Offline-quantized W8A16 FP8 in the compressed-tensors float-quantized format: weights in float8_e4m3fn with per-output-channel symmetric scales, activations kept in bf16/fp16.

How it was quantized

For each linear layer, every output row gets its own scale starting from amax / 448, refined by an MSE clip search over ~9 clip fractions (0.8–1.0× amax) picking the lowest-error scale per row. Weights are then quantized q = e4m3(w / scale) with round-to-nearest and saturation. This per-channel + clipping scheme gives better SNR than vLLM's online per-tensor --quantization fp8 path.

Only 2D linear projection weights are quantized (attention q/k/v/o, MLP gate/up/down). Embeddings, norms, lm_head, routers/experts, and the vision tower stay in bf16 and are listed in the checkpoint's ignore list, so vLLM leaves them untouched.

Supported vLLM platforms

Intel XPU — XPUW8A16FP8LinearKernel (the intended target).
NVIDIA CUDA (SM75+, i.e. Turing and newer) — HummingFP8ScaledMMLinearKernel when the humming package is installed, otherwise MarlinFP8ScaledMMLinearKernel.
Not supported: ROCm, CPU, TPU — vLLM has no W8A16-FP8 kernel for these backends yet, so loading will fail with a "no kernel" error.

hoborific/Glistening-Gem-31B-v2.1-W8A16-FP8

Glistening-Gem-31B-v2.1-W8A16-FP8

Format

How it was quantized

Supported vLLM platforms

hoborific/
Glistening-Gem-31B-v2.1-W8A16-FP8