sm80

DeepSeek-V4.1-Flash inference on NVIDIA Ampere with vLLM, SM80 patches, reproducible throughput and coding-agent benchmarks. Tested on 8x A800 80GB.

GLM-5.3-Flash (320B MoE) served with vLLM on 8x CMP 170HX (SM80, 64 GB, PCIe Gen2 x4): pipeline-parallel 8, NVFP4 weights, DFlash2 speculative decoding, 1M context. Patches, launch config and benchmarks.

vLLM for NVIDIA CMP 170HX (Ampere sm_80): fast, repeatable model serving on one card or several

Qwen3.8-Flash-Next on 3x CMP170HX (SM80): NVMe-streamed PLE, FP8 KV, 1M context (with NVMe offload)

Reproducible vLLM deployment for GLM-5.3-Flash on 8x NVIDIA A800 SM80 GPUs

Capacity-aware single-GPU SGLang benchmarks with MTP A/B, core/extended context matrices, raw telemetry, model/KV memory capture, and generated Markdown/PDF reports.

TensorFold deployment for GLM-5.3-Flash on 4x NVIDIA CMP 170HX (GA100 sm_80): bit-exact 4-rank TP, shm all-gather, W4A16 affine - complete tested solution (199.4 tok/s structured, exactness maintained)

DeepSeek-V4-Flash-Vision-Exp on NVIDIA CMP 170HX/SM80 with vLLM, PP DSpark and 1M context

Run Qwen3.8-Flash-Next (262K ctx) on ONE sm80 64GB GPU (e.g. CMP 170HX) — vLLM Docker with PLE SSD-offload. Based on klee100/iIIusi0n patch (Apache-2.0)

Measured llama.cpp Q6_K CUDA optimizations for NVIDIA A100 (SM80)

Strict-FP32 CUDA SGEMM with shape-dispatched SM80 cp.async kernels and reproducible A100 evidence

Native dual-GPU long-context inference runtime for Qwen3.8-Flash-Next

From-scratch SM80 (A100/A800) CUDA kernels for GDN (Gated DeltaNet) and QSA (query-key sparse attention). Apache-2.0.