Skip to content

Add Qwen3-VL training adaptation for DFlash #475

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
gq112 wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Add Qwen3-VL training adaptation for DFlash #475

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f80be86
Add Qwen3-VL-8B DFlash training support
ggg-s Feb 27, 2026
c19b963
Update train_dflash.py
ggg-s Feb 28, 2026
aff0de1
fix run_qwen3_vl_8b_dflash_online.sh
ggg-s Feb 28, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
54 changes: 54 additions & 0 deletions configs/qwen3-vl-8b-dflash.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,54 @@
{
"architectures": [
"DFlashDraftModel"
],
"attention_bias": false,
"attention_dropout": 0.0,
"block_size": 16,
"bos_token_id": 151643,
"dflash_config": {
"mask_token_id": 151669,
"target_layer_ids": [
3,
9,
17,
25,
33
]
},
"dtype": "bfloat16",
"eos_token_id": 151645,
"head_dim": 128,
"hidden_act": "silu",
"hidden_size": 4096,
"initializer_range": 0.02,
"intermediate_size": 12288,
"layer_types": [
"full_attention",
"full_attention",
"full_attention",
"full_attention",
"full_attention"
],
"max_position_embeddings": 262144,
"model_type": "qwen3_vl_text",
"num_attention_heads": 32,
"num_hidden_layers": 5,
"num_key_value_heads": 8,
"num_target_layers": 36,
"rms_norm_eps": 1e-06,
"rope_scaling": {
"mrope_interleaved": true,
"mrope_section": [
24,
20,
20
],
"rope_type": "default"
},
"rope_theta": 5000000,
"tie_word_embeddings": false,
"transformers_version": "4.57.1",
"use_cache": true,
"vocab_size": 151936
}
40 changes: 40 additions & 0 deletions examples/run_qwen3_vl_8b_dflash_online.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,40 @@
#!/bin/bash

SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
ROOT_DIR=$(dirname $SCRIPT_DIR)

export TORCHINDUCTOR_CACHE_DIR=$ROOT_DIR/cache/compiled_kernels

NUM_GPUS=${1:-8}
ATTENTION_BACKEND=${2:-flex_attention}
BUILD_DATASET_NUM_PROC=${BUILD_DATASET_NUM_PROC:-16}

torchrun \
--standalone \
--nproc_per_node $NUM_GPUS \
$ROOT_DIR/scripts/train_dflash.py \
--target-model-path Qwen/Qwen3-VL-8B-Instruct \
--draft-model-config $ROOT_DIR/configs/qwen3-vl-8b-dflash.json \
--target-model-backend hf \
--is-vlm \
--trust-remote-code \
--train-data-path $ROOT_DIR/cache/dataset/allava4v-mix-20k_train.localimg_regen.jsonl \
--build-dataset-num-proc $BUILD_DATASET_NUM_PROC \
--min-pixels 50176 \
--max-pixels 802816 \
--output-dir $ROOT_DIR/outputs/qwen3-vl-8b-allava4v20k-dflash \
--cache-dir $ROOT_DIR/cache \
--num-epochs 6 \
--batch-size 2 \
--learning-rate 1e-4 \
--warmup-ratio 0.08 \
--max-grad-norm 1.0 \
--max-length 4096 \
--num-draft-layers 5 \
--chat-template qwen3-vl \
--attention-backend $ATTENTION_BACKEND \
--block-size 16 \
--num-anchors 512 \
--loss-decay-gamma 7.0 \
--log-interval 50 \
--save-interval 1000
Loading