Instructions to use Tohirju/qwen35-9b-ca4-sft-e1-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Tohirju/qwen35-9b-ca4-sft-e1-lora with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("unsloth/Qwen3.5-9B-Base")
model = PeftModel.from_pretrained(base_model, "Tohirju/qwen35-9b-ca4-sft-e1-lora")

Transformers

How to use Tohirju/qwen35-9b-ca4-sft-e1-lora with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Tohirju/qwen35-9b-ca4-sft-e1-lora")

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("Tohirju/qwen35-9b-ca4-sft-e1-lora")
model = AutoModelForMultimodalLM.from_pretrained("Tohirju/qwen35-9b-ca4-sft-e1-lora", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Tohirju/qwen35-9b-ca4-sft-e1-lora with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Tohirju/qwen35-9b-ca4-sft-e1-lora"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Tohirju/qwen35-9b-ca4-sft-e1-lora",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Tohirju/qwen35-9b-ca4-sft-e1-lora

SGLang

How to use Tohirju/qwen35-9b-ca4-sft-e1-lora with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Tohirju/qwen35-9b-ca4-sft-e1-lora" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Tohirju/qwen35-9b-ca4-sft-e1-lora",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Tohirju/qwen35-9b-ca4-sft-e1-lora" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Tohirju/qwen35-9b-ca4-sft-e1-lora",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Tohirju/qwen35-9b-ca4-sft-e1-lora with Docker Model Runner:
```
docker model run hf.co/Tohirju/qwen35-9b-ca4-sft-e1-lora
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

See axolotl config

axolotl version: 0.18.0

# Gemma4-12B CA4 united SFT stage 1 — Axolotl, 4-GPU DDP (GPUs 0-3), PRE-MASKED data.
# Seeds from the CPT WINNER (e2-merged). Recipe per SESSION_HANDOFF §3: stacked merge-per-epoch,
# Gemma peak LR 5e-5 (sft_e1), then sft_e2 = fresh LoRA on merged sft_e1 at 2.5e-5.
# Data: /data/corpus/ca4_sft_gemma_masked (prep_sft_ca4.py: Gemma-templated, prompt-masked labels).
base_model: unsloth/Qwen3.5-9B-Base
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true

datasets:
  - path: /data/corpus/ca4_sft_qwen_masked
    ds_type: arrow
    split: train
    type:                       # EMPTY = pretokenized passthrough (labels already prompt-masked)
sample_packing: false
pad_to_sequence_len: false
sequence_len: 1024
dataset_prepared_path: /data/corpus/prepared_qwen_sft_e1
output_dir: /data/runs/qwen35-9b_ca4_sft_e1

adapter: lora
lora_r: 32
lora_alpha: 32
lora_dropout: 0.0
peft_use_rslora: true
lora_target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
# NOTE: no embed_tokens/lm_head in SFT (factory target_modules.sft) — only CPT trains those.

bf16: true
tf32: true
micro_batch_size: 8
gradient_accumulation_steps: 2          # global batch = 8 * 2 * 4 GPUs = 64 (proven 16x4 equiv)
num_epochs: 1
learning_rate: 2.0e-5              # proven Qwen3.5-9B SFT peak (1e-4 diverged; 2e-5+clip0.5 is the fix)
warmup_ratio: 0.03
max_grad_norm: 0.5
optimizer: adamw_torch_fused
lr_scheduler: cosine
weight_decay: 0.0

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
ddp_find_unused_parameters: false       # plain LoRA, no modules_to_save -> all params used

save_steps: 500
save_total_limit: 4
logging_steps: 1

data/runs/qwen35-9b_ca4_sft_e1

This model is a fine-tuned version of unsloth/Qwen3.5-9B-Base on an unknown dataset.

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 8
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 2
total_train_batch_size: 64
total_eval_batch_size: 32
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 1284
training_steps: 42823

Training results

Framework versions

PEFT 0.19.1
Transformers 5.14.1
Pytorch 2.12.1+cu130
Datasets 4.8.4
Tokenizers 0.22.2

Downloads last month: -

Model tree for Tohirju/qwen35-9b-ca4-sft-e1-lora

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

unsloth/Qwen3.5-9B-Base

Adapter

(6)

this model