080 Ti (sm75) 上 ComfyUI 的适配实录：H3 加速配方、int8 原生算子崩溃加固、前端连线/加载兼容、工作流与依赖排坑。实测数据 + 可复现步骤。

Fork of FlashQLA-SM70-SM75 with the SM75 import/build tweaks used by qwen3-8-27b-dual-2080ti-vllm (branch 2080ti_dual_qwen38-27B)

PyTorch CUDA wheels with sm_75 for AWS g5g (Graviton arm64 / aarch64 + NVIDIA T4G) — fixes 'no kernel image is available' — pip index

Instruction-level Tensor Core micro-kernel engineering on SM75 (RTX 2070), including PTX manual emission, shared-memory staging, and hardware-aware auto-tuning.

Fork of vLLM carrying the SM75 / Qwen3.8 changes used by qwen3-8-27b-dual-2080ti-vllm (branch 2080ti_dual_qwen38-27B); upstream main is kept as the 'upstream' branch

Personal-use vLLM 0.33 + FlashInfer 0.7.3 on Turing SM75: guarded planning, Conda setup, synthetic benchmarks and explicit risks

Qwen3.8-27B 双 RTX 2080 Ti 22GB + NVLink 的 vLLM 部署（上游 main 路线）：SM75 移植、MTP 下保住 FULL cudagraph（单并发解码 54.1→37.5 ms/步）、256K/500K 上下文与 128K tiered offload profile

RelayVSR flash decoder attention fix for sm_70 GPUs (Tesla V100 / Quadro GV100 / Titan V). Expands GQA K/V projections to MHA so PyTorch fused cutlass attention is used. ~1.94x faster decoding.