This model was trained using the Speculators library on a subset of Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered and the train_sft split of HuggingFaceH4/ultrachat_200k. Responses were regenerated by Mistral-Small-4-119B-2603. The model was initialized from a full-attention DFlash checkpoint and finetuned with sliding window attention (window=2048) on all 5 draft layers.

Commands

Using the Speculators library and the helper scripts provided in the repo.

Prepare data

bash

# In virtual environment with speculators installed
python scripts/prepare_data.py \
  --model mistralai/Mistral-Small-4-119B-2603 \
  --data ./regenerated_data.jsonl \
  --output ./output/dflash_mistral_119b_sliding_window \
  --seq-length 8192

Launch vLLM

bash

# In (separate) virtual environment with vllm installed
CUDA_VISIBLE_DEVICES=0,1,2,3 python scripts/launch_vllm.py \
  mistralai/Mistral-Small-4-119B-2603 \
  --target-layer-ids 2 18 33 \
  -- --port 8000 \
  --tensor-parallel-size 4 \
  --max-model-len 8194 \
  --enforce-eager

Launch training

Must be run once vLLM has finished launching and is running in the background.

bash

# In virtual environment with speculators installed
CUDA_VISIBLE_DEVICES=4,5,6,7 torchrun \
  --standalone \
  --nproc_per_node 4 \
  scripts/train.py \
  --verifier-name-or-path mistralai/Mistral-Small-4-119B-2603 \
  --speculator-type dflash \
  --from-pretrained ./output/dflash_mistral_119b_sliding_window/patched_checkpoint \
  --data-path ./output/dflash_mistral_119b_sliding_window \
  --vllm-endpoint http://localhost:8000/v1 \
  --save-path ./output/dflash_mistral_119b_sliding_window/checkpoints \
  --epochs 3 \
  --lr 5e-5 \
  --total-seq-len 8192 \
  --max-anchors 1024 \
  --loss-fn '{"ce": 0.1, "tv": 0.9}' \
  --fsdp-shard \
  --no-sample-from-anchor \
  --on-missing generate \
  --on-generate delete \
  --checkpoint-freq 0.1

Model Specifications


Base Model	mistralai/Mistral-Small-4-119B-2603
Chat Template	mistralai/Mistral-Small-4-119B-2603 (use `/chat/completions` endpoint)
Format	Safetensors
License	Apache 2.0
Validation Hardware	Nvidia H100

Deployment

bash

# Install vLLM from the required PR
pip install git+https://github.com/vllm-project/vllm.git@refs/pull/41880/head

# Deploy with speculative decoding
vllm serve mistralai/Mistral-Small-4-119B-2603 \
    --tensor-parallel-size 4 \
    --max-model-len 16384 \
    --speculative-config '{
        "model": "RedHatAI/Mistral-Small-4-119B-2603.dflash",
        "num_speculative_tokens": 7,
        "method": "dflash"
    }'

Preliminary Evaluations

Per-position token acceptance rates across datasets:

Dataset	Pos 1	Pos 2	Pos 3	Pos 4	Pos 5	Pos 6	Pos 7	Avg Length
HumanEval	80.9%	61.0%	45.7%	33.7%	24.7%	17.8%	12.1%	3.759
math_reasoning	91.7%	77.5%	65.5%	55.0%	45.1%	36.0%	27.4%	4.982
qa	62.6%	35.5%	19.5%	11.1%	6.0%	3.1%	1.4%	2.392
question	67.5%	42.3%	27.1%	17.9%	12.0%	8.0%	5.2%	2.801
rag	65.9%	39.7%	22.8%	13.5%	8.1%	4.2%	2.2%	2.565
summarization	65.1%	36.2%	19.3%	10.7%	5.6%	2.8%	1.0%	2.407
tool_call	66.6%	40.7%	24.9%	15.8%	10.0%	6.1%	3.5%	2.675
translation	71.5%	40.9%	20.1%	9.0%	4.0%	1.5%	0.8%	2.478
writing	67.9%	42.2%	26.6%	17.2%	11.8%	7.8%	5.0%	2.786

Latency Speedup

Speedup comparisons of DFlash speculative decoding vs. Eagle3 at varying request rates on Nvidia H100:

References

Paper: DFlash: Block Diffusion for Flash Speculative Decoding