Spark-X2.6-4B FP8 DSpark Draft
A validation model for Draftfit, a platform for building, converting and evaluating speculative decoding drafts. Use with XHToken/Spark-X2.6-4B-FP8.
Validation
| Metric | Source draft | This draft |
|---|---|---|
| Accepted draft proposals | 38.58% | 40.00% |
Greedy prefix verification, seven proposals, 64 output tokens, 24 fixed validation prompts: eight each from general, code and math. Acceptance excludes target bonus tokens. All validation outputs match the fixed target-only reference. One seed; this previously inspected partition was used for model selection. Full measurement records are in metrics.json.
The draft uses BF16 weights with FP32 norms; FP8 identifies the required target. Measurements use the original quantized target weights and dynamic activation QDQ in a decompressed compressed-tensors reference. Cached generation can differ from full-prefix evaluation. Native packed-kernel parity and production serving speed remain unmeasured.
Use
Load model.safetensors with the included SparkDSpark(official_config()) from spark_model.py and a compatible Spark target adapter. The target revision is pinned in config.json. This reference checkpoint requires custom integration for vLLM, SGLang or MLX.