Instructions to use laion/qwen3-30b-a3b-thinking-opencode-sft-sparse with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use laion/qwen3-30b-a3b-thinking-opencode-sft-sparse with Transformers:

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

pipe = pipeline("text-generation", model="laion/qwen3-30b-a3b-thinking-opencode-sft-sparse")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("laion/qwen3-30b-a3b-thinking-opencode-sft-sparse")
model = AutoModelForCausalLM.from_pretrained("laion/qwen3-30b-a3b-thinking-opencode-sft-sparse", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use laion/qwen3-30b-a3b-thinking-opencode-sft-sparse with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "laion/qwen3-30b-a3b-thinking-opencode-sft-sparse"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "laion/qwen3-30b-a3b-thinking-opencode-sft-sparse",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/laion/qwen3-30b-a3b-thinking-opencode-sft-sparse

SGLang

How to use laion/qwen3-30b-a3b-thinking-opencode-sft-sparse 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 "laion/qwen3-30b-a3b-thinking-opencode-sft-sparse" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "laion/qwen3-30b-a3b-thinking-opencode-sft-sparse",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "laion/qwen3-30b-a3b-thinking-opencode-sft-sparse" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "laion/qwen3-30b-a3b-thinking-opencode-sft-sparse",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use laion/qwen3-30b-a3b-thinking-opencode-sft-sparse with Docker Model Runner:
```
docker model run hf.co/laion/qwen3-30b-a3b-thinking-opencode-sft-sparse
```

qwen3-30b-a3b-thinking-opencode-sft-sparse / README.md

penfever

Upload folder using huggingface_hub

d8b49d9 verified 27 days ago

preview code

Raw

History Blame Contribute Delete

5.74 kB

	---
	library_name: transformers
	tags:
	- generated_from_trainer
	model-index:
	- name: mnt/home/bf996/experiments/densemixer/run2_sparse_out
	results: []
	---

	<!-- This model card has been generated automatically according to the information the Trainer had access to. You
	should probably proofread and complete it, then remove this comment. -->

	[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
	<details><summary>See axolotl config</summary>

