Jacob-2-E4B-prod

Merged + FP8-quantized build of jacob-ml/Jacob-2-E4B for vLLM deployment.

• Base model: google/gemma-4-E4B-it
• LoRA adapter: jacob-ml/Jacob-2-E4B (r=16, alpha=16, language-model attention + MLP projections), merged into the base weights in bf16
• Quantization: FP8 (float8_e4m3fn) per-tensor dynamic weights/activations via llmcompressor in compressed-tensors format; lm_head kept in high precision

Serve with vLLM:

vllm serve jacob-ml/Jacob-2-E4B-prod

vLLM auto-detects the compressed-tensors FP8 config; no extra flags required.