useful-quants/FastContext-1.0-4B-RL-w4a16-g128

FastContext-1.0-4B-RL-w4a16-g128

This repository contains a W4A16 / BF16-activation export of microsoft/FastContext-1.0-4B-RL produced with llmcompressor in compressed-tensors format for vLLM deployment on Ampere-class and newer NVIDIA GPUs.

Quantization Summary

• —Source model: microsoft/FastContext-1.0-4B-RL
• —Source revision: bd530b8e1f9a59b12e77061d9fe1e018f6eabe42
• —Quantization: W4A16
• —Group size: 128
• —Activation dtype: bfloat16
• —Backend: llmcompressor + compressed-tensors
• —Export format: compressed-tensors

Calibration and Accuracy Check

• —Calibration samples: 96
• —Held-out eval samples: 24
• —Max sequence length: 4096
• —Held-out perplexity before quantization: 7.3734
• —Held-out perplexity after quantization: 7.1517
• —Relative perplexity delta: -3.01%

The calibration/eval prompts are repo-exploration and tool-calling flavored so the quantizer sees text that matches the deployed FastContext use case rather than generic chat data.

Local Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "useful-quants/FastContext-1.0-4B-RL-w4a16-g128"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

vllm serve useful-quants/FastContext-1.0-4B-RL-w4a16-g128 --quantization compressed-tensors

vLLM Smoke Test

This export was smoke-tested locally with vllm 0.23.0 after upload.

Launch command used:

python -m vllm.entrypoints.openai.api_server \
  --model /path/to/FastContext-1.0-4B-RL-w4a16-g128 \
  --quantization compressed-tensors \
  --max-model-len 4096 \
  --gpu-memory-utilization 0.7 \
  --enforce-eager \
  --host 127.0.0.1 \
  --port 8011

Observed smoke results:

• —/v1/models returned the model successfully.
• —Prompt Return exactly: calibration-pass returned calibration-pass.
• —Prompt Write one rg command to search for OAuth callback state validation. No explanation. returned a valid rg command.

Files

• —recipe.yaml: exact GPTQ recipe used for this export
• —calibration_text.jsonl: chat-rendered calibration corpus
• —eval_text.jsonl: held-out evaluation corpus
• —quantization_report.json: environment, perplexity, and generation checks
• —vllm_smoke_test.json: exact local vLLM smoke-test command and responses