	axolotl version: `0.17.0.dev0`
	```yaml
	# axolotl SFT — Run 2 (DenseMixer OFF = native sparse) — Qwen3-30B-A3B-Thinking-2507 on opencode traces.
	# Experiment: axolotl-sft-opencode-densemoe (task #16). The CONTROL arm of the paired ablation:
	# BYTE-IDENTICAL to Run 1 (densemixer_run1_opencode.yaml) EXCEPT `dense_mixer: false` + `output_dir`.
	# Control discipline (POLICY §2): SAME θ₀, seed, dataset + data-ORDER (same shared prepared path),
	# seq_len, batch, packing, LR schedule, step count — the ONLY functional diff is dense_mixer true→false
	# (plugin stays loaded; with false its pre_model_load is a no-op → stock sparse top-k MoE forward).
	# Design/rationale: experiments/active/axolotl-sft-opencode-densemoe/{POLICY,STATE}.md.
	#
	# ⚠ LAUNCH PATH = DIRECT `axolotl.cli.train` (NOT hpc.launch — it would strip plugins/dense_mixer/fp8).
	# ⚠ REQUIRES image `mega_final_dm.sqsh` (densemixer==1.0.1 + the tf-5.x port baked in).

	# θ₀ — the SHARED init both runs start from (control discipline). IDENTICAL to Run 1.
	base_model: /mnt/home/bf996/experiments/densemixer/theta0 # Qwen/Qwen3-30B-A3B-Thinking-2507 @ 144afc2f...
	model_type: AutoModelForCausalLM
	trust_remote_code: true

	# === THE one-flag control diff: DenseMixer OFF ===
	# Plugin stays in the stack (identical to Run 1); `dense_mixer: false` makes its pre_model_load a
	# no-op → the model keeps the STOCK sparse top-k Qwen3MoE forward (non-selected experts' router
	# logits get zero task-loss gradient). This is the sparse baseline for the Δθ counterfactual.
	plugins:
	- axolotl.integrations.densemixer.DenseMixerPlugin
	dense_mixer: false # Run 2 = OFF (the ONLY functional diff vs Run 1).

	# opencode SFT dataset — IDENTICAL pinned revision + SHARED prepared path (guarantees same data ORDER).
	datasets:
	- path: /mnt/home/bf996/experiments/densemixer/data_nemotron_code_oracle
	ds_type: parquet
	data_files:
	- /mnt/home/bf996/experiments/densemixer/data_nemotron_code_oracle/data/train-*.parquet
	type: chat_template
	field_messages: conversations
	message_property_mappings:
	role: role
	content: content
	split_thinking: false
	chat_template: chatml
	dataset_prepared_path: /mnt/home/bf996/experiments/densemixer/prepared/run1 # SHARED with Run 1 (same tokens + order)
	val_set_size: 0.0
	dataset_num_proc: 1
	dataloader_num_workers: 2
	dataloader_prefetch_factor: 2

	# === precision — bf16 + flash-attn (IDENTICAL to Run 1) ===
	bf16: true
	fp16: false
	fp8: false # ⚠ MANDATORY EXPLICIT — axolotl 0.17 auto-enables fp8 on sm_100 → nan.
	tf32: false
	attn_implementation: flash_attention_2

	# === memory / compute (IDENTICAL to Run 1) ===
	# ⚠ Blackwell fix (B-only, functionally inert for A): the STOCK Qwen3MoE experts default to the
	# `grouped_mm` kernel -> `torch._grouped_mm`, which is Hopper-only (cc 9.0) and RuntimeErrors on the
	# B200 (cc 10.0) at the first step (job 31707). `eager` uses the per-expert F.linear loop (no
	# grouped_mm) -> works on Blackwell. This is NOT a control confound: A (dense_mixer:true) replaces the
	# whole SparseMoeBlock.forward with the tf-5.x port that accesses expert weights directly and NEVER
	# calls self.experts.forward, so `experts_implementation` is never exercised on A's path — the only
	# FUNCTIONAL A/B difference remains dense (all-expert STE) vs sparse (top-k). Both do per-expert F.linear.
	experts_implementation: eager
	deepspeed: /opt/axolotl/deepspeed_configs/zero3_bf16.json
	gradient_checkpointing: true
	chunked_cross_entropy: true
	sequence_len: 16384
	sample_packing: true

	# === control discipline — IDENTICAL to Run 1 ===
	seed: 42
	micro_batch_size: 1
	gradient_accumulation_steps: 4
	num_epochs: 3.0
	learning_rate: 2.0e-5
	lr_scheduler: cosine
	warmup_ratio: 0.1
	max_grad_norm: 1.0
	optimizer: adamw_torch_fused
	weight_decay: 0.0

	# === checkpoint cadence (IDENTICAL to Run 1) — θ₀ + intermediate + final for the Δθ trajectory ===
	logging_steps: 1
	save_steps: 10
	save_total_limit: 100
	output_dir: /mnt/home/bf996/experiments/densemixer/run2_sparse_out # DISTINCT from Run 1 (not a control var)
	special_tokens: {}

	```

	</details><br>

	# mnt/home/bf996/experiments/densemixer/run2_sparse_out

	This model was trained from scratch on the None 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: 1
	- eval_batch_size: 1
	- seed: 42
	- distributed_type: multi-GPU
	- num_devices: 8
	- gradient_accumulation_steps: 4
	- total_train_batch_size: 32
	- total_eval_batch_size: 8
	- 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: 9
	- training_steps: 96

	### Training results



	### Framework versions

	- Transformers 5.12.1
	- Pytorch 2.8.0a0+5228986c39.nv25.06
	- Datasets 4.8.5
	- Tokenizers 0.23.0-rc0