Qwen3-32B DFlash Draft Model (EagleChat 400K Mix)

A DFlash draft model trained from Qwen3-32B using a EagleChat subset (English 200K + Chinese 200K) to accelerate speculative decoding.

Model Summary

This repository provides a DFlash draft model for Qwen3-32B. The draft model is intended to be used together with the target model in SpecForge, improving throughput (output tokens/sec) under standard speculative verification.

• Base / Target model: Qwen/Qwen3-32B
• Draft model type: DFlash (speculative decoding draft)
• Training data: EagleChat subset (English 200K + Chinese 200K; total ~400K)
• Training hardware: H100
• Primary use case: accelerate inference with DFlash / SpecForge

Training Details

Data

• Dataset: EagleChat subset
• Composition:
  • English: ~200,000 samples
  • Chinese: 200,000 samples
• Total: ~400,000 samples

Procedure

• Epochs: 6
• Sequence length: 4096
• Precision: bf16
• Codebase: SpecForge (DFlash training)

Evaluation

Benchmark settings

• Target model: /models/Qwen3-32B
• Draft model: sx-aicp/qwen3-32b-dflash-en-zh (or local path)
• Max new tokens: 2048
• Attention backend: fa3
• Tensor parallel (tp_size): 4
• device_sm: 90 (H100)
• drop_first_batch: true
• Concurrencies: 1 / 4 / 32 (varies by suite)

Speed Bench Results

Environment: H100 (SM90), tp=4, attention=fa3, max_new_tokens=2048, drop_first_batch=true.

Unified Summary

Format: baseline tok/s → DFlash tok/s; Speedup× / L(acceptance length).

How to Evaluate (z-lab / dflash)

python benchmark_sglang.py \
  --tp-size 4 \
  --target-model /models/Qwen3-32B \
  --draft-model /path/to/draft_model \
  --concurrencies 1,4,32 \
  --dataset-name math500 \
  --attention-backends fa3 \
  --output-md sglang_results.md